OpenCloudOS-Kernel/kernel/rcu/tree.c

5095 lines
166 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0+
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/*
* Read-Copy Update mechanism for mutual exclusion (tree-based version)
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
*
* Copyright IBM Corporation, 2008
*
* Authors: Dipankar Sarma <dipankar@in.ibm.com>
* Manfred Spraul <manfred@colorfullife.com>
* Paul E. McKenney <paulmck@linux.ibm.com>
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
*
* Based on the original work by Paul McKenney <paulmck@linux.ibm.com>
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
* and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
*
* For detailed explanation of Read-Copy Update mechanism see -
* Documentation/RCU
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
*/
#define pr_fmt(fmt) "rcu: " fmt
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
#include <linux/types.h>
#include <linux/kernel.h>
#include <linux/init.h>
#include <linux/spinlock.h>
#include <linux/smp.h>
#include <linux/rcupdate_wait.h>
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
#include <linux/interrupt.h>
#include <linux/sched.h>
#include <linux/sched/debug.h>
debug lockups: Improve lockup detection When debugging a recent lockup bug i found various deficiencies in how our current lockup detection helpers work: - SysRq-L is not very efficient as it uses a workqueue, hence it cannot punch through hard lockups and cannot see through most soft lockups either. - The SysRq-L code depends on the NMI watchdog - which is off by default. - We dont print backtraces from the RCU code's built-in 'RCU state machine is stuck' debug code. This debug code tends to be one of the first (and only) mechanisms that show that a lockup has occured. This patch changes the code so taht we: - Trigger the NMI backtrace code from SysRq-L instead of using a workqueue (which cannot punch through hard lockups) - Trigger print-all-CPU-backtraces from the RCU lockup detection code Also decouple the backtrace printing code from the NMI watchdog: - Dont use variable size cpumasks (it might not be initialized and they are a bit more fragile anyway) - Trigger an NMI immediately via an IPI, instead of waiting for the NMI tick to occur. This is a lot faster and can produce more relevant backtraces. It will also work if the NMI watchdog is disabled. - Dont print the 'dazed and confused' message when we print a backtrace from the NMI - Do a show_regs() plus a dump_stack() to get maximum info out of the dump. Worst-case we get two stacktraces - which is not a big deal. Sometimes, if register content is corrupted, the precise stack walker in show_regs() wont give us a full backtrace - in this case dump_stack() will do it. Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-08-02 17:28:21 +08:00
#include <linux/nmi.h>
rcu: Avoid acquiring rcu_node locks in timer functions This commit switches manipulations of the rcu_node ->wakemask field to atomic operations, which allows rcu_cpu_kthread_timer() to avoid acquiring the rcu_node lock. This should avoid the following lockdep splat reported by Valdis Kletnieks: [ 12.872150] usb 1-4: new high speed USB device number 3 using ehci_hcd [ 12.986667] usb 1-4: New USB device found, idVendor=413c, idProduct=2513 [ 12.986679] usb 1-4: New USB device strings: Mfr=0, Product=0, SerialNumber=0 [ 12.987691] hub 1-4:1.0: USB hub found [ 12.987877] hub 1-4:1.0: 3 ports detected [ 12.996372] input: PS/2 Generic Mouse as /devices/platform/i8042/serio1/input/input10 [ 13.071471] udevadm used greatest stack depth: 3984 bytes left [ 13.172129] [ 13.172130] ======================================================= [ 13.172425] [ INFO: possible circular locking dependency detected ] [ 13.172650] 2.6.39-rc6-mmotm0506 #1 [ 13.172773] ------------------------------------------------------- [ 13.172997] blkid/267 is trying to acquire lock: [ 13.173009] (&p->pi_lock){-.-.-.}, at: [<ffffffff81032d8f>] try_to_wake_up+0x29/0x1aa [ 13.173009] [ 13.173009] but task is already holding lock: [ 13.173009] (rcu_node_level_0){..-...}, at: [<ffffffff810901cc>] rcu_cpu_kthread_timer+0x27/0x58 [ 13.173009] [ 13.173009] which lock already depends on the new lock. [ 13.173009] [ 13.173009] [ 13.173009] the existing dependency chain (in reverse order) is: [ 13.173009] [ 13.173009] -> #2 (rcu_node_level_0){..-...}: [ 13.173009] [<ffffffff810679b9>] check_prevs_add+0x8b/0x104 [ 13.173009] [<ffffffff81067da1>] validate_chain+0x36f/0x3ab [ 13.173009] [<ffffffff8106846b>] __lock_acquire+0x369/0x3e2 [ 13.173009] [<ffffffff81068a0f>] lock_acquire+0xfc/0x14c [ 13.173009] [<ffffffff815697f1>] _raw_spin_lock+0x36/0x45 [ 13.173009] [<ffffffff81090794>] rcu_read_unlock_special+0x8c/0x1d5 [ 13.173009] [<ffffffff8109092c>] __rcu_read_unlock+0x4f/0xd7 [ 13.173009] [<ffffffff81027bd3>] rcu_read_unlock+0x21/0x23 [ 13.173009] [<ffffffff8102cc34>] cpuacct_charge+0x6c/0x75 [ 13.173009] [<ffffffff81030cc6>] update_curr+0x101/0x12e [ 13.173009] [<ffffffff810311d0>] check_preempt_wakeup+0xf7/0x23b [ 13.173009] [<ffffffff8102acb3>] check_preempt_curr+0x2b/0x68 [ 13.173009] [<ffffffff81031d40>] ttwu_do_wakeup+0x76/0x128 [ 13.173009] [<ffffffff81031e49>] ttwu_do_activate.constprop.63+0x57/0x5c [ 13.173009] [<ffffffff81031e96>] scheduler_ipi+0x48/0x5d [ 13.173009] [<ffffffff810177d5>] smp_reschedule_interrupt+0x16/0x18 [ 13.173009] [<ffffffff815710f3>] reschedule_interrupt+0x13/0x20 [ 13.173009] [<ffffffff810b66d1>] rcu_read_unlock+0x21/0x23 [ 13.173009] [<ffffffff810b739c>] find_get_page+0xa9/0xb9 [ 13.173009] [<ffffffff810b8b48>] filemap_fault+0x6a/0x34d [ 13.173009] [<ffffffff810d1a25>] __do_fault+0x54/0x3e6 [ 13.173009] [<ffffffff810d447a>] handle_pte_fault+0x12c/0x1ed [ 13.173009] [<ffffffff810d48f7>] handle_mm_fault+0x1cd/0x1e0 [ 13.173009] [<ffffffff8156cfee>] do_page_fault+0x42d/0x5de [ 13.173009] [<ffffffff8156a75f>] page_fault+0x1f/0x30 [ 13.173009] [ 13.173009] -> #1 (&rq->lock){-.-.-.}: [ 13.173009] [<ffffffff810679b9>] check_prevs_add+0x8b/0x104 [ 13.173009] [<ffffffff81067da1>] validate_chain+0x36f/0x3ab [ 13.173009] [<ffffffff8106846b>] __lock_acquire+0x369/0x3e2 [ 13.173009] [<ffffffff81068a0f>] lock_acquire+0xfc/0x14c [ 13.173009] [<ffffffff815697f1>] _raw_spin_lock+0x36/0x45 [ 13.173009] [<ffffffff81027e19>] __task_rq_lock+0x8b/0xd3 [ 13.173009] [<ffffffff81032f7f>] wake_up_new_task+0x41/0x108 [ 13.173009] [<ffffffff810376c3>] do_fork+0x265/0x33f [ 13.173009] [<ffffffff81007d02>] kernel_thread+0x6b/0x6d [ 13.173009] [<ffffffff8153a9dd>] rest_init+0x21/0xd2 [ 13.173009] [<ffffffff81b1db4f>] start_kernel+0x3bb/0x3c6 [ 13.173009] [<ffffffff81b1d29f>] x86_64_start_reservations+0xaf/0xb3 [ 13.173009] [<ffffffff81b1d393>] x86_64_start_kernel+0xf0/0xf7 [ 13.173009] [ 13.173009] -> #0 (&p->pi_lock){-.-.-.}: [ 13.173009] [<ffffffff81067788>] check_prev_add+0x68/0x20e [ 13.173009] [<ffffffff810679b9>] check_prevs_add+0x8b/0x104 [ 13.173009] [<ffffffff81067da1>] validate_chain+0x36f/0x3ab [ 13.173009] [<ffffffff8106846b>] __lock_acquire+0x369/0x3e2 [ 13.173009] [<ffffffff81068a0f>] lock_acquire+0xfc/0x14c [ 13.173009] [<ffffffff815698ea>] _raw_spin_lock_irqsave+0x44/0x57 [ 13.173009] [<ffffffff81032d8f>] try_to_wake_up+0x29/0x1aa [ 13.173009] [<ffffffff81032f3c>] wake_up_process+0x10/0x12 [ 13.173009] [<ffffffff810901e9>] rcu_cpu_kthread_timer+0x44/0x58 [ 13.173009] [<ffffffff81045286>] call_timer_fn+0xac/0x1e9 [ 13.173009] [<ffffffff8104556d>] run_timer_softirq+0x1aa/0x1f2 [ 13.173009] [<ffffffff8103e487>] __do_softirq+0x109/0x26a [ 13.173009] [<ffffffff8157144c>] call_softirq+0x1c/0x30 [ 13.173009] [<ffffffff81003207>] do_softirq+0x44/0xf1 [ 13.173009] [<ffffffff8103e8b9>] irq_exit+0x58/0xc8 [ 13.173009] [<ffffffff81017f5a>] smp_apic_timer_interrupt+0x79/0x87 [ 13.173009] [<ffffffff81570fd3>] apic_timer_interrupt+0x13/0x20 [ 13.173009] [<ffffffff810bd51a>] get_page_from_freelist+0x2aa/0x310 [ 13.173009] [<ffffffff810bdf03>] __alloc_pages_nodemask+0x178/0x243 [ 13.173009] [<ffffffff8101fe2f>] pte_alloc_one+0x1e/0x3a [ 13.173009] [<ffffffff810d27fe>] __pte_alloc+0x22/0x14b [ 13.173009] [<ffffffff810d48a8>] handle_mm_fault+0x17e/0x1e0 [ 13.173009] [<ffffffff8156cfee>] do_page_fault+0x42d/0x5de [ 13.173009] [<ffffffff8156a75f>] page_fault+0x1f/0x30 [ 13.173009] [ 13.173009] other info that might help us debug this: [ 13.173009] [ 13.173009] Chain exists of: [ 13.173009] &p->pi_lock --> &rq->lock --> rcu_node_level_0 [ 13.173009] [ 13.173009] Possible unsafe locking scenario: [ 13.173009] [ 13.173009] CPU0 CPU1 [ 13.173009] ---- ---- [ 13.173009] lock(rcu_node_level_0); [ 13.173009] lock(&rq->lock); [ 13.173009] lock(rcu_node_level_0); [ 13.173009] lock(&p->pi_lock); [ 13.173009] [ 13.173009] *** DEADLOCK *** [ 13.173009] [ 13.173009] 3 locks held by blkid/267: [ 13.173009] #0: (&mm->mmap_sem){++++++}, at: [<ffffffff8156cdb4>] do_page_fault+0x1f3/0x5de [ 13.173009] #1: (&yield_timer){+.-...}, at: [<ffffffff810451da>] call_timer_fn+0x0/0x1e9 [ 13.173009] #2: (rcu_node_level_0){..-...}, at: [<ffffffff810901cc>] rcu_cpu_kthread_timer+0x27/0x58 [ 13.173009] [ 13.173009] stack backtrace: [ 13.173009] Pid: 267, comm: blkid Not tainted 2.6.39-rc6-mmotm0506 #1 [ 13.173009] Call Trace: [ 13.173009] <IRQ> [<ffffffff8154a529>] print_circular_bug+0xc8/0xd9 [ 13.173009] [<ffffffff81067788>] check_prev_add+0x68/0x20e [ 13.173009] [<ffffffff8100c861>] ? save_stack_trace+0x28/0x46 [ 13.173009] [<ffffffff810679b9>] check_prevs_add+0x8b/0x104 [ 13.173009] [<ffffffff81067da1>] validate_chain+0x36f/0x3ab [ 13.173009] [<ffffffff8106846b>] __lock_acquire+0x369/0x3e2 [ 13.173009] [<ffffffff81032d8f>] ? try_to_wake_up+0x29/0x1aa [ 13.173009] [<ffffffff81068a0f>] lock_acquire+0xfc/0x14c [ 13.173009] [<ffffffff81032d8f>] ? try_to_wake_up+0x29/0x1aa [ 13.173009] [<ffffffff810901a5>] ? rcu_check_quiescent_state+0x82/0x82 [ 13.173009] [<ffffffff815698ea>] _raw_spin_lock_irqsave+0x44/0x57 [ 13.173009] [<ffffffff81032d8f>] ? try_to_wake_up+0x29/0x1aa [ 13.173009] [<ffffffff81032d8f>] try_to_wake_up+0x29/0x1aa [ 13.173009] [<ffffffff810901a5>] ? rcu_check_quiescent_state+0x82/0x82 [ 13.173009] [<ffffffff81032f3c>] wake_up_process+0x10/0x12 [ 13.173009] [<ffffffff810901e9>] rcu_cpu_kthread_timer+0x44/0x58 [ 13.173009] [<ffffffff810901a5>] ? rcu_check_quiescent_state+0x82/0x82 [ 13.173009] [<ffffffff81045286>] call_timer_fn+0xac/0x1e9 [ 13.173009] [<ffffffff810451da>] ? del_timer+0x75/0x75 [ 13.173009] [<ffffffff810901a5>] ? rcu_check_quiescent_state+0x82/0x82 [ 13.173009] [<ffffffff8104556d>] run_timer_softirq+0x1aa/0x1f2 [ 13.173009] [<ffffffff8103e487>] __do_softirq+0x109/0x26a [ 13.173009] [<ffffffff8106365f>] ? tick_dev_program_event+0x37/0xf6 [ 13.173009] [<ffffffff810a0e4a>] ? time_hardirqs_off+0x1b/0x2f [ 13.173009] [<ffffffff8157144c>] call_softirq+0x1c/0x30 [ 13.173009] [<ffffffff81003207>] do_softirq+0x44/0xf1 [ 13.173009] [<ffffffff8103e8b9>] irq_exit+0x58/0xc8 [ 13.173009] [<ffffffff81017f5a>] smp_apic_timer_interrupt+0x79/0x87 [ 13.173009] [<ffffffff81570fd3>] apic_timer_interrupt+0x13/0x20 [ 13.173009] <EOI> [<ffffffff810bd384>] ? get_page_from_freelist+0x114/0x310 [ 13.173009] [<ffffffff810bd51a>] ? get_page_from_freelist+0x2aa/0x310 [ 13.173009] [<ffffffff812220e7>] ? clear_page_c+0x7/0x10 [ 13.173009] [<ffffffff810bd1ef>] ? prep_new_page+0x14c/0x1cd [ 13.173009] [<ffffffff810bd51a>] get_page_from_freelist+0x2aa/0x310 [ 13.173009] [<ffffffff810bdf03>] __alloc_pages_nodemask+0x178/0x243 [ 13.173009] [<ffffffff810d46b9>] ? __pmd_alloc+0x87/0x99 [ 13.173009] [<ffffffff8101fe2f>] pte_alloc_one+0x1e/0x3a [ 13.173009] [<ffffffff810d46b9>] ? __pmd_alloc+0x87/0x99 [ 13.173009] [<ffffffff810d27fe>] __pte_alloc+0x22/0x14b [ 13.173009] [<ffffffff810d48a8>] handle_mm_fault+0x17e/0x1e0 [ 13.173009] [<ffffffff8156cfee>] do_page_fault+0x42d/0x5de [ 13.173009] [<ffffffff810d915f>] ? sys_brk+0x32/0x10c [ 13.173009] [<ffffffff810a0e4a>] ? time_hardirqs_off+0x1b/0x2f [ 13.173009] [<ffffffff81065c4f>] ? trace_hardirqs_off_caller+0x3f/0x9c [ 13.173009] [<ffffffff812235dd>] ? trace_hardirqs_off_thunk+0x3a/0x3c [ 13.173009] [<ffffffff8156a75f>] page_fault+0x1f/0x30 [ 14.010075] usb 5-1: new full speed USB device number 2 using uhci_hcd Reported-by: Valdis Kletnieks <Valdis.Kletnieks@vt.edu> Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2011-05-11 20:41:41 +08:00
#include <linux/atomic.h>
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
#include <linux/bitops.h>
#include <linux/export.h>
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
#include <linux/completion.h>
#include <linux/kmemleak.h>
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
#include <linux/moduleparam.h>
kernel.h: split out panic and oops helpers kernel.h is being used as a dump for all kinds of stuff for a long time. Here is the attempt to start cleaning it up by splitting out panic and oops helpers. There are several purposes of doing this: - dropping dependency in bug.h - dropping a loop by moving out panic_notifier.h - unload kernel.h from something which has its own domain At the same time convert users tree-wide to use new headers, although for the time being include new header back to kernel.h to avoid twisted indirected includes for existing users. [akpm@linux-foundation.org: thread_info.h needs limits.h] [andriy.shevchenko@linux.intel.com: ia64 fix] Link: https://lkml.kernel.org/r/20210520130557.55277-1-andriy.shevchenko@linux.intel.com Link: https://lkml.kernel.org/r/20210511074137.33666-1-andriy.shevchenko@linux.intel.com Signed-off-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com> Reviewed-by: Bjorn Andersson <bjorn.andersson@linaro.org> Co-developed-by: Andrew Morton <akpm@linux-foundation.org> Acked-by: Mike Rapoport <rppt@linux.ibm.com> Acked-by: Corey Minyard <cminyard@mvista.com> Acked-by: Christian Brauner <christian.brauner@ubuntu.com> Acked-by: Arnd Bergmann <arnd@arndb.de> Acked-by: Kees Cook <keescook@chromium.org> Acked-by: Wei Liu <wei.liu@kernel.org> Acked-by: Rasmus Villemoes <linux@rasmusvillemoes.dk> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Acked-by: Sebastian Reichel <sre@kernel.org> Acked-by: Luis Chamberlain <mcgrof@kernel.org> Acked-by: Stephen Boyd <sboyd@kernel.org> Acked-by: Thomas Bogendoerfer <tsbogend@alpha.franken.de> Acked-by: Helge Deller <deller@gmx.de> # parisc Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-07-01 09:54:59 +08:00
#include <linux/panic.h>
#include <linux/panic_notifier.h>
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
#include <linux/percpu.h>
#include <linux/notifier.h>
#include <linux/cpu.h>
#include <linux/mutex.h>
#include <linux/time.h>
#include <linux/kernel_stat.h>
#include <linux/wait.h>
#include <linux/kthread.h>
#include <uapi/linux/sched/types.h>
#include <linux/prefetch.h>
#include <linux/delay.h>
#include <linux/random.h>
#include <linux/trace_events.h>
#include <linux/suspend.h>
#include <linux/ftrace.h>
#include <linux/tick.h>
#include <linux/sysrq.h>
#include <linux/kprobes.h>
#include <linux/gfp.h>
#include <linux/oom.h>
#include <linux/smpboot.h>
#include <linux/jiffies.h>
#include <linux/slab.h>
#include <linux/sched/isolation.h>
#include <linux/sched/clock.h>
#include <linux/vmalloc.h>
#include <linux/mm.h>
rcu: kasan: record and print call_rcu() call stack Patch series "kasan: memorize and print call_rcu stack", v8. This patchset improves KASAN reports by making them to have call_rcu() call stack information. It is useful for programmers to solve use-after-free or double-free memory issue. The KASAN report was as follows(cleaned up slightly): BUG: KASAN: use-after-free in kasan_rcu_reclaim+0x58/0x60 Freed by task 0: kasan_save_stack+0x24/0x50 kasan_set_track+0x24/0x38 kasan_set_free_info+0x18/0x20 __kasan_slab_free+0x10c/0x170 kasan_slab_free+0x10/0x18 kfree+0x98/0x270 kasan_rcu_reclaim+0x1c/0x60 Last call_rcu(): kasan_save_stack+0x24/0x50 kasan_record_aux_stack+0xbc/0xd0 call_rcu+0x8c/0x580 kasan_rcu_uaf+0xf4/0xf8 Generic KASAN will record the last two call_rcu() call stacks and print up to 2 call_rcu() call stacks in KASAN report. it is only suitable for generic KASAN. This feature considers the size of struct kasan_alloc_meta and kasan_free_meta, we try to optimize the structure layout and size, lets it get better memory consumption. [1]https://bugzilla.kernel.org/show_bug.cgi?id=198437 [2]https://groups.google.com/forum/#!searchin/kasan-dev/better$20stack$20traces$20for$20rcu%7Csort:date/kasan-dev/KQsjT_88hDE/7rNUZprRBgAJ This patch (of 4): This feature will record the last two call_rcu() call stacks and prints up to 2 call_rcu() call stacks in KASAN report. When call_rcu() is called, we store the call_rcu() call stack into slub alloc meta-data, so that the KASAN report can print rcu stack. [1]https://bugzilla.kernel.org/show_bug.cgi?id=198437 [2]https://groups.google.com/forum/#!searchin/kasan-dev/better$20stack$20traces$20for$20rcu%7Csort:date/kasan-dev/KQsjT_88hDE/7rNUZprRBgAJ [walter-zh.wu@mediatek.com: build fix] Link: http://lkml.kernel.org/r/20200710162401.23816-1-walter-zh.wu@mediatek.com Suggested-by: Dmitry Vyukov <dvyukov@google.com> Signed-off-by: Walter Wu <walter-zh.wu@mediatek.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Tested-by: Dmitry Vyukov <dvyukov@google.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Reviewed-by: Andrey Konovalov <andreyknvl@google.com> Acked-by: Paul E. McKenney <paulmck@kernel.org> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Alexander Potapenko <glider@google.com> Cc: Josh Triplett <josh@joshtriplett.org> Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Joel Fernandes <joel@joelfernandes.org> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthias Brugger <matthias.bgg@gmail.com> Link: http://lkml.kernel.org/r/20200710162123.23713-1-walter-zh.wu@mediatek.com Link: http://lkml.kernel.org/r/20200601050847.1096-1-walter-zh.wu@mediatek.com Link: http://lkml.kernel.org/r/20200601050927.1153-1-walter-zh.wu@mediatek.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-07 14:24:35 +08:00
#include <linux/kasan.h>
#include <linux/context_tracking.h>
#include "../time/tick-internal.h"
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
#include "tree.h"
#include "rcu.h"
#ifdef MODULE_PARAM_PREFIX
#undef MODULE_PARAM_PREFIX
#endif
#define MODULE_PARAM_PREFIX "rcutree."
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/* Data structures. */
static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, rcu_data) = {
.gpwrap = true,
rcu/nocb: Provide basic callback offloading state machine bits Offloading and de-offloading RCU callback processes must be done carefully. There must never be a time at which callback processing is disabled because the task driving the offloading or de-offloading might be preempted or otherwise stalled at that point in time, which would result in OOM due to calbacks piling up indefinitely. This implies that there will be times during which a given CPU's callbacks might be concurrently invoked by both that CPU's RCU_SOFTIRQ handler (or, equivalently, that CPU's rcuc kthread) and by that CPU's rcuo kthread. This situation could fatally confuse both rcu_barrier() and the CPU-hotplug offlining process, so these must be excluded during any concurrent-callback-invocation period. In addition, during times of concurrent callback invocation, changes to ->cblist must be protected both as needed for RCU_SOFTIRQ and as needed for the rcuo kthread. This commit therefore defines and documents the states for a state machine that coordinates offloading and deoffloading. Cc: Josh Triplett <josh@joshtriplett.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Joel Fernandes <joel@joelfernandes.org> Cc: Neeraj Upadhyay <neeraju@codeaurora.org> Cc: Thomas Gleixner <tglx@linutronix.de> Inspired-by: Paul E. McKenney <paulmck@kernel.org> Tested-by: Boqun Feng <boqun.feng@gmail.com> Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-11-13 20:13:17 +08:00
#ifdef CONFIG_RCU_NOCB_CPU
.cblist.flags = SEGCBLIST_RCU_CORE,
rcu/nocb: Provide basic callback offloading state machine bits Offloading and de-offloading RCU callback processes must be done carefully. There must never be a time at which callback processing is disabled because the task driving the offloading or de-offloading might be preempted or otherwise stalled at that point in time, which would result in OOM due to calbacks piling up indefinitely. This implies that there will be times during which a given CPU's callbacks might be concurrently invoked by both that CPU's RCU_SOFTIRQ handler (or, equivalently, that CPU's rcuc kthread) and by that CPU's rcuo kthread. This situation could fatally confuse both rcu_barrier() and the CPU-hotplug offlining process, so these must be excluded during any concurrent-callback-invocation period. In addition, during times of concurrent callback invocation, changes to ->cblist must be protected both as needed for RCU_SOFTIRQ and as needed for the rcuo kthread. This commit therefore defines and documents the states for a state machine that coordinates offloading and deoffloading. Cc: Josh Triplett <josh@joshtriplett.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Joel Fernandes <joel@joelfernandes.org> Cc: Neeraj Upadhyay <neeraju@codeaurora.org> Cc: Thomas Gleixner <tglx@linutronix.de> Inspired-by: Paul E. McKenney <paulmck@kernel.org> Tested-by: Boqun Feng <boqun.feng@gmail.com> Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-11-13 20:13:17 +08:00
#endif
};
static struct rcu_state rcu_state = {
.level = { &rcu_state.node[0] },
.gp_state = RCU_GP_IDLE,
.gp_seq = (0UL - 300UL) << RCU_SEQ_CTR_SHIFT,
.barrier_mutex = __MUTEX_INITIALIZER(rcu_state.barrier_mutex),
rcu: Make rcu_barrier() no longer block CPU-hotplug operations This commit removes the cpus_read_lock() and cpus_read_unlock() calls from rcu_barrier(), thus allowing CPUs to come and go during the course of rcu_barrier() execution. Posting of the ->barrier_head callbacks does synchronize with portions of RCU's CPU-hotplug notifiers, but these locks are held for short time periods on both sides. Thus, full CPU-hotplug operations could both start and finish during the execution of a given rcu_barrier() invocation. Additional synchronization is provided by a global ->barrier_lock. Since the ->barrier_lock is only used during rcu_barrier() execution and during onlining/offlining a CPU, the contention for this lock should be low. It might be tempting to make use of a per-CPU lock just on general principles, but straightforward attempts to do this have the problems shown below. Initial state: 3 CPUs present, CPU 0 and CPU1 do not have any callback and CPU2 has callbacks. 1. CPU0 calls rcu_barrier(). 2. CPU1 starts offlining for CPU2. CPU1 calls rcutree_migrate_callbacks(). rcu_barrier_entrain() is called from rcutree_migrate_callbacks(), with CPU2's rdp->barrier_lock. It does not entrain ->barrier_head for CPU2, as rcu_barrier() on CPU0 hasn't started the barrier sequence (by calling rcu_seq_start(&rcu_state.barrier_sequence)) yet. 3. CPU0 starts new barrier sequence. It iterates over CPU0 and CPU1, after acquiring their per-cpu ->barrier_lock and finds 0 segcblist length. It updates ->barrier_seq_snap for CPU0 and CPU1 and continues loop iteration to CPU2. for_each_possible_cpu(cpu) { raw_spin_lock_irqsave(&rdp->barrier_lock, flags); if (!rcu_segcblist_n_cbs(&rdp->cblist)) { WRITE_ONCE(rdp->barrier_seq_snap, gseq); raw_spin_unlock_irqrestore(&rdp->barrier_lock, flags); rcu_barrier_trace(TPS("NQ"), cpu, rcu_state.barrier_sequence); continue; } 4. rcutree_migrate_callbacks() completes execution on CPU1. Segcblist len for CPU2 becomes 0. 5. The loop iteration on CPU0, checks rcu_segcblist_n_cbs(&rdp->cblist) for CPU2 and completes the loop iteration after setting ->barrier_seq_snap. 6. As there isn't any ->barrier_head callback entrained; at this point, rcu_barrier() in CPU0 returns. 7. The callbacks, which migrated from CPU2 to CPU1, execute. Straightforward per-CPU locking is also subject to the following race condition noted by Boqun Feng: 1. CPU0 calls rcu_barrier(), starting a new barrier sequence by invoking rcu_seq_start() and init_completion(), but does not yet initialize rcu_state.barrier_cpu_count. 2. CPU1 starts offlining for CPU2, calling rcutree_migrate_callbacks(), which in turn calls rcu_barrier_entrain() holding CPU2's. rdp->barrier_lock. It then entrains ->barrier_head for CPU2 and atomically increments rcu_state.barrier_cpu_count, which is unfortunately not yet initialized to the value 2. 3. The just-entrained RCU callback is invoked. It atomically decrements rcu_state.barrier_cpu_count and sees that it is now zero. This callback therefore invokes complete(). 4. CPU0 continues executing rcu_barrier(), but is not blocked by its call to wait_for_completion(). This results in rcu_barrier() returning before all pre-existing callbacks have been invoked, which is a bug. Therefore, synchronization is provided by rcu_state.barrier_lock, which is also held across the initialization sequence, especially the rcu_seq_start() and the atomic_set() that sets rcu_state.barrier_cpu_count to the value 2. In addition, this lock is held when entraining the rcu_barrier() callback, when deciding whether or not a CPU has callbacks that rcu_barrier() must wait on, when setting the ->qsmaskinitnext for incoming CPUs, and when migrating callbacks from a CPU that is going offline. Reviewed-by: Frederic Weisbecker <frederic@kernel.org> Co-developed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com> Signed-off-by: Neeraj Upadhyay <quic_neeraju@quicinc.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2021-12-15 05:35:17 +08:00
.barrier_lock = __RAW_SPIN_LOCK_UNLOCKED(rcu_state.barrier_lock),
.name = RCU_NAME,
.abbr = RCU_ABBR,
.exp_mutex = __MUTEX_INITIALIZER(rcu_state.exp_mutex),
.exp_wake_mutex = __MUTEX_INITIALIZER(rcu_state.exp_wake_mutex),
.ofl_lock = __ARCH_SPIN_LOCK_UNLOCKED,
};
/* Dump rcu_node combining tree at boot to verify correct setup. */
static bool dump_tree;
module_param(dump_tree, bool, 0444);
/* By default, use RCU_SOFTIRQ instead of rcuc kthreads. */
static bool use_softirq = !IS_ENABLED(CONFIG_PREEMPT_RT);
#ifndef CONFIG_PREEMPT_RT
module_param(use_softirq, bool, 0444);
#endif
/* Control rcu_node-tree auto-balancing at boot time. */
static bool rcu_fanout_exact;
module_param(rcu_fanout_exact, bool, 0444);
/* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
module_param(rcu_fanout_leaf, int, 0444);
int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
/* Number of rcu_nodes at specified level. */
int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
/*
rcu: Narrow early boot window of illegal synchronous grace periods The current preemptible RCU implementation goes through three phases during bootup. In the first phase, there is only one CPU that is running with preemption disabled, so that a no-op is a synchronous grace period. In the second mid-boot phase, the scheduler is running, but RCU has not yet gotten its kthreads spawned (and, for expedited grace periods, workqueues are not yet running. During this time, any attempt to do a synchronous grace period will hang the system (or complain bitterly, depending). In the third and final phase, RCU is fully operational and everything works normally. This has been OK for some time, but there has recently been some synchronous grace periods showing up during the second mid-boot phase. This code worked "by accident" for awhile, but started failing as soon as expedited RCU grace periods switched over to workqueues in commit 8b355e3bc140 ("rcu: Drive expedited grace periods from workqueue"). Note that the code was buggy even before this commit, as it was subject to failure on real-time systems that forced all expedited grace periods to run as normal grace periods (for example, using the rcu_normal ksysfs parameter). The callchain from the failure case is as follows: early_amd_iommu_init() |-> acpi_put_table(ivrs_base); |-> acpi_tb_put_table(table_desc); |-> acpi_tb_invalidate_table(table_desc); |-> acpi_tb_release_table(...) |-> acpi_os_unmap_memory |-> acpi_os_unmap_iomem |-> acpi_os_map_cleanup |-> synchronize_rcu_expedited The kernel showing this callchain was built with CONFIG_PREEMPT_RCU=y, which caused the code to try using workqueues before they were initialized, which did not go well. This commit therefore reworks RCU to permit synchronous grace periods to proceed during this mid-boot phase. This commit is therefore a fix to a regression introduced in v4.9, and is therefore being put forward post-merge-window in v4.10. This commit sets a flag from the existing rcu_scheduler_starting() function which causes all synchronous grace periods to take the expedited path. The expedited path now checks this flag, using the requesting task to drive the expedited grace period forward during the mid-boot phase. Finally, this flag is updated by a core_initcall() function named rcu_exp_runtime_mode(), which causes the runtime codepaths to be used. Note that this arrangement assumes that tasks are not sent POSIX signals (or anything similar) from the time that the first task is spawned through core_initcall() time. Fixes: 8b355e3bc140 ("rcu: Drive expedited grace periods from workqueue") Reported-by: "Zheng, Lv" <lv.zheng@intel.com> Reported-by: Borislav Petkov <bp@alien8.de> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Tested-by: Stan Kain <stan.kain@gmail.com> Tested-by: Ivan <waffolz@hotmail.com> Tested-by: Emanuel Castelo <emanuel.castelo@gmail.com> Tested-by: Bruno Pesavento <bpesavento@infinito.it> Tested-by: Borislav Petkov <bp@suse.de> Tested-by: Frederic Bezies <fredbezies@gmail.com> Cc: <stable@vger.kernel.org> # 4.9.0-
2017-01-10 18:28:26 +08:00
* The rcu_scheduler_active variable is initialized to the value
* RCU_SCHEDULER_INACTIVE and transitions RCU_SCHEDULER_INIT just before the
* first task is spawned. So when this variable is RCU_SCHEDULER_INACTIVE,
* RCU can assume that there is but one task, allowing RCU to (for example)
* optimize synchronize_rcu() to a simple barrier(). When this variable
rcu: Narrow early boot window of illegal synchronous grace periods The current preemptible RCU implementation goes through three phases during bootup. In the first phase, there is only one CPU that is running with preemption disabled, so that a no-op is a synchronous grace period. In the second mid-boot phase, the scheduler is running, but RCU has not yet gotten its kthreads spawned (and, for expedited grace periods, workqueues are not yet running. During this time, any attempt to do a synchronous grace period will hang the system (or complain bitterly, depending). In the third and final phase, RCU is fully operational and everything works normally. This has been OK for some time, but there has recently been some synchronous grace periods showing up during the second mid-boot phase. This code worked "by accident" for awhile, but started failing as soon as expedited RCU grace periods switched over to workqueues in commit 8b355e3bc140 ("rcu: Drive expedited grace periods from workqueue"). Note that the code was buggy even before this commit, as it was subject to failure on real-time systems that forced all expedited grace periods to run as normal grace periods (for example, using the rcu_normal ksysfs parameter). The callchain from the failure case is as follows: early_amd_iommu_init() |-> acpi_put_table(ivrs_base); |-> acpi_tb_put_table(table_desc); |-> acpi_tb_invalidate_table(table_desc); |-> acpi_tb_release_table(...) |-> acpi_os_unmap_memory |-> acpi_os_unmap_iomem |-> acpi_os_map_cleanup |-> synchronize_rcu_expedited The kernel showing this callchain was built with CONFIG_PREEMPT_RCU=y, which caused the code to try using workqueues before they were initialized, which did not go well. This commit therefore reworks RCU to permit synchronous grace periods to proceed during this mid-boot phase. This commit is therefore a fix to a regression introduced in v4.9, and is therefore being put forward post-merge-window in v4.10. This commit sets a flag from the existing rcu_scheduler_starting() function which causes all synchronous grace periods to take the expedited path. The expedited path now checks this flag, using the requesting task to drive the expedited grace period forward during the mid-boot phase. Finally, this flag is updated by a core_initcall() function named rcu_exp_runtime_mode(), which causes the runtime codepaths to be used. Note that this arrangement assumes that tasks are not sent POSIX signals (or anything similar) from the time that the first task is spawned through core_initcall() time. Fixes: 8b355e3bc140 ("rcu: Drive expedited grace periods from workqueue") Reported-by: "Zheng, Lv" <lv.zheng@intel.com> Reported-by: Borislav Petkov <bp@alien8.de> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Tested-by: Stan Kain <stan.kain@gmail.com> Tested-by: Ivan <waffolz@hotmail.com> Tested-by: Emanuel Castelo <emanuel.castelo@gmail.com> Tested-by: Bruno Pesavento <bpesavento@infinito.it> Tested-by: Borislav Petkov <bp@suse.de> Tested-by: Frederic Bezies <fredbezies@gmail.com> Cc: <stable@vger.kernel.org> # 4.9.0-
2017-01-10 18:28:26 +08:00
* is RCU_SCHEDULER_INIT, RCU must actually do all the hard work required
* to detect real grace periods. This variable is also used to suppress
* boot-time false positives from lockdep-RCU error checking. Finally, it
* transitions from RCU_SCHEDULER_INIT to RCU_SCHEDULER_RUNNING after RCU
* is fully initialized, including all of its kthreads having been spawned.
*/
int rcu_scheduler_active __read_mostly;
EXPORT_SYMBOL_GPL(rcu_scheduler_active);
/*
* The rcu_scheduler_fully_active variable transitions from zero to one
* during the early_initcall() processing, which is after the scheduler
* is capable of creating new tasks. So RCU processing (for example,
* creating tasks for RCU priority boosting) must be delayed until after
* rcu_scheduler_fully_active transitions from zero to one. We also
* currently delay invocation of any RCU callbacks until after this point.
*
* It might later prove better for people registering RCU callbacks during
* early boot to take responsibility for these callbacks, but one step at
* a time.
*/
static int rcu_scheduler_fully_active __read_mostly;
static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
unsigned long gps, unsigned long flags);
rcu: Yield simpler The rcu_yield() code is amazing. It's there to avoid starvation of the system when lots of (boosting) work is to be done. Now looking at the code it's functionality is: Make the thread SCHED_OTHER and very nice, i.e. get it out of the way Arm a timer with 2 ticks schedule() Now if the system goes idle the rcu task returns, regains SCHED_FIFO and plugs on. If the systems stays busy the timer fires and wakes a per node kthread which in turn makes the per cpu thread SCHED_FIFO and brings it back on the cpu. For the boosting thread the "make it FIFO" bit is missing and it just runs some magic boost checks. Now this is a lot of code with extra threads and complexity. It's way simpler to let the tasks when they detect overload schedule away for 2 ticks and defer the normal wakeup as long as they are in yielded state and the cpu is not idle. That solves the same problem and the only difference is that when the cpu goes idle it's not guaranteed that the thread returns right away, but it won't be longer out than two ticks, so no harm is done. If that's an issue than it is way simpler just to wake the task from idle as RCU has callbacks there anyway. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Srivatsa S. Bhat <srivatsa.bhat@linux.vnet.ibm.com> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: Namhyung Kim <namhyung@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Link: http://lkml.kernel.org/r/20120716103948.131256723@linutronix.de Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2012-07-16 18:42:35 +08:00
static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
static void invoke_rcu_core(void);
static void rcu_report_exp_rdp(struct rcu_data *rdp);
static void sync_sched_exp_online_cleanup(int cpu);
static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp);
static bool rcu_rdp_is_offloaded(struct rcu_data *rdp);
static bool rcu_rdp_cpu_online(struct rcu_data *rdp);
static bool rcu_init_invoked(void);
static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
rcu/nocb: Add option to opt rcuo kthreads out of RT priority This commit introduces a RCU_NOCB_CPU_CB_BOOST Kconfig option that prevents rcuo kthreads from running at real-time priority, even in kernels built with RCU_BOOST. This capability is important to devices needing low-latency (as in a few milliseconds) response from expedited RCU grace periods, but which are not running a classic real-time workload. On such devices, permitting the rcuo kthreads to run at real-time priority results in unacceptable latencies imposed on the application tasks, which run as SCHED_OTHER. See for example the following trace output: <snip> <...>-60 [006] d..1 2979.028717: rcu_batch_start: rcu_preempt CBs=34619 bl=270 <snip> If that rcuop kthread were permitted to run at real-time SCHED_FIFO priority, it would monopolize its CPU for hundreds of milliseconds while invoking those 34619 RCU callback functions, which would cause an unacceptably long latency spike for many application stacks on Android platforms. However, some existing real-time workloads require that callback invocation run at SCHED_FIFO priority, for example, those running on systems with heavy SCHED_OTHER background loads. (It is the real-time system's administrator's responsibility to make sure that important real-time tasks run at a higher priority than do RCU's kthreads.) Therefore, this new RCU_NOCB_CPU_CB_BOOST Kconfig option defaults to "y" on kernels built with PREEMPT_RT and defaults to "n" otherwise. The effect is to preserve current behavior for real-time systems, but for other systems to allow expedited RCU grace periods to run with real-time priority while continuing to invoke RCU callbacks as SCHED_OTHER. As you would expect, this RCU_NOCB_CPU_CB_BOOST Kconfig option has no effect except on CPUs with offloaded RCU callbacks. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org> Acked-by: Joel Fernandes (Google) <joel@joelfernandes.org> Reviewed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com>
2022-05-11 16:57:03 +08:00
/*
* rcuc/rcub/rcuop kthread realtime priority. The "rcuop"
* real-time priority(enabling/disabling) is controlled by
* the extra CONFIG_RCU_NOCB_CPU_CB_BOOST configuration.
*/
static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
module_param(kthread_prio, int, 0444);
rcu: Control grace-period delays directly from value In a misguided attempt to avoid an #ifdef, the use of the gp_init_delay module parameter was conditioned on the corresponding RCU_TORTURE_TEST_SLOW_INIT Kconfig variable, using IS_ENABLED() at the point of use in the code. This meant that the compiler always saw the delay, which meant that RCU_TORTURE_TEST_SLOW_INIT_DELAY had to be unconditionally defined. This in turn caused "make oldconfig" to ask pointless questions about the value of RCU_TORTURE_TEST_SLOW_INIT_DELAY in cases where it was not even used. This commit avoids these pointless questions by defining gp_init_delay under #ifdef. In one branch, gp_init_delay is initialized to RCU_TORTURE_TEST_SLOW_INIT_DELAY and is also a module parameter (thus allowing boot-time modification), and in the other branch gp_init_delay is a const variable initialized by default to zero. This approach also simplifies the code at the delay point by eliminating the IS_DEFINED(). Because gp_init_delay is constant zero in the no-delay case intended for production use, the "gp_init_delay > 0" check causes the delay to become dead code, as desired in this case. In addition, this commit replaces magic constant "10" with the preprocessor variable PER_RCU_NODE_PERIOD, which controls the number of grace periods that are allowed to elapse at full speed before a delay is inserted. Reported-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2015-04-15 10:33:59 +08:00
/* Delay in jiffies for grace-period initialization delays, debug only. */
static int gp_preinit_delay;
module_param(gp_preinit_delay, int, 0444);
static int gp_init_delay;
module_param(gp_init_delay, int, 0444);
static int gp_cleanup_delay;
module_param(gp_cleanup_delay, int, 0444);
// Add delay to rcu_read_unlock() for strict grace periods.
static int rcu_unlock_delay;
#ifdef CONFIG_RCU_STRICT_GRACE_PERIOD
module_param(rcu_unlock_delay, int, 0444);
#endif
/*
* This rcu parameter is runtime-read-only. It reflects
* a minimum allowed number of objects which can be cached
* per-CPU. Object size is equal to one page. This value
* can be changed at boot time.
*/
rcu/tree: Defer kvfree_rcu() allocation to a clean context The current memmory-allocation interface causes the following difficulties for kvfree_rcu(): a) If built with CONFIG_PROVE_RAW_LOCK_NESTING, the lockdep will complain about violation of the nesting rules, as in "BUG: Invalid wait context". This Kconfig option checks for proper raw_spinlock vs. spinlock nesting, in particular, it is not legal to acquire a spinlock_t while holding a raw_spinlock_t. This is a problem because kfree_rcu() uses raw_spinlock_t whereas the "page allocator" internally deals with spinlock_t to access to its zones. The code also can be broken from higher level of view: <snip> raw_spin_lock(&some_lock); kfree_rcu(some_pointer, some_field_offset); <snip> b) If built with CONFIG_PREEMPT_RT, spinlock_t is converted into sleeplock. This means that invoking the page allocator from atomic contexts results in "BUG: scheduling while atomic". c) Please note that call_rcu() is already invoked from raw atomic context, so it is only reasonable to expaect that kfree_rcu() and kvfree_rcu() will also be called from atomic raw context. This commit therefore defers page allocation to a clean context using the combination of an hrtimer and a workqueue. The hrtimer stage is required in order to avoid deadlocks with the scheduler. This deferred allocation is required only when kvfree_rcu()'s per-CPU page cache is empty. Link: https://lore.kernel.org/lkml/20200630164543.4mdcf6zb4zfclhln@linutronix.de/ Fixes: 3042f83f19be ("rcu: Support reclaim for head-less object") Reported-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-10-30 00:50:04 +08:00
static int rcu_min_cached_objs = 5;
module_param(rcu_min_cached_objs, int, 0444);
// A page shrinker can ask for pages to be freed to make them
// available for other parts of the system. This usually happens
// under low memory conditions, and in that case we should also
// defer page-cache filling for a short time period.
//
// The default value is 5 seconds, which is long enough to reduce
// interference with the shrinker while it asks other systems to
// drain their caches.
static int rcu_delay_page_cache_fill_msec = 5000;
module_param(rcu_delay_page_cache_fill_msec, int, 0444);
/* Retrieve RCU kthreads priority for rcutorture */
int rcu_get_gp_kthreads_prio(void)
{
return kthread_prio;
}
EXPORT_SYMBOL_GPL(rcu_get_gp_kthreads_prio);
/*
* Number of grace periods between delays, normalized by the duration of
* the delay. The longer the delay, the more the grace periods between
* each delay. The reason for this normalization is that it means that,
* for non-zero delays, the overall slowdown of grace periods is constant
* regardless of the duration of the delay. This arrangement balances
* the need for long delays to increase some race probabilities with the
* need for fast grace periods to increase other race probabilities.
*/
#define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays for debugging. */
/*
* Return true if an RCU grace period is in progress. The READ_ONCE()s
* permit this function to be invoked without holding the root rcu_node
* structure's ->lock, but of course results can be subject to change.
*/
static int rcu_gp_in_progress(void)
{
return rcu_seq_state(rcu_seq_current(&rcu_state.gp_seq));
}
/*
* Return the number of callbacks queued on the specified CPU.
* Handles both the nocbs and normal cases.
*/
static long rcu_get_n_cbs_cpu(int cpu)
{
struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
if (rcu_segcblist_is_enabled(&rdp->cblist))
return rcu_segcblist_n_cbs(&rdp->cblist);
return 0;
}
void rcu_softirq_qs(void)
{
rcu_qs();
rcu_preempt_deferred_qs(current);
rcu_tasks_qs(current, false);
}
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/*
* Reset the current CPU's ->dynticks counter to indicate that the
* newly onlined CPU is no longer in an extended quiescent state.
* This will either leave the counter unchanged, or increment it
* to the next non-quiescent value.
*
* The non-atomic test/increment sequence works because the upper bits
* of the ->dynticks counter are manipulated only by the corresponding CPU,
* or when the corresponding CPU is offline.
*/
static void rcu_dynticks_eqs_online(void)
{
if (ct_dynticks() & RCU_DYNTICKS_IDX)
return;
ct_state_inc(RCU_DYNTICKS_IDX);
}
/*
* Snapshot the ->dynticks counter with full ordering so as to allow
* stable comparison of this counter with past and future snapshots.
*/
static int rcu_dynticks_snap(int cpu)
{
rcu: Weaken ->dynticks accesses and updates Accesses to the rcu_data structure's ->dynticks field have always been fully ordered because it was not possible to prove that weaker ordering was safe. However, with the removal of the rcu_eqs_special_set() function and the advent of the Linux-kernel memory model, it is now easy to show that two of the four original full memory barriers can be weakened to acquire and release operations. The remaining pair must remain full memory barriers. This change makes the memory ordering requirements more evident, and it might well also speed up the to-idle and from-idle fastpaths on some architectures. The following litmus test, adapted from one supplied off-list by Frederic Weisbecker, models the RCU grace-period kthread detecting an idle CPU that is concurrently transitioning to non-idle: C dynticks-from-idle { DYNTICKS=0; (* Initially idle. *) } P0(int *X, int *DYNTICKS) { int dynticks; int x; // Idle. dynticks = READ_ONCE(*DYNTICKS); smp_store_release(DYNTICKS, dynticks + 1); smp_mb(); // Now non-idle x = READ_ONCE(*X); } P1(int *X, int *DYNTICKS) { int dynticks; WRITE_ONCE(*X, 1); smp_mb(); dynticks = smp_load_acquire(DYNTICKS); } exists (1:dynticks=0 /\ 0:x=1) Running "herd7 -conf linux-kernel.cfg dynticks-from-idle.litmus" verifies this transition, namely, showing that if the RCU grace-period kthread (P1) sees another CPU as idle (P0), then any memory access prior to the start of the grace period (P1's write to X) will be seen by any RCU read-side critical section following the to-non-idle transition (P0's read from X). This is a straightforward use of full memory barriers to force ordering in a store-buffering (SB) litmus test. The following litmus test, also adapted from the one supplied off-list by Frederic Weisbecker, models the RCU grace-period kthread detecting a non-idle CPU that is concurrently transitioning to idle: C dynticks-into-idle { DYNTICKS=1; (* Initially non-idle. *) } P0(int *X, int *DYNTICKS) { int dynticks; // Non-idle. WRITE_ONCE(*X, 1); dynticks = READ_ONCE(*DYNTICKS); smp_store_release(DYNTICKS, dynticks + 1); smp_mb(); // Now idle. } P1(int *X, int *DYNTICKS) { int x; int dynticks; smp_mb(); dynticks = smp_load_acquire(DYNTICKS); x = READ_ONCE(*X); } exists (1:dynticks=2 /\ 1:x=0) Running "herd7 -conf linux-kernel.cfg dynticks-into-idle.litmus" verifies this transition, namely, showing that if the RCU grace-period kthread (P1) sees another CPU as newly idle (P0), then any pre-idle memory access (P0's write to X) will be seen by any code following the grace period (P1's read from X). This is a simple release-acquire pair forcing ordering in a message-passing (MP) litmus test. Of course, if the grace-period kthread detects the CPU as non-idle, it will refrain from reporting a quiescent state on behalf of that CPU, so there are no ordering requirements from the grace-period kthread in that case. However, other subsystems call rcu_is_idle_cpu() to check for CPUs being non-idle from an RCU perspective. That case is also verified by the above litmus tests with the proviso that the sense of the low-order bit of the DYNTICKS counter be inverted. Unfortunately, on x86 smp_mb() is as expensive as a cache-local atomic increment. This commit therefore weakens only the read from ->dynticks. However, the updates are abstracted into a rcu_dynticks_inc() function to ease any future changes that might be needed. [ paulmck: Apply Linus Torvalds feedback. ] Link: https://lore.kernel.org/lkml/20210721202127.2129660-4-paulmck@kernel.org/ Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Acked-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2021-05-20 08:25:42 +08:00
smp_mb(); // Fundamental RCU ordering guarantee.
return ct_dynticks_cpu_acquire(cpu);
}
/*
* Return true if the snapshot returned from rcu_dynticks_snap()
* indicates that RCU is in an extended quiescent state.
*/
static bool rcu_dynticks_in_eqs(int snap)
{
return !(snap & RCU_DYNTICKS_IDX);
}
/*
* Return true if the CPU corresponding to the specified rcu_data
* structure has spent some time in an extended quiescent state since
* rcu_dynticks_snap() returned the specified snapshot.
*/
static bool rcu_dynticks_in_eqs_since(struct rcu_data *rdp, int snap)
{
return snap != rcu_dynticks_snap(rdp->cpu);
}
rcu-tasks: Avoid IPIing userspace/idle tasks if kernel is so built Systems running CPU-bound real-time task do not want IPIs sent to CPUs executing nohz_full userspace tasks. Battery-powered systems don't want IPIs sent to idle CPUs in low-power mode. Unfortunately, RCU tasks trace can and will send such IPIs in some cases. Both of these situations occur only when the target CPU is in RCU dyntick-idle mode, in other words, when RCU is not watching the target CPU. This suggests that CPUs in dyntick-idle mode should use memory barriers in outermost invocations of rcu_read_lock_trace() and rcu_read_unlock_trace(), which would allow the RCU tasks trace grace period to directly read out the target CPU's read-side state. One challenge is that RCU tasks trace is not targeting a specific CPU, but rather a task. And that task could switch from one CPU to another at any time. This commit therefore uses try_invoke_on_locked_down_task() and checks for task_curr() in trc_inspect_reader_notrunning(). When this condition holds, the target task is running and cannot move. If CONFIG_TASKS_TRACE_RCU_READ_MB=y, the new rcu_dynticks_zero_in_eqs() function can be used to check if the specified integer (in this case, t->trc_reader_nesting) is zero while the target CPU remains in that same dyntick-idle sojourn. If so, the target task is in a quiescent state. If not, trc_read_check_handler() must indicate failure so that the grace-period kthread can take appropriate action or retry after an appropriate delay, as the case may be. With this change, given CONFIG_TASKS_TRACE_RCU_READ_MB=y, if a given CPU remains idle or a given task continues executing in nohz_full mode, the RCU tasks trace grace-period kthread will detect this without the need to send an IPI. Suggested-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-03-20 06:33:12 +08:00
/*
* Return true if the referenced integer is zero while the specified
* CPU remains within a single extended quiescent state.
*/
bool rcu_dynticks_zero_in_eqs(int cpu, int *vp)
{
int snap;
// If not quiescent, force back to earlier extended quiescent state.
snap = ct_dynticks_cpu(cpu) & ~RCU_DYNTICKS_IDX;
rcu-tasks: Avoid IPIing userspace/idle tasks if kernel is so built Systems running CPU-bound real-time task do not want IPIs sent to CPUs executing nohz_full userspace tasks. Battery-powered systems don't want IPIs sent to idle CPUs in low-power mode. Unfortunately, RCU tasks trace can and will send such IPIs in some cases. Both of these situations occur only when the target CPU is in RCU dyntick-idle mode, in other words, when RCU is not watching the target CPU. This suggests that CPUs in dyntick-idle mode should use memory barriers in outermost invocations of rcu_read_lock_trace() and rcu_read_unlock_trace(), which would allow the RCU tasks trace grace period to directly read out the target CPU's read-side state. One challenge is that RCU tasks trace is not targeting a specific CPU, but rather a task. And that task could switch from one CPU to another at any time. This commit therefore uses try_invoke_on_locked_down_task() and checks for task_curr() in trc_inspect_reader_notrunning(). When this condition holds, the target task is running and cannot move. If CONFIG_TASKS_TRACE_RCU_READ_MB=y, the new rcu_dynticks_zero_in_eqs() function can be used to check if the specified integer (in this case, t->trc_reader_nesting) is zero while the target CPU remains in that same dyntick-idle sojourn. If so, the target task is in a quiescent state. If not, trc_read_check_handler() must indicate failure so that the grace-period kthread can take appropriate action or retry after an appropriate delay, as the case may be. With this change, given CONFIG_TASKS_TRACE_RCU_READ_MB=y, if a given CPU remains idle or a given task continues executing in nohz_full mode, the RCU tasks trace grace-period kthread will detect this without the need to send an IPI. Suggested-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-03-20 06:33:12 +08:00
smp_rmb(); // Order ->dynticks and *vp reads.
if (READ_ONCE(*vp))
return false; // Non-zero, so report failure;
smp_rmb(); // Order *vp read and ->dynticks re-read.
// If still in the same extended quiescent state, we are good!
return snap == ct_dynticks_cpu(cpu);
}
rcu: Reduce overhead of cond_resched() checks for RCU Commit ac1bea85781e (Make cond_resched() report RCU quiescent states) fixed a problem where a CPU looping in the kernel with but one runnable task would give RCU CPU stall warnings, even if the in-kernel loop contained cond_resched() calls. Unfortunately, in so doing, it introduced performance regressions in Anton Blanchard's will-it-scale "open1" test. The problem appears to be not so much the increased cond_resched() path length as an increase in the rate at which grace periods complete, which increased per-update grace-period overhead. This commit takes a different approach to fixing this bug, mainly by moving the RCU-visible quiescent state from cond_resched() to rcu_note_context_switch(), and by further reducing the check to a simple non-zero test of a single per-CPU variable. However, this approach requires that the force-quiescent-state processing send resched IPIs to the offending CPUs. These will be sent only once the grace period has reached an age specified by the boot/sysfs parameter rcutree.jiffies_till_sched_qs, or once the grace period reaches an age halfway to the point at which RCU CPU stall warnings will be emitted, whichever comes first. Reported-by: Dave Hansen <dave.hansen@intel.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Christoph Lameter <cl@gentwo.org> Cc: Mike Galbraith <umgwanakikbuti@gmail.com> Cc: Eric Dumazet <eric.dumazet@gmail.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org> [ paulmck: Made rcu_momentary_dyntick_idle() as suggested by the ktest build robot. Also fixed smp_mb() comment as noted by Oleg Nesterov. ] Merge with e552592e (Reduce overhead of cond_resched() checks for RCU) Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2014-06-21 07:49:01 +08:00
/*
* Let the RCU core know that this CPU has gone through the scheduler,
* which is a quiescent state. This is called when the need for a
* quiescent state is urgent, so we burn an atomic operation and full
* memory barriers to let the RCU core know about it, regardless of what
* this CPU might (or might not) do in the near future.
*
* We inform the RCU core by emulating a zero-duration dyntick-idle period.
*
* The caller must have disabled interrupts and must not be idle.
rcu: Reduce overhead of cond_resched() checks for RCU Commit ac1bea85781e (Make cond_resched() report RCU quiescent states) fixed a problem where a CPU looping in the kernel with but one runnable task would give RCU CPU stall warnings, even if the in-kernel loop contained cond_resched() calls. Unfortunately, in so doing, it introduced performance regressions in Anton Blanchard's will-it-scale "open1" test. The problem appears to be not so much the increased cond_resched() path length as an increase in the rate at which grace periods complete, which increased per-update grace-period overhead. This commit takes a different approach to fixing this bug, mainly by moving the RCU-visible quiescent state from cond_resched() to rcu_note_context_switch(), and by further reducing the check to a simple non-zero test of a single per-CPU variable. However, this approach requires that the force-quiescent-state processing send resched IPIs to the offending CPUs. These will be sent only once the grace period has reached an age specified by the boot/sysfs parameter rcutree.jiffies_till_sched_qs, or once the grace period reaches an age halfway to the point at which RCU CPU stall warnings will be emitted, whichever comes first. Reported-by: Dave Hansen <dave.hansen@intel.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Christoph Lameter <cl@gentwo.org> Cc: Mike Galbraith <umgwanakikbuti@gmail.com> Cc: Eric Dumazet <eric.dumazet@gmail.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org> [ paulmck: Made rcu_momentary_dyntick_idle() as suggested by the ktest build robot. Also fixed smp_mb() comment as noted by Oleg Nesterov. ] Merge with e552592e (Reduce overhead of cond_resched() checks for RCU) Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2014-06-21 07:49:01 +08:00
*/
notrace void rcu_momentary_dyntick_idle(void)
rcu: Reduce overhead of cond_resched() checks for RCU Commit ac1bea85781e (Make cond_resched() report RCU quiescent states) fixed a problem where a CPU looping in the kernel with but one runnable task would give RCU CPU stall warnings, even if the in-kernel loop contained cond_resched() calls. Unfortunately, in so doing, it introduced performance regressions in Anton Blanchard's will-it-scale "open1" test. The problem appears to be not so much the increased cond_resched() path length as an increase in the rate at which grace periods complete, which increased per-update grace-period overhead. This commit takes a different approach to fixing this bug, mainly by moving the RCU-visible quiescent state from cond_resched() to rcu_note_context_switch(), and by further reducing the check to a simple non-zero test of a single per-CPU variable. However, this approach requires that the force-quiescent-state processing send resched IPIs to the offending CPUs. These will be sent only once the grace period has reached an age specified by the boot/sysfs parameter rcutree.jiffies_till_sched_qs, or once the grace period reaches an age halfway to the point at which RCU CPU stall warnings will be emitted, whichever comes first. Reported-by: Dave Hansen <dave.hansen@intel.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Christoph Lameter <cl@gentwo.org> Cc: Mike Galbraith <umgwanakikbuti@gmail.com> Cc: Eric Dumazet <eric.dumazet@gmail.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org> [ paulmck: Made rcu_momentary_dyntick_idle() as suggested by the ktest build robot. Also fixed smp_mb() comment as noted by Oleg Nesterov. ] Merge with e552592e (Reduce overhead of cond_resched() checks for RCU) Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2014-06-21 07:49:01 +08:00
{
rcu: Weaken ->dynticks accesses and updates Accesses to the rcu_data structure's ->dynticks field have always been fully ordered because it was not possible to prove that weaker ordering was safe. However, with the removal of the rcu_eqs_special_set() function and the advent of the Linux-kernel memory model, it is now easy to show that two of the four original full memory barriers can be weakened to acquire and release operations. The remaining pair must remain full memory barriers. This change makes the memory ordering requirements more evident, and it might well also speed up the to-idle and from-idle fastpaths on some architectures. The following litmus test, adapted from one supplied off-list by Frederic Weisbecker, models the RCU grace-period kthread detecting an idle CPU that is concurrently transitioning to non-idle: C dynticks-from-idle { DYNTICKS=0; (* Initially idle. *) } P0(int *X, int *DYNTICKS) { int dynticks; int x; // Idle. dynticks = READ_ONCE(*DYNTICKS); smp_store_release(DYNTICKS, dynticks + 1); smp_mb(); // Now non-idle x = READ_ONCE(*X); } P1(int *X, int *DYNTICKS) { int dynticks; WRITE_ONCE(*X, 1); smp_mb(); dynticks = smp_load_acquire(DYNTICKS); } exists (1:dynticks=0 /\ 0:x=1) Running "herd7 -conf linux-kernel.cfg dynticks-from-idle.litmus" verifies this transition, namely, showing that if the RCU grace-period kthread (P1) sees another CPU as idle (P0), then any memory access prior to the start of the grace period (P1's write to X) will be seen by any RCU read-side critical section following the to-non-idle transition (P0's read from X). This is a straightforward use of full memory barriers to force ordering in a store-buffering (SB) litmus test. The following litmus test, also adapted from the one supplied off-list by Frederic Weisbecker, models the RCU grace-period kthread detecting a non-idle CPU that is concurrently transitioning to idle: C dynticks-into-idle { DYNTICKS=1; (* Initially non-idle. *) } P0(int *X, int *DYNTICKS) { int dynticks; // Non-idle. WRITE_ONCE(*X, 1); dynticks = READ_ONCE(*DYNTICKS); smp_store_release(DYNTICKS, dynticks + 1); smp_mb(); // Now idle. } P1(int *X, int *DYNTICKS) { int x; int dynticks; smp_mb(); dynticks = smp_load_acquire(DYNTICKS); x = READ_ONCE(*X); } exists (1:dynticks=2 /\ 1:x=0) Running "herd7 -conf linux-kernel.cfg dynticks-into-idle.litmus" verifies this transition, namely, showing that if the RCU grace-period kthread (P1) sees another CPU as newly idle (P0), then any pre-idle memory access (P0's write to X) will be seen by any code following the grace period (P1's read from X). This is a simple release-acquire pair forcing ordering in a message-passing (MP) litmus test. Of course, if the grace-period kthread detects the CPU as non-idle, it will refrain from reporting a quiescent state on behalf of that CPU, so there are no ordering requirements from the grace-period kthread in that case. However, other subsystems call rcu_is_idle_cpu() to check for CPUs being non-idle from an RCU perspective. That case is also verified by the above litmus tests with the proviso that the sense of the low-order bit of the DYNTICKS counter be inverted. Unfortunately, on x86 smp_mb() is as expensive as a cache-local atomic increment. This commit therefore weakens only the read from ->dynticks. However, the updates are abstracted into a rcu_dynticks_inc() function to ease any future changes that might be needed. [ paulmck: Apply Linus Torvalds feedback. ] Link: https://lore.kernel.org/lkml/20210721202127.2129660-4-paulmck@kernel.org/ Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Acked-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2021-05-20 08:25:42 +08:00
int seq;
raw_cpu_write(rcu_data.rcu_need_heavy_qs, false);
seq = ct_state_inc(2 * RCU_DYNTICKS_IDX);
/* It is illegal to call this from idle state. */
WARN_ON_ONCE(!(seq & RCU_DYNTICKS_IDX));
rcu: Defer reporting RCU-preempt quiescent states when disabled This commit defers reporting of RCU-preempt quiescent states at rcu_read_unlock_special() time when any of interrupts, softirq, or preemption are disabled. These deferred quiescent states are reported at a later RCU_SOFTIRQ, context switch, idle entry, or CPU-hotplug offline operation. Of course, if another RCU read-side critical section has started in the meantime, the reporting of the quiescent state will be further deferred. This also means that disabling preemption, interrupts, and/or softirqs will act as an RCU-preempt read-side critical section. This is enforced by checking preempt_count() as needed. Some special cases must be handled on an ad-hoc basis, for example, context switch is a quiescent state even though both the scheduler and do_exit() disable preemption. In these cases, additional calls to rcu_preempt_deferred_qs() override the preemption disabling. Similar logic overrides disabled interrupts in rcu_preempt_check_callbacks() because in this case the quiescent state happened just before the corresponding scheduling-clock interrupt. In theory, this change lifts a long-standing restriction that required that if interrupts were disabled across a call to rcu_read_unlock() that the matching rcu_read_lock() also be contained within that interrupts-disabled region of code. Because the reporting of the corresponding RCU-preempt quiescent state is now deferred until after interrupts have been enabled, it is no longer possible for this situation to result in deadlocks involving the scheduler's runqueue and priority-inheritance locks. This may allow some code simplification that might reduce interrupt latency a bit. Unfortunately, in practice this would also defer deboosting a low-priority task that had been subjected to RCU priority boosting, so real-time-response considerations might well force this restriction to remain in place. Because RCU-preempt grace periods are now blocked not only by RCU read-side critical sections, but also by disabling of interrupts, preemption, and softirqs, it will be possible to eliminate RCU-bh and RCU-sched in favor of RCU-preempt in CONFIG_PREEMPT=y kernels. This may require some additional plumbing to provide the network denial-of-service guarantees that have been traditionally provided by RCU-bh. Once these are in place, CONFIG_PREEMPT=n kernels will be able to fold RCU-bh into RCU-sched. This would mean that all kernels would have but one flavor of RCU, which would open the door to significant code cleanup. Moving to a single flavor of RCU would also have the beneficial effect of reducing the NOCB kthreads by at least a factor of two. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Apply rcu_read_unlock_special() preempt_count() feedback from Joel Fernandes. ] [ paulmck: Adjust rcu_eqs_enter() call to rcu_preempt_deferred_qs() in response to bug reports from kbuild test robot. ] [ paulmck: Fix bug located by kbuild test robot involving recursion via rcu_preempt_deferred_qs(). ]
2018-06-22 03:50:01 +08:00
rcu_preempt_deferred_qs(current);
rcu: Reduce overhead of cond_resched() checks for RCU Commit ac1bea85781e (Make cond_resched() report RCU quiescent states) fixed a problem where a CPU looping in the kernel with but one runnable task would give RCU CPU stall warnings, even if the in-kernel loop contained cond_resched() calls. Unfortunately, in so doing, it introduced performance regressions in Anton Blanchard's will-it-scale "open1" test. The problem appears to be not so much the increased cond_resched() path length as an increase in the rate at which grace periods complete, which increased per-update grace-period overhead. This commit takes a different approach to fixing this bug, mainly by moving the RCU-visible quiescent state from cond_resched() to rcu_note_context_switch(), and by further reducing the check to a simple non-zero test of a single per-CPU variable. However, this approach requires that the force-quiescent-state processing send resched IPIs to the offending CPUs. These will be sent only once the grace period has reached an age specified by the boot/sysfs parameter rcutree.jiffies_till_sched_qs, or once the grace period reaches an age halfway to the point at which RCU CPU stall warnings will be emitted, whichever comes first. Reported-by: Dave Hansen <dave.hansen@intel.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Christoph Lameter <cl@gentwo.org> Cc: Mike Galbraith <umgwanakikbuti@gmail.com> Cc: Eric Dumazet <eric.dumazet@gmail.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org> [ paulmck: Made rcu_momentary_dyntick_idle() as suggested by the ktest build robot. Also fixed smp_mb() comment as noted by Oleg Nesterov. ] Merge with e552592e (Reduce overhead of cond_resched() checks for RCU) Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2014-06-21 07:49:01 +08:00
}
EXPORT_SYMBOL_GPL(rcu_momentary_dyntick_idle);
rcu: Reduce overhead of cond_resched() checks for RCU Commit ac1bea85781e (Make cond_resched() report RCU quiescent states) fixed a problem where a CPU looping in the kernel with but one runnable task would give RCU CPU stall warnings, even if the in-kernel loop contained cond_resched() calls. Unfortunately, in so doing, it introduced performance regressions in Anton Blanchard's will-it-scale "open1" test. The problem appears to be not so much the increased cond_resched() path length as an increase in the rate at which grace periods complete, which increased per-update grace-period overhead. This commit takes a different approach to fixing this bug, mainly by moving the RCU-visible quiescent state from cond_resched() to rcu_note_context_switch(), and by further reducing the check to a simple non-zero test of a single per-CPU variable. However, this approach requires that the force-quiescent-state processing send resched IPIs to the offending CPUs. These will be sent only once the grace period has reached an age specified by the boot/sysfs parameter rcutree.jiffies_till_sched_qs, or once the grace period reaches an age halfway to the point at which RCU CPU stall warnings will be emitted, whichever comes first. Reported-by: Dave Hansen <dave.hansen@intel.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Christoph Lameter <cl@gentwo.org> Cc: Mike Galbraith <umgwanakikbuti@gmail.com> Cc: Eric Dumazet <eric.dumazet@gmail.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org> [ paulmck: Made rcu_momentary_dyntick_idle() as suggested by the ktest build robot. Also fixed smp_mb() comment as noted by Oleg Nesterov. ] Merge with e552592e (Reduce overhead of cond_resched() checks for RCU) Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2014-06-21 07:49:01 +08:00
/**
* rcu_is_cpu_rrupt_from_idle - see if 'interrupted' from idle
rcu: Don't disable preemption for Tiny and Tree RCU readers Because preempt_disable() maps to barrier() for non-debug builds, it forces the compiler to spill and reload registers. Because Tree RCU and Tiny RCU now only appear in CONFIG_PREEMPT=n builds, these barrier() instances generate needless extra code for each instance of rcu_read_lock() and rcu_read_unlock(). This extra code slows down Tree RCU and bloats Tiny RCU. This commit therefore removes the preempt_disable() and preempt_enable() from the non-preemptible implementations of __rcu_read_lock() and __rcu_read_unlock(), respectively. However, for debug purposes, preempt_disable() and preempt_enable() are still invoked if CONFIG_PREEMPT_COUNT=y, because this allows detection of sleeping inside atomic sections in non-preemptible kernels. However, Tiny and Tree RCU operates by coalescing all RCU read-side critical sections on a given CPU that lie between successive quiescent states. It is therefore necessary to compensate for removing barriers from __rcu_read_lock() and __rcu_read_unlock() by adding them to a couple of the RCU functions invoked during quiescent states, namely to rcu_all_qs() and rcu_note_context_switch(). However, note that the latter is more paranoia than necessity, at least until link-time optimizations become more aggressive. This is based on an earlier patch by Paul E. McKenney, fixing a bug encountered in kernels built with CONFIG_PREEMPT=n and CONFIG_PREEMPT_COUNT=y. Signed-off-by: Boqun Feng <boqun.feng@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2015-07-31 07:55:38 +08:00
*
rcu: Add checks for dynticks counters in rcu_is_cpu_rrupt_from_idle() It would be good to combine the dynticks and dynticks_nesting counters in order to simplify the code. Unfortunately, there are concerns about usermode upcalls appearing to RCU as half of an interrupt, as Byungchul learned [1]. The "half" in "half interrupt" is due to an unpaired rcu_irq_enter(): Normally, each rcu_irq_enter() has a later call to rcu_irq_exit(). Out of an abundance of caution, Paul added warnings [2] in the RCU code which if not fired by 2021 will be interpreted as meaning that this half-interrupt scenario cannot happen any more, thus permitting simplification of this code. In the meantime, this commit makes the following changes: (1) Combining these two counters requires that rcu_rrupt_from_idle() is invoked only from hard-interrupt contexts as discussed here [3]. This commit therefore adds the required lockdep_assert_in_irq() to check this constraint. (2) Furthermore, rcu_rrupt_from_idle() is not explicit about how it is using the counters which can lead to weird future bugs. This commit therefore adds comments indicating the meaning and use of each counter. (3) Lastly, this commit checks for counter underflows as another check that half interrupts don't occur. (Previously, the function would simply return true upon underflow.) All these checks checks are NOOPs if PROVE_LOCKING (and thus PROVE_RCU) are disabled. [1] https://lore.kernel.org/patchwork/patch/952349/ [2] Commit e11ec65cc8d6 ("rcu: Add warning to detect half-interrupts") [3] https://lore.kernel.org/lkml/20190312150514.GB249405@google.com/ Cc: byungchul.park@lge.com Cc: kernel-team@android.com Cc: rcu@vger.kernel.org Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
2019-03-27 03:24:09 +08:00
* If the current CPU is idle and running at a first-level (not nested)
* interrupt, or directly, from idle, return true.
*
* The caller must have at least disabled IRQs.
*/
static int rcu_is_cpu_rrupt_from_idle(void)
{
long nesting;
/*
* Usually called from the tick; but also used from smp_function_call()
* for expedited grace periods. This latter can result in running from
* the idle task, instead of an actual IPI.
*/
lockdep_assert_irqs_disabled();
rcu: Add checks for dynticks counters in rcu_is_cpu_rrupt_from_idle() It would be good to combine the dynticks and dynticks_nesting counters in order to simplify the code. Unfortunately, there are concerns about usermode upcalls appearing to RCU as half of an interrupt, as Byungchul learned [1]. The "half" in "half interrupt" is due to an unpaired rcu_irq_enter(): Normally, each rcu_irq_enter() has a later call to rcu_irq_exit(). Out of an abundance of caution, Paul added warnings [2] in the RCU code which if not fired by 2021 will be interpreted as meaning that this half-interrupt scenario cannot happen any more, thus permitting simplification of this code. In the meantime, this commit makes the following changes: (1) Combining these two counters requires that rcu_rrupt_from_idle() is invoked only from hard-interrupt contexts as discussed here [3]. This commit therefore adds the required lockdep_assert_in_irq() to check this constraint. (2) Furthermore, rcu_rrupt_from_idle() is not explicit about how it is using the counters which can lead to weird future bugs. This commit therefore adds comments indicating the meaning and use of each counter. (3) Lastly, this commit checks for counter underflows as another check that half interrupts don't occur. (Previously, the function would simply return true upon underflow.) All these checks checks are NOOPs if PROVE_LOCKING (and thus PROVE_RCU) are disabled. [1] https://lore.kernel.org/patchwork/patch/952349/ [2] Commit e11ec65cc8d6 ("rcu: Add warning to detect half-interrupts") [3] https://lore.kernel.org/lkml/20190312150514.GB249405@google.com/ Cc: byungchul.park@lge.com Cc: kernel-team@android.com Cc: rcu@vger.kernel.org Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
2019-03-27 03:24:09 +08:00
/* Check for counter underflows */
RCU_LOCKDEP_WARN(ct_dynticks_nesting() < 0,
rcu: Add checks for dynticks counters in rcu_is_cpu_rrupt_from_idle() It would be good to combine the dynticks and dynticks_nesting counters in order to simplify the code. Unfortunately, there are concerns about usermode upcalls appearing to RCU as half of an interrupt, as Byungchul learned [1]. The "half" in "half interrupt" is due to an unpaired rcu_irq_enter(): Normally, each rcu_irq_enter() has a later call to rcu_irq_exit(). Out of an abundance of caution, Paul added warnings [2] in the RCU code which if not fired by 2021 will be interpreted as meaning that this half-interrupt scenario cannot happen any more, thus permitting simplification of this code. In the meantime, this commit makes the following changes: (1) Combining these two counters requires that rcu_rrupt_from_idle() is invoked only from hard-interrupt contexts as discussed here [3]. This commit therefore adds the required lockdep_assert_in_irq() to check this constraint. (2) Furthermore, rcu_rrupt_from_idle() is not explicit about how it is using the counters which can lead to weird future bugs. This commit therefore adds comments indicating the meaning and use of each counter. (3) Lastly, this commit checks for counter underflows as another check that half interrupts don't occur. (Previously, the function would simply return true upon underflow.) All these checks checks are NOOPs if PROVE_LOCKING (and thus PROVE_RCU) are disabled. [1] https://lore.kernel.org/patchwork/patch/952349/ [2] Commit e11ec65cc8d6 ("rcu: Add warning to detect half-interrupts") [3] https://lore.kernel.org/lkml/20190312150514.GB249405@google.com/ Cc: byungchul.park@lge.com Cc: kernel-team@android.com Cc: rcu@vger.kernel.org Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
2019-03-27 03:24:09 +08:00
"RCU dynticks_nesting counter underflow!");
RCU_LOCKDEP_WARN(ct_dynticks_nmi_nesting() <= 0,
rcu: Add checks for dynticks counters in rcu_is_cpu_rrupt_from_idle() It would be good to combine the dynticks and dynticks_nesting counters in order to simplify the code. Unfortunately, there are concerns about usermode upcalls appearing to RCU as half of an interrupt, as Byungchul learned [1]. The "half" in "half interrupt" is due to an unpaired rcu_irq_enter(): Normally, each rcu_irq_enter() has a later call to rcu_irq_exit(). Out of an abundance of caution, Paul added warnings [2] in the RCU code which if not fired by 2021 will be interpreted as meaning that this half-interrupt scenario cannot happen any more, thus permitting simplification of this code. In the meantime, this commit makes the following changes: (1) Combining these two counters requires that rcu_rrupt_from_idle() is invoked only from hard-interrupt contexts as discussed here [3]. This commit therefore adds the required lockdep_assert_in_irq() to check this constraint. (2) Furthermore, rcu_rrupt_from_idle() is not explicit about how it is using the counters which can lead to weird future bugs. This commit therefore adds comments indicating the meaning and use of each counter. (3) Lastly, this commit checks for counter underflows as another check that half interrupts don't occur. (Previously, the function would simply return true upon underflow.) All these checks checks are NOOPs if PROVE_LOCKING (and thus PROVE_RCU) are disabled. [1] https://lore.kernel.org/patchwork/patch/952349/ [2] Commit e11ec65cc8d6 ("rcu: Add warning to detect half-interrupts") [3] https://lore.kernel.org/lkml/20190312150514.GB249405@google.com/ Cc: byungchul.park@lge.com Cc: kernel-team@android.com Cc: rcu@vger.kernel.org Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
2019-03-27 03:24:09 +08:00
"RCU dynticks_nmi_nesting counter underflow/zero!");
/* Are we at first interrupt nesting level? */
nesting = ct_dynticks_nmi_nesting();
if (nesting > 1)
rcu: Add checks for dynticks counters in rcu_is_cpu_rrupt_from_idle() It would be good to combine the dynticks and dynticks_nesting counters in order to simplify the code. Unfortunately, there are concerns about usermode upcalls appearing to RCU as half of an interrupt, as Byungchul learned [1]. The "half" in "half interrupt" is due to an unpaired rcu_irq_enter(): Normally, each rcu_irq_enter() has a later call to rcu_irq_exit(). Out of an abundance of caution, Paul added warnings [2] in the RCU code which if not fired by 2021 will be interpreted as meaning that this half-interrupt scenario cannot happen any more, thus permitting simplification of this code. In the meantime, this commit makes the following changes: (1) Combining these two counters requires that rcu_rrupt_from_idle() is invoked only from hard-interrupt contexts as discussed here [3]. This commit therefore adds the required lockdep_assert_in_irq() to check this constraint. (2) Furthermore, rcu_rrupt_from_idle() is not explicit about how it is using the counters which can lead to weird future bugs. This commit therefore adds comments indicating the meaning and use of each counter. (3) Lastly, this commit checks for counter underflows as another check that half interrupts don't occur. (Previously, the function would simply return true upon underflow.) All these checks checks are NOOPs if PROVE_LOCKING (and thus PROVE_RCU) are disabled. [1] https://lore.kernel.org/patchwork/patch/952349/ [2] Commit e11ec65cc8d6 ("rcu: Add warning to detect half-interrupts") [3] https://lore.kernel.org/lkml/20190312150514.GB249405@google.com/ Cc: byungchul.park@lge.com Cc: kernel-team@android.com Cc: rcu@vger.kernel.org Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
2019-03-27 03:24:09 +08:00
return false;
/*
* If we're not in an interrupt, we must be in the idle task!
*/
WARN_ON_ONCE(!nesting && !is_idle_task(current));
rcu: Add checks for dynticks counters in rcu_is_cpu_rrupt_from_idle() It would be good to combine the dynticks and dynticks_nesting counters in order to simplify the code. Unfortunately, there are concerns about usermode upcalls appearing to RCU as half of an interrupt, as Byungchul learned [1]. The "half" in "half interrupt" is due to an unpaired rcu_irq_enter(): Normally, each rcu_irq_enter() has a later call to rcu_irq_exit(). Out of an abundance of caution, Paul added warnings [2] in the RCU code which if not fired by 2021 will be interpreted as meaning that this half-interrupt scenario cannot happen any more, thus permitting simplification of this code. In the meantime, this commit makes the following changes: (1) Combining these two counters requires that rcu_rrupt_from_idle() is invoked only from hard-interrupt contexts as discussed here [3]. This commit therefore adds the required lockdep_assert_in_irq() to check this constraint. (2) Furthermore, rcu_rrupt_from_idle() is not explicit about how it is using the counters which can lead to weird future bugs. This commit therefore adds comments indicating the meaning and use of each counter. (3) Lastly, this commit checks for counter underflows as another check that half interrupts don't occur. (Previously, the function would simply return true upon underflow.) All these checks checks are NOOPs if PROVE_LOCKING (and thus PROVE_RCU) are disabled. [1] https://lore.kernel.org/patchwork/patch/952349/ [2] Commit e11ec65cc8d6 ("rcu: Add warning to detect half-interrupts") [3] https://lore.kernel.org/lkml/20190312150514.GB249405@google.com/ Cc: byungchul.park@lge.com Cc: kernel-team@android.com Cc: rcu@vger.kernel.org Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
2019-03-27 03:24:09 +08:00
/* Does CPU appear to be idle from an RCU standpoint? */
return ct_dynticks_nesting() == 0;
}
#define DEFAULT_RCU_BLIMIT (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ? 1000 : 10)
// Maximum callbacks per rcu_do_batch ...
#define DEFAULT_MAX_RCU_BLIMIT 10000 // ... even during callback flood.
static long blimit = DEFAULT_RCU_BLIMIT;
#define DEFAULT_RCU_QHIMARK 10000 // If this many pending, ignore blimit.
static long qhimark = DEFAULT_RCU_QHIMARK;
#define DEFAULT_RCU_QLOMARK 100 // Once only this many pending, use blimit.
static long qlowmark = DEFAULT_RCU_QLOMARK;
#define DEFAULT_RCU_QOVLD_MULT 2
#define DEFAULT_RCU_QOVLD (DEFAULT_RCU_QOVLD_MULT * DEFAULT_RCU_QHIMARK)
static long qovld = DEFAULT_RCU_QOVLD; // If this many pending, hammer QS.
static long qovld_calc = -1; // No pre-initialization lock acquisitions!
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
module_param(blimit, long, 0444);
module_param(qhimark, long, 0444);
module_param(qlowmark, long, 0444);
module_param(qovld, long, 0444);
static ulong jiffies_till_first_fqs = IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ? 0 : ULONG_MAX;
static ulong jiffies_till_next_fqs = ULONG_MAX;
static bool rcu_kick_kthreads;
static int rcu_divisor = 7;
module_param(rcu_divisor, int, 0644);
/* Force an exit from rcu_do_batch() after 3 milliseconds. */
static long rcu_resched_ns = 3 * NSEC_PER_MSEC;
module_param(rcu_resched_ns, long, 0644);
rcu: Control grace-period duration from sysfs Although almost everyone is well-served by the defaults, some uses of RCU benefit from shorter grace periods, while others benefit more from the greater efficiency provided by longer grace periods. Situations requiring a large number of grace periods to elapse (and wireshark startup has been called out as an example of this) are helped by lower-latency grace periods. Furthermore, in some embedded applications, people are willing to accept a small degradation in update efficiency (due to there being more of the shorter grace-period operations) in order to gain the lower latency. In contrast, those few systems with thousands of CPUs need longer grace periods because the CPU overhead of a grace period rises roughly linearly with the number of CPUs. Such systems normally do not make much use of facilities that require large numbers of grace periods to elapse, so this is a good tradeoff. Therefore, this commit allows the durations to be controlled from sysfs. There are two sysfs parameters, one named "jiffies_till_first_fqs" that specifies the delay in jiffies from the end of grace-period initialization until the first attempt to force quiescent states, and the other named "jiffies_till_next_fqs" that specifies the delay (again in jiffies) between subsequent attempts to force quiescent states. They both default to three jiffies, which is compatible with the old hard-coded behavior. At some future time, it may be possible to automatically increase the grace-period length with the number of CPUs, but we do not yet have sufficient data to do a good job. Preliminary data indicates that we should add an addiitonal jiffy to each of the delays for every 200 CPUs in the system, but more experimentation is needed. For now, the number of systems with more than 1,000 CPUs is small enough that this can be relegated to boot-time hand tuning. Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org>
2012-06-27 11:45:57 +08:00
/*
* How long the grace period must be before we start recruiting
* quiescent-state help from rcu_note_context_switch().
*/
static ulong jiffies_till_sched_qs = ULONG_MAX;
module_param(jiffies_till_sched_qs, ulong, 0444);
static ulong jiffies_to_sched_qs; /* See adjust_jiffies_till_sched_qs(). */
module_param(jiffies_to_sched_qs, ulong, 0444); /* Display only! */
/*
* Make sure that we give the grace-period kthread time to detect any
* idle CPUs before taking active measures to force quiescent states.
* However, don't go below 100 milliseconds, adjusted upwards for really
* large systems.
*/
static void adjust_jiffies_till_sched_qs(void)
{
unsigned long j;
/* If jiffies_till_sched_qs was specified, respect the request. */
if (jiffies_till_sched_qs != ULONG_MAX) {
WRITE_ONCE(jiffies_to_sched_qs, jiffies_till_sched_qs);
return;
}
/* Otherwise, set to third fqs scan, but bound below on large system. */
j = READ_ONCE(jiffies_till_first_fqs) +
2 * READ_ONCE(jiffies_till_next_fqs);
if (j < HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV)
j = HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
pr_info("RCU calculated value of scheduler-enlistment delay is %ld jiffies.\n", j);
WRITE_ONCE(jiffies_to_sched_qs, j);
}
static int param_set_first_fqs_jiffies(const char *val, const struct kernel_param *kp)
{
ulong j;
int ret = kstrtoul(val, 0, &j);
if (!ret) {
WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : j);
adjust_jiffies_till_sched_qs();
}
return ret;
}
static int param_set_next_fqs_jiffies(const char *val, const struct kernel_param *kp)
{
ulong j;
int ret = kstrtoul(val, 0, &j);
if (!ret) {
WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : (j ?: 1));
adjust_jiffies_till_sched_qs();
}
return ret;
}
static const struct kernel_param_ops first_fqs_jiffies_ops = {
.set = param_set_first_fqs_jiffies,
.get = param_get_ulong,
};
static const struct kernel_param_ops next_fqs_jiffies_ops = {
.set = param_set_next_fqs_jiffies,
.get = param_get_ulong,
};
module_param_cb(jiffies_till_first_fqs, &first_fqs_jiffies_ops, &jiffies_till_first_fqs, 0644);
module_param_cb(jiffies_till_next_fqs, &next_fqs_jiffies_ops, &jiffies_till_next_fqs, 0644);
module_param(rcu_kick_kthreads, bool, 0644);
rcu: Control grace-period duration from sysfs Although almost everyone is well-served by the defaults, some uses of RCU benefit from shorter grace periods, while others benefit more from the greater efficiency provided by longer grace periods. Situations requiring a large number of grace periods to elapse (and wireshark startup has been called out as an example of this) are helped by lower-latency grace periods. Furthermore, in some embedded applications, people are willing to accept a small degradation in update efficiency (due to there being more of the shorter grace-period operations) in order to gain the lower latency. In contrast, those few systems with thousands of CPUs need longer grace periods because the CPU overhead of a grace period rises roughly linearly with the number of CPUs. Such systems normally do not make much use of facilities that require large numbers of grace periods to elapse, so this is a good tradeoff. Therefore, this commit allows the durations to be controlled from sysfs. There are two sysfs parameters, one named "jiffies_till_first_fqs" that specifies the delay in jiffies from the end of grace-period initialization until the first attempt to force quiescent states, and the other named "jiffies_till_next_fqs" that specifies the delay (again in jiffies) between subsequent attempts to force quiescent states. They both default to three jiffies, which is compatible with the old hard-coded behavior. At some future time, it may be possible to automatically increase the grace-period length with the number of CPUs, but we do not yet have sufficient data to do a good job. Preliminary data indicates that we should add an addiitonal jiffy to each of the delays for every 200 CPUs in the system, but more experimentation is needed. For now, the number of systems with more than 1,000 CPUs is small enough that this can be relegated to boot-time hand tuning. Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org>
2012-06-27 11:45:57 +08:00
static void force_qs_rnp(int (*f)(struct rcu_data *rdp));
static int rcu_pending(int user);
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/*
* Return the number of RCU GPs completed thus far for debug & stats.
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
*/
unsigned long rcu_get_gp_seq(void)
{
return READ_ONCE(rcu_state.gp_seq);
}
EXPORT_SYMBOL_GPL(rcu_get_gp_seq);
/*
* Return the number of RCU expedited batches completed thus far for
* debug & stats. Odd numbers mean that a batch is in progress, even
* numbers mean idle. The value returned will thus be roughly double
* the cumulative batches since boot.
*/
unsigned long rcu_exp_batches_completed(void)
{
return rcu_state.expedited_sequence;
}
EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
/*
* Return the root node of the rcu_state structure.
*/
static struct rcu_node *rcu_get_root(void)
{
return &rcu_state.node[0];
}
/*
* Send along grace-period-related data for rcutorture diagnostics.
*/
void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
unsigned long *gp_seq)
{
switch (test_type) {
case RCU_FLAVOR:
*flags = READ_ONCE(rcu_state.gp_flags);
*gp_seq = rcu_seq_current(&rcu_state.gp_seq);
break;
default:
break;
}
}
EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
#if defined(CONFIG_NO_HZ_FULL) && (!defined(CONFIG_GENERIC_ENTRY) || !defined(CONFIG_KVM_XFER_TO_GUEST_WORK))
/*
* An empty function that will trigger a reschedule on
* IRQ tail once IRQs get re-enabled on userspace/guest resume.
*/
static void late_wakeup_func(struct irq_work *work)
{
}
static DEFINE_PER_CPU(struct irq_work, late_wakeup_work) =
IRQ_WORK_INIT(late_wakeup_func);
/*
* If either:
*
* 1) the task is about to enter in guest mode and $ARCH doesn't support KVM generic work
* 2) the task is about to enter in user mode and $ARCH doesn't support generic entry.
*
* In these cases the late RCU wake ups aren't supported in the resched loops and our
* last resort is to fire a local irq_work that will trigger a reschedule once IRQs
* get re-enabled again.
*/
noinstr void rcu_irq_work_resched(void)
{
struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
if (IS_ENABLED(CONFIG_GENERIC_ENTRY) && !(current->flags & PF_VCPU))
return;
if (IS_ENABLED(CONFIG_KVM_XFER_TO_GUEST_WORK) && (current->flags & PF_VCPU))
return;
instrumentation_begin();
if (do_nocb_deferred_wakeup(rdp) && need_resched()) {
irq_work_queue(this_cpu_ptr(&late_wakeup_work));
}
instrumentation_end();
}
#endif /* #if defined(CONFIG_NO_HZ_FULL) && (!defined(CONFIG_GENERIC_ENTRY) || !defined(CONFIG_KVM_XFER_TO_GUEST_WORK)) */
#ifdef CONFIG_PROVE_RCU
/**
* rcu_irq_exit_check_preempt - Validate that scheduling is possible
*/
void rcu_irq_exit_check_preempt(void)
{
lockdep_assert_irqs_disabled();
RCU_LOCKDEP_WARN(ct_dynticks_nesting() <= 0,
"RCU dynticks_nesting counter underflow/zero!");
RCU_LOCKDEP_WARN(ct_dynticks_nmi_nesting() !=
DYNTICK_IRQ_NONIDLE,
"Bad RCU dynticks_nmi_nesting counter\n");
RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(),
"RCU in extended quiescent state!");
}
#endif /* #ifdef CONFIG_PROVE_RCU */
#ifdef CONFIG_NO_HZ_FULL
/**
* __rcu_irq_enter_check_tick - Enable scheduler tick on CPU if RCU needs it.
*
* The scheduler tick is not normally enabled when CPUs enter the kernel
* from nohz_full userspace execution. After all, nohz_full userspace
* execution is an RCU quiescent state and the time executing in the kernel
* is quite short. Except of course when it isn't. And it is not hard to
* cause a large system to spend tens of seconds or even minutes looping
* in the kernel, which can cause a number of problems, include RCU CPU
* stall warnings.
*
* Therefore, if a nohz_full CPU fails to report a quiescent state
* in a timely manner, the RCU grace-period kthread sets that CPU's
* ->rcu_urgent_qs flag with the expectation that the next interrupt or
* exception will invoke this function, which will turn on the scheduler
* tick, which will enable RCU to detect that CPU's quiescent states,
* for example, due to cond_resched() calls in CONFIG_PREEMPT=n kernels.
* The tick will be disabled once a quiescent state is reported for
* this CPU.
*
* Of course, in carefully tuned systems, there might never be an
* interrupt or exception. In that case, the RCU grace-period kthread
* will eventually cause one to happen. However, in less carefully
* controlled environments, this function allows RCU to get what it
* needs without creating otherwise useless interruptions.
*/
void __rcu_irq_enter_check_tick(void)
{
struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
// If we're here from NMI there's nothing to do.
if (in_nmi())
return;
RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(),
"Illegal rcu_irq_enter_check_tick() from extended quiescent state");
if (!tick_nohz_full_cpu(rdp->cpu) ||
!READ_ONCE(rdp->rcu_urgent_qs) ||
READ_ONCE(rdp->rcu_forced_tick)) {
// RCU doesn't need nohz_full help from this CPU, or it is
// already getting that help.
return;
}
// We get here only when not in an extended quiescent state and
// from interrupts (as opposed to NMIs). Therefore, (1) RCU is
// already watching and (2) The fact that we are in an interrupt
// handler and that the rcu_node lock is an irq-disabled lock
// prevents self-deadlock. So we can safely recheck under the lock.
// Note that the nohz_full state currently cannot change.
raw_spin_lock_rcu_node(rdp->mynode);
if (READ_ONCE(rdp->rcu_urgent_qs) && !rdp->rcu_forced_tick) {
// A nohz_full CPU is in the kernel and RCU needs a
// quiescent state. Turn on the tick!
WRITE_ONCE(rdp->rcu_forced_tick, true);
tick_dep_set_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
}
raw_spin_unlock_rcu_node(rdp->mynode);
}
rcu: Avoid stack overflow due to __rcu_irq_enter_check_tick() being kprobe-ed Registering a kprobe on __rcu_irq_enter_check_tick() can cause kernel stack overflow as shown below. This issue can be reproduced by enabling CONFIG_NO_HZ_FULL and booting the kernel with argument "nohz_full=", and then giving the following commands at the shell prompt: # cd /sys/kernel/tracing/ # echo 'p:mp1 __rcu_irq_enter_check_tick' >> kprobe_events # echo 1 > events/kprobes/enable This commit therefore adds __rcu_irq_enter_check_tick() to the kprobes blacklist using NOKPROBE_SYMBOL(). Insufficient stack space to handle exception! ESR: 0x00000000f2000004 -- BRK (AArch64) FAR: 0x0000ffffccf3e510 Task stack: [0xffff80000ad30000..0xffff80000ad38000] IRQ stack: [0xffff800008050000..0xffff800008058000] Overflow stack: [0xffff089c36f9f310..0xffff089c36fa0310] CPU: 5 PID: 190 Comm: bash Not tainted 6.2.0-rc2-00320-g1f5abbd77e2c #19 Hardware name: linux,dummy-virt (DT) pstate: 400003c5 (nZcv DAIF -PAN -UAO -TCO -DIT -SSBS BTYPE=--) pc : __rcu_irq_enter_check_tick+0x0/0x1b8 lr : ct_nmi_enter+0x11c/0x138 sp : ffff80000ad30080 x29: ffff80000ad30080 x28: ffff089c82e20000 x27: 0000000000000000 x26: 0000000000000000 x25: ffff089c02a8d100 x24: 0000000000000000 x23: 00000000400003c5 x22: 0000ffffccf3e510 x21: ffff089c36fae148 x20: ffff80000ad30120 x19: ffffa8da8fcce148 x18: 0000000000000000 x17: 0000000000000000 x16: 0000000000000000 x15: ffffa8da8e44ea6c x14: ffffa8da8e44e968 x13: ffffa8da8e03136c x12: 1fffe113804d6809 x11: ffff6113804d6809 x10: 0000000000000a60 x9 : dfff800000000000 x8 : ffff089c026b404f x7 : 00009eec7fb297f7 x6 : 0000000000000001 x5 : ffff80000ad30120 x4 : dfff800000000000 x3 : ffffa8da8e3016f4 x2 : 0000000000000003 x1 : 0000000000000000 x0 : 0000000000000000 Kernel panic - not syncing: kernel stack overflow CPU: 5 PID: 190 Comm: bash Not tainted 6.2.0-rc2-00320-g1f5abbd77e2c #19 Hardware name: linux,dummy-virt (DT) Call trace: dump_backtrace+0xf8/0x108 show_stack+0x20/0x30 dump_stack_lvl+0x68/0x84 dump_stack+0x1c/0x38 panic+0x214/0x404 add_taint+0x0/0xf8 panic_bad_stack+0x144/0x160 handle_bad_stack+0x38/0x58 __bad_stack+0x78/0x7c __rcu_irq_enter_check_tick+0x0/0x1b8 arm64_enter_el1_dbg.isra.0+0x14/0x20 el1_dbg+0x2c/0x90 el1h_64_sync_handler+0xcc/0xe8 el1h_64_sync+0x64/0x68 __rcu_irq_enter_check_tick+0x0/0x1b8 arm64_enter_el1_dbg.isra.0+0x14/0x20 el1_dbg+0x2c/0x90 el1h_64_sync_handler+0xcc/0xe8 el1h_64_sync+0x64/0x68 __rcu_irq_enter_check_tick+0x0/0x1b8 arm64_enter_el1_dbg.isra.0+0x14/0x20 el1_dbg+0x2c/0x90 el1h_64_sync_handler+0xcc/0xe8 el1h_64_sync+0x64/0x68 __rcu_irq_enter_check_tick+0x0/0x1b8 [...] el1_dbg+0x2c/0x90 el1h_64_sync_handler+0xcc/0xe8 el1h_64_sync+0x64/0x68 __rcu_irq_enter_check_tick+0x0/0x1b8 arm64_enter_el1_dbg.isra.0+0x14/0x20 el1_dbg+0x2c/0x90 el1h_64_sync_handler+0xcc/0xe8 el1h_64_sync+0x64/0x68 __rcu_irq_enter_check_tick+0x0/0x1b8 arm64_enter_el1_dbg.isra.0+0x14/0x20 el1_dbg+0x2c/0x90 el1h_64_sync_handler+0xcc/0xe8 el1h_64_sync+0x64/0x68 __rcu_irq_enter_check_tick+0x0/0x1b8 el1_interrupt+0x28/0x60 el1h_64_irq_handler+0x18/0x28 el1h_64_irq+0x64/0x68 __ftrace_set_clr_event_nolock+0x98/0x198 __ftrace_set_clr_event+0x58/0x80 system_enable_write+0x144/0x178 vfs_write+0x174/0x738 ksys_write+0xd0/0x188 __arm64_sys_write+0x4c/0x60 invoke_syscall+0x64/0x180 el0_svc_common.constprop.0+0x84/0x160 do_el0_svc+0x48/0xe8 el0_svc+0x34/0xd0 el0t_64_sync_handler+0xb8/0xc0 el0t_64_sync+0x190/0x194 SMP: stopping secondary CPUs Kernel Offset: 0x28da86000000 from 0xffff800008000000 PHYS_OFFSET: 0xfffff76600000000 CPU features: 0x00000,01a00100,0000421b Memory Limit: none Acked-by: Joel Fernandes (Google) <joel@joelfernandes.org> Link: https://lore.kernel.org/all/20221119040049.795065-1-zhengyejian1@huawei.com/ Fixes: aaf2bc50df1f ("rcu: Abstract out rcu_irq_enter_check_tick() from rcu_nmi_enter()") Signed-off-by: Zheng Yejian <zhengyejian1@huawei.com> Cc: stable@vger.kernel.org Signed-off-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org>
2023-01-06 15:09:34 +08:00
NOKPROBE_SYMBOL(__rcu_irq_enter_check_tick);
#endif /* CONFIG_NO_HZ_FULL */
/*
* Check to see if any future non-offloaded RCU-related work will need
* to be done by the current CPU, even if none need be done immediately,
* returning 1 if so. This function is part of the RCU implementation;
* it is -not- an exported member of the RCU API. This is used by
* the idle-entry code to figure out whether it is safe to disable the
* scheduler-clock interrupt.
*
* Just check whether or not this CPU has non-offloaded RCU callbacks
* queued.
*/
int rcu_needs_cpu(void)
{
return !rcu_segcblist_empty(&this_cpu_ptr(&rcu_data)->cblist) &&
!rcu_rdp_is_offloaded(this_cpu_ptr(&rcu_data));
}
/*
* If any sort of urgency was applied to the current CPU (for example,
* the scheduler-clock interrupt was enabled on a nohz_full CPU) in order
* to get to a quiescent state, disable it.
*/
static void rcu_disable_urgency_upon_qs(struct rcu_data *rdp)
{
raw_lockdep_assert_held_rcu_node(rdp->mynode);
WRITE_ONCE(rdp->rcu_urgent_qs, false);
WRITE_ONCE(rdp->rcu_need_heavy_qs, false);
if (tick_nohz_full_cpu(rdp->cpu) && rdp->rcu_forced_tick) {
tick_dep_clear_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
WRITE_ONCE(rdp->rcu_forced_tick, false);
}
}
/**
* rcu_is_watching - RCU read-side critical sections permitted on current CPU?
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
*
* Return @true if RCU is watching the running CPU and @false otherwise.
* An @true return means that this CPU can safely enter RCU read-side
* critical sections.
*
* Although calls to rcu_is_watching() from most parts of the kernel
* will return @true, there are important exceptions. For example, if the
* current CPU is deep within its idle loop, in kernel entry/exit code,
* or offline, rcu_is_watching() will return @false.
*
* Make notrace because it can be called by the internal functions of
* ftrace, and making this notrace removes unnecessary recursion calls.
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
*/
notrace bool rcu_is_watching(void)
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
{
bool ret;
preempt_disable_notrace();
ret = !rcu_dynticks_curr_cpu_in_eqs();
preempt_enable_notrace();
return ret;
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
}
EXPORT_SYMBOL_GPL(rcu_is_watching);
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/*
* If a holdout task is actually running, request an urgent quiescent
* state from its CPU. This is unsynchronized, so migrations can cause
* the request to go to the wrong CPU. Which is OK, all that will happen
* is that the CPU's next context switch will be a bit slower and next
* time around this task will generate another request.
*/
void rcu_request_urgent_qs_task(struct task_struct *t)
{
int cpu;
barrier();
cpu = task_cpu(t);
if (!task_curr(t))
return; /* This task is not running on that CPU. */
smp_store_release(per_cpu_ptr(&rcu_data.rcu_urgent_qs, cpu), true);
}
/*
* When trying to report a quiescent state on behalf of some other CPU,
* it is our responsibility to check for and handle potential overflow
* of the rcu_node ->gp_seq counter with respect to the rcu_data counters.
* After all, the CPU might be in deep idle state, and thus executing no
* code whatsoever.
*/
static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp)
{
raw_lockdep_assert_held_rcu_node(rnp);
if (ULONG_CMP_LT(rcu_seq_current(&rdp->gp_seq) + ULONG_MAX / 4,
rnp->gp_seq))
WRITE_ONCE(rdp->gpwrap, true);
if (ULONG_CMP_LT(rdp->rcu_iw_gp_seq + ULONG_MAX / 4, rnp->gp_seq))
rdp->rcu_iw_gp_seq = rnp->gp_seq + ULONG_MAX / 4;
}
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/*
* Snapshot the specified CPU's dynticks counter so that we can later
* credit them with an implicit quiescent state. Return 1 if this CPU
* is in dynticks idle mode, which is an extended quiescent state.
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
*/
static int dyntick_save_progress_counter(struct rcu_data *rdp)
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
{
rdp->dynticks_snap = rcu_dynticks_snap(rdp->cpu);
if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
rcu_gpnum_ovf(rdp->mynode, rdp);
return 1;
}
return 0;
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
}
/*
* Returns positive if the specified CPU has passed through a quiescent state
* by virtue of being in or having passed through an dynticks idle state since
* the last call to dyntick_save_progress_counter() for this same CPU, or by
* virtue of having been offline.
*
* Returns negative if the specified CPU needs a force resched.
*
* Returns zero otherwise.
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
*/
static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
{
unsigned long jtsq;
int ret = 0;
struct rcu_node *rnp = rdp->mynode;
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/*
* If the CPU passed through or entered a dynticks idle phase with
* no active irq/NMI handlers, then we can safely pretend that the CPU
* already acknowledged the request to pass through a quiescent
* state. Either way, that CPU cannot possibly be in an RCU
* read-side critical section that started before the beginning
* of the current RCU grace period.
*/
if (rcu_dynticks_in_eqs_since(rdp, rdp->dynticks_snap)) {
trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
rcu_gpnum_ovf(rnp, rdp);
return 1;
}
/*
* Complain if a CPU that is considered to be offline from RCU's
* perspective has not yet reported a quiescent state. After all,
* the offline CPU should have reported a quiescent state during
* the CPU-offline process, or, failing that, by rcu_gp_init()
* if it ran concurrently with either the CPU going offline or the
* last task on a leaf rcu_node structure exiting its RCU read-side
* critical section while all CPUs corresponding to that structure
* are offline. This added warning detects bugs in any of these
* code paths.
*
* The rcu_node structure's ->lock is held here, which excludes
* the relevant portions the CPU-hotplug code, the grace-period
* initialization code, and the rcu_read_unlock() code paths.
*
* For more detail, please refer to the "Hotplug CPU" section
* of RCU's Requirements documentation.
*/
if (WARN_ON_ONCE(!rcu_rdp_cpu_online(rdp))) {
struct rcu_node *rnp1;
pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
__func__, rnp->grplo, rnp->grphi, rnp->level,
(long)rnp->gp_seq, (long)rnp->completedqs);
for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx ->rcu_gp_init_mask %#lx\n",
__func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext, rnp1->rcu_gp_init_mask);
pr_info("%s %d: %c online: %ld(%d) offline: %ld(%d)\n",
__func__, rdp->cpu, ".o"[rcu_rdp_cpu_online(rdp)],
(long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
(long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
return 1; /* Break things loose after complaining. */
}
rcu: Kick adaptive-ticks CPUs that are holding up RCU grace periods Adaptive-ticks CPUs inform RCU when they enter kernel mode, but they do not necessarily turn the scheduler-clock tick back on. This state of affairs could result in RCU waiting on an adaptive-ticks CPU running for an extended period in kernel mode. Such a CPU will never run the RCU state machine, and could therefore indefinitely extend the RCU state machine, sooner or later resulting in an OOM condition. This patch, inspired by an earlier patch by Frederic Weisbecker, therefore causes RCU's force-quiescent-state processing to check for this condition and to send an IPI to CPUs that remain in that state for too long. "Too long" currently means about three jiffies by default, which is quite some time for a CPU to remain in the kernel without blocking. The rcu_tree.jiffies_till_first_fqs and rcutree.jiffies_till_next_fqs sysfs variables may be used to tune "too long" if needed. Reported-by: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org> Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com> Cc: Chris Metcalf <cmetcalf@tilera.com> Cc: Christoph Lameter <cl@linux.com> Cc: Geoff Levand <geoff@infradead.org> Cc: Gilad Ben Yossef <gilad@benyossef.com> Cc: Hakan Akkan <hakanakkan@gmail.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Kevin Hilman <khilman@linaro.org> Cc: Li Zhong <zhong@linux.vnet.ibm.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de>
2013-04-13 07:19:10 +08:00
/*
rcu: Reduce overhead of cond_resched() checks for RCU Commit ac1bea85781e (Make cond_resched() report RCU quiescent states) fixed a problem where a CPU looping in the kernel with but one runnable task would give RCU CPU stall warnings, even if the in-kernel loop contained cond_resched() calls. Unfortunately, in so doing, it introduced performance regressions in Anton Blanchard's will-it-scale "open1" test. The problem appears to be not so much the increased cond_resched() path length as an increase in the rate at which grace periods complete, which increased per-update grace-period overhead. This commit takes a different approach to fixing this bug, mainly by moving the RCU-visible quiescent state from cond_resched() to rcu_note_context_switch(), and by further reducing the check to a simple non-zero test of a single per-CPU variable. However, this approach requires that the force-quiescent-state processing send resched IPIs to the offending CPUs. These will be sent only once the grace period has reached an age specified by the boot/sysfs parameter rcutree.jiffies_till_sched_qs, or once the grace period reaches an age halfway to the point at which RCU CPU stall warnings will be emitted, whichever comes first. Reported-by: Dave Hansen <dave.hansen@intel.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Christoph Lameter <cl@gentwo.org> Cc: Mike Galbraith <umgwanakikbuti@gmail.com> Cc: Eric Dumazet <eric.dumazet@gmail.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org> [ paulmck: Made rcu_momentary_dyntick_idle() as suggested by the ktest build robot. Also fixed smp_mb() comment as noted by Oleg Nesterov. ] Merge with e552592e (Reduce overhead of cond_resched() checks for RCU) Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2014-06-21 07:49:01 +08:00
* A CPU running for an extended time within the kernel can
* delay RCU grace periods: (1) At age jiffies_to_sched_qs,
* set .rcu_urgent_qs, (2) At age 2*jiffies_to_sched_qs, set
* both .rcu_need_heavy_qs and .rcu_urgent_qs. Note that the
* unsynchronized assignments to the per-CPU rcu_need_heavy_qs
* variable are safe because the assignments are repeated if this
* CPU failed to pass through a quiescent state. This code
* also checks .jiffies_resched in case jiffies_to_sched_qs
* is set way high.
*/
jtsq = READ_ONCE(jiffies_to_sched_qs);
if (!READ_ONCE(rdp->rcu_need_heavy_qs) &&
(time_after(jiffies, rcu_state.gp_start + jtsq * 2) ||
time_after(jiffies, rcu_state.jiffies_resched) ||
rcu_state.cbovld)) {
WRITE_ONCE(rdp->rcu_need_heavy_qs, true);
/* Store rcu_need_heavy_qs before rcu_urgent_qs. */
smp_store_release(&rdp->rcu_urgent_qs, true);
} else if (time_after(jiffies, rcu_state.gp_start + jtsq)) {
WRITE_ONCE(rdp->rcu_urgent_qs, true);
}
/*
* NO_HZ_FULL CPUs can run in-kernel without rcu_sched_clock_irq!
* The above code handles this, but only for straight cond_resched().
* And some in-kernel loops check need_resched() before calling
* cond_resched(), which defeats the above code for CPUs that are
* running in-kernel with scheduling-clock interrupts disabled.
* So hit them over the head with the resched_cpu() hammer!
*/
if (tick_nohz_full_cpu(rdp->cpu) &&
(time_after(jiffies, READ_ONCE(rdp->last_fqs_resched) + jtsq * 3) ||
rcu_state.cbovld)) {
WRITE_ONCE(rdp->rcu_urgent_qs, true);
WRITE_ONCE(rdp->last_fqs_resched, jiffies);
ret = -1;
}
/*
* If more than halfway to RCU CPU stall-warning time, invoke
* resched_cpu() more frequently to try to loosen things up a bit.
* Also check to see if the CPU is getting hammered with interrupts,
* but only once per grace period, just to keep the IPIs down to
* a dull roar.
*/
if (time_after(jiffies, rcu_state.jiffies_resched)) {
if (time_after(jiffies,
READ_ONCE(rdp->last_fqs_resched) + jtsq)) {
WRITE_ONCE(rdp->last_fqs_resched, jiffies);
ret = -1;
}
if (IS_ENABLED(CONFIG_IRQ_WORK) &&
!rdp->rcu_iw_pending && rdp->rcu_iw_gp_seq != rnp->gp_seq &&
(rnp->ffmask & rdp->grpmask)) {
rdp->rcu_iw_pending = true;
rdp->rcu_iw_gp_seq = rnp->gp_seq;
irq_work_queue_on(&rdp->rcu_iw, rdp->cpu);
}
rcu: Add RCU stall diagnosis information Because RCU CPU stall warnings are driven from the scheduling-clock interrupt handler, a workload consisting of a very large number of short-duration hardware interrupts can result in misleading stall-warning messages. On systems supporting only a single level of interrupts, that is, where interrupts handlers cannot be interrupted, this can produce misleading diagnostics. The stack traces will show the innocent-bystander interrupted task, not the interrupts that are at the very least exacerbating the stall. This situation can be improved by displaying the number of interrupts and the CPU time that they have consumed. Diagnosing other types of stalls can be eased by also providing the count of softirqs and the CPU time that they consumed as well as the number of context switches and the task-level CPU time consumed. Consider the following output given this change: rcu: INFO: rcu_preempt self-detected stall on CPU rcu: 0-....: (1250 ticks this GP) <omitted> rcu: hardirqs softirqs csw/system rcu: number: 624 45 0 rcu: cputime: 69 1 2425 ==> 2500(ms) This output shows that the number of hard and soft interrupts is small, there are no context switches, and the system takes up a lot of time. This indicates that the current task is looping with preemption disabled. The impact on system performance is negligible because snapshot is recorded only once for all continuous RCU stalls. This added debugging information is suppressed by default and can be enabled by building the kernel with CONFIG_RCU_CPU_STALL_CPUTIME=y or by booting with rcupdate.rcu_cpu_stall_cputime=1. Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com> Reviewed-by: Mukesh Ojha <quic_mojha@quicinc.com> Reviewed-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2022-11-19 17:25:06 +08:00
if (rcu_cpu_stall_cputime && rdp->snap_record.gp_seq != rdp->gp_seq) {
int cpu = rdp->cpu;
struct rcu_snap_record *rsrp;
struct kernel_cpustat *kcsp;
kcsp = &kcpustat_cpu(cpu);
rsrp = &rdp->snap_record;
rsrp->cputime_irq = kcpustat_field(kcsp, CPUTIME_IRQ, cpu);
rsrp->cputime_softirq = kcpustat_field(kcsp, CPUTIME_SOFTIRQ, cpu);
rsrp->cputime_system = kcpustat_field(kcsp, CPUTIME_SYSTEM, cpu);
rsrp->nr_hardirqs = kstat_cpu_irqs_sum(rdp->cpu);
rsrp->nr_softirqs = kstat_cpu_softirqs_sum(rdp->cpu);
rsrp->nr_csw = nr_context_switches_cpu(rdp->cpu);
rsrp->jiffies = jiffies;
rsrp->gp_seq = rdp->gp_seq;
}
}
return ret;
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
}
/* Trace-event wrapper function for trace_rcu_future_grace_period. */
static void trace_rcu_this_gp(struct rcu_node *rnp, struct rcu_data *rdp,
unsigned long gp_seq_req, const char *s)
{
trace_rcu_future_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq),
gp_seq_req, rnp->level,
rnp->grplo, rnp->grphi, s);
}
/*
* rcu_start_this_gp - Request the start of a particular grace period
* @rnp_start: The leaf node of the CPU from which to start.
* @rdp: The rcu_data corresponding to the CPU from which to start.
* @gp_seq_req: The gp_seq of the grace period to start.
*
* Start the specified grace period, as needed to handle newly arrived
* callbacks. The required future grace periods are recorded in each
* rcu_node structure's ->gp_seq_needed field. Returns true if there
rcu: Make callers awaken grace-period kthread The rcu_start_gp_advanced() function currently uses irq_work_queue() to defer wakeups of the RCU grace-period kthread. This deferring is necessary to avoid RCU-scheduler deadlocks involving the rcu_node structure's lock, meaning that RCU cannot call any of the scheduler's wake-up functions while holding one of these locks. Unfortunately, the second and subsequent calls to irq_work_queue() are ignored, and the first call will be ignored (aside from queuing the work item) if the scheduler-clock tick is turned off. This is OK for many uses, especially those where irq_work_queue() is called from an interrupt or softirq handler, because in those cases the scheduler-clock-tick state will be re-evaluated, which will turn the scheduler-clock tick back on. On the next tick, any deferred work will then be processed. However, this strategy does not always work for RCU, which can be invoked at process level from idle CPUs. In this case, the tick might never be turned back on, indefinitely defering a grace-period start request. Note that the RCU CPU stall detector cannot see this condition, because there is no RCU grace period in progress. Therefore, we can (and do!) see long tens-of-seconds stalls in grace-period handling. In theory, we could see a full grace-period hang, but rcutorture testing to date has seen only the tens-of-seconds stalls. Event tracing demonstrates that irq_work_queue() is being called repeatedly to no effect during these stalls: The "newreq" event appears repeatedly from a task that is not one of the grace-period kthreads. In theory, irq_work_queue() might be fixed to avoid this sort of issue, but RCU's requirements are unusual and it is quite straightforward to pass wake-up responsibility up through RCU's call chain, so that the wakeup happens when the offending locks are released. This commit therefore makes this change. The rcu_start_gp_advanced(), rcu_start_future_gp(), rcu_accelerate_cbs(), rcu_advance_cbs(), __note_gp_changes(), and rcu_start_gp() functions now return a boolean which indicates when a wake-up is needed. A new rcu_gp_kthread_wake() does the wakeup when it is necessary and safe to do so: No self-wakes, no wake-ups if the ->gp_flags field indicates there is no need (as in someone else did the wake-up before we got around to it), and no wake-ups before the grace-period kthread has been created. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org>
2014-03-12 04:02:16 +08:00
* is reason to awaken the grace-period kthread.
*
* The caller must hold the specified rcu_node structure's ->lock, which
* is why the caller is responsible for waking the grace-period kthread.
*
* Returns true if the GP thread needs to be awakened else false.
*/
static bool rcu_start_this_gp(struct rcu_node *rnp_start, struct rcu_data *rdp,
unsigned long gp_seq_req)
{
rcu: Make callers awaken grace-period kthread The rcu_start_gp_advanced() function currently uses irq_work_queue() to defer wakeups of the RCU grace-period kthread. This deferring is necessary to avoid RCU-scheduler deadlocks involving the rcu_node structure's lock, meaning that RCU cannot call any of the scheduler's wake-up functions while holding one of these locks. Unfortunately, the second and subsequent calls to irq_work_queue() are ignored, and the first call will be ignored (aside from queuing the work item) if the scheduler-clock tick is turned off. This is OK for many uses, especially those where irq_work_queue() is called from an interrupt or softirq handler, because in those cases the scheduler-clock-tick state will be re-evaluated, which will turn the scheduler-clock tick back on. On the next tick, any deferred work will then be processed. However, this strategy does not always work for RCU, which can be invoked at process level from idle CPUs. In this case, the tick might never be turned back on, indefinitely defering a grace-period start request. Note that the RCU CPU stall detector cannot see this condition, because there is no RCU grace period in progress. Therefore, we can (and do!) see long tens-of-seconds stalls in grace-period handling. In theory, we could see a full grace-period hang, but rcutorture testing to date has seen only the tens-of-seconds stalls. Event tracing demonstrates that irq_work_queue() is being called repeatedly to no effect during these stalls: The "newreq" event appears repeatedly from a task that is not one of the grace-period kthreads. In theory, irq_work_queue() might be fixed to avoid this sort of issue, but RCU's requirements are unusual and it is quite straightforward to pass wake-up responsibility up through RCU's call chain, so that the wakeup happens when the offending locks are released. This commit therefore makes this change. The rcu_start_gp_advanced(), rcu_start_future_gp(), rcu_accelerate_cbs(), rcu_advance_cbs(), __note_gp_changes(), and rcu_start_gp() functions now return a boolean which indicates when a wake-up is needed. A new rcu_gp_kthread_wake() does the wakeup when it is necessary and safe to do so: No self-wakes, no wake-ups if the ->gp_flags field indicates there is no need (as in someone else did the wake-up before we got around to it), and no wake-ups before the grace-period kthread has been created. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org>
2014-03-12 04:02:16 +08:00
bool ret = false;
struct rcu_node *rnp;
/*
* Use funnel locking to either acquire the root rcu_node
* structure's lock or bail out if the need for this grace period
* has already been recorded -- or if that grace period has in
* fact already started. If there is already a grace period in
* progress in a non-leaf node, no recording is needed because the
* end of the grace period will scan the leaf rcu_node structures.
* Note that rnp_start->lock must not be released.
*/
raw_lockdep_assert_held_rcu_node(rnp_start);
trace_rcu_this_gp(rnp_start, rdp, gp_seq_req, TPS("Startleaf"));
for (rnp = rnp_start; 1; rnp = rnp->parent) {
if (rnp != rnp_start)
raw_spin_lock_rcu_node(rnp);
if (ULONG_CMP_GE(rnp->gp_seq_needed, gp_seq_req) ||
rcu_seq_started(&rnp->gp_seq, gp_seq_req) ||
(rnp != rnp_start &&
rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))) {
trace_rcu_this_gp(rnp, rdp, gp_seq_req,
TPS("Prestarted"));
goto unlock_out;
}
WRITE_ONCE(rnp->gp_seq_needed, gp_seq_req);
if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq))) {
/*
* We just marked the leaf or internal node, and a
* grace period is in progress, which means that
* rcu_gp_cleanup() will see the marking. Bail to
* reduce contention.
*/
trace_rcu_this_gp(rnp_start, rdp, gp_seq_req,
TPS("Startedleaf"));
goto unlock_out;
}
if (rnp != rnp_start && rnp->parent != NULL)
raw_spin_unlock_rcu_node(rnp);
if (!rnp->parent)
break; /* At root, and perhaps also leaf. */
}
/* If GP already in progress, just leave, otherwise start one. */
if (rcu_gp_in_progress()) {
trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedleafroot"));
goto unlock_out;
}
trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedroot"));
WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags | RCU_GP_FLAG_INIT);
WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
if (!READ_ONCE(rcu_state.gp_kthread)) {
trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("NoGPkthread"));
goto unlock_out;
}
trace_rcu_grace_period(rcu_state.name, data_race(rcu_state.gp_seq), TPS("newreq"));
ret = true; /* Caller must wake GP kthread. */
unlock_out:
/* Push furthest requested GP to leaf node and rcu_data structure. */
if (ULONG_CMP_LT(gp_seq_req, rnp->gp_seq_needed)) {
WRITE_ONCE(rnp_start->gp_seq_needed, rnp->gp_seq_needed);
WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed);
}
if (rnp != rnp_start)
raw_spin_unlock_rcu_node(rnp);
rcu: Make callers awaken grace-period kthread The rcu_start_gp_advanced() function currently uses irq_work_queue() to defer wakeups of the RCU grace-period kthread. This deferring is necessary to avoid RCU-scheduler deadlocks involving the rcu_node structure's lock, meaning that RCU cannot call any of the scheduler's wake-up functions while holding one of these locks. Unfortunately, the second and subsequent calls to irq_work_queue() are ignored, and the first call will be ignored (aside from queuing the work item) if the scheduler-clock tick is turned off. This is OK for many uses, especially those where irq_work_queue() is called from an interrupt or softirq handler, because in those cases the scheduler-clock-tick state will be re-evaluated, which will turn the scheduler-clock tick back on. On the next tick, any deferred work will then be processed. However, this strategy does not always work for RCU, which can be invoked at process level from idle CPUs. In this case, the tick might never be turned back on, indefinitely defering a grace-period start request. Note that the RCU CPU stall detector cannot see this condition, because there is no RCU grace period in progress. Therefore, we can (and do!) see long tens-of-seconds stalls in grace-period handling. In theory, we could see a full grace-period hang, but rcutorture testing to date has seen only the tens-of-seconds stalls. Event tracing demonstrates that irq_work_queue() is being called repeatedly to no effect during these stalls: The "newreq" event appears repeatedly from a task that is not one of the grace-period kthreads. In theory, irq_work_queue() might be fixed to avoid this sort of issue, but RCU's requirements are unusual and it is quite straightforward to pass wake-up responsibility up through RCU's call chain, so that the wakeup happens when the offending locks are released. This commit therefore makes this change. The rcu_start_gp_advanced(), rcu_start_future_gp(), rcu_accelerate_cbs(), rcu_advance_cbs(), __note_gp_changes(), and rcu_start_gp() functions now return a boolean which indicates when a wake-up is needed. A new rcu_gp_kthread_wake() does the wakeup when it is necessary and safe to do so: No self-wakes, no wake-ups if the ->gp_flags field indicates there is no need (as in someone else did the wake-up before we got around to it), and no wake-ups before the grace-period kthread has been created. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org>
2014-03-12 04:02:16 +08:00
return ret;
}
/*
* Clean up any old requests for the just-ended grace period. Also return
* whether any additional grace periods have been requested.
*/
static bool rcu_future_gp_cleanup(struct rcu_node *rnp)
{
bool needmore;
struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
needmore = ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed);
if (!needmore)
rnp->gp_seq_needed = rnp->gp_seq; /* Avoid counter wrap. */
trace_rcu_this_gp(rnp, rdp, rnp->gp_seq,
needmore ? TPS("CleanupMore") : TPS("Cleanup"));
return needmore;
}
rcu: Defer RCU kthreads wakeup when CPU is dying [ Upstream commit e787644caf7628ad3269c1fbd321c3255cf51710 ] When the CPU goes idle for the last time during the CPU down hotplug process, RCU reports a final quiescent state for the current CPU. If this quiescent state propagates up to the top, some tasks may then be woken up to complete the grace period: the main grace period kthread and/or the expedited main workqueue (or kworker). If those kthreads have a SCHED_FIFO policy, the wake up can indirectly arm the RT bandwith timer to the local offline CPU. Since this happens after hrtimers have been migrated at CPUHP_AP_HRTIMERS_DYING stage, the timer gets ignored. Therefore if the RCU kthreads are waiting for RT bandwidth to be available, they may never be actually scheduled. This triggers TREE03 rcutorture hangs: rcu: INFO: rcu_preempt self-detected stall on CPU rcu: 4-...!: (1 GPs behind) idle=9874/1/0x4000000000000000 softirq=0/0 fqs=20 rcuc=21071 jiffies(starved) rcu: (t=21035 jiffies g=938281 q=40787 ncpus=6) rcu: rcu_preempt kthread starved for 20964 jiffies! g938281 f0x0 RCU_GP_WAIT_FQS(5) ->state=0x0 ->cpu=0 rcu: Unless rcu_preempt kthread gets sufficient CPU time, OOM is now expected behavior. rcu: RCU grace-period kthread stack dump: task:rcu_preempt state:R running task stack:14896 pid:14 tgid:14 ppid:2 flags:0x00004000 Call Trace: <TASK> __schedule+0x2eb/0xa80 schedule+0x1f/0x90 schedule_timeout+0x163/0x270 ? __pfx_process_timeout+0x10/0x10 rcu_gp_fqs_loop+0x37c/0x5b0 ? __pfx_rcu_gp_kthread+0x10/0x10 rcu_gp_kthread+0x17c/0x200 kthread+0xde/0x110 ? __pfx_kthread+0x10/0x10 ret_from_fork+0x2b/0x40 ? __pfx_kthread+0x10/0x10 ret_from_fork_asm+0x1b/0x30 </TASK> The situation can't be solved with just unpinning the timer. The hrtimer infrastructure and the nohz heuristics involved in finding the best remote target for an unpinned timer would then also need to handle enqueues from an offline CPU in the most horrendous way. So fix this on the RCU side instead and defer the wake up to an online CPU if it's too late for the local one. Reported-by: Paul E. McKenney <paulmck@kernel.org> Fixes: 5c0930ccaad5 ("hrtimers: Push pending hrtimers away from outgoing CPU earlier") Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Neeraj Upadhyay (AMD) <neeraj.iitr10@gmail.com> Signed-off-by: Sasha Levin <sashal@kernel.org>
2023-12-19 07:19:15 +08:00
static void swake_up_one_online_ipi(void *arg)
{
struct swait_queue_head *wqh = arg;
swake_up_one(wqh);
}
static void swake_up_one_online(struct swait_queue_head *wqh)
{
int cpu = get_cpu();
/*
* If called from rcutree_report_cpu_starting(), wake up
* is dangerous that late in the CPU-down hotplug process. The
* scheduler might queue an ignored hrtimer. Defer the wake up
* to an online CPU instead.
*/
if (unlikely(cpu_is_offline(cpu))) {
int target;
target = cpumask_any_and(housekeeping_cpumask(HK_TYPE_RCU),
cpu_online_mask);
smp_call_function_single(target, swake_up_one_online_ipi,
wqh, 0);
put_cpu();
} else {
put_cpu();
swake_up_one(wqh);
}
}
rcu: Make callers awaken grace-period kthread The rcu_start_gp_advanced() function currently uses irq_work_queue() to defer wakeups of the RCU grace-period kthread. This deferring is necessary to avoid RCU-scheduler deadlocks involving the rcu_node structure's lock, meaning that RCU cannot call any of the scheduler's wake-up functions while holding one of these locks. Unfortunately, the second and subsequent calls to irq_work_queue() are ignored, and the first call will be ignored (aside from queuing the work item) if the scheduler-clock tick is turned off. This is OK for many uses, especially those where irq_work_queue() is called from an interrupt or softirq handler, because in those cases the scheduler-clock-tick state will be re-evaluated, which will turn the scheduler-clock tick back on. On the next tick, any deferred work will then be processed. However, this strategy does not always work for RCU, which can be invoked at process level from idle CPUs. In this case, the tick might never be turned back on, indefinitely defering a grace-period start request. Note that the RCU CPU stall detector cannot see this condition, because there is no RCU grace period in progress. Therefore, we can (and do!) see long tens-of-seconds stalls in grace-period handling. In theory, we could see a full grace-period hang, but rcutorture testing to date has seen only the tens-of-seconds stalls. Event tracing demonstrates that irq_work_queue() is being called repeatedly to no effect during these stalls: The "newreq" event appears repeatedly from a task that is not one of the grace-period kthreads. In theory, irq_work_queue() might be fixed to avoid this sort of issue, but RCU's requirements are unusual and it is quite straightforward to pass wake-up responsibility up through RCU's call chain, so that the wakeup happens when the offending locks are released. This commit therefore makes this change. The rcu_start_gp_advanced(), rcu_start_future_gp(), rcu_accelerate_cbs(), rcu_advance_cbs(), __note_gp_changes(), and rcu_start_gp() functions now return a boolean which indicates when a wake-up is needed. A new rcu_gp_kthread_wake() does the wakeup when it is necessary and safe to do so: No self-wakes, no wake-ups if the ->gp_flags field indicates there is no need (as in someone else did the wake-up before we got around to it), and no wake-ups before the grace-period kthread has been created. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org>
2014-03-12 04:02:16 +08:00
/*
* Awaken the grace-period kthread. Don't do a self-awaken (unless in an
* interrupt or softirq handler, in which case we just might immediately
* sleep upon return, resulting in a grace-period hang), and don't bother
* awakening when there is nothing for the grace-period kthread to do
* (as in several CPUs raced to awaken, we lost), and finally don't try
* to awaken a kthread that has not yet been created. If all those checks
* are passed, track some debug information and awaken.
rcu: Do RCU GP kthread self-wakeup from softirq and interrupt The rcu_gp_kthread_wake() function is invoked when it might be necessary to wake the RCU grace-period kthread. Because self-wakeups are normally a useless waste of CPU cycles, if rcu_gp_kthread_wake() is invoked from this kthread, it naturally refuses to do the wakeup. Unfortunately, natural though it might be, this heuristic fails when rcu_gp_kthread_wake() is invoked from an interrupt or softirq handler that interrupted the grace-period kthread just after the final check of the wait-event condition but just before the schedule() call. In this case, a wakeup is required, even though the call to rcu_gp_kthread_wake() is within the RCU grace-period kthread's context. Failing to provide this wakeup can result in grace periods failing to start, which in turn results in out-of-memory conditions. This race window is quite narrow, but it actually did happen during real testing. It would of course need to be fixed even if it was strictly theoretical in nature. This patch does not Cc stable because it does not apply cleanly to earlier kernel versions. Fixes: 48a7639ce80c ("rcu: Make callers awaken grace-period kthread") Reported-by: "He, Bo" <bo.he@intel.com> Co-developed-by: "Zhang, Jun" <jun.zhang@intel.com> Co-developed-by: "He, Bo" <bo.he@intel.com> Co-developed-by: "xiao, jin" <jin.xiao@intel.com> Co-developed-by: Bai, Jie A <jie.a.bai@intel.com> Signed-off: "Zhang, Jun" <jun.zhang@intel.com> Signed-off: "He, Bo" <bo.he@intel.com> Signed-off: "xiao, jin" <jin.xiao@intel.com> Signed-off: Bai, Jie A <jie.a.bai@intel.com> Signed-off-by: "Zhang, Jun" <jun.zhang@intel.com> [ paulmck: Switch from !in_softirq() to "!in_interrupt() && !in_serving_softirq() to avoid redundant wakeups and to also handle the interrupt-handler scenario as well as the softirq-handler scenario that actually occurred in testing. ] Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com> Link: https://lkml.kernel.org/r/CD6925E8781EFD4D8E11882D20FC406D52A11F61@SHSMSX104.ccr.corp.intel.com
2018-12-18 22:55:01 +08:00
*
* So why do the self-wakeup when in an interrupt or softirq handler
* in the grace-period kthread's context? Because the kthread might have
* been interrupted just as it was going to sleep, and just after the final
* pre-sleep check of the awaken condition. In this case, a wakeup really
* is required, and is therefore supplied.
rcu: Make callers awaken grace-period kthread The rcu_start_gp_advanced() function currently uses irq_work_queue() to defer wakeups of the RCU grace-period kthread. This deferring is necessary to avoid RCU-scheduler deadlocks involving the rcu_node structure's lock, meaning that RCU cannot call any of the scheduler's wake-up functions while holding one of these locks. Unfortunately, the second and subsequent calls to irq_work_queue() are ignored, and the first call will be ignored (aside from queuing the work item) if the scheduler-clock tick is turned off. This is OK for many uses, especially those where irq_work_queue() is called from an interrupt or softirq handler, because in those cases the scheduler-clock-tick state will be re-evaluated, which will turn the scheduler-clock tick back on. On the next tick, any deferred work will then be processed. However, this strategy does not always work for RCU, which can be invoked at process level from idle CPUs. In this case, the tick might never be turned back on, indefinitely defering a grace-period start request. Note that the RCU CPU stall detector cannot see this condition, because there is no RCU grace period in progress. Therefore, we can (and do!) see long tens-of-seconds stalls in grace-period handling. In theory, we could see a full grace-period hang, but rcutorture testing to date has seen only the tens-of-seconds stalls. Event tracing demonstrates that irq_work_queue() is being called repeatedly to no effect during these stalls: The "newreq" event appears repeatedly from a task that is not one of the grace-period kthreads. In theory, irq_work_queue() might be fixed to avoid this sort of issue, but RCU's requirements are unusual and it is quite straightforward to pass wake-up responsibility up through RCU's call chain, so that the wakeup happens when the offending locks are released. This commit therefore makes this change. The rcu_start_gp_advanced(), rcu_start_future_gp(), rcu_accelerate_cbs(), rcu_advance_cbs(), __note_gp_changes(), and rcu_start_gp() functions now return a boolean which indicates when a wake-up is needed. A new rcu_gp_kthread_wake() does the wakeup when it is necessary and safe to do so: No self-wakes, no wake-ups if the ->gp_flags field indicates there is no need (as in someone else did the wake-up before we got around to it), and no wake-ups before the grace-period kthread has been created. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org>
2014-03-12 04:02:16 +08:00
*/
static void rcu_gp_kthread_wake(void)
rcu: Make callers awaken grace-period kthread The rcu_start_gp_advanced() function currently uses irq_work_queue() to defer wakeups of the RCU grace-period kthread. This deferring is necessary to avoid RCU-scheduler deadlocks involving the rcu_node structure's lock, meaning that RCU cannot call any of the scheduler's wake-up functions while holding one of these locks. Unfortunately, the second and subsequent calls to irq_work_queue() are ignored, and the first call will be ignored (aside from queuing the work item) if the scheduler-clock tick is turned off. This is OK for many uses, especially those where irq_work_queue() is called from an interrupt or softirq handler, because in those cases the scheduler-clock-tick state will be re-evaluated, which will turn the scheduler-clock tick back on. On the next tick, any deferred work will then be processed. However, this strategy does not always work for RCU, which can be invoked at process level from idle CPUs. In this case, the tick might never be turned back on, indefinitely defering a grace-period start request. Note that the RCU CPU stall detector cannot see this condition, because there is no RCU grace period in progress. Therefore, we can (and do!) see long tens-of-seconds stalls in grace-period handling. In theory, we could see a full grace-period hang, but rcutorture testing to date has seen only the tens-of-seconds stalls. Event tracing demonstrates that irq_work_queue() is being called repeatedly to no effect during these stalls: The "newreq" event appears repeatedly from a task that is not one of the grace-period kthreads. In theory, irq_work_queue() might be fixed to avoid this sort of issue, but RCU's requirements are unusual and it is quite straightforward to pass wake-up responsibility up through RCU's call chain, so that the wakeup happens when the offending locks are released. This commit therefore makes this change. The rcu_start_gp_advanced(), rcu_start_future_gp(), rcu_accelerate_cbs(), rcu_advance_cbs(), __note_gp_changes(), and rcu_start_gp() functions now return a boolean which indicates when a wake-up is needed. A new rcu_gp_kthread_wake() does the wakeup when it is necessary and safe to do so: No self-wakes, no wake-ups if the ->gp_flags field indicates there is no need (as in someone else did the wake-up before we got around to it), and no wake-ups before the grace-period kthread has been created. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org>
2014-03-12 04:02:16 +08:00
{
struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
if ((current == t && !in_hardirq() && !in_serving_softirq()) ||
!READ_ONCE(rcu_state.gp_flags) || !t)
rcu: Make callers awaken grace-period kthread The rcu_start_gp_advanced() function currently uses irq_work_queue() to defer wakeups of the RCU grace-period kthread. This deferring is necessary to avoid RCU-scheduler deadlocks involving the rcu_node structure's lock, meaning that RCU cannot call any of the scheduler's wake-up functions while holding one of these locks. Unfortunately, the second and subsequent calls to irq_work_queue() are ignored, and the first call will be ignored (aside from queuing the work item) if the scheduler-clock tick is turned off. This is OK for many uses, especially those where irq_work_queue() is called from an interrupt or softirq handler, because in those cases the scheduler-clock-tick state will be re-evaluated, which will turn the scheduler-clock tick back on. On the next tick, any deferred work will then be processed. However, this strategy does not always work for RCU, which can be invoked at process level from idle CPUs. In this case, the tick might never be turned back on, indefinitely defering a grace-period start request. Note that the RCU CPU stall detector cannot see this condition, because there is no RCU grace period in progress. Therefore, we can (and do!) see long tens-of-seconds stalls in grace-period handling. In theory, we could see a full grace-period hang, but rcutorture testing to date has seen only the tens-of-seconds stalls. Event tracing demonstrates that irq_work_queue() is being called repeatedly to no effect during these stalls: The "newreq" event appears repeatedly from a task that is not one of the grace-period kthreads. In theory, irq_work_queue() might be fixed to avoid this sort of issue, but RCU's requirements are unusual and it is quite straightforward to pass wake-up responsibility up through RCU's call chain, so that the wakeup happens when the offending locks are released. This commit therefore makes this change. The rcu_start_gp_advanced(), rcu_start_future_gp(), rcu_accelerate_cbs(), rcu_advance_cbs(), __note_gp_changes(), and rcu_start_gp() functions now return a boolean which indicates when a wake-up is needed. A new rcu_gp_kthread_wake() does the wakeup when it is necessary and safe to do so: No self-wakes, no wake-ups if the ->gp_flags field indicates there is no need (as in someone else did the wake-up before we got around to it), and no wake-ups before the grace-period kthread has been created. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org>
2014-03-12 04:02:16 +08:00
return;
WRITE_ONCE(rcu_state.gp_wake_time, jiffies);
WRITE_ONCE(rcu_state.gp_wake_seq, READ_ONCE(rcu_state.gp_seq));
rcu: Defer RCU kthreads wakeup when CPU is dying [ Upstream commit e787644caf7628ad3269c1fbd321c3255cf51710 ] When the CPU goes idle for the last time during the CPU down hotplug process, RCU reports a final quiescent state for the current CPU. If this quiescent state propagates up to the top, some tasks may then be woken up to complete the grace period: the main grace period kthread and/or the expedited main workqueue (or kworker). If those kthreads have a SCHED_FIFO policy, the wake up can indirectly arm the RT bandwith timer to the local offline CPU. Since this happens after hrtimers have been migrated at CPUHP_AP_HRTIMERS_DYING stage, the timer gets ignored. Therefore if the RCU kthreads are waiting for RT bandwidth to be available, they may never be actually scheduled. This triggers TREE03 rcutorture hangs: rcu: INFO: rcu_preempt self-detected stall on CPU rcu: 4-...!: (1 GPs behind) idle=9874/1/0x4000000000000000 softirq=0/0 fqs=20 rcuc=21071 jiffies(starved) rcu: (t=21035 jiffies g=938281 q=40787 ncpus=6) rcu: rcu_preempt kthread starved for 20964 jiffies! g938281 f0x0 RCU_GP_WAIT_FQS(5) ->state=0x0 ->cpu=0 rcu: Unless rcu_preempt kthread gets sufficient CPU time, OOM is now expected behavior. rcu: RCU grace-period kthread stack dump: task:rcu_preempt state:R running task stack:14896 pid:14 tgid:14 ppid:2 flags:0x00004000 Call Trace: <TASK> __schedule+0x2eb/0xa80 schedule+0x1f/0x90 schedule_timeout+0x163/0x270 ? __pfx_process_timeout+0x10/0x10 rcu_gp_fqs_loop+0x37c/0x5b0 ? __pfx_rcu_gp_kthread+0x10/0x10 rcu_gp_kthread+0x17c/0x200 kthread+0xde/0x110 ? __pfx_kthread+0x10/0x10 ret_from_fork+0x2b/0x40 ? __pfx_kthread+0x10/0x10 ret_from_fork_asm+0x1b/0x30 </TASK> The situation can't be solved with just unpinning the timer. The hrtimer infrastructure and the nohz heuristics involved in finding the best remote target for an unpinned timer would then also need to handle enqueues from an offline CPU in the most horrendous way. So fix this on the RCU side instead and defer the wake up to an online CPU if it's too late for the local one. Reported-by: Paul E. McKenney <paulmck@kernel.org> Fixes: 5c0930ccaad5 ("hrtimers: Push pending hrtimers away from outgoing CPU earlier") Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Neeraj Upadhyay (AMD) <neeraj.iitr10@gmail.com> Signed-off-by: Sasha Levin <sashal@kernel.org>
2023-12-19 07:19:15 +08:00
swake_up_one_online(&rcu_state.gp_wq);
rcu: Make callers awaken grace-period kthread The rcu_start_gp_advanced() function currently uses irq_work_queue() to defer wakeups of the RCU grace-period kthread. This deferring is necessary to avoid RCU-scheduler deadlocks involving the rcu_node structure's lock, meaning that RCU cannot call any of the scheduler's wake-up functions while holding one of these locks. Unfortunately, the second and subsequent calls to irq_work_queue() are ignored, and the first call will be ignored (aside from queuing the work item) if the scheduler-clock tick is turned off. This is OK for many uses, especially those where irq_work_queue() is called from an interrupt or softirq handler, because in those cases the scheduler-clock-tick state will be re-evaluated, which will turn the scheduler-clock tick back on. On the next tick, any deferred work will then be processed. However, this strategy does not always work for RCU, which can be invoked at process level from idle CPUs. In this case, the tick might never be turned back on, indefinitely defering a grace-period start request. Note that the RCU CPU stall detector cannot see this condition, because there is no RCU grace period in progress. Therefore, we can (and do!) see long tens-of-seconds stalls in grace-period handling. In theory, we could see a full grace-period hang, but rcutorture testing to date has seen only the tens-of-seconds stalls. Event tracing demonstrates that irq_work_queue() is being called repeatedly to no effect during these stalls: The "newreq" event appears repeatedly from a task that is not one of the grace-period kthreads. In theory, irq_work_queue() might be fixed to avoid this sort of issue, but RCU's requirements are unusual and it is quite straightforward to pass wake-up responsibility up through RCU's call chain, so that the wakeup happens when the offending locks are released. This commit therefore makes this change. The rcu_start_gp_advanced(), rcu_start_future_gp(), rcu_accelerate_cbs(), rcu_advance_cbs(), __note_gp_changes(), and rcu_start_gp() functions now return a boolean which indicates when a wake-up is needed. A new rcu_gp_kthread_wake() does the wakeup when it is necessary and safe to do so: No self-wakes, no wake-ups if the ->gp_flags field indicates there is no need (as in someone else did the wake-up before we got around to it), and no wake-ups before the grace-period kthread has been created. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org>
2014-03-12 04:02:16 +08:00
}
rcu: Tag callback lists with corresponding grace-period number Currently, callbacks are advanced each time the corresponding CPU notices a change in its leaf rcu_node structure's ->completed value (this value counts grace-period completions). This approach has worked quite well, but with the advent of RCU_FAST_NO_HZ, we cannot count on a given CPU seeing all the grace-period completions. When a CPU misses a grace-period completion that occurs while it is in dyntick-idle mode, this will delay invocation of its callbacks. In addition, acceleration of callbacks (when RCU realizes that a given callback need only wait until the end of the next grace period, rather than having to wait for a partial grace period followed by a full grace period) must be carried out extremely carefully. Insufficient acceleration will result in unnecessarily long grace-period latencies, while excessive acceleration will result in premature callback invocation. Changes that involve this tradeoff are therefore among the most nerve-wracking changes to RCU. This commit therefore explicitly tags groups of callbacks with the number of the grace period that they are waiting for. This means that callback-advancement and callback-acceleration functions are idempotent, so that excessive acceleration will merely waste a few CPU cycles. This also allows a CPU to take full advantage of any grace periods that have elapsed while it has been in dyntick-idle mode. It should also enable simulataneous simplifications to and optimizations of RCU_FAST_NO_HZ. Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2012-12-04 05:52:00 +08:00
/*
* If there is room, assign a ->gp_seq number to any callbacks on this
* CPU that have not already been assigned. Also accelerate any callbacks
* that were previously assigned a ->gp_seq number that has since proven
* to be too conservative, which can happen if callbacks get assigned a
* ->gp_seq number while RCU is idle, but with reference to a non-root
* rcu_node structure. This function is idempotent, so it does not hurt
* to call it repeatedly. Returns an flag saying that we should awaken
* the RCU grace-period kthread.
rcu: Tag callback lists with corresponding grace-period number Currently, callbacks are advanced each time the corresponding CPU notices a change in its leaf rcu_node structure's ->completed value (this value counts grace-period completions). This approach has worked quite well, but with the advent of RCU_FAST_NO_HZ, we cannot count on a given CPU seeing all the grace-period completions. When a CPU misses a grace-period completion that occurs while it is in dyntick-idle mode, this will delay invocation of its callbacks. In addition, acceleration of callbacks (when RCU realizes that a given callback need only wait until the end of the next grace period, rather than having to wait for a partial grace period followed by a full grace period) must be carried out extremely carefully. Insufficient acceleration will result in unnecessarily long grace-period latencies, while excessive acceleration will result in premature callback invocation. Changes that involve this tradeoff are therefore among the most nerve-wracking changes to RCU. This commit therefore explicitly tags groups of callbacks with the number of the grace period that they are waiting for. This means that callback-advancement and callback-acceleration functions are idempotent, so that excessive acceleration will merely waste a few CPU cycles. This also allows a CPU to take full advantage of any grace periods that have elapsed while it has been in dyntick-idle mode. It should also enable simulataneous simplifications to and optimizations of RCU_FAST_NO_HZ. Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2012-12-04 05:52:00 +08:00
*
* The caller must hold rnp->lock with interrupts disabled.
*/
static bool rcu_accelerate_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
rcu: Tag callback lists with corresponding grace-period number Currently, callbacks are advanced each time the corresponding CPU notices a change in its leaf rcu_node structure's ->completed value (this value counts grace-period completions). This approach has worked quite well, but with the advent of RCU_FAST_NO_HZ, we cannot count on a given CPU seeing all the grace-period completions. When a CPU misses a grace-period completion that occurs while it is in dyntick-idle mode, this will delay invocation of its callbacks. In addition, acceleration of callbacks (when RCU realizes that a given callback need only wait until the end of the next grace period, rather than having to wait for a partial grace period followed by a full grace period) must be carried out extremely carefully. Insufficient acceleration will result in unnecessarily long grace-period latencies, while excessive acceleration will result in premature callback invocation. Changes that involve this tradeoff are therefore among the most nerve-wracking changes to RCU. This commit therefore explicitly tags groups of callbacks with the number of the grace period that they are waiting for. This means that callback-advancement and callback-acceleration functions are idempotent, so that excessive acceleration will merely waste a few CPU cycles. This also allows a CPU to take full advantage of any grace periods that have elapsed while it has been in dyntick-idle mode. It should also enable simulataneous simplifications to and optimizations of RCU_FAST_NO_HZ. Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2012-12-04 05:52:00 +08:00
{
unsigned long gp_seq_req;
bool ret = false;
rcu: Tag callback lists with corresponding grace-period number Currently, callbacks are advanced each time the corresponding CPU notices a change in its leaf rcu_node structure's ->completed value (this value counts grace-period completions). This approach has worked quite well, but with the advent of RCU_FAST_NO_HZ, we cannot count on a given CPU seeing all the grace-period completions. When a CPU misses a grace-period completion that occurs while it is in dyntick-idle mode, this will delay invocation of its callbacks. In addition, acceleration of callbacks (when RCU realizes that a given callback need only wait until the end of the next grace period, rather than having to wait for a partial grace period followed by a full grace period) must be carried out extremely carefully. Insufficient acceleration will result in unnecessarily long grace-period latencies, while excessive acceleration will result in premature callback invocation. Changes that involve this tradeoff are therefore among the most nerve-wracking changes to RCU. This commit therefore explicitly tags groups of callbacks with the number of the grace period that they are waiting for. This means that callback-advancement and callback-acceleration functions are idempotent, so that excessive acceleration will merely waste a few CPU cycles. This also allows a CPU to take full advantage of any grace periods that have elapsed while it has been in dyntick-idle mode. It should also enable simulataneous simplifications to and optimizations of RCU_FAST_NO_HZ. Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2012-12-04 05:52:00 +08:00
rcu/nocb: Add bypass callback queueing Use of the rcu_data structure's segmented ->cblist for no-CBs CPUs takes advantage of unrelated grace periods, thus reducing the memory footprint in the face of floods of call_rcu() invocations. However, the ->cblist field is a more-complex rcu_segcblist structure which must be protected via locking. Even though there are only three entities which can acquire this lock (the CPU invoking call_rcu(), the no-CBs grace-period kthread, and the no-CBs callbacks kthread), the contention on this lock is excessive under heavy stress. This commit therefore greatly reduces contention by provisioning an rcu_cblist structure field named ->nocb_bypass within the rcu_data structure. Each no-CBs CPU is permitted only a limited number of enqueues onto the ->cblist per jiffy, controlled by a new nocb_nobypass_lim_per_jiffy kernel boot parameter that defaults to about 16 enqueues per millisecond (16 * 1000 / HZ). When that limit is exceeded, the CPU instead enqueues onto the new ->nocb_bypass. The ->nocb_bypass is flushed into the ->cblist every jiffy or when the number of callbacks on ->nocb_bypass exceeds qhimark, whichever happens first. During call_rcu() floods, this flushing is carried out by the CPU during the course of its call_rcu() invocations. However, a CPU could simply stop invoking call_rcu() at any time. The no-CBs grace-period kthread therefore carries out less-aggressive flushing (every few jiffies or when the number of callbacks on ->nocb_bypass exceeds (2 * qhimark), whichever comes first). This means that the no-CBs grace-period kthread cannot be permitted to do unbounded waits while there are callbacks on ->nocb_bypass. A ->nocb_bypass_timer is used to provide the needed wakeups. [ paulmck: Apply Coverity feedback reported by Colin Ian King. ] Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
2019-07-03 07:03:33 +08:00
rcu_lockdep_assert_cblist_protected(rdp);
raw_lockdep_assert_held_rcu_node(rnp);
/* If no pending (not yet ready to invoke) callbacks, nothing to do. */
if (!rcu_segcblist_pend_cbs(&rdp->cblist))
rcu: Make callers awaken grace-period kthread The rcu_start_gp_advanced() function currently uses irq_work_queue() to defer wakeups of the RCU grace-period kthread. This deferring is necessary to avoid RCU-scheduler deadlocks involving the rcu_node structure's lock, meaning that RCU cannot call any of the scheduler's wake-up functions while holding one of these locks. Unfortunately, the second and subsequent calls to irq_work_queue() are ignored, and the first call will be ignored (aside from queuing the work item) if the scheduler-clock tick is turned off. This is OK for many uses, especially those where irq_work_queue() is called from an interrupt or softirq handler, because in those cases the scheduler-clock-tick state will be re-evaluated, which will turn the scheduler-clock tick back on. On the next tick, any deferred work will then be processed. However, this strategy does not always work for RCU, which can be invoked at process level from idle CPUs. In this case, the tick might never be turned back on, indefinitely defering a grace-period start request. Note that the RCU CPU stall detector cannot see this condition, because there is no RCU grace period in progress. Therefore, we can (and do!) see long tens-of-seconds stalls in grace-period handling. In theory, we could see a full grace-period hang, but rcutorture testing to date has seen only the tens-of-seconds stalls. Event tracing demonstrates that irq_work_queue() is being called repeatedly to no effect during these stalls: The "newreq" event appears repeatedly from a task that is not one of the grace-period kthreads. In theory, irq_work_queue() might be fixed to avoid this sort of issue, but RCU's requirements are unusual and it is quite straightforward to pass wake-up responsibility up through RCU's call chain, so that the wakeup happens when the offending locks are released. This commit therefore makes this change. The rcu_start_gp_advanced(), rcu_start_future_gp(), rcu_accelerate_cbs(), rcu_advance_cbs(), __note_gp_changes(), and rcu_start_gp() functions now return a boolean which indicates when a wake-up is needed. A new rcu_gp_kthread_wake() does the wakeup when it is necessary and safe to do so: No self-wakes, no wake-ups if the ->gp_flags field indicates there is no need (as in someone else did the wake-up before we got around to it), and no wake-ups before the grace-period kthread has been created. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org>
2014-03-12 04:02:16 +08:00
return false;
rcu: Tag callback lists with corresponding grace-period number Currently, callbacks are advanced each time the corresponding CPU notices a change in its leaf rcu_node structure's ->completed value (this value counts grace-period completions). This approach has worked quite well, but with the advent of RCU_FAST_NO_HZ, we cannot count on a given CPU seeing all the grace-period completions. When a CPU misses a grace-period completion that occurs while it is in dyntick-idle mode, this will delay invocation of its callbacks. In addition, acceleration of callbacks (when RCU realizes that a given callback need only wait until the end of the next grace period, rather than having to wait for a partial grace period followed by a full grace period) must be carried out extremely carefully. Insufficient acceleration will result in unnecessarily long grace-period latencies, while excessive acceleration will result in premature callback invocation. Changes that involve this tradeoff are therefore among the most nerve-wracking changes to RCU. This commit therefore explicitly tags groups of callbacks with the number of the grace period that they are waiting for. This means that callback-advancement and callback-acceleration functions are idempotent, so that excessive acceleration will merely waste a few CPU cycles. This also allows a CPU to take full advantage of any grace periods that have elapsed while it has been in dyntick-idle mode. It should also enable simulataneous simplifications to and optimizations of RCU_FAST_NO_HZ. Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2012-12-04 05:52:00 +08:00
trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbPreAcc"));
rcu: Tag callback lists with corresponding grace-period number Currently, callbacks are advanced each time the corresponding CPU notices a change in its leaf rcu_node structure's ->completed value (this value counts grace-period completions). This approach has worked quite well, but with the advent of RCU_FAST_NO_HZ, we cannot count on a given CPU seeing all the grace-period completions. When a CPU misses a grace-period completion that occurs while it is in dyntick-idle mode, this will delay invocation of its callbacks. In addition, acceleration of callbacks (when RCU realizes that a given callback need only wait until the end of the next grace period, rather than having to wait for a partial grace period followed by a full grace period) must be carried out extremely carefully. Insufficient acceleration will result in unnecessarily long grace-period latencies, while excessive acceleration will result in premature callback invocation. Changes that involve this tradeoff are therefore among the most nerve-wracking changes to RCU. This commit therefore explicitly tags groups of callbacks with the number of the grace period that they are waiting for. This means that callback-advancement and callback-acceleration functions are idempotent, so that excessive acceleration will merely waste a few CPU cycles. This also allows a CPU to take full advantage of any grace periods that have elapsed while it has been in dyntick-idle mode. It should also enable simulataneous simplifications to and optimizations of RCU_FAST_NO_HZ. Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2012-12-04 05:52:00 +08:00
/*
* Callbacks are often registered with incomplete grace-period
* information. Something about the fact that getting exact
* information requires acquiring a global lock... RCU therefore
* makes a conservative estimate of the grace period number at which
* a given callback will become ready to invoke. The following
* code checks this estimate and improves it when possible, thus
* accelerating callback invocation to an earlier grace-period
* number.
rcu: Tag callback lists with corresponding grace-period number Currently, callbacks are advanced each time the corresponding CPU notices a change in its leaf rcu_node structure's ->completed value (this value counts grace-period completions). This approach has worked quite well, but with the advent of RCU_FAST_NO_HZ, we cannot count on a given CPU seeing all the grace-period completions. When a CPU misses a grace-period completion that occurs while it is in dyntick-idle mode, this will delay invocation of its callbacks. In addition, acceleration of callbacks (when RCU realizes that a given callback need only wait until the end of the next grace period, rather than having to wait for a partial grace period followed by a full grace period) must be carried out extremely carefully. Insufficient acceleration will result in unnecessarily long grace-period latencies, while excessive acceleration will result in premature callback invocation. Changes that involve this tradeoff are therefore among the most nerve-wracking changes to RCU. This commit therefore explicitly tags groups of callbacks with the number of the grace period that they are waiting for. This means that callback-advancement and callback-acceleration functions are idempotent, so that excessive acceleration will merely waste a few CPU cycles. This also allows a CPU to take full advantage of any grace periods that have elapsed while it has been in dyntick-idle mode. It should also enable simulataneous simplifications to and optimizations of RCU_FAST_NO_HZ. Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2012-12-04 05:52:00 +08:00
*/
gp_seq_req = rcu_seq_snap(&rcu_state.gp_seq);
if (rcu_segcblist_accelerate(&rdp->cblist, gp_seq_req))
ret = rcu_start_this_gp(rnp, rdp, gp_seq_req);
/* Trace depending on how much we were able to accelerate. */
if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
trace_rcu_grace_period(rcu_state.name, gp_seq_req, TPS("AccWaitCB"));
else
trace_rcu_grace_period(rcu_state.name, gp_seq_req, TPS("AccReadyCB"));
trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbPostAcc"));
rcu: Make callers awaken grace-period kthread The rcu_start_gp_advanced() function currently uses irq_work_queue() to defer wakeups of the RCU grace-period kthread. This deferring is necessary to avoid RCU-scheduler deadlocks involving the rcu_node structure's lock, meaning that RCU cannot call any of the scheduler's wake-up functions while holding one of these locks. Unfortunately, the second and subsequent calls to irq_work_queue() are ignored, and the first call will be ignored (aside from queuing the work item) if the scheduler-clock tick is turned off. This is OK for many uses, especially those where irq_work_queue() is called from an interrupt or softirq handler, because in those cases the scheduler-clock-tick state will be re-evaluated, which will turn the scheduler-clock tick back on. On the next tick, any deferred work will then be processed. However, this strategy does not always work for RCU, which can be invoked at process level from idle CPUs. In this case, the tick might never be turned back on, indefinitely defering a grace-period start request. Note that the RCU CPU stall detector cannot see this condition, because there is no RCU grace period in progress. Therefore, we can (and do!) see long tens-of-seconds stalls in grace-period handling. In theory, we could see a full grace-period hang, but rcutorture testing to date has seen only the tens-of-seconds stalls. Event tracing demonstrates that irq_work_queue() is being called repeatedly to no effect during these stalls: The "newreq" event appears repeatedly from a task that is not one of the grace-period kthreads. In theory, irq_work_queue() might be fixed to avoid this sort of issue, but RCU's requirements are unusual and it is quite straightforward to pass wake-up responsibility up through RCU's call chain, so that the wakeup happens when the offending locks are released. This commit therefore makes this change. The rcu_start_gp_advanced(), rcu_start_future_gp(), rcu_accelerate_cbs(), rcu_advance_cbs(), __note_gp_changes(), and rcu_start_gp() functions now return a boolean which indicates when a wake-up is needed. A new rcu_gp_kthread_wake() does the wakeup when it is necessary and safe to do so: No self-wakes, no wake-ups if the ->gp_flags field indicates there is no need (as in someone else did the wake-up before we got around to it), and no wake-ups before the grace-period kthread has been created. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org>
2014-03-12 04:02:16 +08:00
return ret;
rcu: Tag callback lists with corresponding grace-period number Currently, callbacks are advanced each time the corresponding CPU notices a change in its leaf rcu_node structure's ->completed value (this value counts grace-period completions). This approach has worked quite well, but with the advent of RCU_FAST_NO_HZ, we cannot count on a given CPU seeing all the grace-period completions. When a CPU misses a grace-period completion that occurs while it is in dyntick-idle mode, this will delay invocation of its callbacks. In addition, acceleration of callbacks (when RCU realizes that a given callback need only wait until the end of the next grace period, rather than having to wait for a partial grace period followed by a full grace period) must be carried out extremely carefully. Insufficient acceleration will result in unnecessarily long grace-period latencies, while excessive acceleration will result in premature callback invocation. Changes that involve this tradeoff are therefore among the most nerve-wracking changes to RCU. This commit therefore explicitly tags groups of callbacks with the number of the grace period that they are waiting for. This means that callback-advancement and callback-acceleration functions are idempotent, so that excessive acceleration will merely waste a few CPU cycles. This also allows a CPU to take full advantage of any grace periods that have elapsed while it has been in dyntick-idle mode. It should also enable simulataneous simplifications to and optimizations of RCU_FAST_NO_HZ. Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2012-12-04 05:52:00 +08:00
}
/*
* Similar to rcu_accelerate_cbs(), but does not require that the leaf
* rcu_node structure's ->lock be held. It consults the cached value
* of ->gp_seq_needed in the rcu_data structure, and if that indicates
* that a new grace-period request be made, invokes rcu_accelerate_cbs()
* while holding the leaf rcu_node structure's ->lock.
*/
static void rcu_accelerate_cbs_unlocked(struct rcu_node *rnp,
struct rcu_data *rdp)
{
unsigned long c;
bool needwake;
rcu/nocb: Add bypass callback queueing Use of the rcu_data structure's segmented ->cblist for no-CBs CPUs takes advantage of unrelated grace periods, thus reducing the memory footprint in the face of floods of call_rcu() invocations. However, the ->cblist field is a more-complex rcu_segcblist structure which must be protected via locking. Even though there are only three entities which can acquire this lock (the CPU invoking call_rcu(), the no-CBs grace-period kthread, and the no-CBs callbacks kthread), the contention on this lock is excessive under heavy stress. This commit therefore greatly reduces contention by provisioning an rcu_cblist structure field named ->nocb_bypass within the rcu_data structure. Each no-CBs CPU is permitted only a limited number of enqueues onto the ->cblist per jiffy, controlled by a new nocb_nobypass_lim_per_jiffy kernel boot parameter that defaults to about 16 enqueues per millisecond (16 * 1000 / HZ). When that limit is exceeded, the CPU instead enqueues onto the new ->nocb_bypass. The ->nocb_bypass is flushed into the ->cblist every jiffy or when the number of callbacks on ->nocb_bypass exceeds qhimark, whichever happens first. During call_rcu() floods, this flushing is carried out by the CPU during the course of its call_rcu() invocations. However, a CPU could simply stop invoking call_rcu() at any time. The no-CBs grace-period kthread therefore carries out less-aggressive flushing (every few jiffies or when the number of callbacks on ->nocb_bypass exceeds (2 * qhimark), whichever comes first). This means that the no-CBs grace-period kthread cannot be permitted to do unbounded waits while there are callbacks on ->nocb_bypass. A ->nocb_bypass_timer is used to provide the needed wakeups. [ paulmck: Apply Coverity feedback reported by Colin Ian King. ] Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
2019-07-03 07:03:33 +08:00
rcu_lockdep_assert_cblist_protected(rdp);
c = rcu_seq_snap(&rcu_state.gp_seq);
if (!READ_ONCE(rdp->gpwrap) && ULONG_CMP_GE(rdp->gp_seq_needed, c)) {
/* Old request still live, so mark recent callbacks. */
(void)rcu_segcblist_accelerate(&rdp->cblist, c);
return;
}
raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
needwake = rcu_accelerate_cbs(rnp, rdp);
raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
if (needwake)
rcu_gp_kthread_wake();
}
rcu: Tag callback lists with corresponding grace-period number Currently, callbacks are advanced each time the corresponding CPU notices a change in its leaf rcu_node structure's ->completed value (this value counts grace-period completions). This approach has worked quite well, but with the advent of RCU_FAST_NO_HZ, we cannot count on a given CPU seeing all the grace-period completions. When a CPU misses a grace-period completion that occurs while it is in dyntick-idle mode, this will delay invocation of its callbacks. In addition, acceleration of callbacks (when RCU realizes that a given callback need only wait until the end of the next grace period, rather than having to wait for a partial grace period followed by a full grace period) must be carried out extremely carefully. Insufficient acceleration will result in unnecessarily long grace-period latencies, while excessive acceleration will result in premature callback invocation. Changes that involve this tradeoff are therefore among the most nerve-wracking changes to RCU. This commit therefore explicitly tags groups of callbacks with the number of the grace period that they are waiting for. This means that callback-advancement and callback-acceleration functions are idempotent, so that excessive acceleration will merely waste a few CPU cycles. This also allows a CPU to take full advantage of any grace periods that have elapsed while it has been in dyntick-idle mode. It should also enable simulataneous simplifications to and optimizations of RCU_FAST_NO_HZ. Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2012-12-04 05:52:00 +08:00
/*
* Move any callbacks whose grace period has completed to the
* RCU_DONE_TAIL sublist, then compact the remaining sublists and
* assign ->gp_seq numbers to any callbacks in the RCU_NEXT_TAIL
rcu: Tag callback lists with corresponding grace-period number Currently, callbacks are advanced each time the corresponding CPU notices a change in its leaf rcu_node structure's ->completed value (this value counts grace-period completions). This approach has worked quite well, but with the advent of RCU_FAST_NO_HZ, we cannot count on a given CPU seeing all the grace-period completions. When a CPU misses a grace-period completion that occurs while it is in dyntick-idle mode, this will delay invocation of its callbacks. In addition, acceleration of callbacks (when RCU realizes that a given callback need only wait until the end of the next grace period, rather than having to wait for a partial grace period followed by a full grace period) must be carried out extremely carefully. Insufficient acceleration will result in unnecessarily long grace-period latencies, while excessive acceleration will result in premature callback invocation. Changes that involve this tradeoff are therefore among the most nerve-wracking changes to RCU. This commit therefore explicitly tags groups of callbacks with the number of the grace period that they are waiting for. This means that callback-advancement and callback-acceleration functions are idempotent, so that excessive acceleration will merely waste a few CPU cycles. This also allows a CPU to take full advantage of any grace periods that have elapsed while it has been in dyntick-idle mode. It should also enable simulataneous simplifications to and optimizations of RCU_FAST_NO_HZ. Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2012-12-04 05:52:00 +08:00
* sublist. This function is idempotent, so it does not hurt to
* invoke it repeatedly. As long as it is not invoked -too- often...
rcu: Make callers awaken grace-period kthread The rcu_start_gp_advanced() function currently uses irq_work_queue() to defer wakeups of the RCU grace-period kthread. This deferring is necessary to avoid RCU-scheduler deadlocks involving the rcu_node structure's lock, meaning that RCU cannot call any of the scheduler's wake-up functions while holding one of these locks. Unfortunately, the second and subsequent calls to irq_work_queue() are ignored, and the first call will be ignored (aside from queuing the work item) if the scheduler-clock tick is turned off. This is OK for many uses, especially those where irq_work_queue() is called from an interrupt or softirq handler, because in those cases the scheduler-clock-tick state will be re-evaluated, which will turn the scheduler-clock tick back on. On the next tick, any deferred work will then be processed. However, this strategy does not always work for RCU, which can be invoked at process level from idle CPUs. In this case, the tick might never be turned back on, indefinitely defering a grace-period start request. Note that the RCU CPU stall detector cannot see this condition, because there is no RCU grace period in progress. Therefore, we can (and do!) see long tens-of-seconds stalls in grace-period handling. In theory, we could see a full grace-period hang, but rcutorture testing to date has seen only the tens-of-seconds stalls. Event tracing demonstrates that irq_work_queue() is being called repeatedly to no effect during these stalls: The "newreq" event appears repeatedly from a task that is not one of the grace-period kthreads. In theory, irq_work_queue() might be fixed to avoid this sort of issue, but RCU's requirements are unusual and it is quite straightforward to pass wake-up responsibility up through RCU's call chain, so that the wakeup happens when the offending locks are released. This commit therefore makes this change. The rcu_start_gp_advanced(), rcu_start_future_gp(), rcu_accelerate_cbs(), rcu_advance_cbs(), __note_gp_changes(), and rcu_start_gp() functions now return a boolean which indicates when a wake-up is needed. A new rcu_gp_kthread_wake() does the wakeup when it is necessary and safe to do so: No self-wakes, no wake-ups if the ->gp_flags field indicates there is no need (as in someone else did the wake-up before we got around to it), and no wake-ups before the grace-period kthread has been created. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org>
2014-03-12 04:02:16 +08:00
* Returns true if the RCU grace-period kthread needs to be awakened.
rcu: Tag callback lists with corresponding grace-period number Currently, callbacks are advanced each time the corresponding CPU notices a change in its leaf rcu_node structure's ->completed value (this value counts grace-period completions). This approach has worked quite well, but with the advent of RCU_FAST_NO_HZ, we cannot count on a given CPU seeing all the grace-period completions. When a CPU misses a grace-period completion that occurs while it is in dyntick-idle mode, this will delay invocation of its callbacks. In addition, acceleration of callbacks (when RCU realizes that a given callback need only wait until the end of the next grace period, rather than having to wait for a partial grace period followed by a full grace period) must be carried out extremely carefully. Insufficient acceleration will result in unnecessarily long grace-period latencies, while excessive acceleration will result in premature callback invocation. Changes that involve this tradeoff are therefore among the most nerve-wracking changes to RCU. This commit therefore explicitly tags groups of callbacks with the number of the grace period that they are waiting for. This means that callback-advancement and callback-acceleration functions are idempotent, so that excessive acceleration will merely waste a few CPU cycles. This also allows a CPU to take full advantage of any grace periods that have elapsed while it has been in dyntick-idle mode. It should also enable simulataneous simplifications to and optimizations of RCU_FAST_NO_HZ. Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2012-12-04 05:52:00 +08:00
*
* The caller must hold rnp->lock with interrupts disabled.
*/
static bool rcu_advance_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
rcu: Tag callback lists with corresponding grace-period number Currently, callbacks are advanced each time the corresponding CPU notices a change in its leaf rcu_node structure's ->completed value (this value counts grace-period completions). This approach has worked quite well, but with the advent of RCU_FAST_NO_HZ, we cannot count on a given CPU seeing all the grace-period completions. When a CPU misses a grace-period completion that occurs while it is in dyntick-idle mode, this will delay invocation of its callbacks. In addition, acceleration of callbacks (when RCU realizes that a given callback need only wait until the end of the next grace period, rather than having to wait for a partial grace period followed by a full grace period) must be carried out extremely carefully. Insufficient acceleration will result in unnecessarily long grace-period latencies, while excessive acceleration will result in premature callback invocation. Changes that involve this tradeoff are therefore among the most nerve-wracking changes to RCU. This commit therefore explicitly tags groups of callbacks with the number of the grace period that they are waiting for. This means that callback-advancement and callback-acceleration functions are idempotent, so that excessive acceleration will merely waste a few CPU cycles. This also allows a CPU to take full advantage of any grace periods that have elapsed while it has been in dyntick-idle mode. It should also enable simulataneous simplifications to and optimizations of RCU_FAST_NO_HZ. Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2012-12-04 05:52:00 +08:00
{
rcu/nocb: Add bypass callback queueing Use of the rcu_data structure's segmented ->cblist for no-CBs CPUs takes advantage of unrelated grace periods, thus reducing the memory footprint in the face of floods of call_rcu() invocations. However, the ->cblist field is a more-complex rcu_segcblist structure which must be protected via locking. Even though there are only three entities which can acquire this lock (the CPU invoking call_rcu(), the no-CBs grace-period kthread, and the no-CBs callbacks kthread), the contention on this lock is excessive under heavy stress. This commit therefore greatly reduces contention by provisioning an rcu_cblist structure field named ->nocb_bypass within the rcu_data structure. Each no-CBs CPU is permitted only a limited number of enqueues onto the ->cblist per jiffy, controlled by a new nocb_nobypass_lim_per_jiffy kernel boot parameter that defaults to about 16 enqueues per millisecond (16 * 1000 / HZ). When that limit is exceeded, the CPU instead enqueues onto the new ->nocb_bypass. The ->nocb_bypass is flushed into the ->cblist every jiffy or when the number of callbacks on ->nocb_bypass exceeds qhimark, whichever happens first. During call_rcu() floods, this flushing is carried out by the CPU during the course of its call_rcu() invocations. However, a CPU could simply stop invoking call_rcu() at any time. The no-CBs grace-period kthread therefore carries out less-aggressive flushing (every few jiffies or when the number of callbacks on ->nocb_bypass exceeds (2 * qhimark), whichever comes first). This means that the no-CBs grace-period kthread cannot be permitted to do unbounded waits while there are callbacks on ->nocb_bypass. A ->nocb_bypass_timer is used to provide the needed wakeups. [ paulmck: Apply Coverity feedback reported by Colin Ian King. ] Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
2019-07-03 07:03:33 +08:00
rcu_lockdep_assert_cblist_protected(rdp);
raw_lockdep_assert_held_rcu_node(rnp);
/* If no pending (not yet ready to invoke) callbacks, nothing to do. */
if (!rcu_segcblist_pend_cbs(&rdp->cblist))
rcu: Make callers awaken grace-period kthread The rcu_start_gp_advanced() function currently uses irq_work_queue() to defer wakeups of the RCU grace-period kthread. This deferring is necessary to avoid RCU-scheduler deadlocks involving the rcu_node structure's lock, meaning that RCU cannot call any of the scheduler's wake-up functions while holding one of these locks. Unfortunately, the second and subsequent calls to irq_work_queue() are ignored, and the first call will be ignored (aside from queuing the work item) if the scheduler-clock tick is turned off. This is OK for many uses, especially those where irq_work_queue() is called from an interrupt or softirq handler, because in those cases the scheduler-clock-tick state will be re-evaluated, which will turn the scheduler-clock tick back on. On the next tick, any deferred work will then be processed. However, this strategy does not always work for RCU, which can be invoked at process level from idle CPUs. In this case, the tick might never be turned back on, indefinitely defering a grace-period start request. Note that the RCU CPU stall detector cannot see this condition, because there is no RCU grace period in progress. Therefore, we can (and do!) see long tens-of-seconds stalls in grace-period handling. In theory, we could see a full grace-period hang, but rcutorture testing to date has seen only the tens-of-seconds stalls. Event tracing demonstrates that irq_work_queue() is being called repeatedly to no effect during these stalls: The "newreq" event appears repeatedly from a task that is not one of the grace-period kthreads. In theory, irq_work_queue() might be fixed to avoid this sort of issue, but RCU's requirements are unusual and it is quite straightforward to pass wake-up responsibility up through RCU's call chain, so that the wakeup happens when the offending locks are released. This commit therefore makes this change. The rcu_start_gp_advanced(), rcu_start_future_gp(), rcu_accelerate_cbs(), rcu_advance_cbs(), __note_gp_changes(), and rcu_start_gp() functions now return a boolean which indicates when a wake-up is needed. A new rcu_gp_kthread_wake() does the wakeup when it is necessary and safe to do so: No self-wakes, no wake-ups if the ->gp_flags field indicates there is no need (as in someone else did the wake-up before we got around to it), and no wake-ups before the grace-period kthread has been created. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org>
2014-03-12 04:02:16 +08:00
return false;
rcu: Tag callback lists with corresponding grace-period number Currently, callbacks are advanced each time the corresponding CPU notices a change in its leaf rcu_node structure's ->completed value (this value counts grace-period completions). This approach has worked quite well, but with the advent of RCU_FAST_NO_HZ, we cannot count on a given CPU seeing all the grace-period completions. When a CPU misses a grace-period completion that occurs while it is in dyntick-idle mode, this will delay invocation of its callbacks. In addition, acceleration of callbacks (when RCU realizes that a given callback need only wait until the end of the next grace period, rather than having to wait for a partial grace period followed by a full grace period) must be carried out extremely carefully. Insufficient acceleration will result in unnecessarily long grace-period latencies, while excessive acceleration will result in premature callback invocation. Changes that involve this tradeoff are therefore among the most nerve-wracking changes to RCU. This commit therefore explicitly tags groups of callbacks with the number of the grace period that they are waiting for. This means that callback-advancement and callback-acceleration functions are idempotent, so that excessive acceleration will merely waste a few CPU cycles. This also allows a CPU to take full advantage of any grace periods that have elapsed while it has been in dyntick-idle mode. It should also enable simulataneous simplifications to and optimizations of RCU_FAST_NO_HZ. Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2012-12-04 05:52:00 +08:00
/*
* Find all callbacks whose ->gp_seq numbers indicate that they
rcu: Tag callback lists with corresponding grace-period number Currently, callbacks are advanced each time the corresponding CPU notices a change in its leaf rcu_node structure's ->completed value (this value counts grace-period completions). This approach has worked quite well, but with the advent of RCU_FAST_NO_HZ, we cannot count on a given CPU seeing all the grace-period completions. When a CPU misses a grace-period completion that occurs while it is in dyntick-idle mode, this will delay invocation of its callbacks. In addition, acceleration of callbacks (when RCU realizes that a given callback need only wait until the end of the next grace period, rather than having to wait for a partial grace period followed by a full grace period) must be carried out extremely carefully. Insufficient acceleration will result in unnecessarily long grace-period latencies, while excessive acceleration will result in premature callback invocation. Changes that involve this tradeoff are therefore among the most nerve-wracking changes to RCU. This commit therefore explicitly tags groups of callbacks with the number of the grace period that they are waiting for. This means that callback-advancement and callback-acceleration functions are idempotent, so that excessive acceleration will merely waste a few CPU cycles. This also allows a CPU to take full advantage of any grace periods that have elapsed while it has been in dyntick-idle mode. It should also enable simulataneous simplifications to and optimizations of RCU_FAST_NO_HZ. Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2012-12-04 05:52:00 +08:00
* are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
*/
rcu_segcblist_advance(&rdp->cblist, rnp->gp_seq);
rcu: Tag callback lists with corresponding grace-period number Currently, callbacks are advanced each time the corresponding CPU notices a change in its leaf rcu_node structure's ->completed value (this value counts grace-period completions). This approach has worked quite well, but with the advent of RCU_FAST_NO_HZ, we cannot count on a given CPU seeing all the grace-period completions. When a CPU misses a grace-period completion that occurs while it is in dyntick-idle mode, this will delay invocation of its callbacks. In addition, acceleration of callbacks (when RCU realizes that a given callback need only wait until the end of the next grace period, rather than having to wait for a partial grace period followed by a full grace period) must be carried out extremely carefully. Insufficient acceleration will result in unnecessarily long grace-period latencies, while excessive acceleration will result in premature callback invocation. Changes that involve this tradeoff are therefore among the most nerve-wracking changes to RCU. This commit therefore explicitly tags groups of callbacks with the number of the grace period that they are waiting for. This means that callback-advancement and callback-acceleration functions are idempotent, so that excessive acceleration will merely waste a few CPU cycles. This also allows a CPU to take full advantage of any grace periods that have elapsed while it has been in dyntick-idle mode. It should also enable simulataneous simplifications to and optimizations of RCU_FAST_NO_HZ. Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2012-12-04 05:52:00 +08:00
/* Classify any remaining callbacks. */
return rcu_accelerate_cbs(rnp, rdp);
rcu: Tag callback lists with corresponding grace-period number Currently, callbacks are advanced each time the corresponding CPU notices a change in its leaf rcu_node structure's ->completed value (this value counts grace-period completions). This approach has worked quite well, but with the advent of RCU_FAST_NO_HZ, we cannot count on a given CPU seeing all the grace-period completions. When a CPU misses a grace-period completion that occurs while it is in dyntick-idle mode, this will delay invocation of its callbacks. In addition, acceleration of callbacks (when RCU realizes that a given callback need only wait until the end of the next grace period, rather than having to wait for a partial grace period followed by a full grace period) must be carried out extremely carefully. Insufficient acceleration will result in unnecessarily long grace-period latencies, while excessive acceleration will result in premature callback invocation. Changes that involve this tradeoff are therefore among the most nerve-wracking changes to RCU. This commit therefore explicitly tags groups of callbacks with the number of the grace period that they are waiting for. This means that callback-advancement and callback-acceleration functions are idempotent, so that excessive acceleration will merely waste a few CPU cycles. This also allows a CPU to take full advantage of any grace periods that have elapsed while it has been in dyntick-idle mode. It should also enable simulataneous simplifications to and optimizations of RCU_FAST_NO_HZ. Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2012-12-04 05:52:00 +08:00
}
/*
* Move and classify callbacks, but only if doing so won't require
* that the RCU grace-period kthread be awakened.
*/
static void __maybe_unused rcu_advance_cbs_nowake(struct rcu_node *rnp,
struct rcu_data *rdp)
{
rcu/nocb: Add bypass callback queueing Use of the rcu_data structure's segmented ->cblist for no-CBs CPUs takes advantage of unrelated grace periods, thus reducing the memory footprint in the face of floods of call_rcu() invocations. However, the ->cblist field is a more-complex rcu_segcblist structure which must be protected via locking. Even though there are only three entities which can acquire this lock (the CPU invoking call_rcu(), the no-CBs grace-period kthread, and the no-CBs callbacks kthread), the contention on this lock is excessive under heavy stress. This commit therefore greatly reduces contention by provisioning an rcu_cblist structure field named ->nocb_bypass within the rcu_data structure. Each no-CBs CPU is permitted only a limited number of enqueues onto the ->cblist per jiffy, controlled by a new nocb_nobypass_lim_per_jiffy kernel boot parameter that defaults to about 16 enqueues per millisecond (16 * 1000 / HZ). When that limit is exceeded, the CPU instead enqueues onto the new ->nocb_bypass. The ->nocb_bypass is flushed into the ->cblist every jiffy or when the number of callbacks on ->nocb_bypass exceeds qhimark, whichever happens first. During call_rcu() floods, this flushing is carried out by the CPU during the course of its call_rcu() invocations. However, a CPU could simply stop invoking call_rcu() at any time. The no-CBs grace-period kthread therefore carries out less-aggressive flushing (every few jiffies or when the number of callbacks on ->nocb_bypass exceeds (2 * qhimark), whichever comes first). This means that the no-CBs grace-period kthread cannot be permitted to do unbounded waits while there are callbacks on ->nocb_bypass. A ->nocb_bypass_timer is used to provide the needed wakeups. [ paulmck: Apply Coverity feedback reported by Colin Ian King. ] Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
2019-07-03 07:03:33 +08:00
rcu_lockdep_assert_cblist_protected(rdp);
if (!rcu_seq_state(rcu_seq_current(&rnp->gp_seq)) || !raw_spin_trylock_rcu_node(rnp))
return;
// The grace period cannot end while we hold the rcu_node lock.
if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))
WARN_ON_ONCE(rcu_advance_cbs(rnp, rdp));
raw_spin_unlock_rcu_node(rnp);
}
/*
* In CONFIG_RCU_STRICT_GRACE_PERIOD=y kernels, attempt to generate a
* quiescent state. This is intended to be invoked when the CPU notices
* a new grace period.
*/
static void rcu_strict_gp_check_qs(void)
{
if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) {
rcu_read_lock();
rcu_read_unlock();
}
}
/*
* Update CPU-local rcu_data state to record the beginnings and ends of
* grace periods. The caller must hold the ->lock of the leaf rcu_node
* structure corresponding to the current CPU, and must have irqs disabled.
rcu: Make callers awaken grace-period kthread The rcu_start_gp_advanced() function currently uses irq_work_queue() to defer wakeups of the RCU grace-period kthread. This deferring is necessary to avoid RCU-scheduler deadlocks involving the rcu_node structure's lock, meaning that RCU cannot call any of the scheduler's wake-up functions while holding one of these locks. Unfortunately, the second and subsequent calls to irq_work_queue() are ignored, and the first call will be ignored (aside from queuing the work item) if the scheduler-clock tick is turned off. This is OK for many uses, especially those where irq_work_queue() is called from an interrupt or softirq handler, because in those cases the scheduler-clock-tick state will be re-evaluated, which will turn the scheduler-clock tick back on. On the next tick, any deferred work will then be processed. However, this strategy does not always work for RCU, which can be invoked at process level from idle CPUs. In this case, the tick might never be turned back on, indefinitely defering a grace-period start request. Note that the RCU CPU stall detector cannot see this condition, because there is no RCU grace period in progress. Therefore, we can (and do!) see long tens-of-seconds stalls in grace-period handling. In theory, we could see a full grace-period hang, but rcutorture testing to date has seen only the tens-of-seconds stalls. Event tracing demonstrates that irq_work_queue() is being called repeatedly to no effect during these stalls: The "newreq" event appears repeatedly from a task that is not one of the grace-period kthreads. In theory, irq_work_queue() might be fixed to avoid this sort of issue, but RCU's requirements are unusual and it is quite straightforward to pass wake-up responsibility up through RCU's call chain, so that the wakeup happens when the offending locks are released. This commit therefore makes this change. The rcu_start_gp_advanced(), rcu_start_future_gp(), rcu_accelerate_cbs(), rcu_advance_cbs(), __note_gp_changes(), and rcu_start_gp() functions now return a boolean which indicates when a wake-up is needed. A new rcu_gp_kthread_wake() does the wakeup when it is necessary and safe to do so: No self-wakes, no wake-ups if the ->gp_flags field indicates there is no need (as in someone else did the wake-up before we got around to it), and no wake-ups before the grace-period kthread has been created. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org>
2014-03-12 04:02:16 +08:00
* Returns true if the grace-period kthread needs to be awakened.
*/
static bool __note_gp_changes(struct rcu_node *rnp, struct rcu_data *rdp)
{
bool ret = false;
bool need_qs;
const bool offloaded = rcu_rdp_is_offloaded(rdp);
rcu: Make callers awaken grace-period kthread The rcu_start_gp_advanced() function currently uses irq_work_queue() to defer wakeups of the RCU grace-period kthread. This deferring is necessary to avoid RCU-scheduler deadlocks involving the rcu_node structure's lock, meaning that RCU cannot call any of the scheduler's wake-up functions while holding one of these locks. Unfortunately, the second and subsequent calls to irq_work_queue() are ignored, and the first call will be ignored (aside from queuing the work item) if the scheduler-clock tick is turned off. This is OK for many uses, especially those where irq_work_queue() is called from an interrupt or softirq handler, because in those cases the scheduler-clock-tick state will be re-evaluated, which will turn the scheduler-clock tick back on. On the next tick, any deferred work will then be processed. However, this strategy does not always work for RCU, which can be invoked at process level from idle CPUs. In this case, the tick might never be turned back on, indefinitely defering a grace-period start request. Note that the RCU CPU stall detector cannot see this condition, because there is no RCU grace period in progress. Therefore, we can (and do!) see long tens-of-seconds stalls in grace-period handling. In theory, we could see a full grace-period hang, but rcutorture testing to date has seen only the tens-of-seconds stalls. Event tracing demonstrates that irq_work_queue() is being called repeatedly to no effect during these stalls: The "newreq" event appears repeatedly from a task that is not one of the grace-period kthreads. In theory, irq_work_queue() might be fixed to avoid this sort of issue, but RCU's requirements are unusual and it is quite straightforward to pass wake-up responsibility up through RCU's call chain, so that the wakeup happens when the offending locks are released. This commit therefore makes this change. The rcu_start_gp_advanced(), rcu_start_future_gp(), rcu_accelerate_cbs(), rcu_advance_cbs(), __note_gp_changes(), and rcu_start_gp() functions now return a boolean which indicates when a wake-up is needed. A new rcu_gp_kthread_wake() does the wakeup when it is necessary and safe to do so: No self-wakes, no wake-ups if the ->gp_flags field indicates there is no need (as in someone else did the wake-up before we got around to it), and no wake-ups before the grace-period kthread has been created. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org>
2014-03-12 04:02:16 +08:00
raw_lockdep_assert_held_rcu_node(rnp);
if (rdp->gp_seq == rnp->gp_seq)
return false; /* Nothing to do. */
/* Handle the ends of any preceding grace periods first. */
if (rcu_seq_completed_gp(rdp->gp_seq, rnp->gp_seq) ||
unlikely(READ_ONCE(rdp->gpwrap))) {
if (!offloaded)
ret = rcu_advance_cbs(rnp, rdp); /* Advance CBs. */
rdp->core_needs_qs = false;
trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuend"));
} else {
if (!offloaded)
ret = rcu_accelerate_cbs(rnp, rdp); /* Recent CBs. */
if (rdp->core_needs_qs)
rdp->core_needs_qs = !!(rnp->qsmask & rdp->grpmask);
}
/* Now handle the beginnings of any new-to-this-CPU grace periods. */
if (rcu_seq_new_gp(rdp->gp_seq, rnp->gp_seq) ||
unlikely(READ_ONCE(rdp->gpwrap))) {
/*
* If the current grace period is waiting for this CPU,
* set up to detect a quiescent state, otherwise don't
* go looking for one.
*/
trace_rcu_grace_period(rcu_state.name, rnp->gp_seq, TPS("cpustart"));
need_qs = !!(rnp->qsmask & rdp->grpmask);
rdp->cpu_no_qs.b.norm = need_qs;
rdp->core_needs_qs = need_qs;
zero_cpu_stall_ticks(rdp);
}
rdp->gp_seq = rnp->gp_seq; /* Remember new grace-period state. */
if (ULONG_CMP_LT(rdp->gp_seq_needed, rnp->gp_seq_needed) || rdp->gpwrap)
WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed);
if (IS_ENABLED(CONFIG_PROVE_RCU) && READ_ONCE(rdp->gpwrap))
WRITE_ONCE(rdp->last_sched_clock, jiffies);
WRITE_ONCE(rdp->gpwrap, false);
rcu_gpnum_ovf(rnp, rdp);
rcu: Make callers awaken grace-period kthread The rcu_start_gp_advanced() function currently uses irq_work_queue() to defer wakeups of the RCU grace-period kthread. This deferring is necessary to avoid RCU-scheduler deadlocks involving the rcu_node structure's lock, meaning that RCU cannot call any of the scheduler's wake-up functions while holding one of these locks. Unfortunately, the second and subsequent calls to irq_work_queue() are ignored, and the first call will be ignored (aside from queuing the work item) if the scheduler-clock tick is turned off. This is OK for many uses, especially those where irq_work_queue() is called from an interrupt or softirq handler, because in those cases the scheduler-clock-tick state will be re-evaluated, which will turn the scheduler-clock tick back on. On the next tick, any deferred work will then be processed. However, this strategy does not always work for RCU, which can be invoked at process level from idle CPUs. In this case, the tick might never be turned back on, indefinitely defering a grace-period start request. Note that the RCU CPU stall detector cannot see this condition, because there is no RCU grace period in progress. Therefore, we can (and do!) see long tens-of-seconds stalls in grace-period handling. In theory, we could see a full grace-period hang, but rcutorture testing to date has seen only the tens-of-seconds stalls. Event tracing demonstrates that irq_work_queue() is being called repeatedly to no effect during these stalls: The "newreq" event appears repeatedly from a task that is not one of the grace-period kthreads. In theory, irq_work_queue() might be fixed to avoid this sort of issue, but RCU's requirements are unusual and it is quite straightforward to pass wake-up responsibility up through RCU's call chain, so that the wakeup happens when the offending locks are released. This commit therefore makes this change. The rcu_start_gp_advanced(), rcu_start_future_gp(), rcu_accelerate_cbs(), rcu_advance_cbs(), __note_gp_changes(), and rcu_start_gp() functions now return a boolean which indicates when a wake-up is needed. A new rcu_gp_kthread_wake() does the wakeup when it is necessary and safe to do so: No self-wakes, no wake-ups if the ->gp_flags field indicates there is no need (as in someone else did the wake-up before we got around to it), and no wake-ups before the grace-period kthread has been created. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org>
2014-03-12 04:02:16 +08:00
return ret;
}
static void note_gp_changes(struct rcu_data *rdp)
{
unsigned long flags;
rcu: Make callers awaken grace-period kthread The rcu_start_gp_advanced() function currently uses irq_work_queue() to defer wakeups of the RCU grace-period kthread. This deferring is necessary to avoid RCU-scheduler deadlocks involving the rcu_node structure's lock, meaning that RCU cannot call any of the scheduler's wake-up functions while holding one of these locks. Unfortunately, the second and subsequent calls to irq_work_queue() are ignored, and the first call will be ignored (aside from queuing the work item) if the scheduler-clock tick is turned off. This is OK for many uses, especially those where irq_work_queue() is called from an interrupt or softirq handler, because in those cases the scheduler-clock-tick state will be re-evaluated, which will turn the scheduler-clock tick back on. On the next tick, any deferred work will then be processed. However, this strategy does not always work for RCU, which can be invoked at process level from idle CPUs. In this case, the tick might never be turned back on, indefinitely defering a grace-period start request. Note that the RCU CPU stall detector cannot see this condition, because there is no RCU grace period in progress. Therefore, we can (and do!) see long tens-of-seconds stalls in grace-period handling. In theory, we could see a full grace-period hang, but rcutorture testing to date has seen only the tens-of-seconds stalls. Event tracing demonstrates that irq_work_queue() is being called repeatedly to no effect during these stalls: The "newreq" event appears repeatedly from a task that is not one of the grace-period kthreads. In theory, irq_work_queue() might be fixed to avoid this sort of issue, but RCU's requirements are unusual and it is quite straightforward to pass wake-up responsibility up through RCU's call chain, so that the wakeup happens when the offending locks are released. This commit therefore makes this change. The rcu_start_gp_advanced(), rcu_start_future_gp(), rcu_accelerate_cbs(), rcu_advance_cbs(), __note_gp_changes(), and rcu_start_gp() functions now return a boolean which indicates when a wake-up is needed. A new rcu_gp_kthread_wake() does the wakeup when it is necessary and safe to do so: No self-wakes, no wake-ups if the ->gp_flags field indicates there is no need (as in someone else did the wake-up before we got around to it), and no wake-ups before the grace-period kthread has been created. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org>
2014-03-12 04:02:16 +08:00
bool needwake;
struct rcu_node *rnp;
local_irq_save(flags);
rnp = rdp->mynode;
if ((rdp->gp_seq == rcu_seq_current(&rnp->gp_seq) &&
!unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
!raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
local_irq_restore(flags);
return;
}
needwake = __note_gp_changes(rnp, rdp);
raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
rcu_strict_gp_check_qs();
rcu: Make callers awaken grace-period kthread The rcu_start_gp_advanced() function currently uses irq_work_queue() to defer wakeups of the RCU grace-period kthread. This deferring is necessary to avoid RCU-scheduler deadlocks involving the rcu_node structure's lock, meaning that RCU cannot call any of the scheduler's wake-up functions while holding one of these locks. Unfortunately, the second and subsequent calls to irq_work_queue() are ignored, and the first call will be ignored (aside from queuing the work item) if the scheduler-clock tick is turned off. This is OK for many uses, especially those where irq_work_queue() is called from an interrupt or softirq handler, because in those cases the scheduler-clock-tick state will be re-evaluated, which will turn the scheduler-clock tick back on. On the next tick, any deferred work will then be processed. However, this strategy does not always work for RCU, which can be invoked at process level from idle CPUs. In this case, the tick might never be turned back on, indefinitely defering a grace-period start request. Note that the RCU CPU stall detector cannot see this condition, because there is no RCU grace period in progress. Therefore, we can (and do!) see long tens-of-seconds stalls in grace-period handling. In theory, we could see a full grace-period hang, but rcutorture testing to date has seen only the tens-of-seconds stalls. Event tracing demonstrates that irq_work_queue() is being called repeatedly to no effect during these stalls: The "newreq" event appears repeatedly from a task that is not one of the grace-period kthreads. In theory, irq_work_queue() might be fixed to avoid this sort of issue, but RCU's requirements are unusual and it is quite straightforward to pass wake-up responsibility up through RCU's call chain, so that the wakeup happens when the offending locks are released. This commit therefore makes this change. The rcu_start_gp_advanced(), rcu_start_future_gp(), rcu_accelerate_cbs(), rcu_advance_cbs(), __note_gp_changes(), and rcu_start_gp() functions now return a boolean which indicates when a wake-up is needed. A new rcu_gp_kthread_wake() does the wakeup when it is necessary and safe to do so: No self-wakes, no wake-ups if the ->gp_flags field indicates there is no need (as in someone else did the wake-up before we got around to it), and no wake-ups before the grace-period kthread has been created. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org>
2014-03-12 04:02:16 +08:00
if (needwake)
rcu_gp_kthread_wake();
}
static atomic_t *rcu_gp_slow_suppress;
/* Register a counter to suppress debugging grace-period delays. */
void rcu_gp_slow_register(atomic_t *rgssp)
{
WARN_ON_ONCE(rcu_gp_slow_suppress);
WRITE_ONCE(rcu_gp_slow_suppress, rgssp);
}
EXPORT_SYMBOL_GPL(rcu_gp_slow_register);
/* Unregister a counter, with NULL for not caring which. */
void rcu_gp_slow_unregister(atomic_t *rgssp)
{
WARN_ON_ONCE(rgssp && rgssp != rcu_gp_slow_suppress);
WRITE_ONCE(rcu_gp_slow_suppress, NULL);
}
EXPORT_SYMBOL_GPL(rcu_gp_slow_unregister);
static bool rcu_gp_slow_is_suppressed(void)
{
atomic_t *rgssp = READ_ONCE(rcu_gp_slow_suppress);
return rgssp && atomic_read(rgssp);
}
static void rcu_gp_slow(int delay)
{
if (!rcu_gp_slow_is_suppressed() && delay > 0 &&
!(rcu_seq_ctr(rcu_state.gp_seq) % (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
schedule_timeout_idle(delay);
}
static unsigned long sleep_duration;
/* Allow rcutorture to stall the grace-period kthread. */
void rcu_gp_set_torture_wait(int duration)
{
if (IS_ENABLED(CONFIG_RCU_TORTURE_TEST) && duration > 0)
WRITE_ONCE(sleep_duration, duration);
}
EXPORT_SYMBOL_GPL(rcu_gp_set_torture_wait);
/* Actually implement the aforementioned wait. */
static void rcu_gp_torture_wait(void)
{
unsigned long duration;
if (!IS_ENABLED(CONFIG_RCU_TORTURE_TEST))
return;
duration = xchg(&sleep_duration, 0UL);
if (duration > 0) {
pr_alert("%s: Waiting %lu jiffies\n", __func__, duration);
schedule_timeout_idle(duration);
pr_alert("%s: Wait complete\n", __func__);
}
}
/*
* Handler for on_each_cpu() to invoke the target CPU's RCU core
* processing.
*/
static void rcu_strict_gp_boundary(void *unused)
{
invoke_rcu_core();
}
// Make the polled API aware of the beginning of a grace period.
static void rcu_poll_gp_seq_start(unsigned long *snap)
{
struct rcu_node *rnp = rcu_get_root();
if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE)
raw_lockdep_assert_held_rcu_node(rnp);
// If RCU was idle, note beginning of GP.
if (!rcu_seq_state(rcu_state.gp_seq_polled))
rcu_seq_start(&rcu_state.gp_seq_polled);
// Either way, record current state.
*snap = rcu_state.gp_seq_polled;
}
// Make the polled API aware of the end of a grace period.
static void rcu_poll_gp_seq_end(unsigned long *snap)
{
struct rcu_node *rnp = rcu_get_root();
if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE)
raw_lockdep_assert_held_rcu_node(rnp);
// If the previously noted GP is still in effect, record the
// end of that GP. Either way, zero counter to avoid counter-wrap
// problems.
if (*snap && *snap == rcu_state.gp_seq_polled) {
rcu_seq_end(&rcu_state.gp_seq_polled);
rcu_state.gp_seq_polled_snap = 0;
rcu: Make polled grace-period API account for expedited grace periods Currently, this code could splat: oldstate = get_state_synchronize_rcu(); synchronize_rcu_expedited(); WARN_ON_ONCE(!poll_state_synchronize_rcu(oldstate)); This situation is counter-intuitive and user-unfriendly. After all, there really was a perfectly valid full grace period right after the call to get_state_synchronize_rcu(), so why shouldn't poll_state_synchronize_rcu() know about it? This commit therefore makes the polled grace-period API aware of expedited grace periods in addition to the normal grace periods that it is already aware of. With this change, the above code is guaranteed not to splat. Please note that the above code can still splat due to counter wrap on the one hand and situations involving partially overlapping normal/expedited grace periods on the other. On 64-bit systems, the second is of course much more likely than the first. It is possible to modify this approach to prevent overlapping grace periods from causing splats, but only at the expense of greatly increasing the probability of counter wrap, as in within milliseconds on 32-bit systems and within minutes on 64-bit systems. This commit is in preparation for polled expedited grace periods. Link: https://lore.kernel.org/all/20220121142454.1994916-1-bfoster@redhat.com/ Link: https://docs.google.com/document/d/1RNKWW9jQyfjxw2E8dsXVTdvZYh0HnYeSHDKog9jhdN8/edit?usp=sharing Cc: Brian Foster <bfoster@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Ian Kent <raven@themaw.net> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2022-04-14 21:56:35 +08:00
rcu_state.gp_seq_polled_exp_snap = 0;
} else {
*snap = 0;
}
}
// Make the polled API aware of the beginning of a grace period, but
// where caller does not hold the root rcu_node structure's lock.
static void rcu_poll_gp_seq_start_unlocked(unsigned long *snap)
{
unsigned long flags;
struct rcu_node *rnp = rcu_get_root();
if (rcu_init_invoked()) {
if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE)
lockdep_assert_irqs_enabled();
raw_spin_lock_irqsave_rcu_node(rnp, flags);
}
rcu_poll_gp_seq_start(snap);
if (rcu_init_invoked())
raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
}
// Make the polled API aware of the end of a grace period, but where
// caller does not hold the root rcu_node structure's lock.
static void rcu_poll_gp_seq_end_unlocked(unsigned long *snap)
{
unsigned long flags;
struct rcu_node *rnp = rcu_get_root();
if (rcu_init_invoked()) {
if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE)
lockdep_assert_irqs_enabled();
raw_spin_lock_irqsave_rcu_node(rnp, flags);
}
rcu_poll_gp_seq_end(snap);
if (rcu_init_invoked())
raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
}
/*
* Initialize a new grace period. Return false if no grace period required.
*/
static noinline_for_stack bool rcu_gp_init(void)
{
rcu: Fix grace-period hangs from mid-init task resume Without special fail-safe quiescent-state-propagation checks, grace-period hangs can result from the following scenario: 1. A task running on a given CPU is preempted in its RCU read-side critical section. 2. That CPU goes offline, and there are now no online CPUs corresponding to that CPU's leaf rcu_node structure. 3. The rcu_gp_init() function does the first phase of grace-period initialization, and sets the aforementioned leaf rcu_node structure's ->qsmaskinit field to all zeroes. Because there is a blocked task, it does not propagate the zeroing of either ->qsmaskinit or ->qsmaskinitnext up the rcu_node tree. 4. The task resumes on some other CPU and exits its critical section. There is no grace period in progress, so the resulting quiescent state is not reported up the tree. 5. The rcu_gp_init() function does the second phase of grace-period initialization, which results in the leaf rcu_node structure being initialized to expect no further quiescent states, but with that structure's parent expecting a quiescent-state report. The parent will never receive a quiescent state from this leaf rcu_node structure, so the grace period will hang, resulting in RCU CPU stall warnings. It would be good to get rid of the special fail-safe quiescent-state propagation checks. This commit therefore checks the leaf rcu_node structure's ->wait_blkd_tasks field during grace-period initialization. If this flag is set, the rcu_report_qs_rnp() is invoked to immediately report the possible quiescent state. While in the neighborhood, this commit also report quiescent states for any CPUs that went offline between the two phases of grace-period initialization, thus reducing grace-period delays and hopefully eventually allowing removal of offline-CPU checks from the force-quiescent-state code path. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2018-05-08 00:34:17 +08:00
unsigned long flags;
rcu: Process offlining and onlining only at grace-period start Races between CPU hotplug and grace periods can be difficult to resolve, so the ->onoff_mutex is used to exclude the two events. Unfortunately, this means that it is impossible for an outgoing CPU to perform the last bits of its offlining from its last pass through the idle loop, because sleeplocks cannot be acquired in that context. This commit avoids these problems by buffering online and offline events in a new ->qsmaskinitnext field in the leaf rcu_node structures. When a grace period starts, the events accumulated in this mask are applied to the ->qsmaskinit field, and, if needed, up the rcu_node tree. The special case of all CPUs corresponding to a given leaf rcu_node structure being offline while there are still elements in that structure's ->blkd_tasks list is handled using a new ->wait_blkd_tasks field. In this case, propagating the offline bits up the tree is deferred until the beginning of the grace period after all of the tasks have exited their RCU read-side critical sections and removed themselves from the list, at which point the ->wait_blkd_tasks flag is cleared. If one of that leaf rcu_node structure's CPUs comes back online before the list empties, then the ->wait_blkd_tasks flag is simply cleared. This of course means that RCU's notion of which CPUs are offline can be out of date. This is OK because RCU need only wait on CPUs that were online at the time that the grace period started. In addition, RCU's force-quiescent-state actions will handle the case where a CPU goes offline after the grace period starts. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2015-01-24 13:52:37 +08:00
unsigned long oldmask;
rcu: Fix grace-period hangs from mid-init task resume Without special fail-safe quiescent-state-propagation checks, grace-period hangs can result from the following scenario: 1. A task running on a given CPU is preempted in its RCU read-side critical section. 2. That CPU goes offline, and there are now no online CPUs corresponding to that CPU's leaf rcu_node structure. 3. The rcu_gp_init() function does the first phase of grace-period initialization, and sets the aforementioned leaf rcu_node structure's ->qsmaskinit field to all zeroes. Because there is a blocked task, it does not propagate the zeroing of either ->qsmaskinit or ->qsmaskinitnext up the rcu_node tree. 4. The task resumes on some other CPU and exits its critical section. There is no grace period in progress, so the resulting quiescent state is not reported up the tree. 5. The rcu_gp_init() function does the second phase of grace-period initialization, which results in the leaf rcu_node structure being initialized to expect no further quiescent states, but with that structure's parent expecting a quiescent-state report. The parent will never receive a quiescent state from this leaf rcu_node structure, so the grace period will hang, resulting in RCU CPU stall warnings. It would be good to get rid of the special fail-safe quiescent-state propagation checks. This commit therefore checks the leaf rcu_node structure's ->wait_blkd_tasks field during grace-period initialization. If this flag is set, the rcu_report_qs_rnp() is invoked to immediately report the possible quiescent state. While in the neighborhood, this commit also report quiescent states for any CPUs that went offline between the two phases of grace-period initialization, thus reducing grace-period delays and hopefully eventually allowing removal of offline-CPU checks from the force-quiescent-state code path. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2018-05-08 00:34:17 +08:00
unsigned long mask;
struct rcu_data *rdp;
struct rcu_node *rnp = rcu_get_root();
WRITE_ONCE(rcu_state.gp_activity, jiffies);
raw_spin_lock_irq_rcu_node(rnp);
if (!READ_ONCE(rcu_state.gp_flags)) {
/* Spurious wakeup, tell caller to go back to sleep. */
raw_spin_unlock_irq_rcu_node(rnp);
return false;
}
WRITE_ONCE(rcu_state.gp_flags, 0); /* Clear all flags: New GP. */
if (WARN_ON_ONCE(rcu_gp_in_progress())) {
/*
* Grace period already in progress, don't start another.
* Not supposed to be able to happen.
*/
raw_spin_unlock_irq_rcu_node(rnp);
return false;
}
/* Advance to a new grace period and initialize state. */
record_gp_stall_check_time();
/* Record GP times before starting GP, hence rcu_seq_start(). */
rcu_seq_start(&rcu_state.gp_seq);
ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq);
trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start"));
rcu_poll_gp_seq_start(&rcu_state.gp_seq_polled_snap);
raw_spin_unlock_irq_rcu_node(rnp);
rcu: Process offlining and onlining only at grace-period start Races between CPU hotplug and grace periods can be difficult to resolve, so the ->onoff_mutex is used to exclude the two events. Unfortunately, this means that it is impossible for an outgoing CPU to perform the last bits of its offlining from its last pass through the idle loop, because sleeplocks cannot be acquired in that context. This commit avoids these problems by buffering online and offline events in a new ->qsmaskinitnext field in the leaf rcu_node structures. When a grace period starts, the events accumulated in this mask are applied to the ->qsmaskinit field, and, if needed, up the rcu_node tree. The special case of all CPUs corresponding to a given leaf rcu_node structure being offline while there are still elements in that structure's ->blkd_tasks list is handled using a new ->wait_blkd_tasks field. In this case, propagating the offline bits up the tree is deferred until the beginning of the grace period after all of the tasks have exited their RCU read-side critical sections and removed themselves from the list, at which point the ->wait_blkd_tasks flag is cleared. If one of that leaf rcu_node structure's CPUs comes back online before the list empties, then the ->wait_blkd_tasks flag is simply cleared. This of course means that RCU's notion of which CPUs are offline can be out of date. This is OK because RCU need only wait on CPUs that were online at the time that the grace period started. In addition, RCU's force-quiescent-state actions will handle the case where a CPU goes offline after the grace period starts. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2015-01-24 13:52:37 +08:00
/*
* Apply per-leaf buffered online and offline operations to
* the rcu_node tree. Note that this new grace period need not
* wait for subsequent online CPUs, and that RCU hooks in the CPU
* offlining path, when combined with checks in this function,
* will handle CPUs that are currently going offline or that will
* go offline later. Please also refer to "Hotplug CPU" section
* of RCU's Requirements documentation.
rcu: Process offlining and onlining only at grace-period start Races between CPU hotplug and grace periods can be difficult to resolve, so the ->onoff_mutex is used to exclude the two events. Unfortunately, this means that it is impossible for an outgoing CPU to perform the last bits of its offlining from its last pass through the idle loop, because sleeplocks cannot be acquired in that context. This commit avoids these problems by buffering online and offline events in a new ->qsmaskinitnext field in the leaf rcu_node structures. When a grace period starts, the events accumulated in this mask are applied to the ->qsmaskinit field, and, if needed, up the rcu_node tree. The special case of all CPUs corresponding to a given leaf rcu_node structure being offline while there are still elements in that structure's ->blkd_tasks list is handled using a new ->wait_blkd_tasks field. In this case, propagating the offline bits up the tree is deferred until the beginning of the grace period after all of the tasks have exited their RCU read-side critical sections and removed themselves from the list, at which point the ->wait_blkd_tasks flag is cleared. If one of that leaf rcu_node structure's CPUs comes back online before the list empties, then the ->wait_blkd_tasks flag is simply cleared. This of course means that RCU's notion of which CPUs are offline can be out of date. This is OK because RCU need only wait on CPUs that were online at the time that the grace period started. In addition, RCU's force-quiescent-state actions will handle the case where a CPU goes offline after the grace period starts. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2015-01-24 13:52:37 +08:00
*/
rcu: Check and report missed fqs timer wakeup on RCU stall For a new grace period request, the RCU GP kthread transitions through following states: a. [RCU_GP_WAIT_GPS] -> [RCU_GP_DONE_GPS] The RCU_GP_WAIT_GPS state is where the GP kthread waits for a request for a new GP. Once it receives a request (for example, when a new RCU callback is queued), the GP kthread transitions to RCU_GP_DONE_GPS. b. [RCU_GP_DONE_GPS] -> [RCU_GP_ONOFF] Grace period initialization starts in rcu_gp_init(), which records the start of new GP in rcu_state.gp_seq and transitions to RCU_GP_ONOFF. c. [RCU_GP_ONOFF] -> [RCU_GP_INIT] The purpose of the RCU_GP_ONOFF state is to apply the online/offline information that was buffered for any CPUs that recently came online or went offline. This state is maintained in per-leaf rcu_node bitmasks, with the buffered state in ->qsmaskinitnext and the state for the upcoming GP in ->qsmaskinit. At the end of this RCU_GP_ONOFF state, each bit in ->qsmaskinit will correspond to a CPU that must pass through a quiescent state before the upcoming grace period is allowed to complete. However, a leaf rcu_node structure with an all-zeroes ->qsmaskinit cannot necessarily be ignored. In preemptible RCU, there might well be tasks still in RCU read-side critical sections that were first preempted while running on one of the CPUs managed by this structure. Such tasks will be queued on this structure's ->blkd_tasks list. Only after this list fully drains can this leaf rcu_node structure be ignored, and even then only if none of its CPUs have come back online in the meantime. Once that happens, the ->qsmaskinit masks further up the tree will be updated to exclude this leaf rcu_node structure. Once the ->qsmaskinitnext and ->qsmaskinit fields have been updated as needed, the GP kthread transitions to RCU_GP_INIT. d. [RCU_GP_INIT] -> [RCU_GP_WAIT_FQS] The purpose of the RCU_GP_INIT state is to copy each ->qsmaskinit to the ->qsmask field within each rcu_node structure. This copying is done breadth-first from the root to the leaves. Why not just copy directly from ->qsmaskinitnext to ->qsmask? Because the ->qsmaskinitnext masks can change in the meantime as additional CPUs come online or go offline. Such changes would result in inconsistencies in the ->qsmask fields up and down the tree, which could in turn result in too-short grace periods or grace-period hangs. These issues are avoided by snapshotting the leaf rcu_node structures' ->qsmaskinitnext fields into their ->qsmaskinit counterparts, generating a consistent set of ->qsmaskinit fields throughout the tree, and only then copying these consistent ->qsmaskinit fields to their ->qsmask counterparts. Once this initialization step is complete, the GP kthread transitions to RCU_GP_WAIT_FQS, where it waits to do a force-quiescent-state scan on the one hand or for the end of the grace period on the other. e. [RCU_GP_WAIT_FQS] -> [RCU_GP_DOING_FQS] The RCU_GP_WAIT_FQS state waits for one of three things: (1) An explicit request to do a force-quiescent-state scan, (2) The end of the grace period, or (3) A short interval of time, after which it will do a force-quiescent-state (FQS) scan. The explicit request can come from rcutorture or from any CPU that has too many RCU callbacks queued (see the qhimark kernel parameter and the RCU_GP_FLAG_OVLD flag). The aforementioned "short period of time" is specified by the jiffies_till_first_fqs boot parameter for a given grace period's first FQS scan and by the jiffies_till_next_fqs for later FQS scans. Either way, once the wait is over, the GP kthread transitions to RCU_GP_DOING_FQS. f. [RCU_GP_DOING_FQS] -> [RCU_GP_CLEANUP] The RCU_GP_DOING_FQS state performs an FQS scan. Each such scan carries out two functions for any CPU whose bit is still set in its leaf rcu_node structure's ->qsmask field, that is, for any CPU that has not yet reported a quiescent state for the current grace period: i. Report quiescent states on behalf of CPUs that have been observed to be idle (from an RCU perspective) since the beginning of the grace period. ii. If the current grace period is too old, take various actions to encourage holdout CPUs to pass through quiescent states, including enlisting the aid of any calls to cond_resched() and might_sleep(), and even including IPIing the holdout CPUs. These checks are skipped for any leaf rcu_node structure with a all-zero ->qsmask field, however such structures are subject to RCU priority boosting if there are tasks on a given structure blocking the current grace period. The end of the grace period is detected when the root rcu_node structure's ->qsmask is zero and when there are no longer any preempted tasks blocking the current grace period. (No, this last check is not redundant. To see this, consider an rcu_node tree having exactly one structure that serves as both root and leaf.) Once the end of the grace period is detected, the GP kthread transitions to RCU_GP_CLEANUP. g. [RCU_GP_CLEANUP] -> [RCU_GP_CLEANED] The RCU_GP_CLEANUP state marks the end of grace period by updating the rcu_state structure's ->gp_seq field and also all rcu_node structures' ->gp_seq field. As before, the rcu_node tree is traversed in breadth first order. Once this update is complete, the GP kthread transitions to the RCU_GP_CLEANED state. i. [RCU_GP_CLEANED] -> [RCU_GP_INIT] Once in the RCU_GP_CLEANED state, the GP kthread immediately transitions into the RCU_GP_INIT state. j. The role of timers. If there is at least one idle CPU, and if timers are not firing, the transition from RCU_GP_DOING_FQS to RCU_GP_CLEANUP will never happen. Timers can fail to fire for a number of reasons, including issues in timer configuration, issues in the timer framework, and failure to handle softirqs (for example, when there is a storm of interrupts). Whatever the reason, if the timers fail to fire, the GP kthread will never be awakened, resulting in RCU CPU stall warnings and eventually in OOM. However, an RCU CPU stall warning has a large number of potential causes, as documented in Documentation/RCU/stallwarn.rst. This commit therefore adds analysis to the RCU CPU stall-warning code to emit an additional message if the cause of the stall is likely to be timer failure. Signed-off-by: Neeraj Upadhyay <neeraju@codeaurora.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-11-17 00:06:00 +08:00
WRITE_ONCE(rcu_state.gp_state, RCU_GP_ONOFF);
/* Exclude CPU hotplug operations. */
rcu_for_each_leaf_node(rnp) {
local_irq_save(flags);
arch_spin_lock(&rcu_state.ofl_lock);
raw_spin_lock_rcu_node(rnp);
rcu: Process offlining and onlining only at grace-period start Races between CPU hotplug and grace periods can be difficult to resolve, so the ->onoff_mutex is used to exclude the two events. Unfortunately, this means that it is impossible for an outgoing CPU to perform the last bits of its offlining from its last pass through the idle loop, because sleeplocks cannot be acquired in that context. This commit avoids these problems by buffering online and offline events in a new ->qsmaskinitnext field in the leaf rcu_node structures. When a grace period starts, the events accumulated in this mask are applied to the ->qsmaskinit field, and, if needed, up the rcu_node tree. The special case of all CPUs corresponding to a given leaf rcu_node structure being offline while there are still elements in that structure's ->blkd_tasks list is handled using a new ->wait_blkd_tasks field. In this case, propagating the offline bits up the tree is deferred until the beginning of the grace period after all of the tasks have exited their RCU read-side critical sections and removed themselves from the list, at which point the ->wait_blkd_tasks flag is cleared. If one of that leaf rcu_node structure's CPUs comes back online before the list empties, then the ->wait_blkd_tasks flag is simply cleared. This of course means that RCU's notion of which CPUs are offline can be out of date. This is OK because RCU need only wait on CPUs that were online at the time that the grace period started. In addition, RCU's force-quiescent-state actions will handle the case where a CPU goes offline after the grace period starts. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2015-01-24 13:52:37 +08:00
if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
!rnp->wait_blkd_tasks) {
/* Nothing to do on this leaf rcu_node structure. */
raw_spin_unlock_rcu_node(rnp);
arch_spin_unlock(&rcu_state.ofl_lock);
local_irq_restore(flags);
rcu: Process offlining and onlining only at grace-period start Races between CPU hotplug and grace periods can be difficult to resolve, so the ->onoff_mutex is used to exclude the two events. Unfortunately, this means that it is impossible for an outgoing CPU to perform the last bits of its offlining from its last pass through the idle loop, because sleeplocks cannot be acquired in that context. This commit avoids these problems by buffering online and offline events in a new ->qsmaskinitnext field in the leaf rcu_node structures. When a grace period starts, the events accumulated in this mask are applied to the ->qsmaskinit field, and, if needed, up the rcu_node tree. The special case of all CPUs corresponding to a given leaf rcu_node structure being offline while there are still elements in that structure's ->blkd_tasks list is handled using a new ->wait_blkd_tasks field. In this case, propagating the offline bits up the tree is deferred until the beginning of the grace period after all of the tasks have exited their RCU read-side critical sections and removed themselves from the list, at which point the ->wait_blkd_tasks flag is cleared. If one of that leaf rcu_node structure's CPUs comes back online before the list empties, then the ->wait_blkd_tasks flag is simply cleared. This of course means that RCU's notion of which CPUs are offline can be out of date. This is OK because RCU need only wait on CPUs that were online at the time that the grace period started. In addition, RCU's force-quiescent-state actions will handle the case where a CPU goes offline after the grace period starts. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2015-01-24 13:52:37 +08:00
continue;
}
/* Record old state, apply changes to ->qsmaskinit field. */
oldmask = rnp->qsmaskinit;
rnp->qsmaskinit = rnp->qsmaskinitnext;
/* If zero-ness of ->qsmaskinit changed, propagate up tree. */
if (!oldmask != !rnp->qsmaskinit) {
rcu: Clean up handling of tasks blocked across full-rcu_node offline Commit 0aa04b055e71 ("rcu: Process offlining and onlining only at grace-period start") deferred handling of CPU-hotplug events until the start of the next grace period, but consider the following sequence of events: 1. A task is preempted within an RCU-preempt read-side critical section. 2. The CPU that this task was running on goes offline, along with all other CPUs sharing the corresponding leaf rcu_node structure. 3. The task resumes execution. 4. One of those CPUs comes back online before a new grace period starts. In step 2, the code in the next rcu_gp_init() invocation will (correctly) defer removing the leaf rcu_node structure from the upper-level bitmasks, and will (correctly) set that structure's ->wait_blkd_tasks field. During the ensuing interval, RCU will (correctly) track the tasks preempted on that structure because they must block any subsequent grace period. In step 3, the code in rcu_read_unlock_special() will (correctly) remove the task from the leaf rcu_node structure. From this point forward, RCU need not pay attention to this structure, at least not until one of the corresponding CPUs comes back online. In step 4, the code in the next rcu_gp_init() invocation will (incorrectly) invoke rcu_init_new_rnp(). This is incorrect because the corresponding rcu_cleanup_dead_rnp() was never invoked. This is nevertheless harmless because the upper-level bits are still set. So, no harm, no foul, right? At least, all is well until a little further into rcu_gp_init() invocation, which will notice that there are no longer any tasks blocked on the leaf rcu_node structure, conclude that there is no longer anything left over from step 2's offline operation, and will therefore invoke rcu_cleanup_dead_rnp(). But this invocation of rcu_cleanup_dead_rnp() is for the beginning of the earlier offline interval, and the previous invocation of rcu_init_new_rnp() is for the end of that same interval. That is right, they are invoked out of order. That cannot be good, can it? It turns out that this is not a (correctness!) problem because rcu_cleanup_dead_rnp() checks to see if any of the corresponding CPUs are online, and refuses to do anything if so. In other words, in the case where rcu_init_new_rnp() and rcu_cleanup_dead_rnp() execute out of order, they both have no effect. But this is at best an accident waiting to happen. This commit therefore adds logic to rcu_gp_init() so that rcu_init_new_rnp() and rcu_cleanup_dead_rnp() are always invoked in order, and so that neither are invoked at all in cases where RCU had to pay attention to the leaf rcu_node structure during the entire time that all corresponding CPUs were offline. And, while in the area, this commit reduces confusion by using formal parameters rather than local variables that just happen to have the same value at that particular point in the code. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2018-05-03 03:49:21 +08:00
if (!oldmask) { /* First online CPU for rcu_node. */
if (!rnp->wait_blkd_tasks) /* Ever offline? */
rcu_init_new_rnp(rnp);
} else if (rcu_preempt_has_tasks(rnp)) {
rnp->wait_blkd_tasks = true; /* blocked tasks */
} else { /* Last offline CPU and can propagate. */
rcu: Process offlining and onlining only at grace-period start Races between CPU hotplug and grace periods can be difficult to resolve, so the ->onoff_mutex is used to exclude the two events. Unfortunately, this means that it is impossible for an outgoing CPU to perform the last bits of its offlining from its last pass through the idle loop, because sleeplocks cannot be acquired in that context. This commit avoids these problems by buffering online and offline events in a new ->qsmaskinitnext field in the leaf rcu_node structures. When a grace period starts, the events accumulated in this mask are applied to the ->qsmaskinit field, and, if needed, up the rcu_node tree. The special case of all CPUs corresponding to a given leaf rcu_node structure being offline while there are still elements in that structure's ->blkd_tasks list is handled using a new ->wait_blkd_tasks field. In this case, propagating the offline bits up the tree is deferred until the beginning of the grace period after all of the tasks have exited their RCU read-side critical sections and removed themselves from the list, at which point the ->wait_blkd_tasks flag is cleared. If one of that leaf rcu_node structure's CPUs comes back online before the list empties, then the ->wait_blkd_tasks flag is simply cleared. This of course means that RCU's notion of which CPUs are offline can be out of date. This is OK because RCU need only wait on CPUs that were online at the time that the grace period started. In addition, RCU's force-quiescent-state actions will handle the case where a CPU goes offline after the grace period starts. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2015-01-24 13:52:37 +08:00
rcu_cleanup_dead_rnp(rnp);
rcu: Clean up handling of tasks blocked across full-rcu_node offline Commit 0aa04b055e71 ("rcu: Process offlining and onlining only at grace-period start") deferred handling of CPU-hotplug events until the start of the next grace period, but consider the following sequence of events: 1. A task is preempted within an RCU-preempt read-side critical section. 2. The CPU that this task was running on goes offline, along with all other CPUs sharing the corresponding leaf rcu_node structure. 3. The task resumes execution. 4. One of those CPUs comes back online before a new grace period starts. In step 2, the code in the next rcu_gp_init() invocation will (correctly) defer removing the leaf rcu_node structure from the upper-level bitmasks, and will (correctly) set that structure's ->wait_blkd_tasks field. During the ensuing interval, RCU will (correctly) track the tasks preempted on that structure because they must block any subsequent grace period. In step 3, the code in rcu_read_unlock_special() will (correctly) remove the task from the leaf rcu_node structure. From this point forward, RCU need not pay attention to this structure, at least not until one of the corresponding CPUs comes back online. In step 4, the code in the next rcu_gp_init() invocation will (incorrectly) invoke rcu_init_new_rnp(). This is incorrect because the corresponding rcu_cleanup_dead_rnp() was never invoked. This is nevertheless harmless because the upper-level bits are still set. So, no harm, no foul, right? At least, all is well until a little further into rcu_gp_init() invocation, which will notice that there are no longer any tasks blocked on the leaf rcu_node structure, conclude that there is no longer anything left over from step 2's offline operation, and will therefore invoke rcu_cleanup_dead_rnp(). But this invocation of rcu_cleanup_dead_rnp() is for the beginning of the earlier offline interval, and the previous invocation of rcu_init_new_rnp() is for the end of that same interval. That is right, they are invoked out of order. That cannot be good, can it? It turns out that this is not a (correctness!) problem because rcu_cleanup_dead_rnp() checks to see if any of the corresponding CPUs are online, and refuses to do anything if so. In other words, in the case where rcu_init_new_rnp() and rcu_cleanup_dead_rnp() execute out of order, they both have no effect. But this is at best an accident waiting to happen. This commit therefore adds logic to rcu_gp_init() so that rcu_init_new_rnp() and rcu_cleanup_dead_rnp() are always invoked in order, and so that neither are invoked at all in cases where RCU had to pay attention to the leaf rcu_node structure during the entire time that all corresponding CPUs were offline. And, while in the area, this commit reduces confusion by using formal parameters rather than local variables that just happen to have the same value at that particular point in the code. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2018-05-03 03:49:21 +08:00
}
rcu: Process offlining and onlining only at grace-period start Races between CPU hotplug and grace periods can be difficult to resolve, so the ->onoff_mutex is used to exclude the two events. Unfortunately, this means that it is impossible for an outgoing CPU to perform the last bits of its offlining from its last pass through the idle loop, because sleeplocks cannot be acquired in that context. This commit avoids these problems by buffering online and offline events in a new ->qsmaskinitnext field in the leaf rcu_node structures. When a grace period starts, the events accumulated in this mask are applied to the ->qsmaskinit field, and, if needed, up the rcu_node tree. The special case of all CPUs corresponding to a given leaf rcu_node structure being offline while there are still elements in that structure's ->blkd_tasks list is handled using a new ->wait_blkd_tasks field. In this case, propagating the offline bits up the tree is deferred until the beginning of the grace period after all of the tasks have exited their RCU read-side critical sections and removed themselves from the list, at which point the ->wait_blkd_tasks flag is cleared. If one of that leaf rcu_node structure's CPUs comes back online before the list empties, then the ->wait_blkd_tasks flag is simply cleared. This of course means that RCU's notion of which CPUs are offline can be out of date. This is OK because RCU need only wait on CPUs that were online at the time that the grace period started. In addition, RCU's force-quiescent-state actions will handle the case where a CPU goes offline after the grace period starts. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2015-01-24 13:52:37 +08:00
}
/*
* If all waited-on tasks from prior grace period are
* done, and if all this rcu_node structure's CPUs are
* still offline, propagate up the rcu_node tree and
* clear ->wait_blkd_tasks. Otherwise, if one of this
* rcu_node structure's CPUs has since come back online,
rcu: Clean up handling of tasks blocked across full-rcu_node offline Commit 0aa04b055e71 ("rcu: Process offlining and onlining only at grace-period start") deferred handling of CPU-hotplug events until the start of the next grace period, but consider the following sequence of events: 1. A task is preempted within an RCU-preempt read-side critical section. 2. The CPU that this task was running on goes offline, along with all other CPUs sharing the corresponding leaf rcu_node structure. 3. The task resumes execution. 4. One of those CPUs comes back online before a new grace period starts. In step 2, the code in the next rcu_gp_init() invocation will (correctly) defer removing the leaf rcu_node structure from the upper-level bitmasks, and will (correctly) set that structure's ->wait_blkd_tasks field. During the ensuing interval, RCU will (correctly) track the tasks preempted on that structure because they must block any subsequent grace period. In step 3, the code in rcu_read_unlock_special() will (correctly) remove the task from the leaf rcu_node structure. From this point forward, RCU need not pay attention to this structure, at least not until one of the corresponding CPUs comes back online. In step 4, the code in the next rcu_gp_init() invocation will (incorrectly) invoke rcu_init_new_rnp(). This is incorrect because the corresponding rcu_cleanup_dead_rnp() was never invoked. This is nevertheless harmless because the upper-level bits are still set. So, no harm, no foul, right? At least, all is well until a little further into rcu_gp_init() invocation, which will notice that there are no longer any tasks blocked on the leaf rcu_node structure, conclude that there is no longer anything left over from step 2's offline operation, and will therefore invoke rcu_cleanup_dead_rnp(). But this invocation of rcu_cleanup_dead_rnp() is for the beginning of the earlier offline interval, and the previous invocation of rcu_init_new_rnp() is for the end of that same interval. That is right, they are invoked out of order. That cannot be good, can it? It turns out that this is not a (correctness!) problem because rcu_cleanup_dead_rnp() checks to see if any of the corresponding CPUs are online, and refuses to do anything if so. In other words, in the case where rcu_init_new_rnp() and rcu_cleanup_dead_rnp() execute out of order, they both have no effect. But this is at best an accident waiting to happen. This commit therefore adds logic to rcu_gp_init() so that rcu_init_new_rnp() and rcu_cleanup_dead_rnp() are always invoked in order, and so that neither are invoked at all in cases where RCU had to pay attention to the leaf rcu_node structure during the entire time that all corresponding CPUs were offline. And, while in the area, this commit reduces confusion by using formal parameters rather than local variables that just happen to have the same value at that particular point in the code. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2018-05-03 03:49:21 +08:00
* simply clear ->wait_blkd_tasks.
rcu: Process offlining and onlining only at grace-period start Races between CPU hotplug and grace periods can be difficult to resolve, so the ->onoff_mutex is used to exclude the two events. Unfortunately, this means that it is impossible for an outgoing CPU to perform the last bits of its offlining from its last pass through the idle loop, because sleeplocks cannot be acquired in that context. This commit avoids these problems by buffering online and offline events in a new ->qsmaskinitnext field in the leaf rcu_node structures. When a grace period starts, the events accumulated in this mask are applied to the ->qsmaskinit field, and, if needed, up the rcu_node tree. The special case of all CPUs corresponding to a given leaf rcu_node structure being offline while there are still elements in that structure's ->blkd_tasks list is handled using a new ->wait_blkd_tasks field. In this case, propagating the offline bits up the tree is deferred until the beginning of the grace period after all of the tasks have exited their RCU read-side critical sections and removed themselves from the list, at which point the ->wait_blkd_tasks flag is cleared. If one of that leaf rcu_node structure's CPUs comes back online before the list empties, then the ->wait_blkd_tasks flag is simply cleared. This of course means that RCU's notion of which CPUs are offline can be out of date. This is OK because RCU need only wait on CPUs that were online at the time that the grace period started. In addition, RCU's force-quiescent-state actions will handle the case where a CPU goes offline after the grace period starts. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2015-01-24 13:52:37 +08:00
*/
if (rnp->wait_blkd_tasks &&
rcu: Clean up handling of tasks blocked across full-rcu_node offline Commit 0aa04b055e71 ("rcu: Process offlining and onlining only at grace-period start") deferred handling of CPU-hotplug events until the start of the next grace period, but consider the following sequence of events: 1. A task is preempted within an RCU-preempt read-side critical section. 2. The CPU that this task was running on goes offline, along with all other CPUs sharing the corresponding leaf rcu_node structure. 3. The task resumes execution. 4. One of those CPUs comes back online before a new grace period starts. In step 2, the code in the next rcu_gp_init() invocation will (correctly) defer removing the leaf rcu_node structure from the upper-level bitmasks, and will (correctly) set that structure's ->wait_blkd_tasks field. During the ensuing interval, RCU will (correctly) track the tasks preempted on that structure because they must block any subsequent grace period. In step 3, the code in rcu_read_unlock_special() will (correctly) remove the task from the leaf rcu_node structure. From this point forward, RCU need not pay attention to this structure, at least not until one of the corresponding CPUs comes back online. In step 4, the code in the next rcu_gp_init() invocation will (incorrectly) invoke rcu_init_new_rnp(). This is incorrect because the corresponding rcu_cleanup_dead_rnp() was never invoked. This is nevertheless harmless because the upper-level bits are still set. So, no harm, no foul, right? At least, all is well until a little further into rcu_gp_init() invocation, which will notice that there are no longer any tasks blocked on the leaf rcu_node structure, conclude that there is no longer anything left over from step 2's offline operation, and will therefore invoke rcu_cleanup_dead_rnp(). But this invocation of rcu_cleanup_dead_rnp() is for the beginning of the earlier offline interval, and the previous invocation of rcu_init_new_rnp() is for the end of that same interval. That is right, they are invoked out of order. That cannot be good, can it? It turns out that this is not a (correctness!) problem because rcu_cleanup_dead_rnp() checks to see if any of the corresponding CPUs are online, and refuses to do anything if so. In other words, in the case where rcu_init_new_rnp() and rcu_cleanup_dead_rnp() execute out of order, they both have no effect. But this is at best an accident waiting to happen. This commit therefore adds logic to rcu_gp_init() so that rcu_init_new_rnp() and rcu_cleanup_dead_rnp() are always invoked in order, and so that neither are invoked at all in cases where RCU had to pay attention to the leaf rcu_node structure during the entire time that all corresponding CPUs were offline. And, while in the area, this commit reduces confusion by using formal parameters rather than local variables that just happen to have the same value at that particular point in the code. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2018-05-03 03:49:21 +08:00
(!rcu_preempt_has_tasks(rnp) || rnp->qsmaskinit)) {
rcu: Process offlining and onlining only at grace-period start Races between CPU hotplug and grace periods can be difficult to resolve, so the ->onoff_mutex is used to exclude the two events. Unfortunately, this means that it is impossible for an outgoing CPU to perform the last bits of its offlining from its last pass through the idle loop, because sleeplocks cannot be acquired in that context. This commit avoids these problems by buffering online and offline events in a new ->qsmaskinitnext field in the leaf rcu_node structures. When a grace period starts, the events accumulated in this mask are applied to the ->qsmaskinit field, and, if needed, up the rcu_node tree. The special case of all CPUs corresponding to a given leaf rcu_node structure being offline while there are still elements in that structure's ->blkd_tasks list is handled using a new ->wait_blkd_tasks field. In this case, propagating the offline bits up the tree is deferred until the beginning of the grace period after all of the tasks have exited their RCU read-side critical sections and removed themselves from the list, at which point the ->wait_blkd_tasks flag is cleared. If one of that leaf rcu_node structure's CPUs comes back online before the list empties, then the ->wait_blkd_tasks flag is simply cleared. This of course means that RCU's notion of which CPUs are offline can be out of date. This is OK because RCU need only wait on CPUs that were online at the time that the grace period started. In addition, RCU's force-quiescent-state actions will handle the case where a CPU goes offline after the grace period starts. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2015-01-24 13:52:37 +08:00
rnp->wait_blkd_tasks = false;
rcu: Clean up handling of tasks blocked across full-rcu_node offline Commit 0aa04b055e71 ("rcu: Process offlining and onlining only at grace-period start") deferred handling of CPU-hotplug events until the start of the next grace period, but consider the following sequence of events: 1. A task is preempted within an RCU-preempt read-side critical section. 2. The CPU that this task was running on goes offline, along with all other CPUs sharing the corresponding leaf rcu_node structure. 3. The task resumes execution. 4. One of those CPUs comes back online before a new grace period starts. In step 2, the code in the next rcu_gp_init() invocation will (correctly) defer removing the leaf rcu_node structure from the upper-level bitmasks, and will (correctly) set that structure's ->wait_blkd_tasks field. During the ensuing interval, RCU will (correctly) track the tasks preempted on that structure because they must block any subsequent grace period. In step 3, the code in rcu_read_unlock_special() will (correctly) remove the task from the leaf rcu_node structure. From this point forward, RCU need not pay attention to this structure, at least not until one of the corresponding CPUs comes back online. In step 4, the code in the next rcu_gp_init() invocation will (incorrectly) invoke rcu_init_new_rnp(). This is incorrect because the corresponding rcu_cleanup_dead_rnp() was never invoked. This is nevertheless harmless because the upper-level bits are still set. So, no harm, no foul, right? At least, all is well until a little further into rcu_gp_init() invocation, which will notice that there are no longer any tasks blocked on the leaf rcu_node structure, conclude that there is no longer anything left over from step 2's offline operation, and will therefore invoke rcu_cleanup_dead_rnp(). But this invocation of rcu_cleanup_dead_rnp() is for the beginning of the earlier offline interval, and the previous invocation of rcu_init_new_rnp() is for the end of that same interval. That is right, they are invoked out of order. That cannot be good, can it? It turns out that this is not a (correctness!) problem because rcu_cleanup_dead_rnp() checks to see if any of the corresponding CPUs are online, and refuses to do anything if so. In other words, in the case where rcu_init_new_rnp() and rcu_cleanup_dead_rnp() execute out of order, they both have no effect. But this is at best an accident waiting to happen. This commit therefore adds logic to rcu_gp_init() so that rcu_init_new_rnp() and rcu_cleanup_dead_rnp() are always invoked in order, and so that neither are invoked at all in cases where RCU had to pay attention to the leaf rcu_node structure during the entire time that all corresponding CPUs were offline. And, while in the area, this commit reduces confusion by using formal parameters rather than local variables that just happen to have the same value at that particular point in the code. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2018-05-03 03:49:21 +08:00
if (!rnp->qsmaskinit)
rcu_cleanup_dead_rnp(rnp);
rcu: Process offlining and onlining only at grace-period start Races between CPU hotplug and grace periods can be difficult to resolve, so the ->onoff_mutex is used to exclude the two events. Unfortunately, this means that it is impossible for an outgoing CPU to perform the last bits of its offlining from its last pass through the idle loop, because sleeplocks cannot be acquired in that context. This commit avoids these problems by buffering online and offline events in a new ->qsmaskinitnext field in the leaf rcu_node structures. When a grace period starts, the events accumulated in this mask are applied to the ->qsmaskinit field, and, if needed, up the rcu_node tree. The special case of all CPUs corresponding to a given leaf rcu_node structure being offline while there are still elements in that structure's ->blkd_tasks list is handled using a new ->wait_blkd_tasks field. In this case, propagating the offline bits up the tree is deferred until the beginning of the grace period after all of the tasks have exited their RCU read-side critical sections and removed themselves from the list, at which point the ->wait_blkd_tasks flag is cleared. If one of that leaf rcu_node structure's CPUs comes back online before the list empties, then the ->wait_blkd_tasks flag is simply cleared. This of course means that RCU's notion of which CPUs are offline can be out of date. This is OK because RCU need only wait on CPUs that were online at the time that the grace period started. In addition, RCU's force-quiescent-state actions will handle the case where a CPU goes offline after the grace period starts. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2015-01-24 13:52:37 +08:00
}
raw_spin_unlock_rcu_node(rnp);
arch_spin_unlock(&rcu_state.ofl_lock);
local_irq_restore(flags);
rcu: Process offlining and onlining only at grace-period start Races between CPU hotplug and grace periods can be difficult to resolve, so the ->onoff_mutex is used to exclude the two events. Unfortunately, this means that it is impossible for an outgoing CPU to perform the last bits of its offlining from its last pass through the idle loop, because sleeplocks cannot be acquired in that context. This commit avoids these problems by buffering online and offline events in a new ->qsmaskinitnext field in the leaf rcu_node structures. When a grace period starts, the events accumulated in this mask are applied to the ->qsmaskinit field, and, if needed, up the rcu_node tree. The special case of all CPUs corresponding to a given leaf rcu_node structure being offline while there are still elements in that structure's ->blkd_tasks list is handled using a new ->wait_blkd_tasks field. In this case, propagating the offline bits up the tree is deferred until the beginning of the grace period after all of the tasks have exited their RCU read-side critical sections and removed themselves from the list, at which point the ->wait_blkd_tasks flag is cleared. If one of that leaf rcu_node structure's CPUs comes back online before the list empties, then the ->wait_blkd_tasks flag is simply cleared. This of course means that RCU's notion of which CPUs are offline can be out of date. This is OK because RCU need only wait on CPUs that were online at the time that the grace period started. In addition, RCU's force-quiescent-state actions will handle the case where a CPU goes offline after the grace period starts. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2015-01-24 13:52:37 +08:00
}
rcu_gp_slow(gp_preinit_delay); /* Races with CPU hotplug. */
/*
* Set the quiescent-state-needed bits in all the rcu_node
* structures for all currently online CPUs in breadth-first
* order, starting from the root rcu_node structure, relying on the
* layout of the tree within the rcu_state.node[] array. Note that
* other CPUs will access only the leaves of the hierarchy, thus
* seeing that no grace period is in progress, at least until the
* corresponding leaf node has been initialized.
*
* The grace period cannot complete until the initialization
* process finishes, because this kthread handles both.
*/
rcu: Check and report missed fqs timer wakeup on RCU stall For a new grace period request, the RCU GP kthread transitions through following states: a. [RCU_GP_WAIT_GPS] -> [RCU_GP_DONE_GPS] The RCU_GP_WAIT_GPS state is where the GP kthread waits for a request for a new GP. Once it receives a request (for example, when a new RCU callback is queued), the GP kthread transitions to RCU_GP_DONE_GPS. b. [RCU_GP_DONE_GPS] -> [RCU_GP_ONOFF] Grace period initialization starts in rcu_gp_init(), which records the start of new GP in rcu_state.gp_seq and transitions to RCU_GP_ONOFF. c. [RCU_GP_ONOFF] -> [RCU_GP_INIT] The purpose of the RCU_GP_ONOFF state is to apply the online/offline information that was buffered for any CPUs that recently came online or went offline. This state is maintained in per-leaf rcu_node bitmasks, with the buffered state in ->qsmaskinitnext and the state for the upcoming GP in ->qsmaskinit. At the end of this RCU_GP_ONOFF state, each bit in ->qsmaskinit will correspond to a CPU that must pass through a quiescent state before the upcoming grace period is allowed to complete. However, a leaf rcu_node structure with an all-zeroes ->qsmaskinit cannot necessarily be ignored. In preemptible RCU, there might well be tasks still in RCU read-side critical sections that were first preempted while running on one of the CPUs managed by this structure. Such tasks will be queued on this structure's ->blkd_tasks list. Only after this list fully drains can this leaf rcu_node structure be ignored, and even then only if none of its CPUs have come back online in the meantime. Once that happens, the ->qsmaskinit masks further up the tree will be updated to exclude this leaf rcu_node structure. Once the ->qsmaskinitnext and ->qsmaskinit fields have been updated as needed, the GP kthread transitions to RCU_GP_INIT. d. [RCU_GP_INIT] -> [RCU_GP_WAIT_FQS] The purpose of the RCU_GP_INIT state is to copy each ->qsmaskinit to the ->qsmask field within each rcu_node structure. This copying is done breadth-first from the root to the leaves. Why not just copy directly from ->qsmaskinitnext to ->qsmask? Because the ->qsmaskinitnext masks can change in the meantime as additional CPUs come online or go offline. Such changes would result in inconsistencies in the ->qsmask fields up and down the tree, which could in turn result in too-short grace periods or grace-period hangs. These issues are avoided by snapshotting the leaf rcu_node structures' ->qsmaskinitnext fields into their ->qsmaskinit counterparts, generating a consistent set of ->qsmaskinit fields throughout the tree, and only then copying these consistent ->qsmaskinit fields to their ->qsmask counterparts. Once this initialization step is complete, the GP kthread transitions to RCU_GP_WAIT_FQS, where it waits to do a force-quiescent-state scan on the one hand or for the end of the grace period on the other. e. [RCU_GP_WAIT_FQS] -> [RCU_GP_DOING_FQS] The RCU_GP_WAIT_FQS state waits for one of three things: (1) An explicit request to do a force-quiescent-state scan, (2) The end of the grace period, or (3) A short interval of time, after which it will do a force-quiescent-state (FQS) scan. The explicit request can come from rcutorture or from any CPU that has too many RCU callbacks queued (see the qhimark kernel parameter and the RCU_GP_FLAG_OVLD flag). The aforementioned "short period of time" is specified by the jiffies_till_first_fqs boot parameter for a given grace period's first FQS scan and by the jiffies_till_next_fqs for later FQS scans. Either way, once the wait is over, the GP kthread transitions to RCU_GP_DOING_FQS. f. [RCU_GP_DOING_FQS] -> [RCU_GP_CLEANUP] The RCU_GP_DOING_FQS state performs an FQS scan. Each such scan carries out two functions for any CPU whose bit is still set in its leaf rcu_node structure's ->qsmask field, that is, for any CPU that has not yet reported a quiescent state for the current grace period: i. Report quiescent states on behalf of CPUs that have been observed to be idle (from an RCU perspective) since the beginning of the grace period. ii. If the current grace period is too old, take various actions to encourage holdout CPUs to pass through quiescent states, including enlisting the aid of any calls to cond_resched() and might_sleep(), and even including IPIing the holdout CPUs. These checks are skipped for any leaf rcu_node structure with a all-zero ->qsmask field, however such structures are subject to RCU priority boosting if there are tasks on a given structure blocking the current grace period. The end of the grace period is detected when the root rcu_node structure's ->qsmask is zero and when there are no longer any preempted tasks blocking the current grace period. (No, this last check is not redundant. To see this, consider an rcu_node tree having exactly one structure that serves as both root and leaf.) Once the end of the grace period is detected, the GP kthread transitions to RCU_GP_CLEANUP. g. [RCU_GP_CLEANUP] -> [RCU_GP_CLEANED] The RCU_GP_CLEANUP state marks the end of grace period by updating the rcu_state structure's ->gp_seq field and also all rcu_node structures' ->gp_seq field. As before, the rcu_node tree is traversed in breadth first order. Once this update is complete, the GP kthread transitions to the RCU_GP_CLEANED state. i. [RCU_GP_CLEANED] -> [RCU_GP_INIT] Once in the RCU_GP_CLEANED state, the GP kthread immediately transitions into the RCU_GP_INIT state. j. The role of timers. If there is at least one idle CPU, and if timers are not firing, the transition from RCU_GP_DOING_FQS to RCU_GP_CLEANUP will never happen. Timers can fail to fire for a number of reasons, including issues in timer configuration, issues in the timer framework, and failure to handle softirqs (for example, when there is a storm of interrupts). Whatever the reason, if the timers fail to fire, the GP kthread will never be awakened, resulting in RCU CPU stall warnings and eventually in OOM. However, an RCU CPU stall warning has a large number of potential causes, as documented in Documentation/RCU/stallwarn.rst. This commit therefore adds analysis to the RCU CPU stall-warning code to emit an additional message if the cause of the stall is likely to be timer failure. Signed-off-by: Neeraj Upadhyay <neeraju@codeaurora.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-11-17 00:06:00 +08:00
WRITE_ONCE(rcu_state.gp_state, RCU_GP_INIT);
rcu_for_each_node_breadth_first(rnp) {
rcu_gp_slow(gp_init_delay);
rcu: Fix grace-period hangs from mid-init task resume Without special fail-safe quiescent-state-propagation checks, grace-period hangs can result from the following scenario: 1. A task running on a given CPU is preempted in its RCU read-side critical section. 2. That CPU goes offline, and there are now no online CPUs corresponding to that CPU's leaf rcu_node structure. 3. The rcu_gp_init() function does the first phase of grace-period initialization, and sets the aforementioned leaf rcu_node structure's ->qsmaskinit field to all zeroes. Because there is a blocked task, it does not propagate the zeroing of either ->qsmaskinit or ->qsmaskinitnext up the rcu_node tree. 4. The task resumes on some other CPU and exits its critical section. There is no grace period in progress, so the resulting quiescent state is not reported up the tree. 5. The rcu_gp_init() function does the second phase of grace-period initialization, which results in the leaf rcu_node structure being initialized to expect no further quiescent states, but with that structure's parent expecting a quiescent-state report. The parent will never receive a quiescent state from this leaf rcu_node structure, so the grace period will hang, resulting in RCU CPU stall warnings. It would be good to get rid of the special fail-safe quiescent-state propagation checks. This commit therefore checks the leaf rcu_node structure's ->wait_blkd_tasks field during grace-period initialization. If this flag is set, the rcu_report_qs_rnp() is invoked to immediately report the possible quiescent state. While in the neighborhood, this commit also report quiescent states for any CPUs that went offline between the two phases of grace-period initialization, thus reducing grace-period delays and hopefully eventually allowing removal of offline-CPU checks from the force-quiescent-state code path. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2018-05-08 00:34:17 +08:00
raw_spin_lock_irqsave_rcu_node(rnp, flags);
rdp = this_cpu_ptr(&rcu_data);
rcu_preempt_check_blocked_tasks(rnp);
rnp->qsmask = rnp->qsmaskinit;
WRITE_ONCE(rnp->gp_seq, rcu_state.gp_seq);
if (rnp == rdp->mynode)
(void)__note_gp_changes(rnp, rdp);
rcu_preempt_boost_start_gp(rnp);
trace_rcu_grace_period_init(rcu_state.name, rnp->gp_seq,
rnp->level, rnp->grplo,
rnp->grphi, rnp->qsmask);
rcu: Fix grace-period hangs from mid-init task resume Without special fail-safe quiescent-state-propagation checks, grace-period hangs can result from the following scenario: 1. A task running on a given CPU is preempted in its RCU read-side critical section. 2. That CPU goes offline, and there are now no online CPUs corresponding to that CPU's leaf rcu_node structure. 3. The rcu_gp_init() function does the first phase of grace-period initialization, and sets the aforementioned leaf rcu_node structure's ->qsmaskinit field to all zeroes. Because there is a blocked task, it does not propagate the zeroing of either ->qsmaskinit or ->qsmaskinitnext up the rcu_node tree. 4. The task resumes on some other CPU and exits its critical section. There is no grace period in progress, so the resulting quiescent state is not reported up the tree. 5. The rcu_gp_init() function does the second phase of grace-period initialization, which results in the leaf rcu_node structure being initialized to expect no further quiescent states, but with that structure's parent expecting a quiescent-state report. The parent will never receive a quiescent state from this leaf rcu_node structure, so the grace period will hang, resulting in RCU CPU stall warnings. It would be good to get rid of the special fail-safe quiescent-state propagation checks. This commit therefore checks the leaf rcu_node structure's ->wait_blkd_tasks field during grace-period initialization. If this flag is set, the rcu_report_qs_rnp() is invoked to immediately report the possible quiescent state. While in the neighborhood, this commit also report quiescent states for any CPUs that went offline between the two phases of grace-period initialization, thus reducing grace-period delays and hopefully eventually allowing removal of offline-CPU checks from the force-quiescent-state code path. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2018-05-08 00:34:17 +08:00
/* Quiescent states for tasks on any now-offline CPUs. */
mask = rnp->qsmask & ~rnp->qsmaskinitnext;
rnp->rcu_gp_init_mask = mask;
rcu: Fix grace-period hangs from mid-init task resume Without special fail-safe quiescent-state-propagation checks, grace-period hangs can result from the following scenario: 1. A task running on a given CPU is preempted in its RCU read-side critical section. 2. That CPU goes offline, and there are now no online CPUs corresponding to that CPU's leaf rcu_node structure. 3. The rcu_gp_init() function does the first phase of grace-period initialization, and sets the aforementioned leaf rcu_node structure's ->qsmaskinit field to all zeroes. Because there is a blocked task, it does not propagate the zeroing of either ->qsmaskinit or ->qsmaskinitnext up the rcu_node tree. 4. The task resumes on some other CPU and exits its critical section. There is no grace period in progress, so the resulting quiescent state is not reported up the tree. 5. The rcu_gp_init() function does the second phase of grace-period initialization, which results in the leaf rcu_node structure being initialized to expect no further quiescent states, but with that structure's parent expecting a quiescent-state report. The parent will never receive a quiescent state from this leaf rcu_node structure, so the grace period will hang, resulting in RCU CPU stall warnings. It would be good to get rid of the special fail-safe quiescent-state propagation checks. This commit therefore checks the leaf rcu_node structure's ->wait_blkd_tasks field during grace-period initialization. If this flag is set, the rcu_report_qs_rnp() is invoked to immediately report the possible quiescent state. While in the neighborhood, this commit also report quiescent states for any CPUs that went offline between the two phases of grace-period initialization, thus reducing grace-period delays and hopefully eventually allowing removal of offline-CPU checks from the force-quiescent-state code path. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2018-05-08 00:34:17 +08:00
if ((mask || rnp->wait_blkd_tasks) && rcu_is_leaf_node(rnp))
rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
rcu: Fix grace-period hangs from mid-init task resume Without special fail-safe quiescent-state-propagation checks, grace-period hangs can result from the following scenario: 1. A task running on a given CPU is preempted in its RCU read-side critical section. 2. That CPU goes offline, and there are now no online CPUs corresponding to that CPU's leaf rcu_node structure. 3. The rcu_gp_init() function does the first phase of grace-period initialization, and sets the aforementioned leaf rcu_node structure's ->qsmaskinit field to all zeroes. Because there is a blocked task, it does not propagate the zeroing of either ->qsmaskinit or ->qsmaskinitnext up the rcu_node tree. 4. The task resumes on some other CPU and exits its critical section. There is no grace period in progress, so the resulting quiescent state is not reported up the tree. 5. The rcu_gp_init() function does the second phase of grace-period initialization, which results in the leaf rcu_node structure being initialized to expect no further quiescent states, but with that structure's parent expecting a quiescent-state report. The parent will never receive a quiescent state from this leaf rcu_node structure, so the grace period will hang, resulting in RCU CPU stall warnings. It would be good to get rid of the special fail-safe quiescent-state propagation checks. This commit therefore checks the leaf rcu_node structure's ->wait_blkd_tasks field during grace-period initialization. If this flag is set, the rcu_report_qs_rnp() is invoked to immediately report the possible quiescent state. While in the neighborhood, this commit also report quiescent states for any CPUs that went offline between the two phases of grace-period initialization, thus reducing grace-period delays and hopefully eventually allowing removal of offline-CPU checks from the force-quiescent-state code path. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2018-05-08 00:34:17 +08:00
else
raw_spin_unlock_irq_rcu_node(rnp);
rcu: Rename cond_resched_rcu_qs() to cond_resched_tasks_rcu_qs() Commit e31d28b6ab8f ("trace: Eliminate cond_resched_rcu_qs() in favor of cond_resched()") substituted cond_resched() for the earlier call to cond_resched_rcu_qs(). However, the new-age cond_resched() does not do anything to help RCU-tasks grace periods because (1) RCU-tasks is only enabled when CONFIG_PREEMPT=y and (2) cond_resched() is a complete no-op when preemption is enabled. This situation results in hangs when running the trace benchmarks. A number of potential fixes were discussed on LKML (https://lkml.kernel.org/r/20180224151240.0d63a059@vmware.local.home), including making cond_resched() not be a no-op; making cond_resched() not be a no-op, but only when running tracing benchmarks; reverting the aforementioned commit (which works because cond_resched_rcu_qs() does provide an RCU-tasks quiescent state; and adding a call to the scheduler/RCU rcu_note_voluntary_context_switch() function. All were deemed unsatisfactory, either due to added cond_resched() overhead or due to magic functions inviting cargo culting. This commit renames cond_resched_rcu_qs() to cond_resched_tasks_rcu_qs(), which provides a clear hint as to what this function is doing and why and where it should be used, and then replaces the call to cond_resched() with cond_resched_tasks_rcu_qs() in the trace benchmark's benchmark_event_kthread() function. Reported-by: Steven Rostedt <rostedt@goodmis.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Tested-by: Nicholas Piggin <npiggin@gmail.com>
2018-03-03 08:35:27 +08:00
cond_resched_tasks_rcu_qs();
WRITE_ONCE(rcu_state.gp_activity, jiffies);
}
// If strict, make all CPUs aware of new grace period.
if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
on_each_cpu(rcu_strict_gp_boundary, NULL, 0);
return true;
}
/*
* Helper function for swait_event_idle_exclusive() wakeup at force-quiescent-state
rcu: Use idle versions of swait to make idle-hack clear These RCU waits were set to use interruptible waits to avoid the kthreads contributing to system load average, even though they are not interruptible as they are spawned from a kthread. Use the new TASK_IDLE swaits which makes our goal clear, and removes confusion about these paths possibly being interruptible -- they are not. When the system is idle the RCU grace-period kthread will spend all its time blocked inside the swait_event_interruptible(). If the interruptible() was not used, then this kthread would contribute to the load average. This means that an idle system would have a load average of 2 (or 3 if PREEMPT=y), rather than the load average of 0 that almost fifty years of UNIX has conditioned sysadmins to expect. The same argument applies to swait_event_interruptible_timeout() use. The RCU grace-period kthread spends its time blocked inside this call while waiting for grace periods to complete. In particular, if there was only one busy CPU, but that CPU was frequently invoking call_rcu(), then the RCU grace-period kthread would spend almost all its time blocked inside the swait_event_interruptible_timeout(). This would mean that the load average would be 2 rather than the expected 1 for the single busy CPU. Acked-by: "Eric W. Biederman" <ebiederm@xmission.com> Tested-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Luis R. Rodriguez <mcgrof@kernel.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2017-06-21 05:45:47 +08:00
* time.
*/
static bool rcu_gp_fqs_check_wake(int *gfp)
{
struct rcu_node *rnp = rcu_get_root();
// If under overload conditions, force an immediate FQS scan.
if (*gfp & RCU_GP_FLAG_OVLD)
return true;
// Someone like call_rcu() requested a force-quiescent-state scan.
*gfp = READ_ONCE(rcu_state.gp_flags);
if (*gfp & RCU_GP_FLAG_FQS)
return true;
// The current grace period has completed.
if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
return true;
return false;
}
/*
* Do one round of quiescent-state forcing.
*/
static void rcu_gp_fqs(bool first_time)
{
rcu/tree: Defer setting of jiffies during stall reset commit b96e7a5fa0ba9cda32888e04f8f4bac42d49a7f8 upstream. There are instances where rcu_cpu_stall_reset() is called when jiffies did not get a chance to update for a long time. Before jiffies is updated, the CPU stall detector can go off triggering false-positives where a just-started grace period appears to be ages old. In the past, we disabled stall detection in rcu_cpu_stall_reset() however this got changed [1]. This is resulting in false-positives in KGDB usecase [2]. Fix this by deferring the update of jiffies to the third run of the FQS loop. This is more robust, as, even if rcu_cpu_stall_reset() is called just before jiffies is read, we would end up pushing out the jiffies read by 3 more FQS loops. Meanwhile the CPU stall detection will be delayed and we will not get any false positives. [1] https://lore.kernel.org/all/20210521155624.174524-2-senozhatsky@chromium.org/ [2] https://lore.kernel.org/all/20230814020045.51950-2-chenhuacai@loongson.cn/ Tested with rcutorture.cpu_stall option as well to verify stall behavior with/without patch. Tested-by: Huacai Chen <chenhuacai@loongson.cn> Reported-by: Binbin Zhou <zhoubinbin@loongson.cn> Closes: https://lore.kernel.org/all/20230814020045.51950-2-chenhuacai@loongson.cn/ Suggested-by: Paul McKenney <paulmck@kernel.org> Cc: Sergey Senozhatsky <senozhatsky@chromium.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: stable@vger.kernel.org Fixes: a80be428fbc1 ("rcu: Do not disable GP stall detection in rcu_cpu_stall_reset()") Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2023-09-05 08:02:11 +08:00
int nr_fqs = READ_ONCE(rcu_state.nr_fqs_jiffies_stall);
struct rcu_node *rnp = rcu_get_root();
WRITE_ONCE(rcu_state.gp_activity, jiffies);
WRITE_ONCE(rcu_state.n_force_qs, rcu_state.n_force_qs + 1);
rcu/tree: Defer setting of jiffies during stall reset commit b96e7a5fa0ba9cda32888e04f8f4bac42d49a7f8 upstream. There are instances where rcu_cpu_stall_reset() is called when jiffies did not get a chance to update for a long time. Before jiffies is updated, the CPU stall detector can go off triggering false-positives where a just-started grace period appears to be ages old. In the past, we disabled stall detection in rcu_cpu_stall_reset() however this got changed [1]. This is resulting in false-positives in KGDB usecase [2]. Fix this by deferring the update of jiffies to the third run of the FQS loop. This is more robust, as, even if rcu_cpu_stall_reset() is called just before jiffies is read, we would end up pushing out the jiffies read by 3 more FQS loops. Meanwhile the CPU stall detection will be delayed and we will not get any false positives. [1] https://lore.kernel.org/all/20210521155624.174524-2-senozhatsky@chromium.org/ [2] https://lore.kernel.org/all/20230814020045.51950-2-chenhuacai@loongson.cn/ Tested with rcutorture.cpu_stall option as well to verify stall behavior with/without patch. Tested-by: Huacai Chen <chenhuacai@loongson.cn> Reported-by: Binbin Zhou <zhoubinbin@loongson.cn> Closes: https://lore.kernel.org/all/20230814020045.51950-2-chenhuacai@loongson.cn/ Suggested-by: Paul McKenney <paulmck@kernel.org> Cc: Sergey Senozhatsky <senozhatsky@chromium.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: stable@vger.kernel.org Fixes: a80be428fbc1 ("rcu: Do not disable GP stall detection in rcu_cpu_stall_reset()") Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2023-09-05 08:02:11 +08:00
WARN_ON_ONCE(nr_fqs > 3);
/* Only countdown nr_fqs for stall purposes if jiffies moves. */
if (nr_fqs) {
if (nr_fqs == 1) {
WRITE_ONCE(rcu_state.jiffies_stall,
jiffies + rcu_jiffies_till_stall_check());
}
WRITE_ONCE(rcu_state.nr_fqs_jiffies_stall, --nr_fqs);
}
if (first_time) {
/* Collect dyntick-idle snapshots. */
force_qs_rnp(dyntick_save_progress_counter);
} else {
/* Handle dyntick-idle and offline CPUs. */
force_qs_rnp(rcu_implicit_dynticks_qs);
}
/* Clear flag to prevent immediate re-entry. */
if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
raw_spin_lock_irq_rcu_node(rnp);
WRITE_ONCE(rcu_state.gp_flags,
READ_ONCE(rcu_state.gp_flags) & ~RCU_GP_FLAG_FQS);
raw_spin_unlock_irq_rcu_node(rnp);
}
}
/*
* Loop doing repeated quiescent-state forcing until the grace period ends.
*/
static noinline_for_stack void rcu_gp_fqs_loop(void)
{
bool first_gp_fqs = true;
int gf = 0;
unsigned long j;
int ret;
struct rcu_node *rnp = rcu_get_root();
j = READ_ONCE(jiffies_till_first_fqs);
if (rcu_state.cbovld)
gf = RCU_GP_FLAG_OVLD;
ret = 0;
for (;;) {
if (rcu_state.cbovld) {
j = (j + 2) / 3;
if (j <= 0)
j = 1;
}
if (!ret || time_before(jiffies + j, rcu_state.jiffies_force_qs)) {
rcu: Check and report missed fqs timer wakeup on RCU stall For a new grace period request, the RCU GP kthread transitions through following states: a. [RCU_GP_WAIT_GPS] -> [RCU_GP_DONE_GPS] The RCU_GP_WAIT_GPS state is where the GP kthread waits for a request for a new GP. Once it receives a request (for example, when a new RCU callback is queued), the GP kthread transitions to RCU_GP_DONE_GPS. b. [RCU_GP_DONE_GPS] -> [RCU_GP_ONOFF] Grace period initialization starts in rcu_gp_init(), which records the start of new GP in rcu_state.gp_seq and transitions to RCU_GP_ONOFF. c. [RCU_GP_ONOFF] -> [RCU_GP_INIT] The purpose of the RCU_GP_ONOFF state is to apply the online/offline information that was buffered for any CPUs that recently came online or went offline. This state is maintained in per-leaf rcu_node bitmasks, with the buffered state in ->qsmaskinitnext and the state for the upcoming GP in ->qsmaskinit. At the end of this RCU_GP_ONOFF state, each bit in ->qsmaskinit will correspond to a CPU that must pass through a quiescent state before the upcoming grace period is allowed to complete. However, a leaf rcu_node structure with an all-zeroes ->qsmaskinit cannot necessarily be ignored. In preemptible RCU, there might well be tasks still in RCU read-side critical sections that were first preempted while running on one of the CPUs managed by this structure. Such tasks will be queued on this structure's ->blkd_tasks list. Only after this list fully drains can this leaf rcu_node structure be ignored, and even then only if none of its CPUs have come back online in the meantime. Once that happens, the ->qsmaskinit masks further up the tree will be updated to exclude this leaf rcu_node structure. Once the ->qsmaskinitnext and ->qsmaskinit fields have been updated as needed, the GP kthread transitions to RCU_GP_INIT. d. [RCU_GP_INIT] -> [RCU_GP_WAIT_FQS] The purpose of the RCU_GP_INIT state is to copy each ->qsmaskinit to the ->qsmask field within each rcu_node structure. This copying is done breadth-first from the root to the leaves. Why not just copy directly from ->qsmaskinitnext to ->qsmask? Because the ->qsmaskinitnext masks can change in the meantime as additional CPUs come online or go offline. Such changes would result in inconsistencies in the ->qsmask fields up and down the tree, which could in turn result in too-short grace periods or grace-period hangs. These issues are avoided by snapshotting the leaf rcu_node structures' ->qsmaskinitnext fields into their ->qsmaskinit counterparts, generating a consistent set of ->qsmaskinit fields throughout the tree, and only then copying these consistent ->qsmaskinit fields to their ->qsmask counterparts. Once this initialization step is complete, the GP kthread transitions to RCU_GP_WAIT_FQS, where it waits to do a force-quiescent-state scan on the one hand or for the end of the grace period on the other. e. [RCU_GP_WAIT_FQS] -> [RCU_GP_DOING_FQS] The RCU_GP_WAIT_FQS state waits for one of three things: (1) An explicit request to do a force-quiescent-state scan, (2) The end of the grace period, or (3) A short interval of time, after which it will do a force-quiescent-state (FQS) scan. The explicit request can come from rcutorture or from any CPU that has too many RCU callbacks queued (see the qhimark kernel parameter and the RCU_GP_FLAG_OVLD flag). The aforementioned "short period of time" is specified by the jiffies_till_first_fqs boot parameter for a given grace period's first FQS scan and by the jiffies_till_next_fqs for later FQS scans. Either way, once the wait is over, the GP kthread transitions to RCU_GP_DOING_FQS. f. [RCU_GP_DOING_FQS] -> [RCU_GP_CLEANUP] The RCU_GP_DOING_FQS state performs an FQS scan. Each such scan carries out two functions for any CPU whose bit is still set in its leaf rcu_node structure's ->qsmask field, that is, for any CPU that has not yet reported a quiescent state for the current grace period: i. Report quiescent states on behalf of CPUs that have been observed to be idle (from an RCU perspective) since the beginning of the grace period. ii. If the current grace period is too old, take various actions to encourage holdout CPUs to pass through quiescent states, including enlisting the aid of any calls to cond_resched() and might_sleep(), and even including IPIing the holdout CPUs. These checks are skipped for any leaf rcu_node structure with a all-zero ->qsmask field, however such structures are subject to RCU priority boosting if there are tasks on a given structure blocking the current grace period. The end of the grace period is detected when the root rcu_node structure's ->qsmask is zero and when there are no longer any preempted tasks blocking the current grace period. (No, this last check is not redundant. To see this, consider an rcu_node tree having exactly one structure that serves as both root and leaf.) Once the end of the grace period is detected, the GP kthread transitions to RCU_GP_CLEANUP. g. [RCU_GP_CLEANUP] -> [RCU_GP_CLEANED] The RCU_GP_CLEANUP state marks the end of grace period by updating the rcu_state structure's ->gp_seq field and also all rcu_node structures' ->gp_seq field. As before, the rcu_node tree is traversed in breadth first order. Once this update is complete, the GP kthread transitions to the RCU_GP_CLEANED state. i. [RCU_GP_CLEANED] -> [RCU_GP_INIT] Once in the RCU_GP_CLEANED state, the GP kthread immediately transitions into the RCU_GP_INIT state. j. The role of timers. If there is at least one idle CPU, and if timers are not firing, the transition from RCU_GP_DOING_FQS to RCU_GP_CLEANUP will never happen. Timers can fail to fire for a number of reasons, including issues in timer configuration, issues in the timer framework, and failure to handle softirqs (for example, when there is a storm of interrupts). Whatever the reason, if the timers fail to fire, the GP kthread will never be awakened, resulting in RCU CPU stall warnings and eventually in OOM. However, an RCU CPU stall warning has a large number of potential causes, as documented in Documentation/RCU/stallwarn.rst. This commit therefore adds analysis to the RCU CPU stall-warning code to emit an additional message if the cause of the stall is likely to be timer failure. Signed-off-by: Neeraj Upadhyay <neeraju@codeaurora.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-11-17 00:06:00 +08:00
WRITE_ONCE(rcu_state.jiffies_force_qs, jiffies + j);
/*
* jiffies_force_qs before RCU_GP_WAIT_FQS state
* update; required for stall checks.
*/
smp_wmb();
WRITE_ONCE(rcu_state.jiffies_kick_kthreads,
jiffies + (j ? 3 * j : 2));
}
trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
TPS("fqswait"));
rcu: Check and report missed fqs timer wakeup on RCU stall For a new grace period request, the RCU GP kthread transitions through following states: a. [RCU_GP_WAIT_GPS] -> [RCU_GP_DONE_GPS] The RCU_GP_WAIT_GPS state is where the GP kthread waits for a request for a new GP. Once it receives a request (for example, when a new RCU callback is queued), the GP kthread transitions to RCU_GP_DONE_GPS. b. [RCU_GP_DONE_GPS] -> [RCU_GP_ONOFF] Grace period initialization starts in rcu_gp_init(), which records the start of new GP in rcu_state.gp_seq and transitions to RCU_GP_ONOFF. c. [RCU_GP_ONOFF] -> [RCU_GP_INIT] The purpose of the RCU_GP_ONOFF state is to apply the online/offline information that was buffered for any CPUs that recently came online or went offline. This state is maintained in per-leaf rcu_node bitmasks, with the buffered state in ->qsmaskinitnext and the state for the upcoming GP in ->qsmaskinit. At the end of this RCU_GP_ONOFF state, each bit in ->qsmaskinit will correspond to a CPU that must pass through a quiescent state before the upcoming grace period is allowed to complete. However, a leaf rcu_node structure with an all-zeroes ->qsmaskinit cannot necessarily be ignored. In preemptible RCU, there might well be tasks still in RCU read-side critical sections that were first preempted while running on one of the CPUs managed by this structure. Such tasks will be queued on this structure's ->blkd_tasks list. Only after this list fully drains can this leaf rcu_node structure be ignored, and even then only if none of its CPUs have come back online in the meantime. Once that happens, the ->qsmaskinit masks further up the tree will be updated to exclude this leaf rcu_node structure. Once the ->qsmaskinitnext and ->qsmaskinit fields have been updated as needed, the GP kthread transitions to RCU_GP_INIT. d. [RCU_GP_INIT] -> [RCU_GP_WAIT_FQS] The purpose of the RCU_GP_INIT state is to copy each ->qsmaskinit to the ->qsmask field within each rcu_node structure. This copying is done breadth-first from the root to the leaves. Why not just copy directly from ->qsmaskinitnext to ->qsmask? Because the ->qsmaskinitnext masks can change in the meantime as additional CPUs come online or go offline. Such changes would result in inconsistencies in the ->qsmask fields up and down the tree, which could in turn result in too-short grace periods or grace-period hangs. These issues are avoided by snapshotting the leaf rcu_node structures' ->qsmaskinitnext fields into their ->qsmaskinit counterparts, generating a consistent set of ->qsmaskinit fields throughout the tree, and only then copying these consistent ->qsmaskinit fields to their ->qsmask counterparts. Once this initialization step is complete, the GP kthread transitions to RCU_GP_WAIT_FQS, where it waits to do a force-quiescent-state scan on the one hand or for the end of the grace period on the other. e. [RCU_GP_WAIT_FQS] -> [RCU_GP_DOING_FQS] The RCU_GP_WAIT_FQS state waits for one of three things: (1) An explicit request to do a force-quiescent-state scan, (2) The end of the grace period, or (3) A short interval of time, after which it will do a force-quiescent-state (FQS) scan. The explicit request can come from rcutorture or from any CPU that has too many RCU callbacks queued (see the qhimark kernel parameter and the RCU_GP_FLAG_OVLD flag). The aforementioned "short period of time" is specified by the jiffies_till_first_fqs boot parameter for a given grace period's first FQS scan and by the jiffies_till_next_fqs for later FQS scans. Either way, once the wait is over, the GP kthread transitions to RCU_GP_DOING_FQS. f. [RCU_GP_DOING_FQS] -> [RCU_GP_CLEANUP] The RCU_GP_DOING_FQS state performs an FQS scan. Each such scan carries out two functions for any CPU whose bit is still set in its leaf rcu_node structure's ->qsmask field, that is, for any CPU that has not yet reported a quiescent state for the current grace period: i. Report quiescent states on behalf of CPUs that have been observed to be idle (from an RCU perspective) since the beginning of the grace period. ii. If the current grace period is too old, take various actions to encourage holdout CPUs to pass through quiescent states, including enlisting the aid of any calls to cond_resched() and might_sleep(), and even including IPIing the holdout CPUs. These checks are skipped for any leaf rcu_node structure with a all-zero ->qsmask field, however such structures are subject to RCU priority boosting if there are tasks on a given structure blocking the current grace period. The end of the grace period is detected when the root rcu_node structure's ->qsmask is zero and when there are no longer any preempted tasks blocking the current grace period. (No, this last check is not redundant. To see this, consider an rcu_node tree having exactly one structure that serves as both root and leaf.) Once the end of the grace period is detected, the GP kthread transitions to RCU_GP_CLEANUP. g. [RCU_GP_CLEANUP] -> [RCU_GP_CLEANED] The RCU_GP_CLEANUP state marks the end of grace period by updating the rcu_state structure's ->gp_seq field and also all rcu_node structures' ->gp_seq field. As before, the rcu_node tree is traversed in breadth first order. Once this update is complete, the GP kthread transitions to the RCU_GP_CLEANED state. i. [RCU_GP_CLEANED] -> [RCU_GP_INIT] Once in the RCU_GP_CLEANED state, the GP kthread immediately transitions into the RCU_GP_INIT state. j. The role of timers. If there is at least one idle CPU, and if timers are not firing, the transition from RCU_GP_DOING_FQS to RCU_GP_CLEANUP will never happen. Timers can fail to fire for a number of reasons, including issues in timer configuration, issues in the timer framework, and failure to handle softirqs (for example, when there is a storm of interrupts). Whatever the reason, if the timers fail to fire, the GP kthread will never be awakened, resulting in RCU CPU stall warnings and eventually in OOM. However, an RCU CPU stall warning has a large number of potential causes, as documented in Documentation/RCU/stallwarn.rst. This commit therefore adds analysis to the RCU CPU stall-warning code to emit an additional message if the cause of the stall is likely to be timer failure. Signed-off-by: Neeraj Upadhyay <neeraju@codeaurora.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-11-17 00:06:00 +08:00
WRITE_ONCE(rcu_state.gp_state, RCU_GP_WAIT_FQS);
(void)swait_event_idle_timeout_exclusive(rcu_state.gp_wq,
rcu_gp_fqs_check_wake(&gf), j);
rcu_gp_torture_wait();
rcu: Check and report missed fqs timer wakeup on RCU stall For a new grace period request, the RCU GP kthread transitions through following states: a. [RCU_GP_WAIT_GPS] -> [RCU_GP_DONE_GPS] The RCU_GP_WAIT_GPS state is where the GP kthread waits for a request for a new GP. Once it receives a request (for example, when a new RCU callback is queued), the GP kthread transitions to RCU_GP_DONE_GPS. b. [RCU_GP_DONE_GPS] -> [RCU_GP_ONOFF] Grace period initialization starts in rcu_gp_init(), which records the start of new GP in rcu_state.gp_seq and transitions to RCU_GP_ONOFF. c. [RCU_GP_ONOFF] -> [RCU_GP_INIT] The purpose of the RCU_GP_ONOFF state is to apply the online/offline information that was buffered for any CPUs that recently came online or went offline. This state is maintained in per-leaf rcu_node bitmasks, with the buffered state in ->qsmaskinitnext and the state for the upcoming GP in ->qsmaskinit. At the end of this RCU_GP_ONOFF state, each bit in ->qsmaskinit will correspond to a CPU that must pass through a quiescent state before the upcoming grace period is allowed to complete. However, a leaf rcu_node structure with an all-zeroes ->qsmaskinit cannot necessarily be ignored. In preemptible RCU, there might well be tasks still in RCU read-side critical sections that were first preempted while running on one of the CPUs managed by this structure. Such tasks will be queued on this structure's ->blkd_tasks list. Only after this list fully drains can this leaf rcu_node structure be ignored, and even then only if none of its CPUs have come back online in the meantime. Once that happens, the ->qsmaskinit masks further up the tree will be updated to exclude this leaf rcu_node structure. Once the ->qsmaskinitnext and ->qsmaskinit fields have been updated as needed, the GP kthread transitions to RCU_GP_INIT. d. [RCU_GP_INIT] -> [RCU_GP_WAIT_FQS] The purpose of the RCU_GP_INIT state is to copy each ->qsmaskinit to the ->qsmask field within each rcu_node structure. This copying is done breadth-first from the root to the leaves. Why not just copy directly from ->qsmaskinitnext to ->qsmask? Because the ->qsmaskinitnext masks can change in the meantime as additional CPUs come online or go offline. Such changes would result in inconsistencies in the ->qsmask fields up and down the tree, which could in turn result in too-short grace periods or grace-period hangs. These issues are avoided by snapshotting the leaf rcu_node structures' ->qsmaskinitnext fields into their ->qsmaskinit counterparts, generating a consistent set of ->qsmaskinit fields throughout the tree, and only then copying these consistent ->qsmaskinit fields to their ->qsmask counterparts. Once this initialization step is complete, the GP kthread transitions to RCU_GP_WAIT_FQS, where it waits to do a force-quiescent-state scan on the one hand or for the end of the grace period on the other. e. [RCU_GP_WAIT_FQS] -> [RCU_GP_DOING_FQS] The RCU_GP_WAIT_FQS state waits for one of three things: (1) An explicit request to do a force-quiescent-state scan, (2) The end of the grace period, or (3) A short interval of time, after which it will do a force-quiescent-state (FQS) scan. The explicit request can come from rcutorture or from any CPU that has too many RCU callbacks queued (see the qhimark kernel parameter and the RCU_GP_FLAG_OVLD flag). The aforementioned "short period of time" is specified by the jiffies_till_first_fqs boot parameter for a given grace period's first FQS scan and by the jiffies_till_next_fqs for later FQS scans. Either way, once the wait is over, the GP kthread transitions to RCU_GP_DOING_FQS. f. [RCU_GP_DOING_FQS] -> [RCU_GP_CLEANUP] The RCU_GP_DOING_FQS state performs an FQS scan. Each such scan carries out two functions for any CPU whose bit is still set in its leaf rcu_node structure's ->qsmask field, that is, for any CPU that has not yet reported a quiescent state for the current grace period: i. Report quiescent states on behalf of CPUs that have been observed to be idle (from an RCU perspective) since the beginning of the grace period. ii. If the current grace period is too old, take various actions to encourage holdout CPUs to pass through quiescent states, including enlisting the aid of any calls to cond_resched() and might_sleep(), and even including IPIing the holdout CPUs. These checks are skipped for any leaf rcu_node structure with a all-zero ->qsmask field, however such structures are subject to RCU priority boosting if there are tasks on a given structure blocking the current grace period. The end of the grace period is detected when the root rcu_node structure's ->qsmask is zero and when there are no longer any preempted tasks blocking the current grace period. (No, this last check is not redundant. To see this, consider an rcu_node tree having exactly one structure that serves as both root and leaf.) Once the end of the grace period is detected, the GP kthread transitions to RCU_GP_CLEANUP. g. [RCU_GP_CLEANUP] -> [RCU_GP_CLEANED] The RCU_GP_CLEANUP state marks the end of grace period by updating the rcu_state structure's ->gp_seq field and also all rcu_node structures' ->gp_seq field. As before, the rcu_node tree is traversed in breadth first order. Once this update is complete, the GP kthread transitions to the RCU_GP_CLEANED state. i. [RCU_GP_CLEANED] -> [RCU_GP_INIT] Once in the RCU_GP_CLEANED state, the GP kthread immediately transitions into the RCU_GP_INIT state. j. The role of timers. If there is at least one idle CPU, and if timers are not firing, the transition from RCU_GP_DOING_FQS to RCU_GP_CLEANUP will never happen. Timers can fail to fire for a number of reasons, including issues in timer configuration, issues in the timer framework, and failure to handle softirqs (for example, when there is a storm of interrupts). Whatever the reason, if the timers fail to fire, the GP kthread will never be awakened, resulting in RCU CPU stall warnings and eventually in OOM. However, an RCU CPU stall warning has a large number of potential causes, as documented in Documentation/RCU/stallwarn.rst. This commit therefore adds analysis to the RCU CPU stall-warning code to emit an additional message if the cause of the stall is likely to be timer failure. Signed-off-by: Neeraj Upadhyay <neeraju@codeaurora.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-11-17 00:06:00 +08:00
WRITE_ONCE(rcu_state.gp_state, RCU_GP_DOING_FQS);
/* Locking provides needed memory barriers. */
/*
* Exit the loop if the root rcu_node structure indicates that the grace period
* has ended, leave the loop. The rcu_preempt_blocked_readers_cgp(rnp) check
* is required only for single-node rcu_node trees because readers blocking
* the current grace period are queued only on leaf rcu_node structures.
* For multi-node trees, checking the root node's ->qsmask suffices, because a
* given root node's ->qsmask bit is cleared only when all CPUs and tasks from
* the corresponding leaf nodes have passed through their quiescent state.
*/
if (!READ_ONCE(rnp->qsmask) &&
!rcu_preempt_blocked_readers_cgp(rnp))
break;
/* If time for quiescent-state forcing, do it. */
if (!time_after(rcu_state.jiffies_force_qs, jiffies) ||
(gf & (RCU_GP_FLAG_FQS | RCU_GP_FLAG_OVLD))) {
trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
TPS("fqsstart"));
rcu_gp_fqs(first_gp_fqs);
gf = 0;
if (first_gp_fqs) {
first_gp_fqs = false;
gf = rcu_state.cbovld ? RCU_GP_FLAG_OVLD : 0;
}
trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
TPS("fqsend"));
cond_resched_tasks_rcu_qs();
WRITE_ONCE(rcu_state.gp_activity, jiffies);
ret = 0; /* Force full wait till next FQS. */
j = READ_ONCE(jiffies_till_next_fqs);
} else {
/* Deal with stray signal. */
cond_resched_tasks_rcu_qs();
WRITE_ONCE(rcu_state.gp_activity, jiffies);
WARN_ON(signal_pending(current));
trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
TPS("fqswaitsig"));
ret = 1; /* Keep old FQS timing. */
j = jiffies;
if (time_after(jiffies, rcu_state.jiffies_force_qs))
j = 1;
else
j = rcu_state.jiffies_force_qs - j;
gf = 0;
}
}
}
/*
* Clean up after the old grace period.
*/
static noinline void rcu_gp_cleanup(void)
{
int cpu;
rcu: Make callers awaken grace-period kthread The rcu_start_gp_advanced() function currently uses irq_work_queue() to defer wakeups of the RCU grace-period kthread. This deferring is necessary to avoid RCU-scheduler deadlocks involving the rcu_node structure's lock, meaning that RCU cannot call any of the scheduler's wake-up functions while holding one of these locks. Unfortunately, the second and subsequent calls to irq_work_queue() are ignored, and the first call will be ignored (aside from queuing the work item) if the scheduler-clock tick is turned off. This is OK for many uses, especially those where irq_work_queue() is called from an interrupt or softirq handler, because in those cases the scheduler-clock-tick state will be re-evaluated, which will turn the scheduler-clock tick back on. On the next tick, any deferred work will then be processed. However, this strategy does not always work for RCU, which can be invoked at process level from idle CPUs. In this case, the tick might never be turned back on, indefinitely defering a grace-period start request. Note that the RCU CPU stall detector cannot see this condition, because there is no RCU grace period in progress. Therefore, we can (and do!) see long tens-of-seconds stalls in grace-period handling. In theory, we could see a full grace-period hang, but rcutorture testing to date has seen only the tens-of-seconds stalls. Event tracing demonstrates that irq_work_queue() is being called repeatedly to no effect during these stalls: The "newreq" event appears repeatedly from a task that is not one of the grace-period kthreads. In theory, irq_work_queue() might be fixed to avoid this sort of issue, but RCU's requirements are unusual and it is quite straightforward to pass wake-up responsibility up through RCU's call chain, so that the wakeup happens when the offending locks are released. This commit therefore makes this change. The rcu_start_gp_advanced(), rcu_start_future_gp(), rcu_accelerate_cbs(), rcu_advance_cbs(), __note_gp_changes(), and rcu_start_gp() functions now return a boolean which indicates when a wake-up is needed. A new rcu_gp_kthread_wake() does the wakeup when it is necessary and safe to do so: No self-wakes, no wake-ups if the ->gp_flags field indicates there is no need (as in someone else did the wake-up before we got around to it), and no wake-ups before the grace-period kthread has been created. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org>
2014-03-12 04:02:16 +08:00
bool needgp = false;
unsigned long gp_duration;
unsigned long new_gp_seq;
bool offloaded;
struct rcu_data *rdp;
struct rcu_node *rnp = rcu_get_root();
rcu: Use simple wait queues where possible in rcutree As of commit dae6e64d2bcfd ("rcu: Introduce proper blocking to no-CBs kthreads GP waits") the RCU subsystem started making use of wait queues. Here we convert all additions of RCU wait queues to use simple wait queues, since they don't need the extra overhead of the full wait queue features. Originally this was done for RT kernels[1], since we would get things like... BUG: sleeping function called from invalid context at kernel/rtmutex.c:659 in_atomic(): 1, irqs_disabled(): 1, pid: 8, name: rcu_preempt Pid: 8, comm: rcu_preempt Not tainted Call Trace: [<ffffffff8106c8d0>] __might_sleep+0xd0/0xf0 [<ffffffff817d77b4>] rt_spin_lock+0x24/0x50 [<ffffffff8106fcf6>] __wake_up+0x36/0x70 [<ffffffff810c4542>] rcu_gp_kthread+0x4d2/0x680 [<ffffffff8105f910>] ? __init_waitqueue_head+0x50/0x50 [<ffffffff810c4070>] ? rcu_gp_fqs+0x80/0x80 [<ffffffff8105eabb>] kthread+0xdb/0xe0 [<ffffffff8106b912>] ? finish_task_switch+0x52/0x100 [<ffffffff817e0754>] kernel_thread_helper+0x4/0x10 [<ffffffff8105e9e0>] ? __init_kthread_worker+0x60/0x60 [<ffffffff817e0750>] ? gs_change+0xb/0xb ...and hence simple wait queues were deployed on RT out of necessity (as simple wait uses a raw lock), but mainline might as well take advantage of the more streamline support as well. [1] This is a carry forward of work from v3.10-rt; the original conversion was by Thomas on an earlier -rt version, and Sebastian extended it to additional post-3.10 added RCU waiters; here I've added a commit log and unified the RCU changes into one, and uprev'd it to match mainline RCU. Signed-off-by: Daniel Wagner <daniel.wagner@bmw-carit.de> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: linux-rt-users@vger.kernel.org Cc: Boqun Feng <boqun.feng@gmail.com> Cc: Marcelo Tosatti <mtosatti@redhat.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Paolo Bonzini <pbonzini@redhat.com> Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> Link: http://lkml.kernel.org/r/1455871601-27484-6-git-send-email-wagi@monom.org Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2016-02-19 16:46:41 +08:00
struct swait_queue_head *sq;
WRITE_ONCE(rcu_state.gp_activity, jiffies);
raw_spin_lock_irq_rcu_node(rnp);
rcu_state.gp_end = jiffies;
gp_duration = rcu_state.gp_end - rcu_state.gp_start;
if (gp_duration > rcu_state.gp_max)
rcu_state.gp_max = gp_duration;
/*
* We know the grace period is complete, but to everyone else
* it appears to still be ongoing. But it is also the case
* that to everyone else it looks like there is nothing that
* they can do to advance the grace period. It is therefore
* safe for us to drop the lock in order to mark the grace
* period as completed in all of the rcu_node structures.
*/
rcu_poll_gp_seq_end(&rcu_state.gp_seq_polled_snap);
raw_spin_unlock_irq_rcu_node(rnp);
rcu: Fix day-zero grace-period initialization/cleanup race The current approach to grace-period initialization is vulnerable to extremely low-probability races. These races stem from the fact that the old grace period is marked completed on the same traversal through the rcu_node structure that is marking the start of the new grace period. This means that some rcu_node structures will believe that the old grace period is still in effect at the same time that other rcu_node structures believe that the new grace period has already started. These sorts of disagreements can result in too-short grace periods, as shown in the following scenario: 1. CPU 0 completes a grace period, but needs an additional grace period, so starts initializing one, initializing all the non-leaf rcu_node structures and the first leaf rcu_node structure. Because CPU 0 is both completing the old grace period and starting a new one, it marks the completion of the old grace period and the start of the new grace period in a single traversal of the rcu_node structures. Therefore, CPUs corresponding to the first rcu_node structure can become aware that the prior grace period has completed, but CPUs corresponding to the other rcu_node structures will see this same prior grace period as still being in progress. 2. CPU 1 passes through a quiescent state, and therefore informs the RCU core. Because its leaf rcu_node structure has already been initialized, this CPU's quiescent state is applied to the new (and only partially initialized) grace period. 3. CPU 1 enters an RCU read-side critical section and acquires a reference to data item A. Note that this CPU believes that its critical section started after the beginning of the new grace period, and therefore will not block this new grace period. 4. CPU 16 exits dyntick-idle mode. Because it was in dyntick-idle mode, other CPUs informed the RCU core of its extended quiescent state for the past several grace periods. This means that CPU 16 is not yet aware that these past grace periods have ended. Assume that CPU 16 corresponds to the second leaf rcu_node structure -- which has not yet been made aware of the new grace period. 5. CPU 16 removes data item A from its enclosing data structure and passes it to call_rcu(), which queues a callback in the RCU_NEXT_TAIL segment of the callback queue. 6. CPU 16 enters the RCU core, possibly because it has taken a scheduling-clock interrupt, or alternatively because it has more than 10,000 callbacks queued. It notes that the second most recent grace period has completed (recall that because it corresponds to the second as-yet-uninitialized rcu_node structure, it cannot yet become aware that the most recent grace period has completed), and therefore advances its callbacks. The callback for data item A is therefore in the RCU_NEXT_READY_TAIL segment of the callback queue. 7. CPU 0 completes initialization of the remaining leaf rcu_node structures for the new grace period, including the structure corresponding to CPU 16. 8. CPU 16 again enters the RCU core, again, possibly because it has taken a scheduling-clock interrupt, or alternatively because it now has more than 10,000 callbacks queued. It notes that the most recent grace period has ended, and therefore advances its callbacks. The callback for data item A is therefore in the RCU_DONE_TAIL segment of the callback queue. 9. All CPUs other than CPU 1 pass through quiescent states. Because CPU 1 already passed through its quiescent state, the new grace period completes. Note that CPU 1 is still in its RCU read-side critical section, still referencing data item A. 10. Suppose that CPU 2 wais the last CPU to pass through a quiescent state for the new grace period, and suppose further that CPU 2 did not have any callbacks queued, therefore not needing an additional grace period. CPU 2 therefore traverses all of the rcu_node structures, marking the new grace period as completed, but does not initialize a new grace period. 11. CPU 16 yet again enters the RCU core, yet again possibly because it has taken a scheduling-clock interrupt, or alternatively because it now has more than 10,000 callbacks queued. It notes that the new grace period has ended, and therefore advances its callbacks. The callback for data item A is therefore in the RCU_DONE_TAIL segment of the callback queue. This means that this callback is now considered ready to be invoked. 12. CPU 16 invokes the callback, freeing data item A while CPU 1 is still referencing it. This scenario represents a day-zero bug for TREE_RCU. This commit therefore ensures that the old grace period is marked completed in all leaf rcu_node structures before a new grace period is marked started in any of them. That said, it would have been insanely difficult to force this race to happen before the grace-period initialization process was preemptible. Therefore, this commit is not a candidate for -stable. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org> Conflicts: kernel/rcutree.c
2012-07-07 22:56:57 +08:00
/*
* Propagate new ->gp_seq value to rcu_node structures so that
* other CPUs don't have to wait until the start of the next grace
* period to process their callbacks. This also avoids some nasty
* RCU grace-period initialization races by forcing the end of
* the current grace period to be completely recorded in all of
* the rcu_node structures before the beginning of the next grace
* period is recorded in any of the rcu_node structures.
rcu: Fix day-zero grace-period initialization/cleanup race The current approach to grace-period initialization is vulnerable to extremely low-probability races. These races stem from the fact that the old grace period is marked completed on the same traversal through the rcu_node structure that is marking the start of the new grace period. This means that some rcu_node structures will believe that the old grace period is still in effect at the same time that other rcu_node structures believe that the new grace period has already started. These sorts of disagreements can result in too-short grace periods, as shown in the following scenario: 1. CPU 0 completes a grace period, but needs an additional grace period, so starts initializing one, initializing all the non-leaf rcu_node structures and the first leaf rcu_node structure. Because CPU 0 is both completing the old grace period and starting a new one, it marks the completion of the old grace period and the start of the new grace period in a single traversal of the rcu_node structures. Therefore, CPUs corresponding to the first rcu_node structure can become aware that the prior grace period has completed, but CPUs corresponding to the other rcu_node structures will see this same prior grace period as still being in progress. 2. CPU 1 passes through a quiescent state, and therefore informs the RCU core. Because its leaf rcu_node structure has already been initialized, this CPU's quiescent state is applied to the new (and only partially initialized) grace period. 3. CPU 1 enters an RCU read-side critical section and acquires a reference to data item A. Note that this CPU believes that its critical section started after the beginning of the new grace period, and therefore will not block this new grace period. 4. CPU 16 exits dyntick-idle mode. Because it was in dyntick-idle mode, other CPUs informed the RCU core of its extended quiescent state for the past several grace periods. This means that CPU 16 is not yet aware that these past grace periods have ended. Assume that CPU 16 corresponds to the second leaf rcu_node structure -- which has not yet been made aware of the new grace period. 5. CPU 16 removes data item A from its enclosing data structure and passes it to call_rcu(), which queues a callback in the RCU_NEXT_TAIL segment of the callback queue. 6. CPU 16 enters the RCU core, possibly because it has taken a scheduling-clock interrupt, or alternatively because it has more than 10,000 callbacks queued. It notes that the second most recent grace period has completed (recall that because it corresponds to the second as-yet-uninitialized rcu_node structure, it cannot yet become aware that the most recent grace period has completed), and therefore advances its callbacks. The callback for data item A is therefore in the RCU_NEXT_READY_TAIL segment of the callback queue. 7. CPU 0 completes initialization of the remaining leaf rcu_node structures for the new grace period, including the structure corresponding to CPU 16. 8. CPU 16 again enters the RCU core, again, possibly because it has taken a scheduling-clock interrupt, or alternatively because it now has more than 10,000 callbacks queued. It notes that the most recent grace period has ended, and therefore advances its callbacks. The callback for data item A is therefore in the RCU_DONE_TAIL segment of the callback queue. 9. All CPUs other than CPU 1 pass through quiescent states. Because CPU 1 already passed through its quiescent state, the new grace period completes. Note that CPU 1 is still in its RCU read-side critical section, still referencing data item A. 10. Suppose that CPU 2 wais the last CPU to pass through a quiescent state for the new grace period, and suppose further that CPU 2 did not have any callbacks queued, therefore not needing an additional grace period. CPU 2 therefore traverses all of the rcu_node structures, marking the new grace period as completed, but does not initialize a new grace period. 11. CPU 16 yet again enters the RCU core, yet again possibly because it has taken a scheduling-clock interrupt, or alternatively because it now has more than 10,000 callbacks queued. It notes that the new grace period has ended, and therefore advances its callbacks. The callback for data item A is therefore in the RCU_DONE_TAIL segment of the callback queue. This means that this callback is now considered ready to be invoked. 12. CPU 16 invokes the callback, freeing data item A while CPU 1 is still referencing it. This scenario represents a day-zero bug for TREE_RCU. This commit therefore ensures that the old grace period is marked completed in all leaf rcu_node structures before a new grace period is marked started in any of them. That said, it would have been insanely difficult to force this race to happen before the grace-period initialization process was preemptible. Therefore, this commit is not a candidate for -stable. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org> Conflicts: kernel/rcutree.c
2012-07-07 22:56:57 +08:00
*/
new_gp_seq = rcu_state.gp_seq;
rcu_seq_end(&new_gp_seq);
rcu_for_each_node_breadth_first(rnp) {
raw_spin_lock_irq_rcu_node(rnp);
if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
dump_blkd_tasks(rnp, 10);
WARN_ON_ONCE(rnp->qsmask);
WRITE_ONCE(rnp->gp_seq, new_gp_seq);
if (!rnp->parent)
smp_mb(); // Order against failing poll_state_synchronize_rcu_full().
rdp = this_cpu_ptr(&rcu_data);
if (rnp == rdp->mynode)
needgp = __note_gp_changes(rnp, rdp) || needgp;
/* smp_mb() provided by prior unlock-lock pair. */
needgp = rcu_future_gp_cleanup(rnp) || needgp;
// Reset overload indication for CPUs no longer overloaded
if (rcu_is_leaf_node(rnp))
for_each_leaf_node_cpu_mask(rnp, cpu, rnp->cbovldmask) {
rdp = per_cpu_ptr(&rcu_data, cpu);
check_cb_ovld_locked(rdp, rnp);
}
rcu: Do not call rcu_nocb_gp_cleanup() while holding rnp->lock rcu_nocb_gp_cleanup() is called while holding rnp->lock. Currently, this is okay because the wake_up_all() in rcu_nocb_gp_cleanup() will not enable the IRQs. lockdep is happy. By switching over using swait this is not true anymore. swake_up_all() enables the IRQs while processing the waiters. __do_softirq() can now run and will eventually call rcu_process_callbacks() which wants to grap nrp->lock. Let's move the rcu_nocb_gp_cleanup() call outside the lock before we switch over to swait. If we would hold the rnp->lock and use swait, lockdep reports following: ================================= [ INFO: inconsistent lock state ] 4.2.0-rc5-00025-g9a73ba0 #136 Not tainted --------------------------------- inconsistent {IN-SOFTIRQ-W} -> {SOFTIRQ-ON-W} usage. rcu_preempt/8 [HC0[0]:SC0[0]:HE1:SE1] takes: (rcu_node_1){+.?...}, at: [<ffffffff811387c7>] rcu_gp_kthread+0xb97/0xeb0 {IN-SOFTIRQ-W} state was registered at: [<ffffffff81109b9f>] __lock_acquire+0xd5f/0x21e0 [<ffffffff8110be0f>] lock_acquire+0xdf/0x2b0 [<ffffffff81841cc9>] _raw_spin_lock_irqsave+0x59/0xa0 [<ffffffff81136991>] rcu_process_callbacks+0x141/0x3c0 [<ffffffff810b1a9d>] __do_softirq+0x14d/0x670 [<ffffffff810b2214>] irq_exit+0x104/0x110 [<ffffffff81844e96>] smp_apic_timer_interrupt+0x46/0x60 [<ffffffff81842e70>] apic_timer_interrupt+0x70/0x80 [<ffffffff810dba66>] rq_attach_root+0xa6/0x100 [<ffffffff810dbc2d>] cpu_attach_domain+0x16d/0x650 [<ffffffff810e4b42>] build_sched_domains+0x942/0xb00 [<ffffffff821777c2>] sched_init_smp+0x509/0x5c1 [<ffffffff821551e3>] kernel_init_freeable+0x172/0x28f [<ffffffff8182cdce>] kernel_init+0xe/0xe0 [<ffffffff8184231f>] ret_from_fork+0x3f/0x70 irq event stamp: 76 hardirqs last enabled at (75): [<ffffffff81841330>] _raw_spin_unlock_irq+0x30/0x60 hardirqs last disabled at (76): [<ffffffff8184116f>] _raw_spin_lock_irq+0x1f/0x90 softirqs last enabled at (0): [<ffffffff810a8df2>] copy_process.part.26+0x602/0x1cf0 softirqs last disabled at (0): [< (null)>] (null) other info that might help us debug this: Possible unsafe locking scenario: CPU0 ---- lock(rcu_node_1); <Interrupt> lock(rcu_node_1); *** DEADLOCK *** 1 lock held by rcu_preempt/8: #0: (rcu_node_1){+.?...}, at: [<ffffffff811387c7>] rcu_gp_kthread+0xb97/0xeb0 stack backtrace: CPU: 0 PID: 8 Comm: rcu_preempt Not tainted 4.2.0-rc5-00025-g9a73ba0 #136 Hardware name: Dell Inc. PowerEdge R820/066N7P, BIOS 2.0.20 01/16/2014 0000000000000000 000000006d7e67d8 ffff881fb081fbd8 ffffffff818379e0 0000000000000000 ffff881fb0812a00 ffff881fb081fc38 ffffffff8110813b 0000000000000000 0000000000000001 ffff881f00000001 ffffffff8102fa4f Call Trace: [<ffffffff818379e0>] dump_stack+0x4f/0x7b [<ffffffff8110813b>] print_usage_bug+0x1db/0x1e0 [<ffffffff8102fa4f>] ? save_stack_trace+0x2f/0x50 [<ffffffff811087ad>] mark_lock+0x66d/0x6e0 [<ffffffff81107790>] ? check_usage_forwards+0x150/0x150 [<ffffffff81108898>] mark_held_locks+0x78/0xa0 [<ffffffff81841330>] ? _raw_spin_unlock_irq+0x30/0x60 [<ffffffff81108a28>] trace_hardirqs_on_caller+0x168/0x220 [<ffffffff81108aed>] trace_hardirqs_on+0xd/0x10 [<ffffffff81841330>] _raw_spin_unlock_irq+0x30/0x60 [<ffffffff810fd1c7>] swake_up_all+0xb7/0xe0 [<ffffffff811386e1>] rcu_gp_kthread+0xab1/0xeb0 [<ffffffff811089bf>] ? trace_hardirqs_on_caller+0xff/0x220 [<ffffffff81841341>] ? _raw_spin_unlock_irq+0x41/0x60 [<ffffffff81137c30>] ? rcu_barrier+0x20/0x20 [<ffffffff810d2014>] kthread+0x104/0x120 [<ffffffff81841330>] ? _raw_spin_unlock_irq+0x30/0x60 [<ffffffff810d1f10>] ? kthread_create_on_node+0x260/0x260 [<ffffffff8184231f>] ret_from_fork+0x3f/0x70 [<ffffffff810d1f10>] ? kthread_create_on_node+0x260/0x260 Signed-off-by: Daniel Wagner <daniel.wagner@bmw-carit.de> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: linux-rt-users@vger.kernel.org Cc: Boqun Feng <boqun.feng@gmail.com> Cc: Marcelo Tosatti <mtosatti@redhat.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Paolo Bonzini <pbonzini@redhat.com> Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> Link: http://lkml.kernel.org/r/1455871601-27484-5-git-send-email-wagi@monom.org Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2016-02-19 16:46:40 +08:00
sq = rcu_nocb_gp_get(rnp);
raw_spin_unlock_irq_rcu_node(rnp);
rcu: Do not call rcu_nocb_gp_cleanup() while holding rnp->lock rcu_nocb_gp_cleanup() is called while holding rnp->lock. Currently, this is okay because the wake_up_all() in rcu_nocb_gp_cleanup() will not enable the IRQs. lockdep is happy. By switching over using swait this is not true anymore. swake_up_all() enables the IRQs while processing the waiters. __do_softirq() can now run and will eventually call rcu_process_callbacks() which wants to grap nrp->lock. Let's move the rcu_nocb_gp_cleanup() call outside the lock before we switch over to swait. If we would hold the rnp->lock and use swait, lockdep reports following: ================================= [ INFO: inconsistent lock state ] 4.2.0-rc5-00025-g9a73ba0 #136 Not tainted --------------------------------- inconsistent {IN-SOFTIRQ-W} -> {SOFTIRQ-ON-W} usage. rcu_preempt/8 [HC0[0]:SC0[0]:HE1:SE1] takes: (rcu_node_1){+.?...}, at: [<ffffffff811387c7>] rcu_gp_kthread+0xb97/0xeb0 {IN-SOFTIRQ-W} state was registered at: [<ffffffff81109b9f>] __lock_acquire+0xd5f/0x21e0 [<ffffffff8110be0f>] lock_acquire+0xdf/0x2b0 [<ffffffff81841cc9>] _raw_spin_lock_irqsave+0x59/0xa0 [<ffffffff81136991>] rcu_process_callbacks+0x141/0x3c0 [<ffffffff810b1a9d>] __do_softirq+0x14d/0x670 [<ffffffff810b2214>] irq_exit+0x104/0x110 [<ffffffff81844e96>] smp_apic_timer_interrupt+0x46/0x60 [<ffffffff81842e70>] apic_timer_interrupt+0x70/0x80 [<ffffffff810dba66>] rq_attach_root+0xa6/0x100 [<ffffffff810dbc2d>] cpu_attach_domain+0x16d/0x650 [<ffffffff810e4b42>] build_sched_domains+0x942/0xb00 [<ffffffff821777c2>] sched_init_smp+0x509/0x5c1 [<ffffffff821551e3>] kernel_init_freeable+0x172/0x28f [<ffffffff8182cdce>] kernel_init+0xe/0xe0 [<ffffffff8184231f>] ret_from_fork+0x3f/0x70 irq event stamp: 76 hardirqs last enabled at (75): [<ffffffff81841330>] _raw_spin_unlock_irq+0x30/0x60 hardirqs last disabled at (76): [<ffffffff8184116f>] _raw_spin_lock_irq+0x1f/0x90 softirqs last enabled at (0): [<ffffffff810a8df2>] copy_process.part.26+0x602/0x1cf0 softirqs last disabled at (0): [< (null)>] (null) other info that might help us debug this: Possible unsafe locking scenario: CPU0 ---- lock(rcu_node_1); <Interrupt> lock(rcu_node_1); *** DEADLOCK *** 1 lock held by rcu_preempt/8: #0: (rcu_node_1){+.?...}, at: [<ffffffff811387c7>] rcu_gp_kthread+0xb97/0xeb0 stack backtrace: CPU: 0 PID: 8 Comm: rcu_preempt Not tainted 4.2.0-rc5-00025-g9a73ba0 #136 Hardware name: Dell Inc. PowerEdge R820/066N7P, BIOS 2.0.20 01/16/2014 0000000000000000 000000006d7e67d8 ffff881fb081fbd8 ffffffff818379e0 0000000000000000 ffff881fb0812a00 ffff881fb081fc38 ffffffff8110813b 0000000000000000 0000000000000001 ffff881f00000001 ffffffff8102fa4f Call Trace: [<ffffffff818379e0>] dump_stack+0x4f/0x7b [<ffffffff8110813b>] print_usage_bug+0x1db/0x1e0 [<ffffffff8102fa4f>] ? save_stack_trace+0x2f/0x50 [<ffffffff811087ad>] mark_lock+0x66d/0x6e0 [<ffffffff81107790>] ? check_usage_forwards+0x150/0x150 [<ffffffff81108898>] mark_held_locks+0x78/0xa0 [<ffffffff81841330>] ? _raw_spin_unlock_irq+0x30/0x60 [<ffffffff81108a28>] trace_hardirqs_on_caller+0x168/0x220 [<ffffffff81108aed>] trace_hardirqs_on+0xd/0x10 [<ffffffff81841330>] _raw_spin_unlock_irq+0x30/0x60 [<ffffffff810fd1c7>] swake_up_all+0xb7/0xe0 [<ffffffff811386e1>] rcu_gp_kthread+0xab1/0xeb0 [<ffffffff811089bf>] ? trace_hardirqs_on_caller+0xff/0x220 [<ffffffff81841341>] ? _raw_spin_unlock_irq+0x41/0x60 [<ffffffff81137c30>] ? rcu_barrier+0x20/0x20 [<ffffffff810d2014>] kthread+0x104/0x120 [<ffffffff81841330>] ? _raw_spin_unlock_irq+0x30/0x60 [<ffffffff810d1f10>] ? kthread_create_on_node+0x260/0x260 [<ffffffff8184231f>] ret_from_fork+0x3f/0x70 [<ffffffff810d1f10>] ? kthread_create_on_node+0x260/0x260 Signed-off-by: Daniel Wagner <daniel.wagner@bmw-carit.de> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: linux-rt-users@vger.kernel.org Cc: Boqun Feng <boqun.feng@gmail.com> Cc: Marcelo Tosatti <mtosatti@redhat.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Paolo Bonzini <pbonzini@redhat.com> Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> Link: http://lkml.kernel.org/r/1455871601-27484-5-git-send-email-wagi@monom.org Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2016-02-19 16:46:40 +08:00
rcu_nocb_gp_cleanup(sq);
rcu: Rename cond_resched_rcu_qs() to cond_resched_tasks_rcu_qs() Commit e31d28b6ab8f ("trace: Eliminate cond_resched_rcu_qs() in favor of cond_resched()") substituted cond_resched() for the earlier call to cond_resched_rcu_qs(). However, the new-age cond_resched() does not do anything to help RCU-tasks grace periods because (1) RCU-tasks is only enabled when CONFIG_PREEMPT=y and (2) cond_resched() is a complete no-op when preemption is enabled. This situation results in hangs when running the trace benchmarks. A number of potential fixes were discussed on LKML (https://lkml.kernel.org/r/20180224151240.0d63a059@vmware.local.home), including making cond_resched() not be a no-op; making cond_resched() not be a no-op, but only when running tracing benchmarks; reverting the aforementioned commit (which works because cond_resched_rcu_qs() does provide an RCU-tasks quiescent state; and adding a call to the scheduler/RCU rcu_note_voluntary_context_switch() function. All were deemed unsatisfactory, either due to added cond_resched() overhead or due to magic functions inviting cargo culting. This commit renames cond_resched_rcu_qs() to cond_resched_tasks_rcu_qs(), which provides a clear hint as to what this function is doing and why and where it should be used, and then replaces the call to cond_resched() with cond_resched_tasks_rcu_qs() in the trace benchmark's benchmark_event_kthread() function. Reported-by: Steven Rostedt <rostedt@goodmis.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Tested-by: Nicholas Piggin <npiggin@gmail.com>
2018-03-03 08:35:27 +08:00
cond_resched_tasks_rcu_qs();
WRITE_ONCE(rcu_state.gp_activity, jiffies);
rcu_gp_slow(gp_cleanup_delay);
}
rnp = rcu_get_root();
raw_spin_lock_irq_rcu_node(rnp); /* GP before ->gp_seq update. */
/* Declare grace period done, trace first to use old GP number. */
trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("end"));
rcu_seq_end(&rcu_state.gp_seq);
ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq);
rcu: Check and report missed fqs timer wakeup on RCU stall For a new grace period request, the RCU GP kthread transitions through following states: a. [RCU_GP_WAIT_GPS] -> [RCU_GP_DONE_GPS] The RCU_GP_WAIT_GPS state is where the GP kthread waits for a request for a new GP. Once it receives a request (for example, when a new RCU callback is queued), the GP kthread transitions to RCU_GP_DONE_GPS. b. [RCU_GP_DONE_GPS] -> [RCU_GP_ONOFF] Grace period initialization starts in rcu_gp_init(), which records the start of new GP in rcu_state.gp_seq and transitions to RCU_GP_ONOFF. c. [RCU_GP_ONOFF] -> [RCU_GP_INIT] The purpose of the RCU_GP_ONOFF state is to apply the online/offline information that was buffered for any CPUs that recently came online or went offline. This state is maintained in per-leaf rcu_node bitmasks, with the buffered state in ->qsmaskinitnext and the state for the upcoming GP in ->qsmaskinit. At the end of this RCU_GP_ONOFF state, each bit in ->qsmaskinit will correspond to a CPU that must pass through a quiescent state before the upcoming grace period is allowed to complete. However, a leaf rcu_node structure with an all-zeroes ->qsmaskinit cannot necessarily be ignored. In preemptible RCU, there might well be tasks still in RCU read-side critical sections that were first preempted while running on one of the CPUs managed by this structure. Such tasks will be queued on this structure's ->blkd_tasks list. Only after this list fully drains can this leaf rcu_node structure be ignored, and even then only if none of its CPUs have come back online in the meantime. Once that happens, the ->qsmaskinit masks further up the tree will be updated to exclude this leaf rcu_node structure. Once the ->qsmaskinitnext and ->qsmaskinit fields have been updated as needed, the GP kthread transitions to RCU_GP_INIT. d. [RCU_GP_INIT] -> [RCU_GP_WAIT_FQS] The purpose of the RCU_GP_INIT state is to copy each ->qsmaskinit to the ->qsmask field within each rcu_node structure. This copying is done breadth-first from the root to the leaves. Why not just copy directly from ->qsmaskinitnext to ->qsmask? Because the ->qsmaskinitnext masks can change in the meantime as additional CPUs come online or go offline. Such changes would result in inconsistencies in the ->qsmask fields up and down the tree, which could in turn result in too-short grace periods or grace-period hangs. These issues are avoided by snapshotting the leaf rcu_node structures' ->qsmaskinitnext fields into their ->qsmaskinit counterparts, generating a consistent set of ->qsmaskinit fields throughout the tree, and only then copying these consistent ->qsmaskinit fields to their ->qsmask counterparts. Once this initialization step is complete, the GP kthread transitions to RCU_GP_WAIT_FQS, where it waits to do a force-quiescent-state scan on the one hand or for the end of the grace period on the other. e. [RCU_GP_WAIT_FQS] -> [RCU_GP_DOING_FQS] The RCU_GP_WAIT_FQS state waits for one of three things: (1) An explicit request to do a force-quiescent-state scan, (2) The end of the grace period, or (3) A short interval of time, after which it will do a force-quiescent-state (FQS) scan. The explicit request can come from rcutorture or from any CPU that has too many RCU callbacks queued (see the qhimark kernel parameter and the RCU_GP_FLAG_OVLD flag). The aforementioned "short period of time" is specified by the jiffies_till_first_fqs boot parameter for a given grace period's first FQS scan and by the jiffies_till_next_fqs for later FQS scans. Either way, once the wait is over, the GP kthread transitions to RCU_GP_DOING_FQS. f. [RCU_GP_DOING_FQS] -> [RCU_GP_CLEANUP] The RCU_GP_DOING_FQS state performs an FQS scan. Each such scan carries out two functions for any CPU whose bit is still set in its leaf rcu_node structure's ->qsmask field, that is, for any CPU that has not yet reported a quiescent state for the current grace period: i. Report quiescent states on behalf of CPUs that have been observed to be idle (from an RCU perspective) since the beginning of the grace period. ii. If the current grace period is too old, take various actions to encourage holdout CPUs to pass through quiescent states, including enlisting the aid of any calls to cond_resched() and might_sleep(), and even including IPIing the holdout CPUs. These checks are skipped for any leaf rcu_node structure with a all-zero ->qsmask field, however such structures are subject to RCU priority boosting if there are tasks on a given structure blocking the current grace period. The end of the grace period is detected when the root rcu_node structure's ->qsmask is zero and when there are no longer any preempted tasks blocking the current grace period. (No, this last check is not redundant. To see this, consider an rcu_node tree having exactly one structure that serves as both root and leaf.) Once the end of the grace period is detected, the GP kthread transitions to RCU_GP_CLEANUP. g. [RCU_GP_CLEANUP] -> [RCU_GP_CLEANED] The RCU_GP_CLEANUP state marks the end of grace period by updating the rcu_state structure's ->gp_seq field and also all rcu_node structures' ->gp_seq field. As before, the rcu_node tree is traversed in breadth first order. Once this update is complete, the GP kthread transitions to the RCU_GP_CLEANED state. i. [RCU_GP_CLEANED] -> [RCU_GP_INIT] Once in the RCU_GP_CLEANED state, the GP kthread immediately transitions into the RCU_GP_INIT state. j. The role of timers. If there is at least one idle CPU, and if timers are not firing, the transition from RCU_GP_DOING_FQS to RCU_GP_CLEANUP will never happen. Timers can fail to fire for a number of reasons, including issues in timer configuration, issues in the timer framework, and failure to handle softirqs (for example, when there is a storm of interrupts). Whatever the reason, if the timers fail to fire, the GP kthread will never be awakened, resulting in RCU CPU stall warnings and eventually in OOM. However, an RCU CPU stall warning has a large number of potential causes, as documented in Documentation/RCU/stallwarn.rst. This commit therefore adds analysis to the RCU CPU stall-warning code to emit an additional message if the cause of the stall is likely to be timer failure. Signed-off-by: Neeraj Upadhyay <neeraju@codeaurora.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-11-17 00:06:00 +08:00
WRITE_ONCE(rcu_state.gp_state, RCU_GP_IDLE);
/* Check for GP requests since above loop. */
rdp = this_cpu_ptr(&rcu_data);
if (!needgp && ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed)) {
trace_rcu_this_gp(rnp, rdp, rnp->gp_seq_needed,
TPS("CleanupMore"));
needgp = true;
}
rcu: Make callers awaken grace-period kthread The rcu_start_gp_advanced() function currently uses irq_work_queue() to defer wakeups of the RCU grace-period kthread. This deferring is necessary to avoid RCU-scheduler deadlocks involving the rcu_node structure's lock, meaning that RCU cannot call any of the scheduler's wake-up functions while holding one of these locks. Unfortunately, the second and subsequent calls to irq_work_queue() are ignored, and the first call will be ignored (aside from queuing the work item) if the scheduler-clock tick is turned off. This is OK for many uses, especially those where irq_work_queue() is called from an interrupt or softirq handler, because in those cases the scheduler-clock-tick state will be re-evaluated, which will turn the scheduler-clock tick back on. On the next tick, any deferred work will then be processed. However, this strategy does not always work for RCU, which can be invoked at process level from idle CPUs. In this case, the tick might never be turned back on, indefinitely defering a grace-period start request. Note that the RCU CPU stall detector cannot see this condition, because there is no RCU grace period in progress. Therefore, we can (and do!) see long tens-of-seconds stalls in grace-period handling. In theory, we could see a full grace-period hang, but rcutorture testing to date has seen only the tens-of-seconds stalls. Event tracing demonstrates that irq_work_queue() is being called repeatedly to no effect during these stalls: The "newreq" event appears repeatedly from a task that is not one of the grace-period kthreads. In theory, irq_work_queue() might be fixed to avoid this sort of issue, but RCU's requirements are unusual and it is quite straightforward to pass wake-up responsibility up through RCU's call chain, so that the wakeup happens when the offending locks are released. This commit therefore makes this change. The rcu_start_gp_advanced(), rcu_start_future_gp(), rcu_accelerate_cbs(), rcu_advance_cbs(), __note_gp_changes(), and rcu_start_gp() functions now return a boolean which indicates when a wake-up is needed. A new rcu_gp_kthread_wake() does the wakeup when it is necessary and safe to do so: No self-wakes, no wake-ups if the ->gp_flags field indicates there is no need (as in someone else did the wake-up before we got around to it), and no wake-ups before the grace-period kthread has been created. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org>
2014-03-12 04:02:16 +08:00
/* Advance CBs to reduce false positives below. */
offloaded = rcu_rdp_is_offloaded(rdp);
if ((offloaded || !rcu_accelerate_cbs(rnp, rdp)) && needgp) {
// We get here if a grace period was needed (“needgp”)
// and the above call to rcu_accelerate_cbs() did not set
// the RCU_GP_FLAG_INIT bit in ->gp_state (which records
// the need for another grace period).  The purpose
// of the “offloaded” check is to avoid invoking
// rcu_accelerate_cbs() on an offloaded CPU because we do not
// hold the ->nocb_lock needed to safely access an offloaded
// ->cblist.  We do not want to acquire that lock because
// it can be heavily contended during callback floods.
WRITE_ONCE(rcu_state.gp_flags, RCU_GP_FLAG_INIT);
WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("newreq"));
} else {
// We get here either if there is no need for an
// additional grace period or if rcu_accelerate_cbs() has
// already set the RCU_GP_FLAG_INIT bit in ->gp_flags. 
// So all we need to do is to clear all of the other
// ->gp_flags bits.
WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags & RCU_GP_FLAG_INIT);
}
raw_spin_unlock_irq_rcu_node(rnp);
// If strict, make all CPUs aware of the end of the old grace period.
if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
on_each_cpu(rcu_strict_gp_boundary, NULL, 0);
}
/*
* Body of kthread that handles grace periods.
*/
static int __noreturn rcu_gp_kthread(void *unused)
{
rcu_bind_gp_kthread();
for (;;) {
/* Handle grace-period start. */
for (;;) {
trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
TPS("reqwait"));
rcu: Check and report missed fqs timer wakeup on RCU stall For a new grace period request, the RCU GP kthread transitions through following states: a. [RCU_GP_WAIT_GPS] -> [RCU_GP_DONE_GPS] The RCU_GP_WAIT_GPS state is where the GP kthread waits for a request for a new GP. Once it receives a request (for example, when a new RCU callback is queued), the GP kthread transitions to RCU_GP_DONE_GPS. b. [RCU_GP_DONE_GPS] -> [RCU_GP_ONOFF] Grace period initialization starts in rcu_gp_init(), which records the start of new GP in rcu_state.gp_seq and transitions to RCU_GP_ONOFF. c. [RCU_GP_ONOFF] -> [RCU_GP_INIT] The purpose of the RCU_GP_ONOFF state is to apply the online/offline information that was buffered for any CPUs that recently came online or went offline. This state is maintained in per-leaf rcu_node bitmasks, with the buffered state in ->qsmaskinitnext and the state for the upcoming GP in ->qsmaskinit. At the end of this RCU_GP_ONOFF state, each bit in ->qsmaskinit will correspond to a CPU that must pass through a quiescent state before the upcoming grace period is allowed to complete. However, a leaf rcu_node structure with an all-zeroes ->qsmaskinit cannot necessarily be ignored. In preemptible RCU, there might well be tasks still in RCU read-side critical sections that were first preempted while running on one of the CPUs managed by this structure. Such tasks will be queued on this structure's ->blkd_tasks list. Only after this list fully drains can this leaf rcu_node structure be ignored, and even then only if none of its CPUs have come back online in the meantime. Once that happens, the ->qsmaskinit masks further up the tree will be updated to exclude this leaf rcu_node structure. Once the ->qsmaskinitnext and ->qsmaskinit fields have been updated as needed, the GP kthread transitions to RCU_GP_INIT. d. [RCU_GP_INIT] -> [RCU_GP_WAIT_FQS] The purpose of the RCU_GP_INIT state is to copy each ->qsmaskinit to the ->qsmask field within each rcu_node structure. This copying is done breadth-first from the root to the leaves. Why not just copy directly from ->qsmaskinitnext to ->qsmask? Because the ->qsmaskinitnext masks can change in the meantime as additional CPUs come online or go offline. Such changes would result in inconsistencies in the ->qsmask fields up and down the tree, which could in turn result in too-short grace periods or grace-period hangs. These issues are avoided by snapshotting the leaf rcu_node structures' ->qsmaskinitnext fields into their ->qsmaskinit counterparts, generating a consistent set of ->qsmaskinit fields throughout the tree, and only then copying these consistent ->qsmaskinit fields to their ->qsmask counterparts. Once this initialization step is complete, the GP kthread transitions to RCU_GP_WAIT_FQS, where it waits to do a force-quiescent-state scan on the one hand or for the end of the grace period on the other. e. [RCU_GP_WAIT_FQS] -> [RCU_GP_DOING_FQS] The RCU_GP_WAIT_FQS state waits for one of three things: (1) An explicit request to do a force-quiescent-state scan, (2) The end of the grace period, or (3) A short interval of time, after which it will do a force-quiescent-state (FQS) scan. The explicit request can come from rcutorture or from any CPU that has too many RCU callbacks queued (see the qhimark kernel parameter and the RCU_GP_FLAG_OVLD flag). The aforementioned "short period of time" is specified by the jiffies_till_first_fqs boot parameter for a given grace period's first FQS scan and by the jiffies_till_next_fqs for later FQS scans. Either way, once the wait is over, the GP kthread transitions to RCU_GP_DOING_FQS. f. [RCU_GP_DOING_FQS] -> [RCU_GP_CLEANUP] The RCU_GP_DOING_FQS state performs an FQS scan. Each such scan carries out two functions for any CPU whose bit is still set in its leaf rcu_node structure's ->qsmask field, that is, for any CPU that has not yet reported a quiescent state for the current grace period: i. Report quiescent states on behalf of CPUs that have been observed to be idle (from an RCU perspective) since the beginning of the grace period. ii. If the current grace period is too old, take various actions to encourage holdout CPUs to pass through quiescent states, including enlisting the aid of any calls to cond_resched() and might_sleep(), and even including IPIing the holdout CPUs. These checks are skipped for any leaf rcu_node structure with a all-zero ->qsmask field, however such structures are subject to RCU priority boosting if there are tasks on a given structure blocking the current grace period. The end of the grace period is detected when the root rcu_node structure's ->qsmask is zero and when there are no longer any preempted tasks blocking the current grace period. (No, this last check is not redundant. To see this, consider an rcu_node tree having exactly one structure that serves as both root and leaf.) Once the end of the grace period is detected, the GP kthread transitions to RCU_GP_CLEANUP. g. [RCU_GP_CLEANUP] -> [RCU_GP_CLEANED] The RCU_GP_CLEANUP state marks the end of grace period by updating the rcu_state structure's ->gp_seq field and also all rcu_node structures' ->gp_seq field. As before, the rcu_node tree is traversed in breadth first order. Once this update is complete, the GP kthread transitions to the RCU_GP_CLEANED state. i. [RCU_GP_CLEANED] -> [RCU_GP_INIT] Once in the RCU_GP_CLEANED state, the GP kthread immediately transitions into the RCU_GP_INIT state. j. The role of timers. If there is at least one idle CPU, and if timers are not firing, the transition from RCU_GP_DOING_FQS to RCU_GP_CLEANUP will never happen. Timers can fail to fire for a number of reasons, including issues in timer configuration, issues in the timer framework, and failure to handle softirqs (for example, when there is a storm of interrupts). Whatever the reason, if the timers fail to fire, the GP kthread will never be awakened, resulting in RCU CPU stall warnings and eventually in OOM. However, an RCU CPU stall warning has a large number of potential causes, as documented in Documentation/RCU/stallwarn.rst. This commit therefore adds analysis to the RCU CPU stall-warning code to emit an additional message if the cause of the stall is likely to be timer failure. Signed-off-by: Neeraj Upadhyay <neeraju@codeaurora.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-11-17 00:06:00 +08:00
WRITE_ONCE(rcu_state.gp_state, RCU_GP_WAIT_GPS);
swait_event_idle_exclusive(rcu_state.gp_wq,
READ_ONCE(rcu_state.gp_flags) &
RCU_GP_FLAG_INIT);
rcu_gp_torture_wait();
rcu: Check and report missed fqs timer wakeup on RCU stall For a new grace period request, the RCU GP kthread transitions through following states: a. [RCU_GP_WAIT_GPS] -> [RCU_GP_DONE_GPS] The RCU_GP_WAIT_GPS state is where the GP kthread waits for a request for a new GP. Once it receives a request (for example, when a new RCU callback is queued), the GP kthread transitions to RCU_GP_DONE_GPS. b. [RCU_GP_DONE_GPS] -> [RCU_GP_ONOFF] Grace period initialization starts in rcu_gp_init(), which records the start of new GP in rcu_state.gp_seq and transitions to RCU_GP_ONOFF. c. [RCU_GP_ONOFF] -> [RCU_GP_INIT] The purpose of the RCU_GP_ONOFF state is to apply the online/offline information that was buffered for any CPUs that recently came online or went offline. This state is maintained in per-leaf rcu_node bitmasks, with the buffered state in ->qsmaskinitnext and the state for the upcoming GP in ->qsmaskinit. At the end of this RCU_GP_ONOFF state, each bit in ->qsmaskinit will correspond to a CPU that must pass through a quiescent state before the upcoming grace period is allowed to complete. However, a leaf rcu_node structure with an all-zeroes ->qsmaskinit cannot necessarily be ignored. In preemptible RCU, there might well be tasks still in RCU read-side critical sections that were first preempted while running on one of the CPUs managed by this structure. Such tasks will be queued on this structure's ->blkd_tasks list. Only after this list fully drains can this leaf rcu_node structure be ignored, and even then only if none of its CPUs have come back online in the meantime. Once that happens, the ->qsmaskinit masks further up the tree will be updated to exclude this leaf rcu_node structure. Once the ->qsmaskinitnext and ->qsmaskinit fields have been updated as needed, the GP kthread transitions to RCU_GP_INIT. d. [RCU_GP_INIT] -> [RCU_GP_WAIT_FQS] The purpose of the RCU_GP_INIT state is to copy each ->qsmaskinit to the ->qsmask field within each rcu_node structure. This copying is done breadth-first from the root to the leaves. Why not just copy directly from ->qsmaskinitnext to ->qsmask? Because the ->qsmaskinitnext masks can change in the meantime as additional CPUs come online or go offline. Such changes would result in inconsistencies in the ->qsmask fields up and down the tree, which could in turn result in too-short grace periods or grace-period hangs. These issues are avoided by snapshotting the leaf rcu_node structures' ->qsmaskinitnext fields into their ->qsmaskinit counterparts, generating a consistent set of ->qsmaskinit fields throughout the tree, and only then copying these consistent ->qsmaskinit fields to their ->qsmask counterparts. Once this initialization step is complete, the GP kthread transitions to RCU_GP_WAIT_FQS, where it waits to do a force-quiescent-state scan on the one hand or for the end of the grace period on the other. e. [RCU_GP_WAIT_FQS] -> [RCU_GP_DOING_FQS] The RCU_GP_WAIT_FQS state waits for one of three things: (1) An explicit request to do a force-quiescent-state scan, (2) The end of the grace period, or (3) A short interval of time, after which it will do a force-quiescent-state (FQS) scan. The explicit request can come from rcutorture or from any CPU that has too many RCU callbacks queued (see the qhimark kernel parameter and the RCU_GP_FLAG_OVLD flag). The aforementioned "short period of time" is specified by the jiffies_till_first_fqs boot parameter for a given grace period's first FQS scan and by the jiffies_till_next_fqs for later FQS scans. Either way, once the wait is over, the GP kthread transitions to RCU_GP_DOING_FQS. f. [RCU_GP_DOING_FQS] -> [RCU_GP_CLEANUP] The RCU_GP_DOING_FQS state performs an FQS scan. Each such scan carries out two functions for any CPU whose bit is still set in its leaf rcu_node structure's ->qsmask field, that is, for any CPU that has not yet reported a quiescent state for the current grace period: i. Report quiescent states on behalf of CPUs that have been observed to be idle (from an RCU perspective) since the beginning of the grace period. ii. If the current grace period is too old, take various actions to encourage holdout CPUs to pass through quiescent states, including enlisting the aid of any calls to cond_resched() and might_sleep(), and even including IPIing the holdout CPUs. These checks are skipped for any leaf rcu_node structure with a all-zero ->qsmask field, however such structures are subject to RCU priority boosting if there are tasks on a given structure blocking the current grace period. The end of the grace period is detected when the root rcu_node structure's ->qsmask is zero and when there are no longer any preempted tasks blocking the current grace period. (No, this last check is not redundant. To see this, consider an rcu_node tree having exactly one structure that serves as both root and leaf.) Once the end of the grace period is detected, the GP kthread transitions to RCU_GP_CLEANUP. g. [RCU_GP_CLEANUP] -> [RCU_GP_CLEANED] The RCU_GP_CLEANUP state marks the end of grace period by updating the rcu_state structure's ->gp_seq field and also all rcu_node structures' ->gp_seq field. As before, the rcu_node tree is traversed in breadth first order. Once this update is complete, the GP kthread transitions to the RCU_GP_CLEANED state. i. [RCU_GP_CLEANED] -> [RCU_GP_INIT] Once in the RCU_GP_CLEANED state, the GP kthread immediately transitions into the RCU_GP_INIT state. j. The role of timers. If there is at least one idle CPU, and if timers are not firing, the transition from RCU_GP_DOING_FQS to RCU_GP_CLEANUP will never happen. Timers can fail to fire for a number of reasons, including issues in timer configuration, issues in the timer framework, and failure to handle softirqs (for example, when there is a storm of interrupts). Whatever the reason, if the timers fail to fire, the GP kthread will never be awakened, resulting in RCU CPU stall warnings and eventually in OOM. However, an RCU CPU stall warning has a large number of potential causes, as documented in Documentation/RCU/stallwarn.rst. This commit therefore adds analysis to the RCU CPU stall-warning code to emit an additional message if the cause of the stall is likely to be timer failure. Signed-off-by: Neeraj Upadhyay <neeraju@codeaurora.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-11-17 00:06:00 +08:00
WRITE_ONCE(rcu_state.gp_state, RCU_GP_DONE_GPS);
/* Locking provides needed memory barrier. */
if (rcu_gp_init())
break;
rcu: Rename cond_resched_rcu_qs() to cond_resched_tasks_rcu_qs() Commit e31d28b6ab8f ("trace: Eliminate cond_resched_rcu_qs() in favor of cond_resched()") substituted cond_resched() for the earlier call to cond_resched_rcu_qs(). However, the new-age cond_resched() does not do anything to help RCU-tasks grace periods because (1) RCU-tasks is only enabled when CONFIG_PREEMPT=y and (2) cond_resched() is a complete no-op when preemption is enabled. This situation results in hangs when running the trace benchmarks. A number of potential fixes were discussed on LKML (https://lkml.kernel.org/r/20180224151240.0d63a059@vmware.local.home), including making cond_resched() not be a no-op; making cond_resched() not be a no-op, but only when running tracing benchmarks; reverting the aforementioned commit (which works because cond_resched_rcu_qs() does provide an RCU-tasks quiescent state; and adding a call to the scheduler/RCU rcu_note_voluntary_context_switch() function. All were deemed unsatisfactory, either due to added cond_resched() overhead or due to magic functions inviting cargo culting. This commit renames cond_resched_rcu_qs() to cond_resched_tasks_rcu_qs(), which provides a clear hint as to what this function is doing and why and where it should be used, and then replaces the call to cond_resched() with cond_resched_tasks_rcu_qs() in the trace benchmark's benchmark_event_kthread() function. Reported-by: Steven Rostedt <rostedt@goodmis.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Tested-by: Nicholas Piggin <npiggin@gmail.com>
2018-03-03 08:35:27 +08:00
cond_resched_tasks_rcu_qs();
WRITE_ONCE(rcu_state.gp_activity, jiffies);
WARN_ON(signal_pending(current));
trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
TPS("reqwaitsig"));
}
/* Handle quiescent-state forcing. */
rcu_gp_fqs_loop();
/* Handle grace-period end. */
rcu: Check and report missed fqs timer wakeup on RCU stall For a new grace period request, the RCU GP kthread transitions through following states: a. [RCU_GP_WAIT_GPS] -> [RCU_GP_DONE_GPS] The RCU_GP_WAIT_GPS state is where the GP kthread waits for a request for a new GP. Once it receives a request (for example, when a new RCU callback is queued), the GP kthread transitions to RCU_GP_DONE_GPS. b. [RCU_GP_DONE_GPS] -> [RCU_GP_ONOFF] Grace period initialization starts in rcu_gp_init(), which records the start of new GP in rcu_state.gp_seq and transitions to RCU_GP_ONOFF. c. [RCU_GP_ONOFF] -> [RCU_GP_INIT] The purpose of the RCU_GP_ONOFF state is to apply the online/offline information that was buffered for any CPUs that recently came online or went offline. This state is maintained in per-leaf rcu_node bitmasks, with the buffered state in ->qsmaskinitnext and the state for the upcoming GP in ->qsmaskinit. At the end of this RCU_GP_ONOFF state, each bit in ->qsmaskinit will correspond to a CPU that must pass through a quiescent state before the upcoming grace period is allowed to complete. However, a leaf rcu_node structure with an all-zeroes ->qsmaskinit cannot necessarily be ignored. In preemptible RCU, there might well be tasks still in RCU read-side critical sections that were first preempted while running on one of the CPUs managed by this structure. Such tasks will be queued on this structure's ->blkd_tasks list. Only after this list fully drains can this leaf rcu_node structure be ignored, and even then only if none of its CPUs have come back online in the meantime. Once that happens, the ->qsmaskinit masks further up the tree will be updated to exclude this leaf rcu_node structure. Once the ->qsmaskinitnext and ->qsmaskinit fields have been updated as needed, the GP kthread transitions to RCU_GP_INIT. d. [RCU_GP_INIT] -> [RCU_GP_WAIT_FQS] The purpose of the RCU_GP_INIT state is to copy each ->qsmaskinit to the ->qsmask field within each rcu_node structure. This copying is done breadth-first from the root to the leaves. Why not just copy directly from ->qsmaskinitnext to ->qsmask? Because the ->qsmaskinitnext masks can change in the meantime as additional CPUs come online or go offline. Such changes would result in inconsistencies in the ->qsmask fields up and down the tree, which could in turn result in too-short grace periods or grace-period hangs. These issues are avoided by snapshotting the leaf rcu_node structures' ->qsmaskinitnext fields into their ->qsmaskinit counterparts, generating a consistent set of ->qsmaskinit fields throughout the tree, and only then copying these consistent ->qsmaskinit fields to their ->qsmask counterparts. Once this initialization step is complete, the GP kthread transitions to RCU_GP_WAIT_FQS, where it waits to do a force-quiescent-state scan on the one hand or for the end of the grace period on the other. e. [RCU_GP_WAIT_FQS] -> [RCU_GP_DOING_FQS] The RCU_GP_WAIT_FQS state waits for one of three things: (1) An explicit request to do a force-quiescent-state scan, (2) The end of the grace period, or (3) A short interval of time, after which it will do a force-quiescent-state (FQS) scan. The explicit request can come from rcutorture or from any CPU that has too many RCU callbacks queued (see the qhimark kernel parameter and the RCU_GP_FLAG_OVLD flag). The aforementioned "short period of time" is specified by the jiffies_till_first_fqs boot parameter for a given grace period's first FQS scan and by the jiffies_till_next_fqs for later FQS scans. Either way, once the wait is over, the GP kthread transitions to RCU_GP_DOING_FQS. f. [RCU_GP_DOING_FQS] -> [RCU_GP_CLEANUP] The RCU_GP_DOING_FQS state performs an FQS scan. Each such scan carries out two functions for any CPU whose bit is still set in its leaf rcu_node structure's ->qsmask field, that is, for any CPU that has not yet reported a quiescent state for the current grace period: i. Report quiescent states on behalf of CPUs that have been observed to be idle (from an RCU perspective) since the beginning of the grace period. ii. If the current grace period is too old, take various actions to encourage holdout CPUs to pass through quiescent states, including enlisting the aid of any calls to cond_resched() and might_sleep(), and even including IPIing the holdout CPUs. These checks are skipped for any leaf rcu_node structure with a all-zero ->qsmask field, however such structures are subject to RCU priority boosting if there are tasks on a given structure blocking the current grace period. The end of the grace period is detected when the root rcu_node structure's ->qsmask is zero and when there are no longer any preempted tasks blocking the current grace period. (No, this last check is not redundant. To see this, consider an rcu_node tree having exactly one structure that serves as both root and leaf.) Once the end of the grace period is detected, the GP kthread transitions to RCU_GP_CLEANUP. g. [RCU_GP_CLEANUP] -> [RCU_GP_CLEANED] The RCU_GP_CLEANUP state marks the end of grace period by updating the rcu_state structure's ->gp_seq field and also all rcu_node structures' ->gp_seq field. As before, the rcu_node tree is traversed in breadth first order. Once this update is complete, the GP kthread transitions to the RCU_GP_CLEANED state. i. [RCU_GP_CLEANED] -> [RCU_GP_INIT] Once in the RCU_GP_CLEANED state, the GP kthread immediately transitions into the RCU_GP_INIT state. j. The role of timers. If there is at least one idle CPU, and if timers are not firing, the transition from RCU_GP_DOING_FQS to RCU_GP_CLEANUP will never happen. Timers can fail to fire for a number of reasons, including issues in timer configuration, issues in the timer framework, and failure to handle softirqs (for example, when there is a storm of interrupts). Whatever the reason, if the timers fail to fire, the GP kthread will never be awakened, resulting in RCU CPU stall warnings and eventually in OOM. However, an RCU CPU stall warning has a large number of potential causes, as documented in Documentation/RCU/stallwarn.rst. This commit therefore adds analysis to the RCU CPU stall-warning code to emit an additional message if the cause of the stall is likely to be timer failure. Signed-off-by: Neeraj Upadhyay <neeraju@codeaurora.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-11-17 00:06:00 +08:00
WRITE_ONCE(rcu_state.gp_state, RCU_GP_CLEANUP);
rcu_gp_cleanup();
rcu: Check and report missed fqs timer wakeup on RCU stall For a new grace period request, the RCU GP kthread transitions through following states: a. [RCU_GP_WAIT_GPS] -> [RCU_GP_DONE_GPS] The RCU_GP_WAIT_GPS state is where the GP kthread waits for a request for a new GP. Once it receives a request (for example, when a new RCU callback is queued), the GP kthread transitions to RCU_GP_DONE_GPS. b. [RCU_GP_DONE_GPS] -> [RCU_GP_ONOFF] Grace period initialization starts in rcu_gp_init(), which records the start of new GP in rcu_state.gp_seq and transitions to RCU_GP_ONOFF. c. [RCU_GP_ONOFF] -> [RCU_GP_INIT] The purpose of the RCU_GP_ONOFF state is to apply the online/offline information that was buffered for any CPUs that recently came online or went offline. This state is maintained in per-leaf rcu_node bitmasks, with the buffered state in ->qsmaskinitnext and the state for the upcoming GP in ->qsmaskinit. At the end of this RCU_GP_ONOFF state, each bit in ->qsmaskinit will correspond to a CPU that must pass through a quiescent state before the upcoming grace period is allowed to complete. However, a leaf rcu_node structure with an all-zeroes ->qsmaskinit cannot necessarily be ignored. In preemptible RCU, there might well be tasks still in RCU read-side critical sections that were first preempted while running on one of the CPUs managed by this structure. Such tasks will be queued on this structure's ->blkd_tasks list. Only after this list fully drains can this leaf rcu_node structure be ignored, and even then only if none of its CPUs have come back online in the meantime. Once that happens, the ->qsmaskinit masks further up the tree will be updated to exclude this leaf rcu_node structure. Once the ->qsmaskinitnext and ->qsmaskinit fields have been updated as needed, the GP kthread transitions to RCU_GP_INIT. d. [RCU_GP_INIT] -> [RCU_GP_WAIT_FQS] The purpose of the RCU_GP_INIT state is to copy each ->qsmaskinit to the ->qsmask field within each rcu_node structure. This copying is done breadth-first from the root to the leaves. Why not just copy directly from ->qsmaskinitnext to ->qsmask? Because the ->qsmaskinitnext masks can change in the meantime as additional CPUs come online or go offline. Such changes would result in inconsistencies in the ->qsmask fields up and down the tree, which could in turn result in too-short grace periods or grace-period hangs. These issues are avoided by snapshotting the leaf rcu_node structures' ->qsmaskinitnext fields into their ->qsmaskinit counterparts, generating a consistent set of ->qsmaskinit fields throughout the tree, and only then copying these consistent ->qsmaskinit fields to their ->qsmask counterparts. Once this initialization step is complete, the GP kthread transitions to RCU_GP_WAIT_FQS, where it waits to do a force-quiescent-state scan on the one hand or for the end of the grace period on the other. e. [RCU_GP_WAIT_FQS] -> [RCU_GP_DOING_FQS] The RCU_GP_WAIT_FQS state waits for one of three things: (1) An explicit request to do a force-quiescent-state scan, (2) The end of the grace period, or (3) A short interval of time, after which it will do a force-quiescent-state (FQS) scan. The explicit request can come from rcutorture or from any CPU that has too many RCU callbacks queued (see the qhimark kernel parameter and the RCU_GP_FLAG_OVLD flag). The aforementioned "short period of time" is specified by the jiffies_till_first_fqs boot parameter for a given grace period's first FQS scan and by the jiffies_till_next_fqs for later FQS scans. Either way, once the wait is over, the GP kthread transitions to RCU_GP_DOING_FQS. f. [RCU_GP_DOING_FQS] -> [RCU_GP_CLEANUP] The RCU_GP_DOING_FQS state performs an FQS scan. Each such scan carries out two functions for any CPU whose bit is still set in its leaf rcu_node structure's ->qsmask field, that is, for any CPU that has not yet reported a quiescent state for the current grace period: i. Report quiescent states on behalf of CPUs that have been observed to be idle (from an RCU perspective) since the beginning of the grace period. ii. If the current grace period is too old, take various actions to encourage holdout CPUs to pass through quiescent states, including enlisting the aid of any calls to cond_resched() and might_sleep(), and even including IPIing the holdout CPUs. These checks are skipped for any leaf rcu_node structure with a all-zero ->qsmask field, however such structures are subject to RCU priority boosting if there are tasks on a given structure blocking the current grace period. The end of the grace period is detected when the root rcu_node structure's ->qsmask is zero and when there are no longer any preempted tasks blocking the current grace period. (No, this last check is not redundant. To see this, consider an rcu_node tree having exactly one structure that serves as both root and leaf.) Once the end of the grace period is detected, the GP kthread transitions to RCU_GP_CLEANUP. g. [RCU_GP_CLEANUP] -> [RCU_GP_CLEANED] The RCU_GP_CLEANUP state marks the end of grace period by updating the rcu_state structure's ->gp_seq field and also all rcu_node structures' ->gp_seq field. As before, the rcu_node tree is traversed in breadth first order. Once this update is complete, the GP kthread transitions to the RCU_GP_CLEANED state. i. [RCU_GP_CLEANED] -> [RCU_GP_INIT] Once in the RCU_GP_CLEANED state, the GP kthread immediately transitions into the RCU_GP_INIT state. j. The role of timers. If there is at least one idle CPU, and if timers are not firing, the transition from RCU_GP_DOING_FQS to RCU_GP_CLEANUP will never happen. Timers can fail to fire for a number of reasons, including issues in timer configuration, issues in the timer framework, and failure to handle softirqs (for example, when there is a storm of interrupts). Whatever the reason, if the timers fail to fire, the GP kthread will never be awakened, resulting in RCU CPU stall warnings and eventually in OOM. However, an RCU CPU stall warning has a large number of potential causes, as documented in Documentation/RCU/stallwarn.rst. This commit therefore adds analysis to the RCU CPU stall-warning code to emit an additional message if the cause of the stall is likely to be timer failure. Signed-off-by: Neeraj Upadhyay <neeraju@codeaurora.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-11-17 00:06:00 +08:00
WRITE_ONCE(rcu_state.gp_state, RCU_GP_CLEANED);
}
}
rcu: Merge preemptable-RCU functionality into hierarchical RCU Create a kernel/rcutree_plugin.h file that contains definitions for preemptable RCU (or, under the #else branch of the #ifdef, empty definitions for the classic non-preemptable semantics). These definitions fit into plugins defined in kernel/rcutree.c for this purpose. This variant of preemptable RCU uses a new algorithm whose read-side expense is roughly that of classic hierarchical RCU under CONFIG_PREEMPT. This new algorithm's update-side expense is similar to that of classic hierarchical RCU, and, in absence of read-side preemption or blocking, is exactly that of classic hierarchical RCU. Perhaps more important, this new algorithm has a much simpler implementation, saving well over 1,000 lines of code compared to mainline's implementation of preemptable RCU, which will hopefully be retired in favor of this new algorithm. The simplifications are obtained by maintaining per-task nesting state for running tasks, and using a simple lock-protected algorithm to handle accounting when tasks block within RCU read-side critical sections, making use of lessons learned while creating numerous user-level RCU implementations over the past 18 months. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: laijs@cn.fujitsu.com Cc: dipankar@in.ibm.com Cc: akpm@linux-foundation.org Cc: mathieu.desnoyers@polymtl.ca Cc: josht@linux.vnet.ibm.com Cc: dvhltc@us.ibm.com Cc: niv@us.ibm.com Cc: peterz@infradead.org Cc: rostedt@goodmis.org LKML-Reference: <12509746134003-git-send-email-> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-08-23 04:56:52 +08:00
/*
* Report a full set of quiescent states to the rcu_state data structure.
* Invoke rcu_gp_kthread_wake() to awaken the grace-period kthread if
* another grace period is required. Whether we wake the grace-period
* kthread or it awakens itself for the next round of quiescent-state
* forcing, that kthread will clean up after the just-completed grace
* period. Note that the caller must hold rnp->lock, which is released
* before return.
rcu: Merge preemptable-RCU functionality into hierarchical RCU Create a kernel/rcutree_plugin.h file that contains definitions for preemptable RCU (or, under the #else branch of the #ifdef, empty definitions for the classic non-preemptable semantics). These definitions fit into plugins defined in kernel/rcutree.c for this purpose. This variant of preemptable RCU uses a new algorithm whose read-side expense is roughly that of classic hierarchical RCU under CONFIG_PREEMPT. This new algorithm's update-side expense is similar to that of classic hierarchical RCU, and, in absence of read-side preemption or blocking, is exactly that of classic hierarchical RCU. Perhaps more important, this new algorithm has a much simpler implementation, saving well over 1,000 lines of code compared to mainline's implementation of preemptable RCU, which will hopefully be retired in favor of this new algorithm. The simplifications are obtained by maintaining per-task nesting state for running tasks, and using a simple lock-protected algorithm to handle accounting when tasks block within RCU read-side critical sections, making use of lessons learned while creating numerous user-level RCU implementations over the past 18 months. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: laijs@cn.fujitsu.com Cc: dipankar@in.ibm.com Cc: akpm@linux-foundation.org Cc: mathieu.desnoyers@polymtl.ca Cc: josht@linux.vnet.ibm.com Cc: dvhltc@us.ibm.com Cc: niv@us.ibm.com Cc: peterz@infradead.org Cc: rostedt@goodmis.org LKML-Reference: <12509746134003-git-send-email-> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-08-23 04:56:52 +08:00
*/
static void rcu_report_qs_rsp(unsigned long flags)
__releases(rcu_get_root()->lock)
rcu: Merge preemptable-RCU functionality into hierarchical RCU Create a kernel/rcutree_plugin.h file that contains definitions for preemptable RCU (or, under the #else branch of the #ifdef, empty definitions for the classic non-preemptable semantics). These definitions fit into plugins defined in kernel/rcutree.c for this purpose. This variant of preemptable RCU uses a new algorithm whose read-side expense is roughly that of classic hierarchical RCU under CONFIG_PREEMPT. This new algorithm's update-side expense is similar to that of classic hierarchical RCU, and, in absence of read-side preemption or blocking, is exactly that of classic hierarchical RCU. Perhaps more important, this new algorithm has a much simpler implementation, saving well over 1,000 lines of code compared to mainline's implementation of preemptable RCU, which will hopefully be retired in favor of this new algorithm. The simplifications are obtained by maintaining per-task nesting state for running tasks, and using a simple lock-protected algorithm to handle accounting when tasks block within RCU read-side critical sections, making use of lessons learned while creating numerous user-level RCU implementations over the past 18 months. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: laijs@cn.fujitsu.com Cc: dipankar@in.ibm.com Cc: akpm@linux-foundation.org Cc: mathieu.desnoyers@polymtl.ca Cc: josht@linux.vnet.ibm.com Cc: dvhltc@us.ibm.com Cc: niv@us.ibm.com Cc: peterz@infradead.org Cc: rostedt@goodmis.org LKML-Reference: <12509746134003-git-send-email-> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-08-23 04:56:52 +08:00
{
raw_lockdep_assert_held_rcu_node(rcu_get_root());
WARN_ON_ONCE(!rcu_gp_in_progress());
WRITE_ONCE(rcu_state.gp_flags,
READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(), flags);
rcu_gp_kthread_wake();
rcu: Merge preemptable-RCU functionality into hierarchical RCU Create a kernel/rcutree_plugin.h file that contains definitions for preemptable RCU (or, under the #else branch of the #ifdef, empty definitions for the classic non-preemptable semantics). These definitions fit into plugins defined in kernel/rcutree.c for this purpose. This variant of preemptable RCU uses a new algorithm whose read-side expense is roughly that of classic hierarchical RCU under CONFIG_PREEMPT. This new algorithm's update-side expense is similar to that of classic hierarchical RCU, and, in absence of read-side preemption or blocking, is exactly that of classic hierarchical RCU. Perhaps more important, this new algorithm has a much simpler implementation, saving well over 1,000 lines of code compared to mainline's implementation of preemptable RCU, which will hopefully be retired in favor of this new algorithm. The simplifications are obtained by maintaining per-task nesting state for running tasks, and using a simple lock-protected algorithm to handle accounting when tasks block within RCU read-side critical sections, making use of lessons learned while creating numerous user-level RCU implementations over the past 18 months. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: laijs@cn.fujitsu.com Cc: dipankar@in.ibm.com Cc: akpm@linux-foundation.org Cc: mathieu.desnoyers@polymtl.ca Cc: josht@linux.vnet.ibm.com Cc: dvhltc@us.ibm.com Cc: niv@us.ibm.com Cc: peterz@infradead.org Cc: rostedt@goodmis.org LKML-Reference: <12509746134003-git-send-email-> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-08-23 04:56:52 +08:00
}
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/*
rcu: Rename "quiet" functions The number of "quiet" functions has grown recently, and the names are no longer very descriptive. The point of all of these functions is to do some portion of the task of reporting a quiescent state, so rename them accordingly: o cpu_quiet() becomes rcu_report_qs_rdp(), which reports a quiescent state to the per-CPU rcu_data structure. If this turns out to be a new quiescent state for this grace period, then rcu_report_qs_rnp() will be invoked to propagate the quiescent state up the rcu_node hierarchy. o cpu_quiet_msk() becomes rcu_report_qs_rnp(), which reports a quiescent state for a given CPU (or possibly a set of CPUs) up the rcu_node hierarchy. o cpu_quiet_msk_finish() becomes rcu_report_qs_rsp(), which reports a full set of quiescent states to the global rcu_state structure. o task_quiet() becomes rcu_report_unblock_qs_rnp(), which reports a quiescent state due to a task exiting an RCU read-side critical section that had previously blocked in that same critical section. As indicated by the new name, this type of quiescent state is reported up the rcu_node hierarchy (using rcu_report_qs_rnp() to do so). Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Acked-by: Josh Triplett <josh@joshtriplett.org> Acked-by: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: dipankar@in.ibm.com Cc: mathieu.desnoyers@polymtl.ca Cc: dvhltc@us.ibm.com Cc: niv@us.ibm.com Cc: peterz@infradead.org Cc: rostedt@goodmis.org Cc: Valdis.Kletnieks@vt.edu Cc: dhowells@redhat.com LKML-Reference: <12597846163698-git-send-email-> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-12-03 04:10:13 +08:00
* Similar to rcu_report_qs_rdp(), for which it is a helper function.
* Allows quiescent states for a group of CPUs to be reported at one go
* to the specified rcu_node structure, though all the CPUs in the group
rcu: Associate quiescent-state reports with grace period As noted in earlier commit logs, CPU hotplug operations running concurrently with grace-period initialization can result in a given leaf rcu_node structure having all CPUs offline and no blocked readers, but with this rcu_node structure nevertheless blocking the current grace period. Therefore, the quiescent-state forcing code now checks for this situation and repairs it. Unfortunately, this checking can result in false positives, for example, when the last task has just removed itself from this leaf rcu_node structure, but has not yet started clearing the ->qsmask bits further up the structure. This means that the grace-period kthread (which forces quiescent states) and some other task might be attempting to concurrently clear these ->qsmask bits. This is usually not a problem: One of these tasks will be the first to acquire the upper-level rcu_node structure's lock and with therefore clear the bit, and the other task, seeing the bit already cleared, will stop trying to clear bits. Sadly, this means that the following unusual sequence of events -can- result in a problem: 1. The grace-period kthread wins, and clears the ->qsmask bits. 2. This is the last thing blocking the current grace period, so that the grace-period kthread clears ->qsmask bits all the way to the root and finds that the root ->qsmask field is now zero. 3. Another grace period is required, so that the grace period kthread initializes it, including setting all the needed qsmask bits. 4. The leaf rcu_node structure (the one that started this whole mess) is blocking this new grace period, either because it has at least one online CPU or because there is at least one task that had blocked within an RCU read-side critical section while running on one of this leaf rcu_node structure's CPUs. (And yes, that CPU might well have gone offline before the grace period in step (3) above started, which can mean that there is a task on the leaf rcu_node structure's ->blkd_tasks list, but ->qsmask equal to zero.) 5. The other kthread didn't get around to trying to clear the upper level ->qsmask bits until all the above had happened. This means that it now sees bits set in the upper-level ->qsmask field, so it proceeds to clear them. Too bad that it is doing so on behalf of a quiescent state that does not apply to the current grace period! This sequence of events can result in the new grace period being too short. It can also result in the new grace period ending before the leaf rcu_node structure's ->qsmask bits have been cleared, which will result in splats during initialization of the next grace period. In addition, it can result in tasks blocking the new grace period still being queued at the start of the next grace period, which will result in other splats. Sasha's testing turned up another of these splats, as did rcutorture testing. (And yes, rcutorture is being adjusted to make these splats show up more quickly. Which probably is having the undesirable side effect of making other problems show up less quickly. Can't have everything!) Reported-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: <stable@vger.kernel.org> # 4.0.x Tested-by: Sasha Levin <sasha.levin@oracle.com>
2015-03-16 00:19:35 +08:00
* must be represented by the same rcu_node structure (which need not be a
* leaf rcu_node structure, though it often will be). The gps parameter
* is the grace-period snapshot, which means that the quiescent states
* are valid only if rnp->gp_seq is equal to gps. That structure's lock
rcu: Associate quiescent-state reports with grace period As noted in earlier commit logs, CPU hotplug operations running concurrently with grace-period initialization can result in a given leaf rcu_node structure having all CPUs offline and no blocked readers, but with this rcu_node structure nevertheless blocking the current grace period. Therefore, the quiescent-state forcing code now checks for this situation and repairs it. Unfortunately, this checking can result in false positives, for example, when the last task has just removed itself from this leaf rcu_node structure, but has not yet started clearing the ->qsmask bits further up the structure. This means that the grace-period kthread (which forces quiescent states) and some other task might be attempting to concurrently clear these ->qsmask bits. This is usually not a problem: One of these tasks will be the first to acquire the upper-level rcu_node structure's lock and with therefore clear the bit, and the other task, seeing the bit already cleared, will stop trying to clear bits. Sadly, this means that the following unusual sequence of events -can- result in a problem: 1. The grace-period kthread wins, and clears the ->qsmask bits. 2. This is the last thing blocking the current grace period, so that the grace-period kthread clears ->qsmask bits all the way to the root and finds that the root ->qsmask field is now zero. 3. Another grace period is required, so that the grace period kthread initializes it, including setting all the needed qsmask bits. 4. The leaf rcu_node structure (the one that started this whole mess) is blocking this new grace period, either because it has at least one online CPU or because there is at least one task that had blocked within an RCU read-side critical section while running on one of this leaf rcu_node structure's CPUs. (And yes, that CPU might well have gone offline before the grace period in step (3) above started, which can mean that there is a task on the leaf rcu_node structure's ->blkd_tasks list, but ->qsmask equal to zero.) 5. The other kthread didn't get around to trying to clear the upper level ->qsmask bits until all the above had happened. This means that it now sees bits set in the upper-level ->qsmask field, so it proceeds to clear them. Too bad that it is doing so on behalf of a quiescent state that does not apply to the current grace period! This sequence of events can result in the new grace period being too short. It can also result in the new grace period ending before the leaf rcu_node structure's ->qsmask bits have been cleared, which will result in splats during initialization of the next grace period. In addition, it can result in tasks blocking the new grace period still being queued at the start of the next grace period, which will result in other splats. Sasha's testing turned up another of these splats, as did rcutorture testing. (And yes, rcutorture is being adjusted to make these splats show up more quickly. Which probably is having the undesirable side effect of making other problems show up less quickly. Can't have everything!) Reported-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: <stable@vger.kernel.org> # 4.0.x Tested-by: Sasha Levin <sasha.levin@oracle.com>
2015-03-16 00:19:35 +08:00
* must be held upon entry, and it is released before return.
rcu: Fix grace-period hangs from mid-init task resume Without special fail-safe quiescent-state-propagation checks, grace-period hangs can result from the following scenario: 1. A task running on a given CPU is preempted in its RCU read-side critical section. 2. That CPU goes offline, and there are now no online CPUs corresponding to that CPU's leaf rcu_node structure. 3. The rcu_gp_init() function does the first phase of grace-period initialization, and sets the aforementioned leaf rcu_node structure's ->qsmaskinit field to all zeroes. Because there is a blocked task, it does not propagate the zeroing of either ->qsmaskinit or ->qsmaskinitnext up the rcu_node tree. 4. The task resumes on some other CPU and exits its critical section. There is no grace period in progress, so the resulting quiescent state is not reported up the tree. 5. The rcu_gp_init() function does the second phase of grace-period initialization, which results in the leaf rcu_node structure being initialized to expect no further quiescent states, but with that structure's parent expecting a quiescent-state report. The parent will never receive a quiescent state from this leaf rcu_node structure, so the grace period will hang, resulting in RCU CPU stall warnings. It would be good to get rid of the special fail-safe quiescent-state propagation checks. This commit therefore checks the leaf rcu_node structure's ->wait_blkd_tasks field during grace-period initialization. If this flag is set, the rcu_report_qs_rnp() is invoked to immediately report the possible quiescent state. While in the neighborhood, this commit also report quiescent states for any CPUs that went offline between the two phases of grace-period initialization, thus reducing grace-period delays and hopefully eventually allowing removal of offline-CPU checks from the force-quiescent-state code path. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2018-05-08 00:34:17 +08:00
*
* As a special case, if mask is zero, the bit-already-cleared check is
* disabled. This allows propagating quiescent state due to resumed tasks
* during grace-period initialization.
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
*/
static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
unsigned long gps, unsigned long flags)
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
__releases(rnp->lock)
{
rcu: Associate quiescent-state reports with grace period As noted in earlier commit logs, CPU hotplug operations running concurrently with grace-period initialization can result in a given leaf rcu_node structure having all CPUs offline and no blocked readers, but with this rcu_node structure nevertheless blocking the current grace period. Therefore, the quiescent-state forcing code now checks for this situation and repairs it. Unfortunately, this checking can result in false positives, for example, when the last task has just removed itself from this leaf rcu_node structure, but has not yet started clearing the ->qsmask bits further up the structure. This means that the grace-period kthread (which forces quiescent states) and some other task might be attempting to concurrently clear these ->qsmask bits. This is usually not a problem: One of these tasks will be the first to acquire the upper-level rcu_node structure's lock and with therefore clear the bit, and the other task, seeing the bit already cleared, will stop trying to clear bits. Sadly, this means that the following unusual sequence of events -can- result in a problem: 1. The grace-period kthread wins, and clears the ->qsmask bits. 2. This is the last thing blocking the current grace period, so that the grace-period kthread clears ->qsmask bits all the way to the root and finds that the root ->qsmask field is now zero. 3. Another grace period is required, so that the grace period kthread initializes it, including setting all the needed qsmask bits. 4. The leaf rcu_node structure (the one that started this whole mess) is blocking this new grace period, either because it has at least one online CPU or because there is at least one task that had blocked within an RCU read-side critical section while running on one of this leaf rcu_node structure's CPUs. (And yes, that CPU might well have gone offline before the grace period in step (3) above started, which can mean that there is a task on the leaf rcu_node structure's ->blkd_tasks list, but ->qsmask equal to zero.) 5. The other kthread didn't get around to trying to clear the upper level ->qsmask bits until all the above had happened. This means that it now sees bits set in the upper-level ->qsmask field, so it proceeds to clear them. Too bad that it is doing so on behalf of a quiescent state that does not apply to the current grace period! This sequence of events can result in the new grace period being too short. It can also result in the new grace period ending before the leaf rcu_node structure's ->qsmask bits have been cleared, which will result in splats during initialization of the next grace period. In addition, it can result in tasks blocking the new grace period still being queued at the start of the next grace period, which will result in other splats. Sasha's testing turned up another of these splats, as did rcutorture testing. (And yes, rcutorture is being adjusted to make these splats show up more quickly. Which probably is having the undesirable side effect of making other problems show up less quickly. Can't have everything!) Reported-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: <stable@vger.kernel.org> # 4.0.x Tested-by: Sasha Levin <sasha.levin@oracle.com>
2015-03-16 00:19:35 +08:00
unsigned long oldmask = 0;
struct rcu_node *rnp_c;
raw_lockdep_assert_held_rcu_node(rnp);
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/* Walk up the rcu_node hierarchy. */
for (;;) {
rcu: Fix grace-period hangs from mid-init task resume Without special fail-safe quiescent-state-propagation checks, grace-period hangs can result from the following scenario: 1. A task running on a given CPU is preempted in its RCU read-side critical section. 2. That CPU goes offline, and there are now no online CPUs corresponding to that CPU's leaf rcu_node structure. 3. The rcu_gp_init() function does the first phase of grace-period initialization, and sets the aforementioned leaf rcu_node structure's ->qsmaskinit field to all zeroes. Because there is a blocked task, it does not propagate the zeroing of either ->qsmaskinit or ->qsmaskinitnext up the rcu_node tree. 4. The task resumes on some other CPU and exits its critical section. There is no grace period in progress, so the resulting quiescent state is not reported up the tree. 5. The rcu_gp_init() function does the second phase of grace-period initialization, which results in the leaf rcu_node structure being initialized to expect no further quiescent states, but with that structure's parent expecting a quiescent-state report. The parent will never receive a quiescent state from this leaf rcu_node structure, so the grace period will hang, resulting in RCU CPU stall warnings. It would be good to get rid of the special fail-safe quiescent-state propagation checks. This commit therefore checks the leaf rcu_node structure's ->wait_blkd_tasks field during grace-period initialization. If this flag is set, the rcu_report_qs_rnp() is invoked to immediately report the possible quiescent state. While in the neighborhood, this commit also report quiescent states for any CPUs that went offline between the two phases of grace-period initialization, thus reducing grace-period delays and hopefully eventually allowing removal of offline-CPU checks from the force-quiescent-state code path. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2018-05-08 00:34:17 +08:00
if ((!(rnp->qsmask & mask) && mask) || rnp->gp_seq != gps) {
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
rcu: Associate quiescent-state reports with grace period As noted in earlier commit logs, CPU hotplug operations running concurrently with grace-period initialization can result in a given leaf rcu_node structure having all CPUs offline and no blocked readers, but with this rcu_node structure nevertheless blocking the current grace period. Therefore, the quiescent-state forcing code now checks for this situation and repairs it. Unfortunately, this checking can result in false positives, for example, when the last task has just removed itself from this leaf rcu_node structure, but has not yet started clearing the ->qsmask bits further up the structure. This means that the grace-period kthread (which forces quiescent states) and some other task might be attempting to concurrently clear these ->qsmask bits. This is usually not a problem: One of these tasks will be the first to acquire the upper-level rcu_node structure's lock and with therefore clear the bit, and the other task, seeing the bit already cleared, will stop trying to clear bits. Sadly, this means that the following unusual sequence of events -can- result in a problem: 1. The grace-period kthread wins, and clears the ->qsmask bits. 2. This is the last thing blocking the current grace period, so that the grace-period kthread clears ->qsmask bits all the way to the root and finds that the root ->qsmask field is now zero. 3. Another grace period is required, so that the grace period kthread initializes it, including setting all the needed qsmask bits. 4. The leaf rcu_node structure (the one that started this whole mess) is blocking this new grace period, either because it has at least one online CPU or because there is at least one task that had blocked within an RCU read-side critical section while running on one of this leaf rcu_node structure's CPUs. (And yes, that CPU might well have gone offline before the grace period in step (3) above started, which can mean that there is a task on the leaf rcu_node structure's ->blkd_tasks list, but ->qsmask equal to zero.) 5. The other kthread didn't get around to trying to clear the upper level ->qsmask bits until all the above had happened. This means that it now sees bits set in the upper-level ->qsmask field, so it proceeds to clear them. Too bad that it is doing so on behalf of a quiescent state that does not apply to the current grace period! This sequence of events can result in the new grace period being too short. It can also result in the new grace period ending before the leaf rcu_node structure's ->qsmask bits have been cleared, which will result in splats during initialization of the next grace period. In addition, it can result in tasks blocking the new grace period still being queued at the start of the next grace period, which will result in other splats. Sasha's testing turned up another of these splats, as did rcutorture testing. (And yes, rcutorture is being adjusted to make these splats show up more quickly. Which probably is having the undesirable side effect of making other problems show up less quickly. Can't have everything!) Reported-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: <stable@vger.kernel.org> # 4.0.x Tested-by: Sasha Levin <sasha.levin@oracle.com>
2015-03-16 00:19:35 +08:00
/*
* Our bit has already been cleared, or the
* relevant grace period is already over, so done.
*/
raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
return;
}
rcu: Associate quiescent-state reports with grace period As noted in earlier commit logs, CPU hotplug operations running concurrently with grace-period initialization can result in a given leaf rcu_node structure having all CPUs offline and no blocked readers, but with this rcu_node structure nevertheless blocking the current grace period. Therefore, the quiescent-state forcing code now checks for this situation and repairs it. Unfortunately, this checking can result in false positives, for example, when the last task has just removed itself from this leaf rcu_node structure, but has not yet started clearing the ->qsmask bits further up the structure. This means that the grace-period kthread (which forces quiescent states) and some other task might be attempting to concurrently clear these ->qsmask bits. This is usually not a problem: One of these tasks will be the first to acquire the upper-level rcu_node structure's lock and with therefore clear the bit, and the other task, seeing the bit already cleared, will stop trying to clear bits. Sadly, this means that the following unusual sequence of events -can- result in a problem: 1. The grace-period kthread wins, and clears the ->qsmask bits. 2. This is the last thing blocking the current grace period, so that the grace-period kthread clears ->qsmask bits all the way to the root and finds that the root ->qsmask field is now zero. 3. Another grace period is required, so that the grace period kthread initializes it, including setting all the needed qsmask bits. 4. The leaf rcu_node structure (the one that started this whole mess) is blocking this new grace period, either because it has at least one online CPU or because there is at least one task that had blocked within an RCU read-side critical section while running on one of this leaf rcu_node structure's CPUs. (And yes, that CPU might well have gone offline before the grace period in step (3) above started, which can mean that there is a task on the leaf rcu_node structure's ->blkd_tasks list, but ->qsmask equal to zero.) 5. The other kthread didn't get around to trying to clear the upper level ->qsmask bits until all the above had happened. This means that it now sees bits set in the upper-level ->qsmask field, so it proceeds to clear them. Too bad that it is doing so on behalf of a quiescent state that does not apply to the current grace period! This sequence of events can result in the new grace period being too short. It can also result in the new grace period ending before the leaf rcu_node structure's ->qsmask bits have been cleared, which will result in splats during initialization of the next grace period. In addition, it can result in tasks blocking the new grace period still being queued at the start of the next grace period, which will result in other splats. Sasha's testing turned up another of these splats, as did rcutorture testing. (And yes, rcutorture is being adjusted to make these splats show up more quickly. Which probably is having the undesirable side effect of making other problems show up less quickly. Can't have everything!) Reported-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: <stable@vger.kernel.org> # 4.0.x Tested-by: Sasha Levin <sasha.levin@oracle.com>
2015-03-16 00:19:35 +08:00
WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
WARN_ON_ONCE(!rcu_is_leaf_node(rnp) &&
rcu_preempt_blocked_readers_cgp(rnp));
WRITE_ONCE(rnp->qsmask, rnp->qsmask & ~mask);
trace_rcu_quiescent_state_report(rcu_state.name, rnp->gp_seq,
mask, rnp->qsmask, rnp->level,
rnp->grplo, rnp->grphi,
!!rnp->gp_tasks);
if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/* Other bits still set at this level, so done. */
raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
return;
}
rnp->completedqs = rnp->gp_seq;
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
mask = rnp->grpmask;
if (rnp->parent == NULL) {
/* No more levels. Exit loop holding root lock. */
break;
}
raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
rnp_c = rnp;
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
rnp = rnp->parent;
raw_spin_lock_irqsave_rcu_node(rnp, flags);
oldmask = READ_ONCE(rnp_c->qsmask);
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
}
/*
* Get here if we are the last CPU to pass through a quiescent
rcu: Rename "quiet" functions The number of "quiet" functions has grown recently, and the names are no longer very descriptive. The point of all of these functions is to do some portion of the task of reporting a quiescent state, so rename them accordingly: o cpu_quiet() becomes rcu_report_qs_rdp(), which reports a quiescent state to the per-CPU rcu_data structure. If this turns out to be a new quiescent state for this grace period, then rcu_report_qs_rnp() will be invoked to propagate the quiescent state up the rcu_node hierarchy. o cpu_quiet_msk() becomes rcu_report_qs_rnp(), which reports a quiescent state for a given CPU (or possibly a set of CPUs) up the rcu_node hierarchy. o cpu_quiet_msk_finish() becomes rcu_report_qs_rsp(), which reports a full set of quiescent states to the global rcu_state structure. o task_quiet() becomes rcu_report_unblock_qs_rnp(), which reports a quiescent state due to a task exiting an RCU read-side critical section that had previously blocked in that same critical section. As indicated by the new name, this type of quiescent state is reported up the rcu_node hierarchy (using rcu_report_qs_rnp() to do so). Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Acked-by: Josh Triplett <josh@joshtriplett.org> Acked-by: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: dipankar@in.ibm.com Cc: mathieu.desnoyers@polymtl.ca Cc: dvhltc@us.ibm.com Cc: niv@us.ibm.com Cc: peterz@infradead.org Cc: rostedt@goodmis.org Cc: Valdis.Kletnieks@vt.edu Cc: dhowells@redhat.com LKML-Reference: <12597846163698-git-send-email-> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-12-03 04:10:13 +08:00
* state for this grace period. Invoke rcu_report_qs_rsp()
rcu: Merge preemptable-RCU functionality into hierarchical RCU Create a kernel/rcutree_plugin.h file that contains definitions for preemptable RCU (or, under the #else branch of the #ifdef, empty definitions for the classic non-preemptable semantics). These definitions fit into plugins defined in kernel/rcutree.c for this purpose. This variant of preemptable RCU uses a new algorithm whose read-side expense is roughly that of classic hierarchical RCU under CONFIG_PREEMPT. This new algorithm's update-side expense is similar to that of classic hierarchical RCU, and, in absence of read-side preemption or blocking, is exactly that of classic hierarchical RCU. Perhaps more important, this new algorithm has a much simpler implementation, saving well over 1,000 lines of code compared to mainline's implementation of preemptable RCU, which will hopefully be retired in favor of this new algorithm. The simplifications are obtained by maintaining per-task nesting state for running tasks, and using a simple lock-protected algorithm to handle accounting when tasks block within RCU read-side critical sections, making use of lessons learned while creating numerous user-level RCU implementations over the past 18 months. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: laijs@cn.fujitsu.com Cc: dipankar@in.ibm.com Cc: akpm@linux-foundation.org Cc: mathieu.desnoyers@polymtl.ca Cc: josht@linux.vnet.ibm.com Cc: dvhltc@us.ibm.com Cc: niv@us.ibm.com Cc: peterz@infradead.org Cc: rostedt@goodmis.org LKML-Reference: <12509746134003-git-send-email-> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-08-23 04:56:52 +08:00
* to clean up and start the next grace period if one is needed.
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
*/
rcu_report_qs_rsp(flags); /* releases rnp->lock. */
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
}
/*
* Record a quiescent state for all tasks that were previously queued
* on the specified rcu_node structure and that were blocking the current
* RCU grace period. The caller must hold the corresponding rnp->lock with
* irqs disabled, and this lock is released upon return, but irqs remain
* disabled.
*/
static void __maybe_unused
rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
__releases(rnp->lock)
{
rcu: Associate quiescent-state reports with grace period As noted in earlier commit logs, CPU hotplug operations running concurrently with grace-period initialization can result in a given leaf rcu_node structure having all CPUs offline and no blocked readers, but with this rcu_node structure nevertheless blocking the current grace period. Therefore, the quiescent-state forcing code now checks for this situation and repairs it. Unfortunately, this checking can result in false positives, for example, when the last task has just removed itself from this leaf rcu_node structure, but has not yet started clearing the ->qsmask bits further up the structure. This means that the grace-period kthread (which forces quiescent states) and some other task might be attempting to concurrently clear these ->qsmask bits. This is usually not a problem: One of these tasks will be the first to acquire the upper-level rcu_node structure's lock and with therefore clear the bit, and the other task, seeing the bit already cleared, will stop trying to clear bits. Sadly, this means that the following unusual sequence of events -can- result in a problem: 1. The grace-period kthread wins, and clears the ->qsmask bits. 2. This is the last thing blocking the current grace period, so that the grace-period kthread clears ->qsmask bits all the way to the root and finds that the root ->qsmask field is now zero. 3. Another grace period is required, so that the grace period kthread initializes it, including setting all the needed qsmask bits. 4. The leaf rcu_node structure (the one that started this whole mess) is blocking this new grace period, either because it has at least one online CPU or because there is at least one task that had blocked within an RCU read-side critical section while running on one of this leaf rcu_node structure's CPUs. (And yes, that CPU might well have gone offline before the grace period in step (3) above started, which can mean that there is a task on the leaf rcu_node structure's ->blkd_tasks list, but ->qsmask equal to zero.) 5. The other kthread didn't get around to trying to clear the upper level ->qsmask bits until all the above had happened. This means that it now sees bits set in the upper-level ->qsmask field, so it proceeds to clear them. Too bad that it is doing so on behalf of a quiescent state that does not apply to the current grace period! This sequence of events can result in the new grace period being too short. It can also result in the new grace period ending before the leaf rcu_node structure's ->qsmask bits have been cleared, which will result in splats during initialization of the next grace period. In addition, it can result in tasks blocking the new grace period still being queued at the start of the next grace period, which will result in other splats. Sasha's testing turned up another of these splats, as did rcutorture testing. (And yes, rcutorture is being adjusted to make these splats show up more quickly. Which probably is having the undesirable side effect of making other problems show up less quickly. Can't have everything!) Reported-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: <stable@vger.kernel.org> # 4.0.x Tested-by: Sasha Levin <sasha.levin@oracle.com>
2015-03-16 00:19:35 +08:00
unsigned long gps;
unsigned long mask;
struct rcu_node *rnp_p;
raw_lockdep_assert_held_rcu_node(rnp);
if (WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT_RCU)) ||
WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)) ||
rnp->qsmask != 0) {
raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
return; /* Still need more quiescent states! */
}
rnp->completedqs = rnp->gp_seq;
rnp_p = rnp->parent;
if (rnp_p == NULL) {
/*
rcu: Yet another fix for preemption and CPU hotplug As noted earlier, the following sequence of events can occur when running PREEMPT_RCU and HOTPLUG_CPU on a system with a multi-level rcu_node combining tree: 1. A group of tasks block on CPUs corresponding to a given leaf rcu_node structure while within RCU read-side critical sections. 2. All CPUs corrsponding to that rcu_node structure go offline. 3. The next grace period starts, but because there are still tasks blocked, the upper-level bits corresponding to this leaf rcu_node structure remain set. 4. All the tasks exit their RCU read-side critical sections and remove themselves from the leaf rcu_node structure's list, leaving it empty. 5. But because there now is code to check for this condition at force-quiescent-state time, the upper bits are cleared and the grace period completes. However, there is another complication that can occur following step 4 above: 4a. The grace period starts, and the leaf rcu_node structure's gp_tasks pointer is set to NULL because there are no tasks blocked on this structure. 4b. One of the CPUs corresponding to the leaf rcu_node structure comes back online. 4b. An endless stream of tasks are preempted within RCU read-side critical sections on this CPU, such that the ->blkd_tasks list is always non-empty. The grace period will never end. This commit therefore makes the force-quiescent-state processing check only for absence of tasks blocking the current grace period rather than absence of tasks altogether. This will cause a quiescent state to be reported if the current leaf rcu_node structure is not blocking the current grace period and its parent thinks that it is, regardless of how RCU managed to get itself into this state. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: <stable@vger.kernel.org> # 4.0.x Tested-by: Sasha Levin <sasha.levin@oracle.com>
2015-03-09 05:52:27 +08:00
* Only one rcu_node structure in the tree, so don't
* try to report up to its nonexistent parent!
*/
rcu_report_qs_rsp(flags);
return;
}
/* Report up the rest of the hierarchy, tracking current ->gp_seq. */
gps = rnp->gp_seq;
mask = rnp->grpmask;
raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */
rcu_report_qs_rnp(mask, rnp_p, gps, flags);
}
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/*
rcu: Rename "quiet" functions The number of "quiet" functions has grown recently, and the names are no longer very descriptive. The point of all of these functions is to do some portion of the task of reporting a quiescent state, so rename them accordingly: o cpu_quiet() becomes rcu_report_qs_rdp(), which reports a quiescent state to the per-CPU rcu_data structure. If this turns out to be a new quiescent state for this grace period, then rcu_report_qs_rnp() will be invoked to propagate the quiescent state up the rcu_node hierarchy. o cpu_quiet_msk() becomes rcu_report_qs_rnp(), which reports a quiescent state for a given CPU (or possibly a set of CPUs) up the rcu_node hierarchy. o cpu_quiet_msk_finish() becomes rcu_report_qs_rsp(), which reports a full set of quiescent states to the global rcu_state structure. o task_quiet() becomes rcu_report_unblock_qs_rnp(), which reports a quiescent state due to a task exiting an RCU read-side critical section that had previously blocked in that same critical section. As indicated by the new name, this type of quiescent state is reported up the rcu_node hierarchy (using rcu_report_qs_rnp() to do so). Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Acked-by: Josh Triplett <josh@joshtriplett.org> Acked-by: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: dipankar@in.ibm.com Cc: mathieu.desnoyers@polymtl.ca Cc: dvhltc@us.ibm.com Cc: niv@us.ibm.com Cc: peterz@infradead.org Cc: rostedt@goodmis.org Cc: Valdis.Kletnieks@vt.edu Cc: dhowells@redhat.com LKML-Reference: <12597846163698-git-send-email-> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-12-03 04:10:13 +08:00
* Record a quiescent state for the specified CPU to that CPU's rcu_data
* structure. This must be called from the specified CPU.
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
*/
static void
rcu_report_qs_rdp(struct rcu_data *rdp)
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
{
unsigned long flags;
unsigned long mask;
bool needacc = false;
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
struct rcu_node *rnp;
WARN_ON_ONCE(rdp->cpu != smp_processor_id());
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
rnp = rdp->mynode;
raw_spin_lock_irqsave_rcu_node(rnp, flags);
if (rdp->cpu_no_qs.b.norm || rdp->gp_seq != rnp->gp_seq ||
rdp->gpwrap) {
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/*
rcu: Simplify quiescent-state accounting There is often a delay between the time that a CPU passes through a quiescent state and the time that this quiescent state is reported to the RCU core. It is quite possible that the grace period ended before the quiescent state could be reported, for example, some other CPU might have deduced that this CPU passed through dyntick-idle mode. It is critically important that quiescent state be counted only against the grace period that was in effect at the time that the quiescent state was detected. Previously, this was handled by recording the number of the last grace period to complete when passing through a quiescent state. The RCU core then checks this number against the current value, and rejects the quiescent state if there is a mismatch. However, one additional possibility must be accounted for, namely that the quiescent state was recorded after the prior grace period completed but before the current grace period started. In this case, the RCU core must reject the quiescent state, but the recorded number will match. This is handled when the CPU becomes aware of a new grace period -- at that point, it invalidates any prior quiescent state. This works, but is a bit indirect. The new approach records the current grace period, and the RCU core checks to see (1) that this is still the current grace period and (2) that this grace period has not yet ended. This approach simplifies reasoning about correctness, and this commit changes over to this new approach. Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2011-06-27 15:17:43 +08:00
* The grace period in which this quiescent state was
* recorded has ended, so don't report it upwards.
* We will instead need a new quiescent state that lies
* within the current grace period.
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
*/
rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
return;
}
mask = rdp->grpmask;
rdp->core_needs_qs = false;
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
if ((rnp->qsmask & mask) == 0) {
raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
} else {
/*
* This GP can't end until cpu checks in, so all of our
* callbacks can be processed during the next GP.
rcu/nocb: Make rcu_core() callbacks acceleration preempt-safe While reporting a quiescent state for a given CPU, rcu_core() takes advantage of the freshly loaded grace period sequence number and the locked rnp to accelerate the callbacks whose sequence number have been assigned a stale value. This action is only necessary when the rdp isn't offloaded, otherwise the NOCB kthreads already take care of the callbacks progression. However the check for the offloaded state is volatile because it is performed outside the IRQs disabled section. It's possible for the offloading process to preempt rcu_core() at that point on PREEMPT_RT. This is dangerous because rcu_core() may end up accelerating callbacks concurrently with NOCB kthreads without appropriate locking. Fix this with moving the offloaded check inside the rnp locking section. Reported-and-tested-by: Valentin Schneider <valentin.schneider@arm.com> Reviewed-by: Valentin Schneider <valentin.schneider@arm.com> Tested-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Cc: Josh Triplett <josh@joshtriplett.org> Cc: Joel Fernandes <joel@joelfernandes.org> Cc: Boqun Feng <boqun.feng@gmail.com> Cc: Neeraj Upadhyay <neeraju@codeaurora.org> Cc: Uladzislau Rezki <urezki@gmail.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2021-10-19 08:08:09 +08:00
*
* NOCB kthreads have their own way to deal with that...
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
*/
if (!rcu_rdp_is_offloaded(rdp)) {
/*
* The current GP has not yet ended, so it
* should not be possible for rcu_accelerate_cbs()
* to return true. So complain, but don't awaken.
*/
WARN_ON_ONCE(rcu_accelerate_cbs(rnp, rdp));
} else if (!rcu_segcblist_completely_offloaded(&rdp->cblist)) {
/*
* ...but NOCB kthreads may miss or delay callbacks acceleration
* if in the middle of a (de-)offloading process.
*/
needacc = true;
}
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
rcu_disable_urgency_upon_qs(rdp);
rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
rcu: Associate quiescent-state reports with grace period As noted in earlier commit logs, CPU hotplug operations running concurrently with grace-period initialization can result in a given leaf rcu_node structure having all CPUs offline and no blocked readers, but with this rcu_node structure nevertheless blocking the current grace period. Therefore, the quiescent-state forcing code now checks for this situation and repairs it. Unfortunately, this checking can result in false positives, for example, when the last task has just removed itself from this leaf rcu_node structure, but has not yet started clearing the ->qsmask bits further up the structure. This means that the grace-period kthread (which forces quiescent states) and some other task might be attempting to concurrently clear these ->qsmask bits. This is usually not a problem: One of these tasks will be the first to acquire the upper-level rcu_node structure's lock and with therefore clear the bit, and the other task, seeing the bit already cleared, will stop trying to clear bits. Sadly, this means that the following unusual sequence of events -can- result in a problem: 1. The grace-period kthread wins, and clears the ->qsmask bits. 2. This is the last thing blocking the current grace period, so that the grace-period kthread clears ->qsmask bits all the way to the root and finds that the root ->qsmask field is now zero. 3. Another grace period is required, so that the grace period kthread initializes it, including setting all the needed qsmask bits. 4. The leaf rcu_node structure (the one that started this whole mess) is blocking this new grace period, either because it has at least one online CPU or because there is at least one task that had blocked within an RCU read-side critical section while running on one of this leaf rcu_node structure's CPUs. (And yes, that CPU might well have gone offline before the grace period in step (3) above started, which can mean that there is a task on the leaf rcu_node structure's ->blkd_tasks list, but ->qsmask equal to zero.) 5. The other kthread didn't get around to trying to clear the upper level ->qsmask bits until all the above had happened. This means that it now sees bits set in the upper-level ->qsmask field, so it proceeds to clear them. Too bad that it is doing so on behalf of a quiescent state that does not apply to the current grace period! This sequence of events can result in the new grace period being too short. It can also result in the new grace period ending before the leaf rcu_node structure's ->qsmask bits have been cleared, which will result in splats during initialization of the next grace period. In addition, it can result in tasks blocking the new grace period still being queued at the start of the next grace period, which will result in other splats. Sasha's testing turned up another of these splats, as did rcutorture testing. (And yes, rcutorture is being adjusted to make these splats show up more quickly. Which probably is having the undesirable side effect of making other problems show up less quickly. Can't have everything!) Reported-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: <stable@vger.kernel.org> # 4.0.x Tested-by: Sasha Levin <sasha.levin@oracle.com>
2015-03-16 00:19:35 +08:00
/* ^^^ Released rnp->lock */
if (needacc) {
rcu_nocb_lock_irqsave(rdp, flags);
rcu_accelerate_cbs_unlocked(rnp, rdp);
rcu_nocb_unlock_irqrestore(rdp, flags);
}
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
}
}
/*
* Check to see if there is a new grace period of which this CPU
* is not yet aware, and if so, set up local rcu_data state for it.
* Otherwise, see if this CPU has just passed through its first
* quiescent state for this grace period, and record that fact if so.
*/
static void
rcu_check_quiescent_state(struct rcu_data *rdp)
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
{
/* Check for grace-period ends and beginnings. */
note_gp_changes(rdp);
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/*
* Does this CPU still need to do its part for current grace period?
* If no, return and let the other CPUs do their part as well.
*/
if (!rdp->core_needs_qs)
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
return;
/*
* Was there a quiescent state since the beginning of the grace
* period? If no, then exit and wait for the next call.
*/
if (rdp->cpu_no_qs.b.norm)
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
return;
rcu: Rename "quiet" functions The number of "quiet" functions has grown recently, and the names are no longer very descriptive. The point of all of these functions is to do some portion of the task of reporting a quiescent state, so rename them accordingly: o cpu_quiet() becomes rcu_report_qs_rdp(), which reports a quiescent state to the per-CPU rcu_data structure. If this turns out to be a new quiescent state for this grace period, then rcu_report_qs_rnp() will be invoked to propagate the quiescent state up the rcu_node hierarchy. o cpu_quiet_msk() becomes rcu_report_qs_rnp(), which reports a quiescent state for a given CPU (or possibly a set of CPUs) up the rcu_node hierarchy. o cpu_quiet_msk_finish() becomes rcu_report_qs_rsp(), which reports a full set of quiescent states to the global rcu_state structure. o task_quiet() becomes rcu_report_unblock_qs_rnp(), which reports a quiescent state due to a task exiting an RCU read-side critical section that had previously blocked in that same critical section. As indicated by the new name, this type of quiescent state is reported up the rcu_node hierarchy (using rcu_report_qs_rnp() to do so). Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Acked-by: Josh Triplett <josh@joshtriplett.org> Acked-by: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: dipankar@in.ibm.com Cc: mathieu.desnoyers@polymtl.ca Cc: dvhltc@us.ibm.com Cc: niv@us.ibm.com Cc: peterz@infradead.org Cc: rostedt@goodmis.org Cc: Valdis.Kletnieks@vt.edu Cc: dhowells@redhat.com LKML-Reference: <12597846163698-git-send-email-> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-12-03 04:10:13 +08:00
/*
* Tell RCU we are done (but rcu_report_qs_rdp() will be the
* judge of that).
*/
rcu_report_qs_rdp(rdp);
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
}
/* Return true if callback-invocation time limit exceeded. */
static bool rcu_do_batch_check_time(long count, long tlimit,
bool jlimit_check, unsigned long jlimit)
{
// Invoke local_clock() only once per 32 consecutive callbacks.
return unlikely(tlimit) &&
(!likely(count & 31) ||
(IS_ENABLED(CONFIG_RCU_DOUBLE_CHECK_CB_TIME) &&
jlimit_check && time_after(jiffies, jlimit))) &&
local_clock() >= tlimit;
}
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/*
* Invoke any RCU callbacks that have made it to the end of their grace
* period. Throttle as specified by rdp->blimit.
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
*/
static void rcu_do_batch(struct rcu_data *rdp)
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
{
long bl;
long count = 0;
int div;
bool __maybe_unused empty;
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
unsigned long flags;
unsigned long jlimit;
bool jlimit_check = false;
long pending;
struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
struct rcu_head *rhp;
long tlimit = 0;
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
rcu: Tag callback lists with corresponding grace-period number Currently, callbacks are advanced each time the corresponding CPU notices a change in its leaf rcu_node structure's ->completed value (this value counts grace-period completions). This approach has worked quite well, but with the advent of RCU_FAST_NO_HZ, we cannot count on a given CPU seeing all the grace-period completions. When a CPU misses a grace-period completion that occurs while it is in dyntick-idle mode, this will delay invocation of its callbacks. In addition, acceleration of callbacks (when RCU realizes that a given callback need only wait until the end of the next grace period, rather than having to wait for a partial grace period followed by a full grace period) must be carried out extremely carefully. Insufficient acceleration will result in unnecessarily long grace-period latencies, while excessive acceleration will result in premature callback invocation. Changes that involve this tradeoff are therefore among the most nerve-wracking changes to RCU. This commit therefore explicitly tags groups of callbacks with the number of the grace period that they are waiting for. This means that callback-advancement and callback-acceleration functions are idempotent, so that excessive acceleration will merely waste a few CPU cycles. This also allows a CPU to take full advantage of any grace periods that have elapsed while it has been in dyntick-idle mode. It should also enable simulataneous simplifications to and optimizations of RCU_FAST_NO_HZ. Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2012-12-04 05:52:00 +08:00
/* If no callbacks are ready, just return. */
if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
trace_rcu_batch_start(rcu_state.name,
rcu_segcblist_n_cbs(&rdp->cblist), 0);
trace_rcu_batch_end(rcu_state.name, 0,
!rcu_segcblist_empty(&rdp->cblist),
need_resched(), is_idle_task(current),
rcu_is_callbacks_kthread(rdp));
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
return;
}
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/*
* Extract the list of ready callbacks, disabling IRQs to prevent
* races with call_rcu() from interrupt handlers. Leave the
* callback counts, as rcu_barrier() needs to be conservative.
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
*/
rcu_nocb_lock_irqsave(rdp, flags);
WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
pending = rcu_segcblist_get_seglen(&rdp->cblist, RCU_DONE_TAIL);
div = READ_ONCE(rcu_divisor);
div = div < 0 ? 7 : div > sizeof(long) * 8 - 2 ? sizeof(long) * 8 - 2 : div;
bl = max(rdp->blimit, pending >> div);
if ((in_serving_softirq() || rdp->rcu_cpu_kthread_status == RCU_KTHREAD_RUNNING) &&
(IS_ENABLED(CONFIG_RCU_DOUBLE_CHECK_CB_TIME) || unlikely(bl > 100))) {
const long npj = NSEC_PER_SEC / HZ;
long rrn = READ_ONCE(rcu_resched_ns);
rrn = rrn < NSEC_PER_MSEC ? NSEC_PER_MSEC : rrn > NSEC_PER_SEC ? NSEC_PER_SEC : rrn;
tlimit = local_clock() + rrn;
jlimit = jiffies + (rrn + npj + 1) / npj;
jlimit_check = true;
}
trace_rcu_batch_start(rcu_state.name,
rcu_segcblist_n_cbs(&rdp->cblist), bl);
rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
rcu/nocb: Check a stable offloaded state to manipulate qlen_last_fqs_check It's not entirely obvious why rdp->qlen_last_fqs_check is updated before processing the queue only on offloaded rdp. There can be different effect to that, either in favour of triggering the force quiescent state path or not. For example: 1) If the number of callbacks has decreased since the last rdp->qlen_last_fqs_check update (because we recently called rcu_do_batch() and we executed below qhimark callbacks) and the number of processed callbacks on a subsequent do_batch() arranges for exceeding qhimark on non-offloaded but not on offloaded setup, then we may spare a later run to the force quiescent state slow path on __call_rcu_nocb_wake(), as compared to the non-offloaded counterpart scenario. Here is such an offloaded scenario instance: qhimark = 1000 rdp->last_qlen_last_fqs_check = 3000 rcu_segcblist_n_cbs(rdp) = 2000 rcu_do_batch() { if (offloaded) rdp->last_qlen_fqs_check = rcu_segcblist_n_cbs(rdp) // 2000 // run 1000 callback rcu_segcblist_n_cbs(rdp) = 1000 // Not updating rdp->qlen_last_fqs_check if (count < rdp->qlen_last_fqs_check - qhimark) rdp->qlen_last_fqs_check = count; } call_rcu() * 1001 { __call_rcu_nocb_wake() { // not taking the fqs slowpath: // rcu_segcblist_n_cbs(rdp) == 2001 // rdp->qlen_last_fqs_check == 2000 // qhimark == 1000 if (len > rdp->qlen_last_fqs_check + qhimark) ... } In the case of a non-offloaded scenario, rdp->qlen_last_fqs_check would be 1000 and the fqs slowpath would have executed. 2) If the number of callbacks has increased since the last rdp->qlen_last_fqs_check update (because we recently queued below qhimark callbacks) and the number of callbacks executed in rcu_do_batch() doesn't exceed qhimark for either offloaded or non-offloaded setup, then it's possible that the offloaded scenario later run the force quiescent state slow path on __call_rcu_nocb_wake() while the non-offloaded doesn't. qhimark = 1000 rdp->last_qlen_last_fqs_check = 3000 rcu_segcblist_n_cbs(rdp) = 2000 rcu_do_batch() { if (offloaded) rdp->last_qlen_last_fqs_check = rcu_segcblist_n_cbs(rdp) // 2000 // run 100 callbacks // concurrent queued 100 rcu_segcblist_n_cbs(rdp) = 2000 // Not updating rdp->qlen_last_fqs_check if (count < rdp->qlen_last_fqs_check - qhimark) rdp->qlen_last_fqs_check = count; } call_rcu() * 1001 { __call_rcu_nocb_wake() { // Taking the fqs slowpath: // rcu_segcblist_n_cbs(rdp) == 3001 // rdp->qlen_last_fqs_check == 2000 // qhimark == 1000 if (len > rdp->qlen_last_fqs_check + qhimark) ... } In the case of a non-offloaded scenario, rdp->qlen_last_fqs_check would be 3000 and the fqs slowpath would have executed. The reason for updating rdp->qlen_last_fqs_check when invoking callbacks for offloaded CPUs is that there is usually no point in waking up either the rcuog or rcuoc kthreads while in this state. After all, both threads are prohibited from indefinite sleeps. The exception is when some huge number of callbacks are enqueued while rcu_do_batch() is in the midst of invoking, in which case interrupting the rcuog kthread's timed sleep might get more callbacks set up for the next grace period. Reported-and-tested-by: Valentin Schneider <valentin.schneider@arm.com> Tested-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Original-patch-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Cc: Valentin Schneider <valentin.schneider@arm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Cc: Josh Triplett <josh@joshtriplett.org> Cc: Joel Fernandes <joel@joelfernandes.org> Cc: Boqun Feng <boqun.feng@gmail.com> Cc: Neeraj Upadhyay <neeraju@codeaurora.org> Cc: Uladzislau Rezki <urezki@gmail.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2021-10-19 08:08:11 +08:00
if (rcu_rdp_is_offloaded(rdp))
rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbDequeued"));
rcu_nocb_unlock_irqrestore(rdp, flags);
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/* Invoke callbacks. */
tick_dep_set_task(current, TICK_DEP_BIT_RCU);
rhp = rcu_cblist_dequeue(&rcl);
for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
rcu_callback_t f;
rcu/tree: Make rcu_do_batch count how many callbacks were executed The rcu_do_batch() function extracts the ready-to-invoke callbacks from the rcu_segcblist located in the ->cblist field of the current CPU's rcu_data structure. These callbacks are first moved to a local (unsegmented) rcu_cblist. The rcu_do_batch() function then uses this rcu_cblist's ->len field to count how many CBs it has invoked, but it does so by counting that field down from zero. Finally, this function negates the value in this ->len field (resulting in a positive number) and subtracts the result from the ->len field of the current CPU's ->cblist field. Except that it is sometimes necessary for rcu_do_batch() to stop invoking callbacks mid-stream, despite there being more ready to invoke, for example, if a high-priority task wakes up. In this case the remaining not-yet-invoked callbacks are requeued back onto the CPU's ->cblist, but remain in the ready-to-invoke segment of that list. As above, the negative of the local rcu_cblist's ->len field is still subtracted from the ->len field of the current CPU's ->cblist field. The design of counting down from 0 is confusing and error-prone, plus use of a positive count will make it easier to provide a uniform and consistent API to deal with the per-segment counts that are added later in this series. For example, rcu_segcblist_extract_done_cbs() can unconditionally populate the resulting unsegmented list's ->len field during extraction. This commit therefore explicitly counts how many callbacks were executed in rcu_do_batch() itself, counting up from zero, and then uses that to update the per-CPU segcb list's ->len field, without relying on the downcounting of rcl->len from zero. Reviewed-by: Frederic Weisbecker <frederic@kernel.org> Reviewed-by: Neeraj Upadhyay <neeraju@codeaurora.org> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-11-03 22:25:57 +08:00
count++;
debug_rcu_head_unqueue(rhp);
rcu_lock_acquire(&rcu_callback_map);
trace_rcu_invoke_callback(rcu_state.name, rhp);
f = rhp->func;
WRITE_ONCE(rhp->func, (rcu_callback_t)0L);
f(rhp);
rcu_lock_release(&rcu_callback_map);
/*
* Stop only if limit reached and CPU has something to do.
*/
if (in_serving_softirq()) {
if (count >= bl && (need_resched() || !is_idle_task(current)))
break;
/*
* Make sure we don't spend too much time here and deprive other
* softirq vectors of CPU cycles.
*/
if (rcu_do_batch_check_time(count, tlimit, jlimit_check, jlimit))
break;
} else {
// In rcuc/rcuoc context, so no worries about
// depriving other softirq vectors of CPU cycles.
local_bh_enable();
lockdep_assert_irqs_enabled();
cond_resched_tasks_rcu_qs();
lockdep_assert_irqs_enabled();
local_bh_disable();
// But rcuc kthreads can delay quiescent-state
// reporting, so check time limits for them.
if (rdp->rcu_cpu_kthread_status == RCU_KTHREAD_RUNNING &&
rcu_do_batch_check_time(count, tlimit, jlimit_check, jlimit)) {
rdp->rcu_cpu_has_work = 1;
break;
}
}
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
}
rcu_nocb_lock_irqsave(rdp, flags);
rdp->n_cbs_invoked += count;
trace_rcu_batch_end(rcu_state.name, count, !!rcl.head, need_resched(),
is_idle_task(current), rcu_is_callbacks_kthread(rdp));
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/* Update counts and requeue any remaining callbacks. */
rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
rcu/tree: Make rcu_do_batch count how many callbacks were executed The rcu_do_batch() function extracts the ready-to-invoke callbacks from the rcu_segcblist located in the ->cblist field of the current CPU's rcu_data structure. These callbacks are first moved to a local (unsegmented) rcu_cblist. The rcu_do_batch() function then uses this rcu_cblist's ->len field to count how many CBs it has invoked, but it does so by counting that field down from zero. Finally, this function negates the value in this ->len field (resulting in a positive number) and subtracts the result from the ->len field of the current CPU's ->cblist field. Except that it is sometimes necessary for rcu_do_batch() to stop invoking callbacks mid-stream, despite there being more ready to invoke, for example, if a high-priority task wakes up. In this case the remaining not-yet-invoked callbacks are requeued back onto the CPU's ->cblist, but remain in the ready-to-invoke segment of that list. As above, the negative of the local rcu_cblist's ->len field is still subtracted from the ->len field of the current CPU's ->cblist field. The design of counting down from 0 is confusing and error-prone, plus use of a positive count will make it easier to provide a uniform and consistent API to deal with the per-segment counts that are added later in this series. For example, rcu_segcblist_extract_done_cbs() can unconditionally populate the resulting unsegmented list's ->len field during extraction. This commit therefore explicitly counts how many callbacks were executed in rcu_do_batch() itself, counting up from zero, and then uses that to update the per-CPU segcb list's ->len field, without relying on the downcounting of rcl->len from zero. Reviewed-by: Frederic Weisbecker <frederic@kernel.org> Reviewed-by: Neeraj Upadhyay <neeraju@codeaurora.org> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-11-03 22:25:57 +08:00
rcu_segcblist_add_len(&rdp->cblist, -count);
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/* Reinstate batch limit if we have worked down the excess. */
count = rcu_segcblist_n_cbs(&rdp->cblist);
if (rdp->blimit >= DEFAULT_MAX_RCU_BLIMIT && count <= qlowmark)
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
rdp->blimit = blimit;
/* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
if (count == 0 && rdp->qlen_last_fqs_check != 0) {
rdp->qlen_last_fqs_check = 0;
rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs);
} else if (count < rdp->qlen_last_fqs_check - qhimark)
rdp->qlen_last_fqs_check = count;
/*
* The following usually indicates a double call_rcu(). To track
* this down, try building with CONFIG_DEBUG_OBJECTS_RCU_HEAD=y.
*/
empty = rcu_segcblist_empty(&rdp->cblist);
WARN_ON_ONCE(count == 0 && !empty);
rcu/nocb: Add bypass callback queueing Use of the rcu_data structure's segmented ->cblist for no-CBs CPUs takes advantage of unrelated grace periods, thus reducing the memory footprint in the face of floods of call_rcu() invocations. However, the ->cblist field is a more-complex rcu_segcblist structure which must be protected via locking. Even though there are only three entities which can acquire this lock (the CPU invoking call_rcu(), the no-CBs grace-period kthread, and the no-CBs callbacks kthread), the contention on this lock is excessive under heavy stress. This commit therefore greatly reduces contention by provisioning an rcu_cblist structure field named ->nocb_bypass within the rcu_data structure. Each no-CBs CPU is permitted only a limited number of enqueues onto the ->cblist per jiffy, controlled by a new nocb_nobypass_lim_per_jiffy kernel boot parameter that defaults to about 16 enqueues per millisecond (16 * 1000 / HZ). When that limit is exceeded, the CPU instead enqueues onto the new ->nocb_bypass. The ->nocb_bypass is flushed into the ->cblist every jiffy or when the number of callbacks on ->nocb_bypass exceeds qhimark, whichever happens first. During call_rcu() floods, this flushing is carried out by the CPU during the course of its call_rcu() invocations. However, a CPU could simply stop invoking call_rcu() at any time. The no-CBs grace-period kthread therefore carries out less-aggressive flushing (every few jiffies or when the number of callbacks on ->nocb_bypass exceeds (2 * qhimark), whichever comes first). This means that the no-CBs grace-period kthread cannot be permitted to do unbounded waits while there are callbacks on ->nocb_bypass. A ->nocb_bypass_timer is used to provide the needed wakeups. [ paulmck: Apply Coverity feedback reported by Colin Ian King. ] Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
2019-07-03 07:03:33 +08:00
WARN_ON_ONCE(!IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
count != 0 && empty);
WARN_ON_ONCE(count == 0 && rcu_segcblist_n_segment_cbs(&rdp->cblist) != 0);
WARN_ON_ONCE(!empty && rcu_segcblist_n_segment_cbs(&rdp->cblist) == 0);
rcu_nocb_unlock_irqrestore(rdp, flags);
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
tick_dep_clear_task(current, TICK_DEP_BIT_RCU);
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
}
/*
* This function is invoked from each scheduling-clock interrupt,
* and checks to see if this CPU is in a non-context-switch quiescent
* state, for example, user mode or idle loop. It also schedules RCU
* core processing. If the current grace period has gone on too long,
* it will ask the scheduler to manufacture a context switch for the sole
* purpose of providing the needed quiescent state.
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
*/
void rcu_sched_clock_irq(int user)
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
{
unsigned long j;
if (IS_ENABLED(CONFIG_PROVE_RCU)) {
j = jiffies;
WARN_ON_ONCE(time_before(j, __this_cpu_read(rcu_data.last_sched_clock)));
__this_cpu_write(rcu_data.last_sched_clock, j);
}
rcu: Have the RCU tracepoints use the tracepoint_string infrastructure Currently, RCU tracepoints save only a pointer to strings in the ring buffer. When displayed via the /sys/kernel/debug/tracing/trace file they are referenced like the printf "%s" that looks at the address in the ring buffer and prints out the string it points too. This requires that the strings are constant and persistent in the kernel. The problem with this is for tools like trace-cmd and perf that read the binary data from the buffers but have no access to the kernel memory to find out what string is represented by the address in the buffer. By using the tracepoint_string infrastructure, the RCU tracepoint strings can be exported such that userspace tools can map the addresses to the strings. # cat /sys/kernel/debug/tracing/printk_formats 0xffffffff81a4a0e8 : "rcu_preempt" 0xffffffff81a4a0f4 : "rcu_bh" 0xffffffff81a4a100 : "rcu_sched" 0xffffffff818437a0 : "cpuqs" 0xffffffff818437a6 : "rcu_sched" 0xffffffff818437a0 : "cpuqs" 0xffffffff818437b0 : "rcu_bh" 0xffffffff818437b7 : "Start context switch" 0xffffffff818437cc : "End context switch" 0xffffffff818437a0 : "cpuqs" [...] Now userspaces tools can display: rcu_utilization: Start context switch rcu_dyntick: Start 1 0 rcu_utilization: End context switch rcu_batch_start: rcu_preempt CBs=0/5 bl=10 rcu_dyntick: End 0 140000000000000 rcu_invoke_callback: rcu_preempt rhp=0xffff880071c0d600 func=proc_i_callback rcu_invoke_callback: rcu_preempt rhp=0xffff880077b5b230 func=__d_free rcu_dyntick: Start 140000000000000 0 rcu_invoke_callback: rcu_preempt rhp=0xffff880077563980 func=file_free_rcu rcu_batch_end: rcu_preempt CBs-invoked=3 idle=>c<>c<>c<>c< rcu_utilization: End RCU core rcu_grace_period: rcu_preempt 9741 start rcu_dyntick: Start 1 0 rcu_dyntick: End 0 140000000000000 rcu_dyntick: Start 140000000000000 0 Instead of: rcu_utilization: ffffffff81843110 rcu_future_grace_period: ffffffff81842f1d 9939 9939 9940 0 0 3 ffffffff81842f32 rcu_batch_start: ffffffff81842f1d CBs=0/4 bl=10 rcu_future_grace_period: ffffffff81842f1d 9939 9939 9940 0 0 3 ffffffff81842f3c rcu_grace_period: ffffffff81842f1d 9939 ffffffff81842f80 rcu_invoke_callback: ffffffff81842f1d rhp=0xffff88007888aac0 func=file_free_rcu rcu_grace_period: ffffffff81842f1d 9939 ffffffff81842f95 rcu_invoke_callback: ffffffff81842f1d rhp=0xffff88006aeb4600 func=proc_i_callback rcu_future_grace_period: ffffffff81842f1d 9939 9939 9940 0 0 3 ffffffff81842f32 rcu_future_grace_period: ffffffff81842f1d 9939 9939 9940 0 0 3 ffffffff81842f3c rcu_invoke_callback: ffffffff81842f1d rhp=0xffff880071cb9fc0 func=__d_free rcu_grace_period: ffffffff81842f1d 9939 ffffffff81842f80 rcu_invoke_callback: ffffffff81842f1d rhp=0xffff88007888ae80 func=file_free_rcu rcu_batch_end: ffffffff81842f1d CBs-invoked=4 idle=>c<>c<>c<>c< rcu_utilization: ffffffff8184311f Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
2013-07-13 05:18:47 +08:00
trace_rcu_utilization(TPS("Start scheduler-tick"));
lockdep_assert_irqs_disabled();
raw_cpu_inc(rcu_data.ticks_this_gp);
/* The load-acquire pairs with the store-release setting to true. */
if (smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
/* Idle and userspace execution already are quiescent states. */
if (!rcu_is_cpu_rrupt_from_idle() && !user) {
set_tsk_need_resched(current);
set_preempt_need_resched();
}
__this_cpu_write(rcu_data.rcu_urgent_qs, false);
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
}
rcu_flavor_sched_clock_irq(user);
if (rcu_pending(user))
invoke_rcu_core();
rcu-tasks: Make RCU Tasks Trace check for userspace execution Userspace execution is a valid quiescent state for RCU Tasks Trace, but the scheduling-clock interrupt does not currently report such quiescent states. Of course, the scheduling-clock interrupt is not strictly speaking userspace execution. However, the only way that this code is not in a quiescent state is if something invoked rcu_read_lock_trace(), and that would be reflected in the ->trc_reader_nesting field in the task_struct structure. Furthermore, this field is checked by rcu_tasks_trace_qs(), which is invoked by rcu_tasks_qs() which is in turn invoked by rcu_note_voluntary_context_switch() in kernels building at least one of the RCU Tasks flavors. It is therefore safe to invoke rcu_tasks_trace_qs() from the rcu_sched_clock_irq(). But rcu_tasks_qs() also invokes rcu_tasks_classic_qs() for RCU Tasks, which lacks the read-side markers provided by RCU Tasks Trace. This raises the possibility that an RCU Tasks grace period could start after the interrupt from userspace execution, but before the call to rcu_sched_clock_irq(). However, it turns out that this is safe because the RCU Tasks grace period waits for an RCU grace period, which will wait for the entire scheduling-clock interrupt handler, including any RCU Tasks read-side critical section that this handler might contain. This commit therefore updates the rcu_sched_clock_irq() function's check for usermode execution and its call to rcu_tasks_classic_qs() to instead check for both usermode execution and interrupt from idle, and to instead call rcu_note_voluntary_context_switch(). This consolidates code and provides more faster RCU Tasks Trace reporting of quiescent states in kernels that do scheduling-clock interrupts for userspace execution. [ paulmck: Consolidate checks into rcu_sched_clock_irq(). ] Signed-off-by: Zqiang <qiang1.zhang@intel.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2022-07-19 12:39:00 +08:00
if (user || rcu_is_cpu_rrupt_from_idle())
rcu_note_voluntary_context_switch(current);
lockdep_assert_irqs_disabled();
rcu: Have the RCU tracepoints use the tracepoint_string infrastructure Currently, RCU tracepoints save only a pointer to strings in the ring buffer. When displayed via the /sys/kernel/debug/tracing/trace file they are referenced like the printf "%s" that looks at the address in the ring buffer and prints out the string it points too. This requires that the strings are constant and persistent in the kernel. The problem with this is for tools like trace-cmd and perf that read the binary data from the buffers but have no access to the kernel memory to find out what string is represented by the address in the buffer. By using the tracepoint_string infrastructure, the RCU tracepoint strings can be exported such that userspace tools can map the addresses to the strings. # cat /sys/kernel/debug/tracing/printk_formats 0xffffffff81a4a0e8 : "rcu_preempt" 0xffffffff81a4a0f4 : "rcu_bh" 0xffffffff81a4a100 : "rcu_sched" 0xffffffff818437a0 : "cpuqs" 0xffffffff818437a6 : "rcu_sched" 0xffffffff818437a0 : "cpuqs" 0xffffffff818437b0 : "rcu_bh" 0xffffffff818437b7 : "Start context switch" 0xffffffff818437cc : "End context switch" 0xffffffff818437a0 : "cpuqs" [...] Now userspaces tools can display: rcu_utilization: Start context switch rcu_dyntick: Start 1 0 rcu_utilization: End context switch rcu_batch_start: rcu_preempt CBs=0/5 bl=10 rcu_dyntick: End 0 140000000000000 rcu_invoke_callback: rcu_preempt rhp=0xffff880071c0d600 func=proc_i_callback rcu_invoke_callback: rcu_preempt rhp=0xffff880077b5b230 func=__d_free rcu_dyntick: Start 140000000000000 0 rcu_invoke_callback: rcu_preempt rhp=0xffff880077563980 func=file_free_rcu rcu_batch_end: rcu_preempt CBs-invoked=3 idle=>c<>c<>c<>c< rcu_utilization: End RCU core rcu_grace_period: rcu_preempt 9741 start rcu_dyntick: Start 1 0 rcu_dyntick: End 0 140000000000000 rcu_dyntick: Start 140000000000000 0 Instead of: rcu_utilization: ffffffff81843110 rcu_future_grace_period: ffffffff81842f1d 9939 9939 9940 0 0 3 ffffffff81842f32 rcu_batch_start: ffffffff81842f1d CBs=0/4 bl=10 rcu_future_grace_period: ffffffff81842f1d 9939 9939 9940 0 0 3 ffffffff81842f3c rcu_grace_period: ffffffff81842f1d 9939 ffffffff81842f80 rcu_invoke_callback: ffffffff81842f1d rhp=0xffff88007888aac0 func=file_free_rcu rcu_grace_period: ffffffff81842f1d 9939 ffffffff81842f95 rcu_invoke_callback: ffffffff81842f1d rhp=0xffff88006aeb4600 func=proc_i_callback rcu_future_grace_period: ffffffff81842f1d 9939 9939 9940 0 0 3 ffffffff81842f32 rcu_future_grace_period: ffffffff81842f1d 9939 9939 9940 0 0 3 ffffffff81842f3c rcu_invoke_callback: ffffffff81842f1d rhp=0xffff880071cb9fc0 func=__d_free rcu_grace_period: ffffffff81842f1d 9939 ffffffff81842f80 rcu_invoke_callback: ffffffff81842f1d rhp=0xffff88007888ae80 func=file_free_rcu rcu_batch_end: ffffffff81842f1d CBs-invoked=4 idle=>c<>c<>c<>c< rcu_utilization: ffffffff8184311f Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
2013-07-13 05:18:47 +08:00
trace_rcu_utilization(TPS("End scheduler-tick"));
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
}
/*
* Scan the leaf rcu_node structures. For each structure on which all
* CPUs have reported a quiescent state and on which there are tasks
* blocking the current grace period, initiate RCU priority boosting.
* Otherwise, invoke the specified function to check dyntick state for
* each CPU that has not yet reported a quiescent state.
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
*/
static void force_qs_rnp(int (*f)(struct rcu_data *rdp))
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
{
int cpu;
unsigned long flags;
struct rcu_node *rnp;
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
rcu_state.cbovld = rcu_state.cbovldnext;
rcu_state.cbovldnext = false;
rcu_for_each_leaf_node(rnp) {
unsigned long mask = 0;
unsigned long rsmask = 0;
rcu: Rename cond_resched_rcu_qs() to cond_resched_tasks_rcu_qs() Commit e31d28b6ab8f ("trace: Eliminate cond_resched_rcu_qs() in favor of cond_resched()") substituted cond_resched() for the earlier call to cond_resched_rcu_qs(). However, the new-age cond_resched() does not do anything to help RCU-tasks grace periods because (1) RCU-tasks is only enabled when CONFIG_PREEMPT=y and (2) cond_resched() is a complete no-op when preemption is enabled. This situation results in hangs when running the trace benchmarks. A number of potential fixes were discussed on LKML (https://lkml.kernel.org/r/20180224151240.0d63a059@vmware.local.home), including making cond_resched() not be a no-op; making cond_resched() not be a no-op, but only when running tracing benchmarks; reverting the aforementioned commit (which works because cond_resched_rcu_qs() does provide an RCU-tasks quiescent state; and adding a call to the scheduler/RCU rcu_note_voluntary_context_switch() function. All were deemed unsatisfactory, either due to added cond_resched() overhead or due to magic functions inviting cargo culting. This commit renames cond_resched_rcu_qs() to cond_resched_tasks_rcu_qs(), which provides a clear hint as to what this function is doing and why and where it should be used, and then replaces the call to cond_resched() with cond_resched_tasks_rcu_qs() in the trace benchmark's benchmark_event_kthread() function. Reported-by: Steven Rostedt <rostedt@goodmis.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Tested-by: Nicholas Piggin <npiggin@gmail.com>
2018-03-03 08:35:27 +08:00
cond_resched_tasks_rcu_qs();
raw_spin_lock_irqsave_rcu_node(rnp, flags);
rcu_state.cbovldnext |= !!rnp->cbovldmask;
if (rnp->qsmask == 0) {
if (rcu_preempt_blocked_readers_cgp(rnp)) {
rcu: Yet another fix for preemption and CPU hotplug As noted earlier, the following sequence of events can occur when running PREEMPT_RCU and HOTPLUG_CPU on a system with a multi-level rcu_node combining tree: 1. A group of tasks block on CPUs corresponding to a given leaf rcu_node structure while within RCU read-side critical sections. 2. All CPUs corrsponding to that rcu_node structure go offline. 3. The next grace period starts, but because there are still tasks blocked, the upper-level bits corresponding to this leaf rcu_node structure remain set. 4. All the tasks exit their RCU read-side critical sections and remove themselves from the leaf rcu_node structure's list, leaving it empty. 5. But because there now is code to check for this condition at force-quiescent-state time, the upper bits are cleared and the grace period completes. However, there is another complication that can occur following step 4 above: 4a. The grace period starts, and the leaf rcu_node structure's gp_tasks pointer is set to NULL because there are no tasks blocked on this structure. 4b. One of the CPUs corresponding to the leaf rcu_node structure comes back online. 4b. An endless stream of tasks are preempted within RCU read-side critical sections on this CPU, such that the ->blkd_tasks list is always non-empty. The grace period will never end. This commit therefore makes the force-quiescent-state processing check only for absence of tasks blocking the current grace period rather than absence of tasks altogether. This will cause a quiescent state to be reported if the current leaf rcu_node structure is not blocking the current grace period and its parent thinks that it is, regardless of how RCU managed to get itself into this state. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: <stable@vger.kernel.org> # 4.0.x Tested-by: Sasha Levin <sasha.levin@oracle.com>
2015-03-09 05:52:27 +08:00
/*
* No point in scanning bits because they
* are all zero. But we might need to
* priority-boost blocked readers.
*/
rcu_initiate_boost(rnp, flags);
/* rcu_initiate_boost() releases rnp->lock */
continue;
}
raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
continue;
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
}
for_each_leaf_node_cpu_mask(rnp, cpu, rnp->qsmask) {
struct rcu_data *rdp;
int ret;
rdp = per_cpu_ptr(&rcu_data, cpu);
ret = f(rdp);
if (ret > 0) {
mask |= rdp->grpmask;
rcu_disable_urgency_upon_qs(rdp);
nohz_full: Add full-system-idle state machine This commit adds the state machine that takes the per-CPU idle data as input and produces a full-system-idle indication as output. This state machine is driven out of RCU's quiescent-state-forcing mechanism, which invokes rcu_sysidle_check_cpu() to collect per-CPU idle state and then rcu_sysidle_report() to drive the state machine. The full-system-idle state is sampled using rcu_sys_is_idle(), which also drives the state machine if RCU is idle (and does so by forcing RCU to become non-idle). This function returns true if all but the timekeeping CPU (tick_do_timer_cpu) are idle and have been idle long enough to avoid memory contention on the full_sysidle_state state variable. The rcu_sysidle_force_exit() may be called externally to reset the state machine back into non-idle state. For large systems the state machine is driven out of RCU's force-quiescent-state logic, which provides good scalability at the price of millisecond-scale latencies on the transition to full-system-idle state. This is not so good for battery-powered systems, which are usually small enough that they don't need to care about scalability, but which do care deeply about energy efficiency. Small systems therefore drive the state machine directly out of the idle-entry code. The number of CPUs in a "small" system is defined by a new NO_HZ_FULL_SYSIDLE_SMALL Kconfig parameter, which defaults to 8. Note that this is a build-time definition. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> [ paulmck: Use true and false for boolean constants per Lai Jiangshan. ] Reviewed-by: Josh Triplett <josh@joshtriplett.org> [ paulmck: Simplify logic and provide better comments for memory barriers, based on review comments and questions by Lai Jiangshan. ]
2013-06-22 07:37:22 +08:00
}
if (ret < 0)
rsmask |= rdp->grpmask;
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
}
if (mask != 0) {
/* Idle/offline CPUs, report (releases rnp->lock). */
rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
rcu: Process offlining and onlining only at grace-period start Races between CPU hotplug and grace periods can be difficult to resolve, so the ->onoff_mutex is used to exclude the two events. Unfortunately, this means that it is impossible for an outgoing CPU to perform the last bits of its offlining from its last pass through the idle loop, because sleeplocks cannot be acquired in that context. This commit avoids these problems by buffering online and offline events in a new ->qsmaskinitnext field in the leaf rcu_node structures. When a grace period starts, the events accumulated in this mask are applied to the ->qsmaskinit field, and, if needed, up the rcu_node tree. The special case of all CPUs corresponding to a given leaf rcu_node structure being offline while there are still elements in that structure's ->blkd_tasks list is handled using a new ->wait_blkd_tasks field. In this case, propagating the offline bits up the tree is deferred until the beginning of the grace period after all of the tasks have exited their RCU read-side critical sections and removed themselves from the list, at which point the ->wait_blkd_tasks flag is cleared. If one of that leaf rcu_node structure's CPUs comes back online before the list empties, then the ->wait_blkd_tasks flag is simply cleared. This of course means that RCU's notion of which CPUs are offline can be out of date. This is OK because RCU need only wait on CPUs that were online at the time that the grace period started. In addition, RCU's force-quiescent-state actions will handle the case where a CPU goes offline after the grace period starts. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2015-01-24 13:52:37 +08:00
} else {
/* Nothing to do here, so just drop the lock. */
raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
}
for_each_leaf_node_cpu_mask(rnp, cpu, rsmask)
resched_cpu(cpu);
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
}
}
/*
* Force quiescent states on reluctant CPUs, and also detect which
* CPUs are in dyntick-idle mode.
*/
void rcu_force_quiescent_state(void)
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
{
unsigned long flags;
bool ret;
struct rcu_node *rnp;
struct rcu_node *rnp_old = NULL;
/* Funnel through hierarchy to reduce memory contention. */
rnp = raw_cpu_read(rcu_data.mynode);
for (; rnp != NULL; rnp = rnp->parent) {
ret = (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) ||
!raw_spin_trylock(&rnp->fqslock);
if (rnp_old != NULL)
raw_spin_unlock(&rnp_old->fqslock);
if (ret)
return;
rnp_old = rnp;
}
/* rnp_old == rcu_get_root(), rnp == NULL. */
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/* Reached the root of the rcu_node tree, acquire lock. */
raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
raw_spin_unlock(&rnp_old->fqslock);
if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
return; /* Someone beat us to it. */
}
WRITE_ONCE(rcu_state.gp_flags,
READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
rcu_gp_kthread_wake();
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
}
EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
// Workqueue handler for an RCU reader for kernels enforcing struct RCU
// grace periods.
static void strict_work_handler(struct work_struct *work)
{
rcu_read_lock();
rcu_read_unlock();
}
/* Perform RCU core processing work for the current CPU. */
static __latent_entropy void rcu_core(void)
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
{
unsigned long flags;
struct rcu_data *rdp = raw_cpu_ptr(&rcu_data);
struct rcu_node *rnp = rdp->mynode;
/*
* On RT rcu_core() can be preempted when IRQs aren't disabled.
* Therefore this function can race with concurrent NOCB (de-)offloading
* on this CPU and the below condition must be considered volatile.
* However if we race with:
*
* _ Offloading: In the worst case we accelerate or process callbacks
* concurrently with NOCB kthreads. We are guaranteed to
* call rcu_nocb_lock() if that happens.
*
* _ Deoffloading: In the worst case we miss callbacks acceleration or
* processing. This is fine because the early stage
* of deoffloading invokes rcu_core() after setting
* SEGCBLIST_RCU_CORE. So we guarantee that we'll process
* what could have been dismissed without the need to wait
* for the next rcu_pending() check in the next jiffy.
*/
const bool do_batch = !rcu_segcblist_completely_offloaded(&rdp->cblist);
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
if (cpu_is_offline(smp_processor_id()))
return;
trace_rcu_utilization(TPS("Start RCU core"));
WARN_ON_ONCE(!rdp->beenonline);
rcu: Defer reporting RCU-preempt quiescent states when disabled This commit defers reporting of RCU-preempt quiescent states at rcu_read_unlock_special() time when any of interrupts, softirq, or preemption are disabled. These deferred quiescent states are reported at a later RCU_SOFTIRQ, context switch, idle entry, or CPU-hotplug offline operation. Of course, if another RCU read-side critical section has started in the meantime, the reporting of the quiescent state will be further deferred. This also means that disabling preemption, interrupts, and/or softirqs will act as an RCU-preempt read-side critical section. This is enforced by checking preempt_count() as needed. Some special cases must be handled on an ad-hoc basis, for example, context switch is a quiescent state even though both the scheduler and do_exit() disable preemption. In these cases, additional calls to rcu_preempt_deferred_qs() override the preemption disabling. Similar logic overrides disabled interrupts in rcu_preempt_check_callbacks() because in this case the quiescent state happened just before the corresponding scheduling-clock interrupt. In theory, this change lifts a long-standing restriction that required that if interrupts were disabled across a call to rcu_read_unlock() that the matching rcu_read_lock() also be contained within that interrupts-disabled region of code. Because the reporting of the corresponding RCU-preempt quiescent state is now deferred until after interrupts have been enabled, it is no longer possible for this situation to result in deadlocks involving the scheduler's runqueue and priority-inheritance locks. This may allow some code simplification that might reduce interrupt latency a bit. Unfortunately, in practice this would also defer deboosting a low-priority task that had been subjected to RCU priority boosting, so real-time-response considerations might well force this restriction to remain in place. Because RCU-preempt grace periods are now blocked not only by RCU read-side critical sections, but also by disabling of interrupts, preemption, and softirqs, it will be possible to eliminate RCU-bh and RCU-sched in favor of RCU-preempt in CONFIG_PREEMPT=y kernels. This may require some additional plumbing to provide the network denial-of-service guarantees that have been traditionally provided by RCU-bh. Once these are in place, CONFIG_PREEMPT=n kernels will be able to fold RCU-bh into RCU-sched. This would mean that all kernels would have but one flavor of RCU, which would open the door to significant code cleanup. Moving to a single flavor of RCU would also have the beneficial effect of reducing the NOCB kthreads by at least a factor of two. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Apply rcu_read_unlock_special() preempt_count() feedback from Joel Fernandes. ] [ paulmck: Adjust rcu_eqs_enter() call to rcu_preempt_deferred_qs() in response to bug reports from kbuild test robot. ] [ paulmck: Fix bug located by kbuild test robot involving recursion via rcu_preempt_deferred_qs(). ]
2018-06-22 03:50:01 +08:00
/* Report any deferred quiescent states if preemption enabled. */
if (IS_ENABLED(CONFIG_PREEMPT_COUNT) && (!(preempt_count() & PREEMPT_MASK))) {
rcu: Defer reporting RCU-preempt quiescent states when disabled This commit defers reporting of RCU-preempt quiescent states at rcu_read_unlock_special() time when any of interrupts, softirq, or preemption are disabled. These deferred quiescent states are reported at a later RCU_SOFTIRQ, context switch, idle entry, or CPU-hotplug offline operation. Of course, if another RCU read-side critical section has started in the meantime, the reporting of the quiescent state will be further deferred. This also means that disabling preemption, interrupts, and/or softirqs will act as an RCU-preempt read-side critical section. This is enforced by checking preempt_count() as needed. Some special cases must be handled on an ad-hoc basis, for example, context switch is a quiescent state even though both the scheduler and do_exit() disable preemption. In these cases, additional calls to rcu_preempt_deferred_qs() override the preemption disabling. Similar logic overrides disabled interrupts in rcu_preempt_check_callbacks() because in this case the quiescent state happened just before the corresponding scheduling-clock interrupt. In theory, this change lifts a long-standing restriction that required that if interrupts were disabled across a call to rcu_read_unlock() that the matching rcu_read_lock() also be contained within that interrupts-disabled region of code. Because the reporting of the corresponding RCU-preempt quiescent state is now deferred until after interrupts have been enabled, it is no longer possible for this situation to result in deadlocks involving the scheduler's runqueue and priority-inheritance locks. This may allow some code simplification that might reduce interrupt latency a bit. Unfortunately, in practice this would also defer deboosting a low-priority task that had been subjected to RCU priority boosting, so real-time-response considerations might well force this restriction to remain in place. Because RCU-preempt grace periods are now blocked not only by RCU read-side critical sections, but also by disabling of interrupts, preemption, and softirqs, it will be possible to eliminate RCU-bh and RCU-sched in favor of RCU-preempt in CONFIG_PREEMPT=y kernels. This may require some additional plumbing to provide the network denial-of-service guarantees that have been traditionally provided by RCU-bh. Once these are in place, CONFIG_PREEMPT=n kernels will be able to fold RCU-bh into RCU-sched. This would mean that all kernels would have but one flavor of RCU, which would open the door to significant code cleanup. Moving to a single flavor of RCU would also have the beneficial effect of reducing the NOCB kthreads by at least a factor of two. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Apply rcu_read_unlock_special() preempt_count() feedback from Joel Fernandes. ] [ paulmck: Adjust rcu_eqs_enter() call to rcu_preempt_deferred_qs() in response to bug reports from kbuild test robot. ] [ paulmck: Fix bug located by kbuild test robot involving recursion via rcu_preempt_deferred_qs(). ]
2018-06-22 03:50:01 +08:00
rcu_preempt_deferred_qs(current);
} else if (rcu_preempt_need_deferred_qs(current)) {
set_tsk_need_resched(current);
set_preempt_need_resched();
}
rcu: Defer reporting RCU-preempt quiescent states when disabled This commit defers reporting of RCU-preempt quiescent states at rcu_read_unlock_special() time when any of interrupts, softirq, or preemption are disabled. These deferred quiescent states are reported at a later RCU_SOFTIRQ, context switch, idle entry, or CPU-hotplug offline operation. Of course, if another RCU read-side critical section has started in the meantime, the reporting of the quiescent state will be further deferred. This also means that disabling preemption, interrupts, and/or softirqs will act as an RCU-preempt read-side critical section. This is enforced by checking preempt_count() as needed. Some special cases must be handled on an ad-hoc basis, for example, context switch is a quiescent state even though both the scheduler and do_exit() disable preemption. In these cases, additional calls to rcu_preempt_deferred_qs() override the preemption disabling. Similar logic overrides disabled interrupts in rcu_preempt_check_callbacks() because in this case the quiescent state happened just before the corresponding scheduling-clock interrupt. In theory, this change lifts a long-standing restriction that required that if interrupts were disabled across a call to rcu_read_unlock() that the matching rcu_read_lock() also be contained within that interrupts-disabled region of code. Because the reporting of the corresponding RCU-preempt quiescent state is now deferred until after interrupts have been enabled, it is no longer possible for this situation to result in deadlocks involving the scheduler's runqueue and priority-inheritance locks. This may allow some code simplification that might reduce interrupt latency a bit. Unfortunately, in practice this would also defer deboosting a low-priority task that had been subjected to RCU priority boosting, so real-time-response considerations might well force this restriction to remain in place. Because RCU-preempt grace periods are now blocked not only by RCU read-side critical sections, but also by disabling of interrupts, preemption, and softirqs, it will be possible to eliminate RCU-bh and RCU-sched in favor of RCU-preempt in CONFIG_PREEMPT=y kernels. This may require some additional plumbing to provide the network denial-of-service guarantees that have been traditionally provided by RCU-bh. Once these are in place, CONFIG_PREEMPT=n kernels will be able to fold RCU-bh into RCU-sched. This would mean that all kernels would have but one flavor of RCU, which would open the door to significant code cleanup. Moving to a single flavor of RCU would also have the beneficial effect of reducing the NOCB kthreads by at least a factor of two. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Apply rcu_read_unlock_special() preempt_count() feedback from Joel Fernandes. ] [ paulmck: Adjust rcu_eqs_enter() call to rcu_preempt_deferred_qs() in response to bug reports from kbuild test robot. ] [ paulmck: Fix bug located by kbuild test robot involving recursion via rcu_preempt_deferred_qs(). ]
2018-06-22 03:50:01 +08:00
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/* Update RCU state based on any recent quiescent states. */
rcu_check_quiescent_state(rdp);
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
rcu: Switch __rcu_process_callbacks() to rcu_accelerate_cbs() The __rcu_process_callbacks() function currently checks to see if the current CPU needs a grace period and also if there is any other reason to kick off a new grace period. This is one of the fail-safe checks that has been rendered unnecessary by the changes that increase the accuracy of rcu_gp_cleanup()'s estimate as to whether another grace period is required. Because this particular fail-safe involved acquiring the root rcu_node structure's ->lock, which has seen excessive contention in real life, this fail-safe needs to go. However, one check must remain, namely the check for newly arrived RCU callbacks that have not yet been associated with a grace period. One might hope that the checks in __note_gp_changes(), which is invoked indirectly from rcu_check_quiescent_state(), would suffice, but this function won't be invoked at all if RCU is idle. It is therefore necessary to replace the fail-safe checks with a simpler check for newly arrived callbacks during an RCU idle period, which is exactly what this commit does. This change removes the final call to rcu_start_gp(), so this function is removed as well. Note that lockless use of cpu_needs_another_gp() is racy, but that these races are harmless in this case. If RCU really is idle, the values will not change, so the return value from cpu_needs_another_gp() will be correct. If RCU is not idle, the resulting redundant call to rcu_accelerate_cbs() will be harmless, and might even have the benefit of reducing grace-period latency a bit. This commit also moves interrupt disabling into the "if" statement to improve real-time response a bit. Reported-by: Nicholas Piggin <npiggin@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Tested-by: Nicholas Piggin <npiggin@gmail.com>
2018-04-12 00:51:20 +08:00
/* No grace period and unregistered callbacks? */
if (!rcu_gp_in_progress() &&
rcu_segcblist_is_enabled(&rdp->cblist) && do_batch) {
rcu_nocb_lock_irqsave(rdp, flags);
if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
rcu_accelerate_cbs_unlocked(rnp, rdp);
rcu_nocb_unlock_irqrestore(rdp, flags);
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
}
rcu_check_gp_start_stall(rnp, rdp, rcu_jiffies_till_stall_check());
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/* If there are callbacks ready, invoke them. */
if (do_batch && rcu_segcblist_ready_cbs(&rdp->cblist) &&
likely(READ_ONCE(rcu_scheduler_fully_active))) {
rcu_do_batch(rdp);
/* Re-invoke RCU core processing if there are callbacks remaining. */
if (rcu_segcblist_ready_cbs(&rdp->cblist))
invoke_rcu_core();
}
rcu: Break call_rcu() deadlock involving scheduler and perf Dave Jones got the following lockdep splat: > ====================================================== > [ INFO: possible circular locking dependency detected ] > 3.12.0-rc3+ #92 Not tainted > ------------------------------------------------------- > trinity-child2/15191 is trying to acquire lock: > (&rdp->nocb_wq){......}, at: [<ffffffff8108ff43>] __wake_up+0x23/0x50 > > but task is already holding lock: > (&ctx->lock){-.-...}, at: [<ffffffff81154c19>] perf_event_exit_task+0x109/0x230 > > which lock already depends on the new lock. > > > the existing dependency chain (in reverse order) is: > > -> #3 (&ctx->lock){-.-...}: > [<ffffffff810cc243>] lock_acquire+0x93/0x200 > [<ffffffff81733f90>] _raw_spin_lock+0x40/0x80 > [<ffffffff811500ff>] __perf_event_task_sched_out+0x2df/0x5e0 > [<ffffffff81091b83>] perf_event_task_sched_out+0x93/0xa0 > [<ffffffff81732052>] __schedule+0x1d2/0xa20 > [<ffffffff81732f30>] preempt_schedule_irq+0x50/0xb0 > [<ffffffff817352b6>] retint_kernel+0x26/0x30 > [<ffffffff813eed04>] tty_flip_buffer_push+0x34/0x50 > [<ffffffff813f0504>] pty_write+0x54/0x60 > [<ffffffff813e900d>] n_tty_write+0x32d/0x4e0 > [<ffffffff813e5838>] tty_write+0x158/0x2d0 > [<ffffffff811c4850>] vfs_write+0xc0/0x1f0 > [<ffffffff811c52cc>] SyS_write+0x4c/0xa0 > [<ffffffff8173d4e4>] tracesys+0xdd/0xe2 > > -> #2 (&rq->lock){-.-.-.}: > [<ffffffff810cc243>] lock_acquire+0x93/0x200 > [<ffffffff81733f90>] _raw_spin_lock+0x40/0x80 > [<ffffffff810980b2>] wake_up_new_task+0xc2/0x2e0 > [<ffffffff81054336>] do_fork+0x126/0x460 > [<ffffffff81054696>] kernel_thread+0x26/0x30 > [<ffffffff8171ff93>] rest_init+0x23/0x140 > [<ffffffff81ee1e4b>] start_kernel+0x3f6/0x403 > [<ffffffff81ee1571>] x86_64_start_reservations+0x2a/0x2c > [<ffffffff81ee1664>] x86_64_start_kernel+0xf1/0xf4 > > -> #1 (&p->pi_lock){-.-.-.}: > [<ffffffff810cc243>] lock_acquire+0x93/0x200 > [<ffffffff8173419b>] _raw_spin_lock_irqsave+0x4b/0x90 > [<ffffffff810979d1>] try_to_wake_up+0x31/0x350 > [<ffffffff81097d62>] default_wake_function+0x12/0x20 > [<ffffffff81084af8>] autoremove_wake_function+0x18/0x40 > [<ffffffff8108ea38>] __wake_up_common+0x58/0x90 > [<ffffffff8108ff59>] __wake_up+0x39/0x50 > [<ffffffff8110d4f8>] __call_rcu_nocb_enqueue+0xa8/0xc0 > [<ffffffff81111450>] __call_rcu+0x140/0x820 > [<ffffffff81111b8d>] call_rcu+0x1d/0x20 > [<ffffffff81093697>] cpu_attach_domain+0x287/0x360 > [<ffffffff81099d7e>] build_sched_domains+0xe5e/0x10a0 > [<ffffffff81efa7fc>] sched_init_smp+0x3b7/0x47a > [<ffffffff81ee1f4e>] kernel_init_freeable+0xf6/0x202 > [<ffffffff817200be>] kernel_init+0xe/0x190 > [<ffffffff8173d22c>] ret_from_fork+0x7c/0xb0 > > -> #0 (&rdp->nocb_wq){......}: > [<ffffffff810cb7ca>] __lock_acquire+0x191a/0x1be0 > [<ffffffff810cc243>] lock_acquire+0x93/0x200 > [<ffffffff8173419b>] _raw_spin_lock_irqsave+0x4b/0x90 > [<ffffffff8108ff43>] __wake_up+0x23/0x50 > [<ffffffff8110d4f8>] __call_rcu_nocb_enqueue+0xa8/0xc0 > [<ffffffff81111450>] __call_rcu+0x140/0x820 > [<ffffffff81111bb0>] kfree_call_rcu+0x20/0x30 > [<ffffffff81149abf>] put_ctx+0x4f/0x70 > [<ffffffff81154c3e>] perf_event_exit_task+0x12e/0x230 > [<ffffffff81056b8d>] do_exit+0x30d/0xcc0 > [<ffffffff8105893c>] do_group_exit+0x4c/0xc0 > [<ffffffff810589c4>] SyS_exit_group+0x14/0x20 > [<ffffffff8173d4e4>] tracesys+0xdd/0xe2 > > other info that might help us debug this: > > Chain exists of: > &rdp->nocb_wq --> &rq->lock --> &ctx->lock > > Possible unsafe locking scenario: > > CPU0 CPU1 > ---- ---- > lock(&ctx->lock); > lock(&rq->lock); > lock(&ctx->lock); > lock(&rdp->nocb_wq); > > *** DEADLOCK *** > > 1 lock held by trinity-child2/15191: > #0: (&ctx->lock){-.-...}, at: [<ffffffff81154c19>] perf_event_exit_task+0x109/0x230 > > stack backtrace: > CPU: 2 PID: 15191 Comm: trinity-child2 Not tainted 3.12.0-rc3+ #92 > ffffffff82565b70 ffff880070c2dbf8 ffffffff8172a363 ffffffff824edf40 > ffff880070c2dc38 ffffffff81726741 ffff880070c2dc90 ffff88022383b1c0 > ffff88022383aac0 0000000000000000 ffff88022383b188 ffff88022383b1c0 > Call Trace: > [<ffffffff8172a363>] dump_stack+0x4e/0x82 > [<ffffffff81726741>] print_circular_bug+0x200/0x20f > [<ffffffff810cb7ca>] __lock_acquire+0x191a/0x1be0 > [<ffffffff810c6439>] ? get_lock_stats+0x19/0x60 > [<ffffffff8100b2f4>] ? native_sched_clock+0x24/0x80 > [<ffffffff810cc243>] lock_acquire+0x93/0x200 > [<ffffffff8108ff43>] ? __wake_up+0x23/0x50 > [<ffffffff8173419b>] _raw_spin_lock_irqsave+0x4b/0x90 > [<ffffffff8108ff43>] ? __wake_up+0x23/0x50 > [<ffffffff8108ff43>] __wake_up+0x23/0x50 > [<ffffffff8110d4f8>] __call_rcu_nocb_enqueue+0xa8/0xc0 > [<ffffffff81111450>] __call_rcu+0x140/0x820 > [<ffffffff8109bc8f>] ? local_clock+0x3f/0x50 > [<ffffffff81111bb0>] kfree_call_rcu+0x20/0x30 > [<ffffffff81149abf>] put_ctx+0x4f/0x70 > [<ffffffff81154c3e>] perf_event_exit_task+0x12e/0x230 > [<ffffffff81056b8d>] do_exit+0x30d/0xcc0 > [<ffffffff810c9af5>] ? trace_hardirqs_on_caller+0x115/0x1e0 > [<ffffffff810c9bcd>] ? trace_hardirqs_on+0xd/0x10 > [<ffffffff8105893c>] do_group_exit+0x4c/0xc0 > [<ffffffff810589c4>] SyS_exit_group+0x14/0x20 > [<ffffffff8173d4e4>] tracesys+0xdd/0xe2 The underlying problem is that perf is invoking call_rcu() with the scheduler locks held, but in NOCB mode, call_rcu() will with high probability invoke the scheduler -- which just might want to use its locks. The reason that call_rcu() needs to invoke the scheduler is to wake up the corresponding rcuo callback-offload kthread, which does the job of starting up a grace period and invoking the callbacks afterwards. One solution (championed on a related problem by Lai Jiangshan) is to simply defer the wakeup to some point where scheduler locks are no longer held. Since we don't want to unnecessarily incur the cost of such deferral, the task before us is threefold: 1. Determine when it is likely that a relevant scheduler lock is held. 2. Defer the wakeup in such cases. 3. Ensure that all deferred wakeups eventually happen, preferably sooner rather than later. We use irqs_disabled_flags() as a proxy for relevant scheduler locks being held. This works because the relevant locks are always acquired with interrupts disabled. We may defer more often than needed, but that is at least safe. The wakeup deferral is tracked via a new field in the per-CPU and per-RCU-flavor rcu_data structure, namely ->nocb_defer_wakeup. This flag is checked by the RCU core processing. The __rcu_pending() function now checks this flag, which causes rcu_check_callbacks() to initiate RCU core processing at each scheduling-clock interrupt where this flag is set. Of course this is not sufficient because scheduling-clock interrupts are often turned off (the things we used to be able to count on!). So the flags are also checked on entry to any state that RCU considers to be idle, which includes both NO_HZ_IDLE idle state and NO_HZ_FULL user-mode-execution state. This approach should allow call_rcu() to be invoked regardless of what locks you might be holding, the key word being "should". Reported-by: Dave Jones <davej@redhat.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org>
2013-10-05 05:33:34 +08:00
/* Do any needed deferred wakeups of rcuo kthreads. */
do_nocb_deferred_wakeup(rdp);
rcu: Have the RCU tracepoints use the tracepoint_string infrastructure Currently, RCU tracepoints save only a pointer to strings in the ring buffer. When displayed via the /sys/kernel/debug/tracing/trace file they are referenced like the printf "%s" that looks at the address in the ring buffer and prints out the string it points too. This requires that the strings are constant and persistent in the kernel. The problem with this is for tools like trace-cmd and perf that read the binary data from the buffers but have no access to the kernel memory to find out what string is represented by the address in the buffer. By using the tracepoint_string infrastructure, the RCU tracepoint strings can be exported such that userspace tools can map the addresses to the strings. # cat /sys/kernel/debug/tracing/printk_formats 0xffffffff81a4a0e8 : "rcu_preempt" 0xffffffff81a4a0f4 : "rcu_bh" 0xffffffff81a4a100 : "rcu_sched" 0xffffffff818437a0 : "cpuqs" 0xffffffff818437a6 : "rcu_sched" 0xffffffff818437a0 : "cpuqs" 0xffffffff818437b0 : "rcu_bh" 0xffffffff818437b7 : "Start context switch" 0xffffffff818437cc : "End context switch" 0xffffffff818437a0 : "cpuqs" [...] Now userspaces tools can display: rcu_utilization: Start context switch rcu_dyntick: Start 1 0 rcu_utilization: End context switch rcu_batch_start: rcu_preempt CBs=0/5 bl=10 rcu_dyntick: End 0 140000000000000 rcu_invoke_callback: rcu_preempt rhp=0xffff880071c0d600 func=proc_i_callback rcu_invoke_callback: rcu_preempt rhp=0xffff880077b5b230 func=__d_free rcu_dyntick: Start 140000000000000 0 rcu_invoke_callback: rcu_preempt rhp=0xffff880077563980 func=file_free_rcu rcu_batch_end: rcu_preempt CBs-invoked=3 idle=>c<>c<>c<>c< rcu_utilization: End RCU core rcu_grace_period: rcu_preempt 9741 start rcu_dyntick: Start 1 0 rcu_dyntick: End 0 140000000000000 rcu_dyntick: Start 140000000000000 0 Instead of: rcu_utilization: ffffffff81843110 rcu_future_grace_period: ffffffff81842f1d 9939 9939 9940 0 0 3 ffffffff81842f32 rcu_batch_start: ffffffff81842f1d CBs=0/4 bl=10 rcu_future_grace_period: ffffffff81842f1d 9939 9939 9940 0 0 3 ffffffff81842f3c rcu_grace_period: ffffffff81842f1d 9939 ffffffff81842f80 rcu_invoke_callback: ffffffff81842f1d rhp=0xffff88007888aac0 func=file_free_rcu rcu_grace_period: ffffffff81842f1d 9939 ffffffff81842f95 rcu_invoke_callback: ffffffff81842f1d rhp=0xffff88006aeb4600 func=proc_i_callback rcu_future_grace_period: ffffffff81842f1d 9939 9939 9940 0 0 3 ffffffff81842f32 rcu_future_grace_period: ffffffff81842f1d 9939 9939 9940 0 0 3 ffffffff81842f3c rcu_invoke_callback: ffffffff81842f1d rhp=0xffff880071cb9fc0 func=__d_free rcu_grace_period: ffffffff81842f1d 9939 ffffffff81842f80 rcu_invoke_callback: ffffffff81842f1d rhp=0xffff88007888ae80 func=file_free_rcu rcu_batch_end: ffffffff81842f1d CBs-invoked=4 idle=>c<>c<>c<>c< rcu_utilization: ffffffff8184311f Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
2013-07-13 05:18:47 +08:00
trace_rcu_utilization(TPS("End RCU core"));
// If strict GPs, schedule an RCU reader in a clean environment.
if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
queue_work_on(rdp->cpu, rcu_gp_wq, &rdp->strict_work);
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
}
static void rcu_core_si(struct softirq_action *h)
{
rcu_core();
}
static void rcu_wake_cond(struct task_struct *t, int status)
{
/*
* If the thread is yielding, only wake it when this
* is invoked from idle
*/
if (t && (status != RCU_KTHREAD_YIELDING || is_idle_task(current)))
wake_up_process(t);
}
static void invoke_rcu_core_kthread(void)
{
struct task_struct *t;
unsigned long flags;
local_irq_save(flags);
__this_cpu_write(rcu_data.rcu_cpu_has_work, 1);
t = __this_cpu_read(rcu_data.rcu_cpu_kthread_task);
if (t != NULL && t != current)
rcu_wake_cond(t, __this_cpu_read(rcu_data.rcu_cpu_kthread_status));
local_irq_restore(flags);
}
/*
* Wake up this CPU's rcuc kthread to do RCU core processing.
*/
static void invoke_rcu_core(void)
rcu: Use softirq to address performance regression Commit a26ac2455ffcf3(rcu: move TREE_RCU from softirq to kthread) introduced performance regression. In an AIM7 test, this commit degraded performance by about 40%. The commit runs rcu callbacks in a kthread instead of softirq. We observed high rate of context switch which is caused by this. Out test system has 64 CPUs and HZ is 1000, so we saw more than 64k context switch per second which is caused by RCU's per-CPU kthread. A trace showed that most of the time the RCU per-CPU kthread doesn't actually handle any callbacks, but instead just does a very small amount of work handling grace periods. This means that RCU's per-CPU kthreads are making the scheduler do quite a bit of work in order to allow a very small amount of RCU-related processing to be done. Alex Shi's analysis determined that this slowdown is due to lock contention within the scheduler. Unfortunately, as Peter Zijlstra points out, the scheduler's real-time semantics require global action, which means that this contention is inherent in real-time scheduling. (Yes, perhaps someone will come up with a workaround -- otherwise, -rt is not going to do well on large SMP systems -- but this patch will work around this issue in the meantime. And "the meantime" might well be forever.) This patch therefore re-introduces softirq processing to RCU, but only for core RCU work. RCU callbacks are still executed in kthread context, so that only a small amount of RCU work runs in softirq context in the common case. This should minimize ksoftirqd execution, allowing us to skip boosting of ksoftirqd for CONFIG_RCU_BOOST=y kernels. Signed-off-by: Shaohua Li <shaohua.li@intel.com> Tested-by: "Alex,Shi" <alex.shi@intel.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2011-06-14 13:26:25 +08:00
{
if (!cpu_online(smp_processor_id()))
return;
if (use_softirq)
raise_softirq(RCU_SOFTIRQ);
else
invoke_rcu_core_kthread();
}
static void rcu_cpu_kthread_park(unsigned int cpu)
{
per_cpu(rcu_data.rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
}
static int rcu_cpu_kthread_should_run(unsigned int cpu)
{
return __this_cpu_read(rcu_data.rcu_cpu_has_work);
}
/*
* Per-CPU kernel thread that invokes RCU callbacks. This replaces
* the RCU softirq used in configurations of RCU that do not support RCU
* priority boosting.
*/
static void rcu_cpu_kthread(unsigned int cpu)
{
unsigned int *statusp = this_cpu_ptr(&rcu_data.rcu_cpu_kthread_status);
char work, *workp = this_cpu_ptr(&rcu_data.rcu_cpu_has_work);
unsigned long *j = this_cpu_ptr(&rcu_data.rcuc_activity);
int spincnt;
trace_rcu_utilization(TPS("Start CPU kthread@rcu_run"));
for (spincnt = 0; spincnt < 10; spincnt++) {
WRITE_ONCE(*j, jiffies);
local_bh_disable();
*statusp = RCU_KTHREAD_RUNNING;
local_irq_disable();
work = *workp;
WRITE_ONCE(*workp, 0);
local_irq_enable();
if (work)
rcu_core();
local_bh_enable();
if (!READ_ONCE(*workp)) {
trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
*statusp = RCU_KTHREAD_WAITING;
return;
}
}
*statusp = RCU_KTHREAD_YIELDING;
trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
schedule_timeout_idle(2);
trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
*statusp = RCU_KTHREAD_WAITING;
WRITE_ONCE(*j, jiffies);
}
static struct smp_hotplug_thread rcu_cpu_thread_spec = {
.store = &rcu_data.rcu_cpu_kthread_task,
.thread_should_run = rcu_cpu_kthread_should_run,
.thread_fn = rcu_cpu_kthread,
.thread_comm = "rcuc/%u",
.setup = rcu_cpu_kthread_setup,
.park = rcu_cpu_kthread_park,
};
/*
* Spawn per-CPU RCU core processing kthreads.
*/
static int __init rcu_spawn_core_kthreads(void)
{
int cpu;
for_each_possible_cpu(cpu)
per_cpu(rcu_data.rcu_cpu_has_work, cpu) = 0;
if (use_softirq)
return 0;
WARN_ONCE(smpboot_register_percpu_thread(&rcu_cpu_thread_spec),
"%s: Could not start rcuc kthread, OOM is now expected behavior\n", __func__);
return 0;
rcu: Use softirq to address performance regression Commit a26ac2455ffcf3(rcu: move TREE_RCU from softirq to kthread) introduced performance regression. In an AIM7 test, this commit degraded performance by about 40%. The commit runs rcu callbacks in a kthread instead of softirq. We observed high rate of context switch which is caused by this. Out test system has 64 CPUs and HZ is 1000, so we saw more than 64k context switch per second which is caused by RCU's per-CPU kthread. A trace showed that most of the time the RCU per-CPU kthread doesn't actually handle any callbacks, but instead just does a very small amount of work handling grace periods. This means that RCU's per-CPU kthreads are making the scheduler do quite a bit of work in order to allow a very small amount of RCU-related processing to be done. Alex Shi's analysis determined that this slowdown is due to lock contention within the scheduler. Unfortunately, as Peter Zijlstra points out, the scheduler's real-time semantics require global action, which means that this contention is inherent in real-time scheduling. (Yes, perhaps someone will come up with a workaround -- otherwise, -rt is not going to do well on large SMP systems -- but this patch will work around this issue in the meantime. And "the meantime" might well be forever.) This patch therefore re-introduces softirq processing to RCU, but only for core RCU work. RCU callbacks are still executed in kthread context, so that only a small amount of RCU work runs in softirq context in the common case. This should minimize ksoftirqd execution, allowing us to skip boosting of ksoftirqd for CONFIG_RCU_BOOST=y kernels. Signed-off-by: Shaohua Li <shaohua.li@intel.com> Tested-by: "Alex,Shi" <alex.shi@intel.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2011-06-14 13:26:25 +08:00
}
/*
* Handle any core-RCU processing required by a call_rcu() invocation.
*/
static void __call_rcu_core(struct rcu_data *rdp, struct rcu_head *head,
unsigned long flags)
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
{
/*
* If called from an extended quiescent state, invoke the RCU
* core in order to force a re-evaluation of RCU's idleness.
*/
if (!rcu_is_watching())
invoke_rcu_core();
/* If interrupts were disabled or CPU offline, don't invoke RCU core. */
if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
return;
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/*
* Force the grace period if too many callbacks or too long waiting.
* Enforce hysteresis, and don't invoke rcu_force_quiescent_state()
* if some other CPU has recently done so. Also, don't bother
* invoking rcu_force_quiescent_state() if the newly enqueued callback
* is the only one waiting for a grace period to complete.
*/
if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
rdp->qlen_last_fqs_check + qhimark)) {
/* Are we ignoring a completed grace period? */
note_gp_changes(rdp);
/* Start a new grace period if one not already started. */
if (!rcu_gp_in_progress()) {
rcu_accelerate_cbs_unlocked(rdp->mynode, rdp);
} else {
/* Give the grace period a kick. */
rdp->blimit = DEFAULT_MAX_RCU_BLIMIT;
if (READ_ONCE(rcu_state.n_force_qs) == rdp->n_force_qs_snap &&
rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
rcu_force_quiescent_state();
rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs);
rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
}
}
}
/*
* RCU callback function to leak a callback.
*/
static void rcu_leak_callback(struct rcu_head *rhp)
{
}
/*
* Check and if necessary update the leaf rcu_node structure's
* ->cbovldmask bit corresponding to the current CPU based on that CPU's
* number of queued RCU callbacks. The caller must hold the leaf rcu_node
* structure's ->lock.
*/
static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp)
{
raw_lockdep_assert_held_rcu_node(rnp);
if (qovld_calc <= 0)
return; // Early boot and wildcard value set.
if (rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc)
WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask | rdp->grpmask);
else
WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask & ~rdp->grpmask);
}
/*
* Check and if necessary update the leaf rcu_node structure's
* ->cbovldmask bit corresponding to the current CPU based on that CPU's
* number of queued RCU callbacks. No locks need be held, but the
* caller must have disabled interrupts.
*
* Note that this function ignores the possibility that there are a lot
* of callbacks all of which have already seen the end of their respective
* grace periods. This omission is due to the need for no-CBs CPUs to
* be holding ->nocb_lock to do this check, which is too heavy for a
* common-case operation.
*/
static void check_cb_ovld(struct rcu_data *rdp)
{
struct rcu_node *const rnp = rdp->mynode;
if (qovld_calc <= 0 ||
((rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc) ==
!!(READ_ONCE(rnp->cbovldmask) & rdp->grpmask)))
return; // Early boot wildcard value or already set correctly.
raw_spin_lock_rcu_node(rnp);
check_cb_ovld_locked(rdp, rnp);
raw_spin_unlock_rcu_node(rnp);
}
static void
__call_rcu_common(struct rcu_head *head, rcu_callback_t func, bool lazy_in)
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
{
static atomic_t doublefrees;
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
unsigned long flags;
bool lazy;
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
struct rcu_data *rdp;
bool was_alldone;
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/* Misaligned rcu_head! */
WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
if (debug_rcu_head_queue(head)) {
/*
* Probable double call_rcu(), so leak the callback.
* Use rcu:rcu_callback trace event to find the previous
* time callback was passed to call_rcu().
*/
if (atomic_inc_return(&doublefrees) < 4) {
pr_err("%s(): Double-freed CB %p->%pS()!!! ", __func__, head, head->func);
mem_dump_obj(head);
}
WRITE_ONCE(head->func, rcu_leak_callback);
return;
}
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
head->func = func;
head->next = NULL;
rcu: Avoid alloc_pages() when recording stack The default kasan_record_aux_stack() calls stack_depot_save() with GFP_NOWAIT, which in turn can then call alloc_pages(GFP_NOWAIT, ...). In general, however, it is not even possible to use either GFP_ATOMIC nor GFP_NOWAIT in certain non-preemptive contexts/RT kernel including raw_spin_locks (see gfp.h and ab00db216c9c7). Fix it by instructing stackdepot to not expand stack storage via alloc_pages() in case it runs out by using kasan_record_aux_stack_noalloc(). Jianwei Hu reported: BUG: sleeping function called from invalid context at kernel/locking/rtmutex.c:969 in_atomic(): 0, irqs_disabled(): 1, non_block: 0, pid: 15319, name: python3 INFO: lockdep is turned off. irq event stamp: 0 hardirqs last enabled at (0): [<0000000000000000>] 0x0 hardirqs last disabled at (0): [<ffffffff856c8b13>] copy_process+0xaf3/0x2590 softirqs last enabled at (0): [<ffffffff856c8b13>] copy_process+0xaf3/0x2590 softirqs last disabled at (0): [<0000000000000000>] 0x0 CPU: 6 PID: 15319 Comm: python3 Tainted: G W O 5.15-rc7-preempt-rt #1 Hardware name: Supermicro SYS-E300-9A-8C/A2SDi-8C-HLN4F, BIOS 1.1b 12/17/2018 Call Trace: show_stack+0x52/0x58 dump_stack+0xa1/0xd6 ___might_sleep.cold+0x11c/0x12d rt_spin_lock+0x3f/0xc0 rmqueue+0x100/0x1460 rmqueue+0x100/0x1460 mark_usage+0x1a0/0x1a0 ftrace_graph_ret_addr+0x2a/0xb0 rmqueue_pcplist.constprop.0+0x6a0/0x6a0 __kasan_check_read+0x11/0x20 __zone_watermark_ok+0x114/0x270 get_page_from_freelist+0x148/0x630 is_module_text_address+0x32/0xa0 __alloc_pages_nodemask+0x2f6/0x790 __alloc_pages_slowpath.constprop.0+0x12d0/0x12d0 create_prof_cpu_mask+0x30/0x30 alloc_pages_current+0xb1/0x150 stack_depot_save+0x39f/0x490 kasan_save_stack+0x42/0x50 kasan_save_stack+0x23/0x50 kasan_record_aux_stack+0xa9/0xc0 __call_rcu+0xff/0x9c0 call_rcu+0xe/0x10 put_object+0x53/0x70 __delete_object+0x7b/0x90 kmemleak_free+0x46/0x70 slab_free_freelist_hook+0xb4/0x160 kfree+0xe5/0x420 kfree_const+0x17/0x30 kobject_cleanup+0xaa/0x230 kobject_put+0x76/0x90 netdev_queue_update_kobjects+0x17d/0x1f0 ... ... ksys_write+0xd9/0x180 __x64_sys_write+0x42/0x50 do_syscall_64+0x38/0x50 entry_SYSCALL_64_after_hwframe+0x44/0xa9 Links: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/include/linux/kasan.h?id=7cb3007ce2da27ec02a1a3211941e7fe6875b642 Fixes: 84109ab58590 ("rcu: Record kvfree_call_rcu() call stack for KASAN") Fixes: 26e760c9a7c8 ("rcu: kasan: record and print call_rcu() call stack") Reported-by: Jianwei Hu <jianwei.hu@windriver.com> Reviewed-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Acked-by: Marco Elver <elver@google.com> Tested-by: Juri Lelli <juri.lelli@redhat.com> Signed-off-by: Jun Miao <jun.miao@intel.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2021-11-16 07:23:02 +08:00
kasan_record_aux_stack_noalloc(head);
local_irq_save(flags);
rdp = this_cpu_ptr(&rcu_data);
lazy = lazy_in && !rcu_async_should_hurry();
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/* Add the callback to our list. */
if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist))) {
// This can trigger due to call_rcu() from offline CPU:
WARN_ON_ONCE(rcu_scheduler_active != RCU_SCHEDULER_INACTIVE);
WARN_ON_ONCE(!rcu_is_watching());
// Very early boot, before rcu_init(). Initialize if needed
// and then drop through to queue the callback.
if (rcu_segcblist_empty(&rdp->cblist))
rcu_segcblist_init(&rdp->cblist);
}
check_cb_ovld(rdp);
if (rcu_nocb_try_bypass(rdp, head, &was_alldone, flags, lazy))
rcu/nocb: Add bypass callback queueing Use of the rcu_data structure's segmented ->cblist for no-CBs CPUs takes advantage of unrelated grace periods, thus reducing the memory footprint in the face of floods of call_rcu() invocations. However, the ->cblist field is a more-complex rcu_segcblist structure which must be protected via locking. Even though there are only three entities which can acquire this lock (the CPU invoking call_rcu(), the no-CBs grace-period kthread, and the no-CBs callbacks kthread), the contention on this lock is excessive under heavy stress. This commit therefore greatly reduces contention by provisioning an rcu_cblist structure field named ->nocb_bypass within the rcu_data structure. Each no-CBs CPU is permitted only a limited number of enqueues onto the ->cblist per jiffy, controlled by a new nocb_nobypass_lim_per_jiffy kernel boot parameter that defaults to about 16 enqueues per millisecond (16 * 1000 / HZ). When that limit is exceeded, the CPU instead enqueues onto the new ->nocb_bypass. The ->nocb_bypass is flushed into the ->cblist every jiffy or when the number of callbacks on ->nocb_bypass exceeds qhimark, whichever happens first. During call_rcu() floods, this flushing is carried out by the CPU during the course of its call_rcu() invocations. However, a CPU could simply stop invoking call_rcu() at any time. The no-CBs grace-period kthread therefore carries out less-aggressive flushing (every few jiffies or when the number of callbacks on ->nocb_bypass exceeds (2 * qhimark), whichever comes first). This means that the no-CBs grace-period kthread cannot be permitted to do unbounded waits while there are callbacks on ->nocb_bypass. A ->nocb_bypass_timer is used to provide the needed wakeups. [ paulmck: Apply Coverity feedback reported by Colin Ian King. ] Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
2019-07-03 07:03:33 +08:00
return; // Enqueued onto ->nocb_bypass, so just leave.
// If no-CBs CPU gets here, rcu_nocb_try_bypass() acquired ->nocb_lock.
rcu_segcblist_enqueue(&rdp->cblist, head);
if (__is_kvfree_rcu_offset((unsigned long)func))
trace_rcu_kvfree_callback(rcu_state.name, head,
(unsigned long)func,
rcu_segcblist_n_cbs(&rdp->cblist));
else
trace_rcu_callback(rcu_state.name, head,
rcu_segcblist_n_cbs(&rdp->cblist));
trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCBQueued"));
/* Go handle any RCU core processing required. */
if (unlikely(rcu_rdp_is_offloaded(rdp))) {
__call_rcu_nocb_wake(rdp, was_alldone, flags); /* unlocks */
} else {
__call_rcu_core(rdp, head, flags);
local_irq_restore(flags);
}
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
}
#ifdef CONFIG_RCU_LAZY
/**
* call_rcu_hurry() - Queue RCU callback for invocation after grace period, and
* flush all lazy callbacks (including the new one) to the main ->cblist while
* doing so.
*
* @head: structure to be used for queueing the RCU updates.
* @func: actual callback function to be invoked after the grace period
*
* The callback function will be invoked some time after a full grace
* period elapses, in other words after all pre-existing RCU read-side
* critical sections have completed.
*
* Use this API instead of call_rcu() if you don't want the callback to be
* invoked after very long periods of time, which can happen on systems without
* memory pressure and on systems which are lightly loaded or mostly idle.
* This function will cause callbacks to be invoked sooner than later at the
* expense of extra power. Other than that, this function is identical to, and
* reuses call_rcu()'s logic. Refer to call_rcu() for more details about memory
* ordering and other functionality.
*/
void call_rcu_hurry(struct rcu_head *head, rcu_callback_t func)
{
return __call_rcu_common(head, func, false);
}
EXPORT_SYMBOL_GPL(call_rcu_hurry);
#endif
/**
* call_rcu() - Queue an RCU callback for invocation after a grace period.
* By default the callbacks are 'lazy' and are kept hidden from the main
* ->cblist to prevent starting of grace periods too soon.
* If you desire grace periods to start very soon, use call_rcu_hurry().
*
* @head: structure to be used for queueing the RCU updates.
* @func: actual callback function to be invoked after the grace period
*
* The callback function will be invoked some time after a full grace
* period elapses, in other words after all pre-existing RCU read-side
* critical sections have completed. However, the callback function
* might well execute concurrently with RCU read-side critical sections
* that started after call_rcu() was invoked.
*
* RCU read-side critical sections are delimited by rcu_read_lock()
* and rcu_read_unlock(), and may be nested. In addition, but only in
* v5.0 and later, regions of code across which interrupts, preemption,
* or softirqs have been disabled also serve as RCU read-side critical
* sections. This includes hardware interrupt handlers, softirq handlers,
* and NMI handlers.
*
* Note that all CPUs must agree that the grace period extended beyond
* all pre-existing RCU read-side critical section. On systems with more
* than one CPU, this means that when "func()" is invoked, each CPU is
* guaranteed to have executed a full memory barrier since the end of its
* last RCU read-side critical section whose beginning preceded the call
* to call_rcu(). It also means that each CPU executing an RCU read-side
* critical section that continues beyond the start of "func()" must have
* executed a memory barrier after the call_rcu() but before the beginning
* of that RCU read-side critical section. Note that these guarantees
* include CPUs that are offline, idle, or executing in user mode, as
* well as CPUs that are executing in the kernel.
*
* Furthermore, if CPU A invoked call_rcu() and CPU B invoked the
* resulting RCU callback function "func()", then both CPU A and CPU B are
* guaranteed to execute a full memory barrier during the time interval
* between the call to call_rcu() and the invocation of "func()" -- even
* if CPU A and CPU B are the same CPU (but again only if the system has
* more than one CPU).
*
* Implementation of these memory-ordering guarantees is described here:
* Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst.
*/
void call_rcu(struct rcu_head *head, rcu_callback_t func)
{
return __call_rcu_common(head, func, IS_ENABLED(CONFIG_RCU_LAZY));
}
EXPORT_SYMBOL_GPL(call_rcu);
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
/* Maximum number of jiffies to wait before draining a batch. */
rcu/kvfree: Update KFREE_DRAIN_JIFFIES interval Currently the monitor work is scheduled with a fixed interval of HZ/20, which is roughly 50 milliseconds. The drawback of this approach is low utilization of the 512 page slots in scenarios with infrequence kvfree_rcu() calls. For example on an Android system: <snip> kworker/3:3-507 [003] .... 470.286305: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000d0f0dde5 nr_records=6 kworker/6:1-76 [006] .... 470.416613: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000ea0d6556 nr_records=1 kworker/6:1-76 [006] .... 470.416625: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000003e025849 nr_records=9 kworker/3:3-507 [003] .... 471.390000: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000815a8713 nr_records=48 kworker/1:1-73 [001] .... 471.725785: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000fda9bf20 nr_records=3 kworker/1:1-73 [001] .... 471.725833: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000a425b67b nr_records=76 kworker/0:4-1411 [000] .... 472.085673: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000007996be9d nr_records=1 kworker/0:4-1411 [000] .... 472.085728: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000d0f0dde5 nr_records=5 kworker/6:1-76 [006] .... 472.260340: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000065630ee4 nr_records=102 <snip> In many cases, out of 512 slots, fewer than 10 were actually used. In order to improve batching and make utilization more efficient this commit sets a drain interval to a fixed 5-seconds interval. Floods are detected when a page fills quickly, and in that case, the reclaim work is re-scheduled for the next scheduling-clock tick (jiffy). After this change: <snip> kworker/7:1-371 [007] .... 5630.725708: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000005ab0ffb3 nr_records=121 kworker/7:1-371 [007] .... 5630.989702: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000060c84761 nr_records=47 kworker/7:1-371 [007] .... 5630.989714: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000000babf308 nr_records=510 kworker/7:1-371 [007] .... 5631.553790: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000bb7bd0ef nr_records=169 kworker/7:1-371 [007] .... 5631.553808: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000044c78753 nr_records=510 kworker/5:6-9428 [005] .... 5631.746102: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000d98519aa nr_records=123 kworker/4:7-9434 [004] .... 5632.001758: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000526c9d44 nr_records=322 kworker/4:7-9434 [004] .... 5632.002073: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000002c6a8afa nr_records=185 kworker/7:1-371 [007] .... 5632.277515: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000007f4a962f nr_records=510 <snip> Here, all but one of the cases, more than one hundreds slots were used, representing an order-of-magnitude improvement. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2022-07-01 00:33:35 +08:00
#define KFREE_DRAIN_JIFFIES (5 * HZ)
#define KFREE_N_BATCHES 2
#define FREE_N_CHANNELS 2
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
/**
* struct kvfree_rcu_bulk_data - single block to store kvfree_rcu() pointers
* @list: List node. All blocks are linked between each other
rcu/kvfree: Use a polled API to speedup a reclaim process Currently all objects placed into a batch wait for a full grace period to elapse after that batch is ready to send to RCU. However, this can unnecessarily delay freeing of the first objects that were added to the batch. After all, several RCU grace periods might have elapsed since those objects were added, and if so, there is no point in further deferring their freeing. This commit therefore adds per-page grace-period snapshots which are obtained from get_state_synchronize_rcu(). When the batch is ready to be passed to call_rcu(), each page's snapshot is checked by passing it to poll_state_synchronize_rcu(). If a given page's RCU grace period has already elapsed, its objects are freed immediately by kvfree_rcu_bulk(). Otherwise, these objects are freed after a call to synchronize_rcu(). This approach requires that the pages be traversed in reverse order, that is, the oldest ones first. Test example: kvm.sh --memory 10G --torture rcuscale --allcpus --duration 1 \ --kconfig CONFIG_NR_CPUS=64 \ --kconfig CONFIG_RCU_NOCB_CPU=y \ --kconfig CONFIG_RCU_NOCB_CPU_DEFAULT_ALL=y \ --kconfig CONFIG_RCU_LAZY=n \ --bootargs "rcuscale.kfree_rcu_test=1 rcuscale.kfree_nthreads=16 \ rcuscale.holdoff=20 rcuscale.kfree_loops=10000 \ torture.disable_onoff_at_boot" --trust-make Before this commit: Total time taken by all kfree'ers: 8535693700 ns, loops: 10000, batches: 1188, memory footprint: 2248MB Total time taken by all kfree'ers: 8466933582 ns, loops: 10000, batches: 1157, memory footprint: 2820MB Total time taken by all kfree'ers: 5375602446 ns, loops: 10000, batches: 1130, memory footprint: 6502MB Total time taken by all kfree'ers: 7523283832 ns, loops: 10000, batches: 1006, memory footprint: 3343MB Total time taken by all kfree'ers: 6459171956 ns, loops: 10000, batches: 1150, memory footprint: 6549MB After this commit: Total time taken by all kfree'ers: 8560060176 ns, loops: 10000, batches: 1787, memory footprint: 61MB Total time taken by all kfree'ers: 8573885501 ns, loops: 10000, batches: 1777, memory footprint: 93MB Total time taken by all kfree'ers: 8320000202 ns, loops: 10000, batches: 1727, memory footprint: 66MB Total time taken by all kfree'ers: 8552718794 ns, loops: 10000, batches: 1790, memory footprint: 75MB Total time taken by all kfree'ers: 8601368792 ns, loops: 10000, batches: 1724, memory footprint: 62MB The reduction in memory footprint is well in excess of an order of magnitude. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2022-11-29 23:58:22 +08:00
* @gp_snap: Snapshot of RCU state for objects placed to this bulk
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
* @nr_records: Number of active pointers in the array
* @records: Array of the kvfree_rcu() pointers
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
*/
struct kvfree_rcu_bulk_data {
struct list_head list;
struct rcu_gp_oldstate gp_snap;
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
unsigned long nr_records;
void *records[];
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
};
/*
* This macro defines how many entries the "records" array
* will contain. It is based on the fact that the size of
* kvfree_rcu_bulk_data structure becomes exactly one page.
*/
#define KVFREE_BULK_MAX_ENTR \
((PAGE_SIZE - sizeof(struct kvfree_rcu_bulk_data)) / sizeof(void *))
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
/**
* struct kfree_rcu_cpu_work - single batch of kfree_rcu() requests
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
* @rcu_work: Let queue_rcu_work() invoke workqueue handler after grace period
* @head_free: List of kfree_rcu() objects waiting for a grace period
* @head_free_gp_snap: Grace-period snapshot to check for attempted premature frees.
* @bulk_head_free: Bulk-List of kvfree_rcu() objects waiting for a grace period
* @krcp: Pointer to @kfree_rcu_cpu structure
*/
struct kfree_rcu_cpu_work {
struct rcu_work rcu_work;
struct rcu_head *head_free;
struct rcu_gp_oldstate head_free_gp_snap;
struct list_head bulk_head_free[FREE_N_CHANNELS];
struct kfree_rcu_cpu *krcp;
};
/**
* struct kfree_rcu_cpu - batch up kfree_rcu() requests for RCU grace period
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
* @head: List of kfree_rcu() objects not yet waiting for a grace period
* @head_gp_snap: Snapshot of RCU state for objects placed to "@head"
* @bulk_head: Bulk-List of kvfree_rcu() objects not yet waiting for a grace period
* @krw_arr: Array of batches of kfree_rcu() objects waiting for a grace period
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
* @lock: Synchronize access to this structure
* @monitor_work: Promote @head to @head_free after KFREE_DRAIN_JIFFIES
* @initialized: The @rcu_work fields have been initialized
* @head_count: Number of objects in rcu_head singular list
* @bulk_count: Number of objects in bulk-list
* @bkvcache:
* A simple cache list that contains objects for reuse purpose.
* In order to save some per-cpu space the list is singular.
* Even though it is lockless an access has to be protected by the
* per-cpu lock.
rcu/tree: Defer kvfree_rcu() allocation to a clean context The current memmory-allocation interface causes the following difficulties for kvfree_rcu(): a) If built with CONFIG_PROVE_RAW_LOCK_NESTING, the lockdep will complain about violation of the nesting rules, as in "BUG: Invalid wait context". This Kconfig option checks for proper raw_spinlock vs. spinlock nesting, in particular, it is not legal to acquire a spinlock_t while holding a raw_spinlock_t. This is a problem because kfree_rcu() uses raw_spinlock_t whereas the "page allocator" internally deals with spinlock_t to access to its zones. The code also can be broken from higher level of view: <snip> raw_spin_lock(&some_lock); kfree_rcu(some_pointer, some_field_offset); <snip> b) If built with CONFIG_PREEMPT_RT, spinlock_t is converted into sleeplock. This means that invoking the page allocator from atomic contexts results in "BUG: scheduling while atomic". c) Please note that call_rcu() is already invoked from raw atomic context, so it is only reasonable to expaect that kfree_rcu() and kvfree_rcu() will also be called from atomic raw context. This commit therefore defers page allocation to a clean context using the combination of an hrtimer and a workqueue. The hrtimer stage is required in order to avoid deadlocks with the scheduler. This deferred allocation is required only when kvfree_rcu()'s per-CPU page cache is empty. Link: https://lore.kernel.org/lkml/20200630164543.4mdcf6zb4zfclhln@linutronix.de/ Fixes: 3042f83f19be ("rcu: Support reclaim for head-less object") Reported-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-10-30 00:50:04 +08:00
* @page_cache_work: A work to refill the cache when it is empty
* @backoff_page_cache_fill: Delay cache refills
rcu/tree: Defer kvfree_rcu() allocation to a clean context The current memmory-allocation interface causes the following difficulties for kvfree_rcu(): a) If built with CONFIG_PROVE_RAW_LOCK_NESTING, the lockdep will complain about violation of the nesting rules, as in "BUG: Invalid wait context". This Kconfig option checks for proper raw_spinlock vs. spinlock nesting, in particular, it is not legal to acquire a spinlock_t while holding a raw_spinlock_t. This is a problem because kfree_rcu() uses raw_spinlock_t whereas the "page allocator" internally deals with spinlock_t to access to its zones. The code also can be broken from higher level of view: <snip> raw_spin_lock(&some_lock); kfree_rcu(some_pointer, some_field_offset); <snip> b) If built with CONFIG_PREEMPT_RT, spinlock_t is converted into sleeplock. This means that invoking the page allocator from atomic contexts results in "BUG: scheduling while atomic". c) Please note that call_rcu() is already invoked from raw atomic context, so it is only reasonable to expaect that kfree_rcu() and kvfree_rcu() will also be called from atomic raw context. This commit therefore defers page allocation to a clean context using the combination of an hrtimer and a workqueue. The hrtimer stage is required in order to avoid deadlocks with the scheduler. This deferred allocation is required only when kvfree_rcu()'s per-CPU page cache is empty. Link: https://lore.kernel.org/lkml/20200630164543.4mdcf6zb4zfclhln@linutronix.de/ Fixes: 3042f83f19be ("rcu: Support reclaim for head-less object") Reported-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-10-30 00:50:04 +08:00
* @work_in_progress: Indicates that page_cache_work is running
* @hrtimer: A hrtimer for scheduling a page_cache_work
* @nr_bkv_objs: number of allocated objects at @bkvcache.
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
*
* This is a per-CPU structure. The reason that it is not included in
* the rcu_data structure is to permit this code to be extracted from
* the RCU files. Such extraction could allow further optimization of
* the interactions with the slab allocators.
*/
struct kfree_rcu_cpu {
// Objects queued on a linked list
// through their rcu_head structures.
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
struct rcu_head *head;
unsigned long head_gp_snap;
atomic_t head_count;
// Objects queued on a bulk-list.
struct list_head bulk_head[FREE_N_CHANNELS];
atomic_t bulk_count[FREE_N_CHANNELS];
struct kfree_rcu_cpu_work krw_arr[KFREE_N_BATCHES];
raw_spinlock_t lock;
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
struct delayed_work monitor_work;
bool initialized;
rcu/tree: Defer kvfree_rcu() allocation to a clean context The current memmory-allocation interface causes the following difficulties for kvfree_rcu(): a) If built with CONFIG_PROVE_RAW_LOCK_NESTING, the lockdep will complain about violation of the nesting rules, as in "BUG: Invalid wait context". This Kconfig option checks for proper raw_spinlock vs. spinlock nesting, in particular, it is not legal to acquire a spinlock_t while holding a raw_spinlock_t. This is a problem because kfree_rcu() uses raw_spinlock_t whereas the "page allocator" internally deals with spinlock_t to access to its zones. The code also can be broken from higher level of view: <snip> raw_spin_lock(&some_lock); kfree_rcu(some_pointer, some_field_offset); <snip> b) If built with CONFIG_PREEMPT_RT, spinlock_t is converted into sleeplock. This means that invoking the page allocator from atomic contexts results in "BUG: scheduling while atomic". c) Please note that call_rcu() is already invoked from raw atomic context, so it is only reasonable to expaect that kfree_rcu() and kvfree_rcu() will also be called from atomic raw context. This commit therefore defers page allocation to a clean context using the combination of an hrtimer and a workqueue. The hrtimer stage is required in order to avoid deadlocks with the scheduler. This deferred allocation is required only when kvfree_rcu()'s per-CPU page cache is empty. Link: https://lore.kernel.org/lkml/20200630164543.4mdcf6zb4zfclhln@linutronix.de/ Fixes: 3042f83f19be ("rcu: Support reclaim for head-less object") Reported-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-10-30 00:50:04 +08:00
struct delayed_work page_cache_work;
atomic_t backoff_page_cache_fill;
rcu/tree: Defer kvfree_rcu() allocation to a clean context The current memmory-allocation interface causes the following difficulties for kvfree_rcu(): a) If built with CONFIG_PROVE_RAW_LOCK_NESTING, the lockdep will complain about violation of the nesting rules, as in "BUG: Invalid wait context". This Kconfig option checks for proper raw_spinlock vs. spinlock nesting, in particular, it is not legal to acquire a spinlock_t while holding a raw_spinlock_t. This is a problem because kfree_rcu() uses raw_spinlock_t whereas the "page allocator" internally deals with spinlock_t to access to its zones. The code also can be broken from higher level of view: <snip> raw_spin_lock(&some_lock); kfree_rcu(some_pointer, some_field_offset); <snip> b) If built with CONFIG_PREEMPT_RT, spinlock_t is converted into sleeplock. This means that invoking the page allocator from atomic contexts results in "BUG: scheduling while atomic". c) Please note that call_rcu() is already invoked from raw atomic context, so it is only reasonable to expaect that kfree_rcu() and kvfree_rcu() will also be called from atomic raw context. This commit therefore defers page allocation to a clean context using the combination of an hrtimer and a workqueue. The hrtimer stage is required in order to avoid deadlocks with the scheduler. This deferred allocation is required only when kvfree_rcu()'s per-CPU page cache is empty. Link: https://lore.kernel.org/lkml/20200630164543.4mdcf6zb4zfclhln@linutronix.de/ Fixes: 3042f83f19be ("rcu: Support reclaim for head-less object") Reported-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-10-30 00:50:04 +08:00
atomic_t work_in_progress;
struct hrtimer hrtimer;
struct llist_head bkvcache;
int nr_bkv_objs;
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
};
static DEFINE_PER_CPU(struct kfree_rcu_cpu, krc) = {
.lock = __RAW_SPIN_LOCK_UNLOCKED(krc.lock),
};
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
static __always_inline void
debug_rcu_bhead_unqueue(struct kvfree_rcu_bulk_data *bhead)
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
{
#ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
int i;
for (i = 0; i < bhead->nr_records; i++)
debug_rcu_head_unqueue((struct rcu_head *)(bhead->records[i]));
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
#endif
}
static inline struct kfree_rcu_cpu *
krc_this_cpu_lock(unsigned long *flags)
{
struct kfree_rcu_cpu *krcp;
local_irq_save(*flags); // For safely calling this_cpu_ptr().
krcp = this_cpu_ptr(&krc);
raw_spin_lock(&krcp->lock);
return krcp;
}
static inline void
krc_this_cpu_unlock(struct kfree_rcu_cpu *krcp, unsigned long flags)
{
raw_spin_unlock_irqrestore(&krcp->lock, flags);
}
static inline struct kvfree_rcu_bulk_data *
get_cached_bnode(struct kfree_rcu_cpu *krcp)
{
if (!krcp->nr_bkv_objs)
return NULL;
WRITE_ONCE(krcp->nr_bkv_objs, krcp->nr_bkv_objs - 1);
return (struct kvfree_rcu_bulk_data *)
llist_del_first(&krcp->bkvcache);
}
static inline bool
put_cached_bnode(struct kfree_rcu_cpu *krcp,
struct kvfree_rcu_bulk_data *bnode)
{
// Check the limit.
if (krcp->nr_bkv_objs >= rcu_min_cached_objs)
return false;
llist_add((struct llist_node *) bnode, &krcp->bkvcache);
WRITE_ONCE(krcp->nr_bkv_objs, krcp->nr_bkv_objs + 1);
return true;
}
static int
drain_page_cache(struct kfree_rcu_cpu *krcp)
{
unsigned long flags;
struct llist_node *page_list, *pos, *n;
int freed = 0;
if (!rcu_min_cached_objs)
return 0;
raw_spin_lock_irqsave(&krcp->lock, flags);
page_list = llist_del_all(&krcp->bkvcache);
WRITE_ONCE(krcp->nr_bkv_objs, 0);
raw_spin_unlock_irqrestore(&krcp->lock, flags);
llist_for_each_safe(pos, n, page_list) {
free_page((unsigned long)pos);
freed++;
}
return freed;
}
static void
kvfree_rcu_bulk(struct kfree_rcu_cpu *krcp,
struct kvfree_rcu_bulk_data *bnode, int idx)
{
unsigned long flags;
int i;
if (!WARN_ON_ONCE(!poll_state_synchronize_rcu_full(&bnode->gp_snap))) {
debug_rcu_bhead_unqueue(bnode);
rcu_lock_acquire(&rcu_callback_map);
if (idx == 0) { // kmalloc() / kfree().
trace_rcu_invoke_kfree_bulk_callback(
rcu_state.name, bnode->nr_records,
bnode->records);
kfree_bulk(bnode->nr_records, bnode->records);
} else { // vmalloc() / vfree().
for (i = 0; i < bnode->nr_records; i++) {
trace_rcu_invoke_kvfree_callback(
rcu_state.name, bnode->records[i], 0);
vfree(bnode->records[i]);
}
}
rcu_lock_release(&rcu_callback_map);
}
raw_spin_lock_irqsave(&krcp->lock, flags);
if (put_cached_bnode(krcp, bnode))
bnode = NULL;
raw_spin_unlock_irqrestore(&krcp->lock, flags);
if (bnode)
free_page((unsigned long) bnode);
cond_resched_tasks_rcu_qs();
}
static void
kvfree_rcu_list(struct rcu_head *head)
{
struct rcu_head *next;
for (; head; head = next) {
void *ptr = (void *) head->func;
unsigned long offset = (void *) head - ptr;
next = head->next;
debug_rcu_head_unqueue((struct rcu_head *)ptr);
rcu_lock_acquire(&rcu_callback_map);
trace_rcu_invoke_kvfree_callback(rcu_state.name, head, offset);
if (!WARN_ON_ONCE(!__is_kvfree_rcu_offset(offset)))
kvfree(ptr);
rcu_lock_release(&rcu_callback_map);
cond_resched_tasks_rcu_qs();
}
}
/*
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
* This function is invoked in workqueue context after a grace period.
* It frees all the objects queued on ->bulk_head_free or ->head_free.
*/
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
static void kfree_rcu_work(struct work_struct *work)
{
unsigned long flags;
struct kvfree_rcu_bulk_data *bnode, *n;
struct list_head bulk_head[FREE_N_CHANNELS];
struct rcu_head *head;
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
struct kfree_rcu_cpu *krcp;
struct kfree_rcu_cpu_work *krwp;
struct rcu_gp_oldstate head_gp_snap;
int i;
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
krwp = container_of(to_rcu_work(work),
rcu/kvfree: Use a polled API to speedup a reclaim process Currently all objects placed into a batch wait for a full grace period to elapse after that batch is ready to send to RCU. However, this can unnecessarily delay freeing of the first objects that were added to the batch. After all, several RCU grace periods might have elapsed since those objects were added, and if so, there is no point in further deferring their freeing. This commit therefore adds per-page grace-period snapshots which are obtained from get_state_synchronize_rcu(). When the batch is ready to be passed to call_rcu(), each page's snapshot is checked by passing it to poll_state_synchronize_rcu(). If a given page's RCU grace period has already elapsed, its objects are freed immediately by kvfree_rcu_bulk(). Otherwise, these objects are freed after a call to synchronize_rcu(). This approach requires that the pages be traversed in reverse order, that is, the oldest ones first. Test example: kvm.sh --memory 10G --torture rcuscale --allcpus --duration 1 \ --kconfig CONFIG_NR_CPUS=64 \ --kconfig CONFIG_RCU_NOCB_CPU=y \ --kconfig CONFIG_RCU_NOCB_CPU_DEFAULT_ALL=y \ --kconfig CONFIG_RCU_LAZY=n \ --bootargs "rcuscale.kfree_rcu_test=1 rcuscale.kfree_nthreads=16 \ rcuscale.holdoff=20 rcuscale.kfree_loops=10000 \ torture.disable_onoff_at_boot" --trust-make Before this commit: Total time taken by all kfree'ers: 8535693700 ns, loops: 10000, batches: 1188, memory footprint: 2248MB Total time taken by all kfree'ers: 8466933582 ns, loops: 10000, batches: 1157, memory footprint: 2820MB Total time taken by all kfree'ers: 5375602446 ns, loops: 10000, batches: 1130, memory footprint: 6502MB Total time taken by all kfree'ers: 7523283832 ns, loops: 10000, batches: 1006, memory footprint: 3343MB Total time taken by all kfree'ers: 6459171956 ns, loops: 10000, batches: 1150, memory footprint: 6549MB After this commit: Total time taken by all kfree'ers: 8560060176 ns, loops: 10000, batches: 1787, memory footprint: 61MB Total time taken by all kfree'ers: 8573885501 ns, loops: 10000, batches: 1777, memory footprint: 93MB Total time taken by all kfree'ers: 8320000202 ns, loops: 10000, batches: 1727, memory footprint: 66MB Total time taken by all kfree'ers: 8552718794 ns, loops: 10000, batches: 1790, memory footprint: 75MB Total time taken by all kfree'ers: 8601368792 ns, loops: 10000, batches: 1724, memory footprint: 62MB The reduction in memory footprint is well in excess of an order of magnitude. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2022-11-29 23:58:22 +08:00
struct kfree_rcu_cpu_work, rcu_work);
krcp = krwp->krcp;
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
raw_spin_lock_irqsave(&krcp->lock, flags);
// Channels 1 and 2.
for (i = 0; i < FREE_N_CHANNELS; i++)
list_replace_init(&krwp->bulk_head_free[i], &bulk_head[i]);
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
// Channel 3.
head = krwp->head_free;
krwp->head_free = NULL;
head_gp_snap = krwp->head_free_gp_snap;
raw_spin_unlock_irqrestore(&krcp->lock, flags);
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
// Handle the first two channels.
for (i = 0; i < FREE_N_CHANNELS; i++) {
rcu/kvfree: Use a polled API to speedup a reclaim process Currently all objects placed into a batch wait for a full grace period to elapse after that batch is ready to send to RCU. However, this can unnecessarily delay freeing of the first objects that were added to the batch. After all, several RCU grace periods might have elapsed since those objects were added, and if so, there is no point in further deferring their freeing. This commit therefore adds per-page grace-period snapshots which are obtained from get_state_synchronize_rcu(). When the batch is ready to be passed to call_rcu(), each page's snapshot is checked by passing it to poll_state_synchronize_rcu(). If a given page's RCU grace period has already elapsed, its objects are freed immediately by kvfree_rcu_bulk(). Otherwise, these objects are freed after a call to synchronize_rcu(). This approach requires that the pages be traversed in reverse order, that is, the oldest ones first. Test example: kvm.sh --memory 10G --torture rcuscale --allcpus --duration 1 \ --kconfig CONFIG_NR_CPUS=64 \ --kconfig CONFIG_RCU_NOCB_CPU=y \ --kconfig CONFIG_RCU_NOCB_CPU_DEFAULT_ALL=y \ --kconfig CONFIG_RCU_LAZY=n \ --bootargs "rcuscale.kfree_rcu_test=1 rcuscale.kfree_nthreads=16 \ rcuscale.holdoff=20 rcuscale.kfree_loops=10000 \ torture.disable_onoff_at_boot" --trust-make Before this commit: Total time taken by all kfree'ers: 8535693700 ns, loops: 10000, batches: 1188, memory footprint: 2248MB Total time taken by all kfree'ers: 8466933582 ns, loops: 10000, batches: 1157, memory footprint: 2820MB Total time taken by all kfree'ers: 5375602446 ns, loops: 10000, batches: 1130, memory footprint: 6502MB Total time taken by all kfree'ers: 7523283832 ns, loops: 10000, batches: 1006, memory footprint: 3343MB Total time taken by all kfree'ers: 6459171956 ns, loops: 10000, batches: 1150, memory footprint: 6549MB After this commit: Total time taken by all kfree'ers: 8560060176 ns, loops: 10000, batches: 1787, memory footprint: 61MB Total time taken by all kfree'ers: 8573885501 ns, loops: 10000, batches: 1777, memory footprint: 93MB Total time taken by all kfree'ers: 8320000202 ns, loops: 10000, batches: 1727, memory footprint: 66MB Total time taken by all kfree'ers: 8552718794 ns, loops: 10000, batches: 1790, memory footprint: 75MB Total time taken by all kfree'ers: 8601368792 ns, loops: 10000, batches: 1724, memory footprint: 62MB The reduction in memory footprint is well in excess of an order of magnitude. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2022-11-29 23:58:22 +08:00
// Start from the tail page, so a GP is likely passed for it.
list_for_each_entry_safe(bnode, n, &bulk_head[i], list)
kvfree_rcu_bulk(krcp, bnode, i);
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
}
/*
* This is used when the "bulk" path can not be used for the
* double-argument of kvfree_rcu(). This happens when the
* page-cache is empty, which means that objects are instead
* queued on a linked list through their rcu_head structures.
* This list is named "Channel 3".
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
*/
if (head && !WARN_ON_ONCE(!poll_state_synchronize_rcu_full(&head_gp_snap)))
kvfree_rcu_list(head);
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
}
static bool
need_offload_krc(struct kfree_rcu_cpu *krcp)
{
int i;
for (i = 0; i < FREE_N_CHANNELS; i++)
if (!list_empty(&krcp->bulk_head[i]))
return true;
return !!READ_ONCE(krcp->head);
}
rcu/kvfree: Avoid freeing new kfree_rcu() memory after old grace period Memory passed to kvfree_rcu() that is to be freed is tracked by a per-CPU kfree_rcu_cpu structure, which in turn contains pointers to kvfree_rcu_bulk_data structures that contain pointers to memory that has not yet been handed to RCU, along with an kfree_rcu_cpu_work structure that tracks the memory that has already been handed to RCU. These structures track three categories of memory: (1) Memory for kfree(), (2) Memory for kvfree(), and (3) Memory for both that arrived during an OOM episode. The first two categories are tracked in a cache-friendly manner involving a dynamically allocated page of pointers (the aforementioned kvfree_rcu_bulk_data structures), while the third uses a simple (but decidedly cache-unfriendly) linked list through the rcu_head structures in each block of memory. On a given CPU, these three categories are handled as a unit, with that CPU's kfree_rcu_cpu_work structure having one pointer for each of the three categories. Clearly, new memory for a given category cannot be placed in the corresponding kfree_rcu_cpu_work structure until any old memory has had its grace period elapse and thus has been removed. And the kfree_rcu_monitor() function does in fact check for this. Except that the kfree_rcu_monitor() function checks these pointers one at a time. This means that if the previous kfree_rcu() memory passed to RCU had only category 1 and the current one has only category 2, the kfree_rcu_monitor() function will send that current category-2 memory along immediately. This can result in memory being freed too soon, that is, out from under unsuspecting RCU readers. To see this, consider the following sequence of events, in which: o Task A on CPU 0 calls rcu_read_lock(), then uses "from_cset", then is preempted. o CPU 1 calls kfree_rcu(cset, rcu_head) in order to free "from_cset" after a later grace period. Except that "from_cset" is freed right after the previous grace period ended, so that "from_cset" is immediately freed. Task A resumes and references "from_cset"'s member, after which nothing good happens. In full detail: CPU 0 CPU 1 ---------------------- ---------------------- count_memcg_event_mm() |rcu_read_lock() <--- |mem_cgroup_from_task() |// css_set_ptr is the "from_cset" mentioned on CPU 1 |css_set_ptr = rcu_dereference((task)->cgroups) |// Hard irq comes, current task is scheduled out. cgroup_attach_task() |cgroup_migrate() |cgroup_migrate_execute() |css_set_move_task(task, from_cset, to_cset, true) |cgroup_move_task(task, to_cset) |rcu_assign_pointer(.., to_cset) |... |cgroup_migrate_finish() |put_css_set_locked(from_cset) |from_cset->refcount return 0 |kfree_rcu(cset, rcu_head) // free from_cset after new gp |add_ptr_to_bulk_krc_lock() |schedule_delayed_work(&krcp->monitor_work, ..) kfree_rcu_monitor() |krcp->bulk_head[0]'s work attached to krwp->bulk_head_free[] |queue_rcu_work(system_wq, &krwp->rcu_work) |if rwork->rcu.work is not in WORK_STRUCT_PENDING_BIT state, |call_rcu(&rwork->rcu, rcu_work_rcufn) <--- request new gp // There is a perious call_rcu(.., rcu_work_rcufn) // gp end, rcu_work_rcufn() is called. rcu_work_rcufn() |__queue_work(.., rwork->wq, &rwork->work); |kfree_rcu_work() |krwp->bulk_head_free[0] bulk is freed before new gp end!!! |The "from_cset" is freed before new gp end. // the task resumes some time later. |css_set_ptr->subsys[(subsys_id) <--- Caused kernel crash, because css_set_ptr is freed. This commit therefore causes kfree_rcu_monitor() to refrain from moving kfree_rcu() memory to the kfree_rcu_cpu_work structure until the RCU grace period has completed for all three categories. v2: Use helper function instead of inserted code block at kfree_rcu_monitor(). Fixes: 34c881745549 ("rcu: Support kfree_bulk() interface in kfree_rcu()") Fixes: 5f3c8d620447 ("rcu/tree: Maintain separate array for vmalloc ptrs") Reported-by: Mukesh Ojha <quic_mojha@quicinc.com> Signed-off-by: Ziwei Dai <ziwei.dai@unisoc.com> Reviewed-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2023-03-31 20:42:09 +08:00
static bool
need_wait_for_krwp_work(struct kfree_rcu_cpu_work *krwp)
{
int i;
for (i = 0; i < FREE_N_CHANNELS; i++)
if (!list_empty(&krwp->bulk_head_free[i]))
return true;
return !!krwp->head_free;
}
static int krc_count(struct kfree_rcu_cpu *krcp)
{
int sum = atomic_read(&krcp->head_count);
int i;
for (i = 0; i < FREE_N_CHANNELS; i++)
sum += atomic_read(&krcp->bulk_count[i]);
return sum;
}
rcu/kvfree: Update KFREE_DRAIN_JIFFIES interval Currently the monitor work is scheduled with a fixed interval of HZ/20, which is roughly 50 milliseconds. The drawback of this approach is low utilization of the 512 page slots in scenarios with infrequence kvfree_rcu() calls. For example on an Android system: <snip> kworker/3:3-507 [003] .... 470.286305: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000d0f0dde5 nr_records=6 kworker/6:1-76 [006] .... 470.416613: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000ea0d6556 nr_records=1 kworker/6:1-76 [006] .... 470.416625: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000003e025849 nr_records=9 kworker/3:3-507 [003] .... 471.390000: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000815a8713 nr_records=48 kworker/1:1-73 [001] .... 471.725785: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000fda9bf20 nr_records=3 kworker/1:1-73 [001] .... 471.725833: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000a425b67b nr_records=76 kworker/0:4-1411 [000] .... 472.085673: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000007996be9d nr_records=1 kworker/0:4-1411 [000] .... 472.085728: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000d0f0dde5 nr_records=5 kworker/6:1-76 [006] .... 472.260340: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000065630ee4 nr_records=102 <snip> In many cases, out of 512 slots, fewer than 10 were actually used. In order to improve batching and make utilization more efficient this commit sets a drain interval to a fixed 5-seconds interval. Floods are detected when a page fills quickly, and in that case, the reclaim work is re-scheduled for the next scheduling-clock tick (jiffy). After this change: <snip> kworker/7:1-371 [007] .... 5630.725708: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000005ab0ffb3 nr_records=121 kworker/7:1-371 [007] .... 5630.989702: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000060c84761 nr_records=47 kworker/7:1-371 [007] .... 5630.989714: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000000babf308 nr_records=510 kworker/7:1-371 [007] .... 5631.553790: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000bb7bd0ef nr_records=169 kworker/7:1-371 [007] .... 5631.553808: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000044c78753 nr_records=510 kworker/5:6-9428 [005] .... 5631.746102: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000d98519aa nr_records=123 kworker/4:7-9434 [004] .... 5632.001758: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000526c9d44 nr_records=322 kworker/4:7-9434 [004] .... 5632.002073: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000002c6a8afa nr_records=185 kworker/7:1-371 [007] .... 5632.277515: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000007f4a962f nr_records=510 <snip> Here, all but one of the cases, more than one hundreds slots were used, representing an order-of-magnitude improvement. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2022-07-01 00:33:35 +08:00
static void
schedule_delayed_monitor_work(struct kfree_rcu_cpu *krcp)
{
long delay, delay_left;
delay = krc_count(krcp) >= KVFREE_BULK_MAX_ENTR ? 1:KFREE_DRAIN_JIFFIES;
rcu/kvfree: Update KFREE_DRAIN_JIFFIES interval Currently the monitor work is scheduled with a fixed interval of HZ/20, which is roughly 50 milliseconds. The drawback of this approach is low utilization of the 512 page slots in scenarios with infrequence kvfree_rcu() calls. For example on an Android system: <snip> kworker/3:3-507 [003] .... 470.286305: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000d0f0dde5 nr_records=6 kworker/6:1-76 [006] .... 470.416613: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000ea0d6556 nr_records=1 kworker/6:1-76 [006] .... 470.416625: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000003e025849 nr_records=9 kworker/3:3-507 [003] .... 471.390000: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000815a8713 nr_records=48 kworker/1:1-73 [001] .... 471.725785: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000fda9bf20 nr_records=3 kworker/1:1-73 [001] .... 471.725833: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000a425b67b nr_records=76 kworker/0:4-1411 [000] .... 472.085673: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000007996be9d nr_records=1 kworker/0:4-1411 [000] .... 472.085728: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000d0f0dde5 nr_records=5 kworker/6:1-76 [006] .... 472.260340: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000065630ee4 nr_records=102 <snip> In many cases, out of 512 slots, fewer than 10 were actually used. In order to improve batching and make utilization more efficient this commit sets a drain interval to a fixed 5-seconds interval. Floods are detected when a page fills quickly, and in that case, the reclaim work is re-scheduled for the next scheduling-clock tick (jiffy). After this change: <snip> kworker/7:1-371 [007] .... 5630.725708: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000005ab0ffb3 nr_records=121 kworker/7:1-371 [007] .... 5630.989702: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000060c84761 nr_records=47 kworker/7:1-371 [007] .... 5630.989714: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000000babf308 nr_records=510 kworker/7:1-371 [007] .... 5631.553790: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000bb7bd0ef nr_records=169 kworker/7:1-371 [007] .... 5631.553808: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000044c78753 nr_records=510 kworker/5:6-9428 [005] .... 5631.746102: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000d98519aa nr_records=123 kworker/4:7-9434 [004] .... 5632.001758: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000526c9d44 nr_records=322 kworker/4:7-9434 [004] .... 5632.002073: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000002c6a8afa nr_records=185 kworker/7:1-371 [007] .... 5632.277515: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000007f4a962f nr_records=510 <snip> Here, all but one of the cases, more than one hundreds slots were used, representing an order-of-magnitude improvement. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2022-07-01 00:33:35 +08:00
if (delayed_work_pending(&krcp->monitor_work)) {
delay_left = krcp->monitor_work.timer.expires - jiffies;
if (delay < delay_left)
mod_delayed_work(system_wq, &krcp->monitor_work, delay);
return;
}
queue_delayed_work(system_wq, &krcp->monitor_work, delay);
}
static void
kvfree_rcu_drain_ready(struct kfree_rcu_cpu *krcp)
{
struct list_head bulk_ready[FREE_N_CHANNELS];
struct kvfree_rcu_bulk_data *bnode, *n;
struct rcu_head *head_ready = NULL;
unsigned long flags;
int i;
raw_spin_lock_irqsave(&krcp->lock, flags);
for (i = 0; i < FREE_N_CHANNELS; i++) {
INIT_LIST_HEAD(&bulk_ready[i]);
list_for_each_entry_safe_reverse(bnode, n, &krcp->bulk_head[i], list) {
if (!poll_state_synchronize_rcu_full(&bnode->gp_snap))
break;
atomic_sub(bnode->nr_records, &krcp->bulk_count[i]);
list_move(&bnode->list, &bulk_ready[i]);
}
}
if (krcp->head && poll_state_synchronize_rcu(krcp->head_gp_snap)) {
head_ready = krcp->head;
atomic_set(&krcp->head_count, 0);
WRITE_ONCE(krcp->head, NULL);
}
raw_spin_unlock_irqrestore(&krcp->lock, flags);
for (i = 0; i < FREE_N_CHANNELS; i++) {
list_for_each_entry_safe(bnode, n, &bulk_ready[i], list)
kvfree_rcu_bulk(krcp, bnode, i);
}
if (head_ready)
kvfree_rcu_list(head_ready);
}
/*
* This function is invoked after the KFREE_DRAIN_JIFFIES timeout.
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
*/
static void kfree_rcu_monitor(struct work_struct *work)
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
{
struct kfree_rcu_cpu *krcp = container_of(work,
struct kfree_rcu_cpu, monitor_work.work);
unsigned long flags;
int i, j;
// Drain ready for reclaim.
kvfree_rcu_drain_ready(krcp);
raw_spin_lock_irqsave(&krcp->lock, flags);
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
// Attempt to start a new batch.
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
for (i = 0; i < KFREE_N_BATCHES; i++) {
struct kfree_rcu_cpu_work *krwp = &(krcp->krw_arr[i]);
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
rcu/kvfree: Avoid freeing new kfree_rcu() memory after old grace period Memory passed to kvfree_rcu() that is to be freed is tracked by a per-CPU kfree_rcu_cpu structure, which in turn contains pointers to kvfree_rcu_bulk_data structures that contain pointers to memory that has not yet been handed to RCU, along with an kfree_rcu_cpu_work structure that tracks the memory that has already been handed to RCU. These structures track three categories of memory: (1) Memory for kfree(), (2) Memory for kvfree(), and (3) Memory for both that arrived during an OOM episode. The first two categories are tracked in a cache-friendly manner involving a dynamically allocated page of pointers (the aforementioned kvfree_rcu_bulk_data structures), while the third uses a simple (but decidedly cache-unfriendly) linked list through the rcu_head structures in each block of memory. On a given CPU, these three categories are handled as a unit, with that CPU's kfree_rcu_cpu_work structure having one pointer for each of the three categories. Clearly, new memory for a given category cannot be placed in the corresponding kfree_rcu_cpu_work structure until any old memory has had its grace period elapse and thus has been removed. And the kfree_rcu_monitor() function does in fact check for this. Except that the kfree_rcu_monitor() function checks these pointers one at a time. This means that if the previous kfree_rcu() memory passed to RCU had only category 1 and the current one has only category 2, the kfree_rcu_monitor() function will send that current category-2 memory along immediately. This can result in memory being freed too soon, that is, out from under unsuspecting RCU readers. To see this, consider the following sequence of events, in which: o Task A on CPU 0 calls rcu_read_lock(), then uses "from_cset", then is preempted. o CPU 1 calls kfree_rcu(cset, rcu_head) in order to free "from_cset" after a later grace period. Except that "from_cset" is freed right after the previous grace period ended, so that "from_cset" is immediately freed. Task A resumes and references "from_cset"'s member, after which nothing good happens. In full detail: CPU 0 CPU 1 ---------------------- ---------------------- count_memcg_event_mm() |rcu_read_lock() <--- |mem_cgroup_from_task() |// css_set_ptr is the "from_cset" mentioned on CPU 1 |css_set_ptr = rcu_dereference((task)->cgroups) |// Hard irq comes, current task is scheduled out. cgroup_attach_task() |cgroup_migrate() |cgroup_migrate_execute() |css_set_move_task(task, from_cset, to_cset, true) |cgroup_move_task(task, to_cset) |rcu_assign_pointer(.., to_cset) |... |cgroup_migrate_finish() |put_css_set_locked(from_cset) |from_cset->refcount return 0 |kfree_rcu(cset, rcu_head) // free from_cset after new gp |add_ptr_to_bulk_krc_lock() |schedule_delayed_work(&krcp->monitor_work, ..) kfree_rcu_monitor() |krcp->bulk_head[0]'s work attached to krwp->bulk_head_free[] |queue_rcu_work(system_wq, &krwp->rcu_work) |if rwork->rcu.work is not in WORK_STRUCT_PENDING_BIT state, |call_rcu(&rwork->rcu, rcu_work_rcufn) <--- request new gp // There is a perious call_rcu(.., rcu_work_rcufn) // gp end, rcu_work_rcufn() is called. rcu_work_rcufn() |__queue_work(.., rwork->wq, &rwork->work); |kfree_rcu_work() |krwp->bulk_head_free[0] bulk is freed before new gp end!!! |The "from_cset" is freed before new gp end. // the task resumes some time later. |css_set_ptr->subsys[(subsys_id) <--- Caused kernel crash, because css_set_ptr is freed. This commit therefore causes kfree_rcu_monitor() to refrain from moving kfree_rcu() memory to the kfree_rcu_cpu_work structure until the RCU grace period has completed for all three categories. v2: Use helper function instead of inserted code block at kfree_rcu_monitor(). Fixes: 34c881745549 ("rcu: Support kfree_bulk() interface in kfree_rcu()") Fixes: 5f3c8d620447 ("rcu/tree: Maintain separate array for vmalloc ptrs") Reported-by: Mukesh Ojha <quic_mojha@quicinc.com> Signed-off-by: Ziwei Dai <ziwei.dai@unisoc.com> Reviewed-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2023-03-31 20:42:09 +08:00
// Try to detach bulk_head or head and attach it, only when
// all channels are free. Any channel is not free means at krwp
// there is on-going rcu work to handle krwp's free business.
if (need_wait_for_krwp_work(krwp))
continue;
rcu/kvfree: Avoid freeing new kfree_rcu() memory after old grace period Memory passed to kvfree_rcu() that is to be freed is tracked by a per-CPU kfree_rcu_cpu structure, which in turn contains pointers to kvfree_rcu_bulk_data structures that contain pointers to memory that has not yet been handed to RCU, along with an kfree_rcu_cpu_work structure that tracks the memory that has already been handed to RCU. These structures track three categories of memory: (1) Memory for kfree(), (2) Memory for kvfree(), and (3) Memory for both that arrived during an OOM episode. The first two categories are tracked in a cache-friendly manner involving a dynamically allocated page of pointers (the aforementioned kvfree_rcu_bulk_data structures), while the third uses a simple (but decidedly cache-unfriendly) linked list through the rcu_head structures in each block of memory. On a given CPU, these three categories are handled as a unit, with that CPU's kfree_rcu_cpu_work structure having one pointer for each of the three categories. Clearly, new memory for a given category cannot be placed in the corresponding kfree_rcu_cpu_work structure until any old memory has had its grace period elapse and thus has been removed. And the kfree_rcu_monitor() function does in fact check for this. Except that the kfree_rcu_monitor() function checks these pointers one at a time. This means that if the previous kfree_rcu() memory passed to RCU had only category 1 and the current one has only category 2, the kfree_rcu_monitor() function will send that current category-2 memory along immediately. This can result in memory being freed too soon, that is, out from under unsuspecting RCU readers. To see this, consider the following sequence of events, in which: o Task A on CPU 0 calls rcu_read_lock(), then uses "from_cset", then is preempted. o CPU 1 calls kfree_rcu(cset, rcu_head) in order to free "from_cset" after a later grace period. Except that "from_cset" is freed right after the previous grace period ended, so that "from_cset" is immediately freed. Task A resumes and references "from_cset"'s member, after which nothing good happens. In full detail: CPU 0 CPU 1 ---------------------- ---------------------- count_memcg_event_mm() |rcu_read_lock() <--- |mem_cgroup_from_task() |// css_set_ptr is the "from_cset" mentioned on CPU 1 |css_set_ptr = rcu_dereference((task)->cgroups) |// Hard irq comes, current task is scheduled out. cgroup_attach_task() |cgroup_migrate() |cgroup_migrate_execute() |css_set_move_task(task, from_cset, to_cset, true) |cgroup_move_task(task, to_cset) |rcu_assign_pointer(.., to_cset) |... |cgroup_migrate_finish() |put_css_set_locked(from_cset) |from_cset->refcount return 0 |kfree_rcu(cset, rcu_head) // free from_cset after new gp |add_ptr_to_bulk_krc_lock() |schedule_delayed_work(&krcp->monitor_work, ..) kfree_rcu_monitor() |krcp->bulk_head[0]'s work attached to krwp->bulk_head_free[] |queue_rcu_work(system_wq, &krwp->rcu_work) |if rwork->rcu.work is not in WORK_STRUCT_PENDING_BIT state, |call_rcu(&rwork->rcu, rcu_work_rcufn) <--- request new gp // There is a perious call_rcu(.., rcu_work_rcufn) // gp end, rcu_work_rcufn() is called. rcu_work_rcufn() |__queue_work(.., rwork->wq, &rwork->work); |kfree_rcu_work() |krwp->bulk_head_free[0] bulk is freed before new gp end!!! |The "from_cset" is freed before new gp end. // the task resumes some time later. |css_set_ptr->subsys[(subsys_id) <--- Caused kernel crash, because css_set_ptr is freed. This commit therefore causes kfree_rcu_monitor() to refrain from moving kfree_rcu() memory to the kfree_rcu_cpu_work structure until the RCU grace period has completed for all three categories. v2: Use helper function instead of inserted code block at kfree_rcu_monitor(). Fixes: 34c881745549 ("rcu: Support kfree_bulk() interface in kfree_rcu()") Fixes: 5f3c8d620447 ("rcu/tree: Maintain separate array for vmalloc ptrs") Reported-by: Mukesh Ojha <quic_mojha@quicinc.com> Signed-off-by: Ziwei Dai <ziwei.dai@unisoc.com> Reviewed-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2023-03-31 20:42:09 +08:00
// kvfree_rcu_drain_ready() might handle this krcp, if so give up.
if (need_offload_krc(krcp)) {
// Channel 1 corresponds to the SLAB-pointer bulk path.
// Channel 2 corresponds to vmalloc-pointer bulk path.
for (j = 0; j < FREE_N_CHANNELS; j++) {
if (list_empty(&krwp->bulk_head_free[j])) {
atomic_set(&krcp->bulk_count[j], 0);
list_replace_init(&krcp->bulk_head[j],
&krwp->bulk_head_free[j]);
}
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
}
// Channel 3 corresponds to both SLAB and vmalloc
// objects queued on the linked list.
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
if (!krwp->head_free) {
krwp->head_free = krcp->head;
get_state_synchronize_rcu_full(&krwp->head_free_gp_snap);
atomic_set(&krcp->head_count, 0);
WRITE_ONCE(krcp->head, NULL);
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
}
// One work is per one batch, so there are three
// "free channels", the batch can handle. It can
// be that the work is in the pending state when
// channels have been detached following by each
// other.
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
queue_rcu_work(system_wq, &krwp->rcu_work);
}
}
raw_spin_unlock_irqrestore(&krcp->lock, flags);
// If there is nothing to detach, it means that our job is
// successfully done here. In case of having at least one
// of the channels that is still busy we should rearm the
// work to repeat an attempt. Because previous batches are
// still in progress.
if (need_offload_krc(krcp))
rcu/kvfree: Update KFREE_DRAIN_JIFFIES interval Currently the monitor work is scheduled with a fixed interval of HZ/20, which is roughly 50 milliseconds. The drawback of this approach is low utilization of the 512 page slots in scenarios with infrequence kvfree_rcu() calls. For example on an Android system: <snip> kworker/3:3-507 [003] .... 470.286305: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000d0f0dde5 nr_records=6 kworker/6:1-76 [006] .... 470.416613: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000ea0d6556 nr_records=1 kworker/6:1-76 [006] .... 470.416625: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000003e025849 nr_records=9 kworker/3:3-507 [003] .... 471.390000: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000815a8713 nr_records=48 kworker/1:1-73 [001] .... 471.725785: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000fda9bf20 nr_records=3 kworker/1:1-73 [001] .... 471.725833: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000a425b67b nr_records=76 kworker/0:4-1411 [000] .... 472.085673: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000007996be9d nr_records=1 kworker/0:4-1411 [000] .... 472.085728: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000d0f0dde5 nr_records=5 kworker/6:1-76 [006] .... 472.260340: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000065630ee4 nr_records=102 <snip> In many cases, out of 512 slots, fewer than 10 were actually used. In order to improve batching and make utilization more efficient this commit sets a drain interval to a fixed 5-seconds interval. Floods are detected when a page fills quickly, and in that case, the reclaim work is re-scheduled for the next scheduling-clock tick (jiffy). After this change: <snip> kworker/7:1-371 [007] .... 5630.725708: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000005ab0ffb3 nr_records=121 kworker/7:1-371 [007] .... 5630.989702: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000060c84761 nr_records=47 kworker/7:1-371 [007] .... 5630.989714: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000000babf308 nr_records=510 kworker/7:1-371 [007] .... 5631.553790: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000bb7bd0ef nr_records=169 kworker/7:1-371 [007] .... 5631.553808: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000044c78753 nr_records=510 kworker/5:6-9428 [005] .... 5631.746102: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000d98519aa nr_records=123 kworker/4:7-9434 [004] .... 5632.001758: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000526c9d44 nr_records=322 kworker/4:7-9434 [004] .... 5632.002073: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000002c6a8afa nr_records=185 kworker/7:1-371 [007] .... 5632.277515: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000007f4a962f nr_records=510 <snip> Here, all but one of the cases, more than one hundreds slots were used, representing an order-of-magnitude improvement. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2022-07-01 00:33:35 +08:00
schedule_delayed_monitor_work(krcp);
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
}
rcu/tree: Defer kvfree_rcu() allocation to a clean context The current memmory-allocation interface causes the following difficulties for kvfree_rcu(): a) If built with CONFIG_PROVE_RAW_LOCK_NESTING, the lockdep will complain about violation of the nesting rules, as in "BUG: Invalid wait context". This Kconfig option checks for proper raw_spinlock vs. spinlock nesting, in particular, it is not legal to acquire a spinlock_t while holding a raw_spinlock_t. This is a problem because kfree_rcu() uses raw_spinlock_t whereas the "page allocator" internally deals with spinlock_t to access to its zones. The code also can be broken from higher level of view: <snip> raw_spin_lock(&some_lock); kfree_rcu(some_pointer, some_field_offset); <snip> b) If built with CONFIG_PREEMPT_RT, spinlock_t is converted into sleeplock. This means that invoking the page allocator from atomic contexts results in "BUG: scheduling while atomic". c) Please note that call_rcu() is already invoked from raw atomic context, so it is only reasonable to expaect that kfree_rcu() and kvfree_rcu() will also be called from atomic raw context. This commit therefore defers page allocation to a clean context using the combination of an hrtimer and a workqueue. The hrtimer stage is required in order to avoid deadlocks with the scheduler. This deferred allocation is required only when kvfree_rcu()'s per-CPU page cache is empty. Link: https://lore.kernel.org/lkml/20200630164543.4mdcf6zb4zfclhln@linutronix.de/ Fixes: 3042f83f19be ("rcu: Support reclaim for head-less object") Reported-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-10-30 00:50:04 +08:00
static enum hrtimer_restart
schedule_page_work_fn(struct hrtimer *t)
{
struct kfree_rcu_cpu *krcp =
container_of(t, struct kfree_rcu_cpu, hrtimer);
queue_delayed_work(system_highpri_wq, &krcp->page_cache_work, 0);
rcu/tree: Defer kvfree_rcu() allocation to a clean context The current memmory-allocation interface causes the following difficulties for kvfree_rcu(): a) If built with CONFIG_PROVE_RAW_LOCK_NESTING, the lockdep will complain about violation of the nesting rules, as in "BUG: Invalid wait context". This Kconfig option checks for proper raw_spinlock vs. spinlock nesting, in particular, it is not legal to acquire a spinlock_t while holding a raw_spinlock_t. This is a problem because kfree_rcu() uses raw_spinlock_t whereas the "page allocator" internally deals with spinlock_t to access to its zones. The code also can be broken from higher level of view: <snip> raw_spin_lock(&some_lock); kfree_rcu(some_pointer, some_field_offset); <snip> b) If built with CONFIG_PREEMPT_RT, spinlock_t is converted into sleeplock. This means that invoking the page allocator from atomic contexts results in "BUG: scheduling while atomic". c) Please note that call_rcu() is already invoked from raw atomic context, so it is only reasonable to expaect that kfree_rcu() and kvfree_rcu() will also be called from atomic raw context. This commit therefore defers page allocation to a clean context using the combination of an hrtimer and a workqueue. The hrtimer stage is required in order to avoid deadlocks with the scheduler. This deferred allocation is required only when kvfree_rcu()'s per-CPU page cache is empty. Link: https://lore.kernel.org/lkml/20200630164543.4mdcf6zb4zfclhln@linutronix.de/ Fixes: 3042f83f19be ("rcu: Support reclaim for head-less object") Reported-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-10-30 00:50:04 +08:00
return HRTIMER_NORESTART;
}
static void fill_page_cache_func(struct work_struct *work)
{
struct kvfree_rcu_bulk_data *bnode;
struct kfree_rcu_cpu *krcp =
container_of(work, struct kfree_rcu_cpu,
page_cache_work.work);
rcu/tree: Defer kvfree_rcu() allocation to a clean context The current memmory-allocation interface causes the following difficulties for kvfree_rcu(): a) If built with CONFIG_PROVE_RAW_LOCK_NESTING, the lockdep will complain about violation of the nesting rules, as in "BUG: Invalid wait context". This Kconfig option checks for proper raw_spinlock vs. spinlock nesting, in particular, it is not legal to acquire a spinlock_t while holding a raw_spinlock_t. This is a problem because kfree_rcu() uses raw_spinlock_t whereas the "page allocator" internally deals with spinlock_t to access to its zones. The code also can be broken from higher level of view: <snip> raw_spin_lock(&some_lock); kfree_rcu(some_pointer, some_field_offset); <snip> b) If built with CONFIG_PREEMPT_RT, spinlock_t is converted into sleeplock. This means that invoking the page allocator from atomic contexts results in "BUG: scheduling while atomic". c) Please note that call_rcu() is already invoked from raw atomic context, so it is only reasonable to expaect that kfree_rcu() and kvfree_rcu() will also be called from atomic raw context. This commit therefore defers page allocation to a clean context using the combination of an hrtimer and a workqueue. The hrtimer stage is required in order to avoid deadlocks with the scheduler. This deferred allocation is required only when kvfree_rcu()'s per-CPU page cache is empty. Link: https://lore.kernel.org/lkml/20200630164543.4mdcf6zb4zfclhln@linutronix.de/ Fixes: 3042f83f19be ("rcu: Support reclaim for head-less object") Reported-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-10-30 00:50:04 +08:00
unsigned long flags;
int nr_pages;
rcu/tree: Defer kvfree_rcu() allocation to a clean context The current memmory-allocation interface causes the following difficulties for kvfree_rcu(): a) If built with CONFIG_PROVE_RAW_LOCK_NESTING, the lockdep will complain about violation of the nesting rules, as in "BUG: Invalid wait context". This Kconfig option checks for proper raw_spinlock vs. spinlock nesting, in particular, it is not legal to acquire a spinlock_t while holding a raw_spinlock_t. This is a problem because kfree_rcu() uses raw_spinlock_t whereas the "page allocator" internally deals with spinlock_t to access to its zones. The code also can be broken from higher level of view: <snip> raw_spin_lock(&some_lock); kfree_rcu(some_pointer, some_field_offset); <snip> b) If built with CONFIG_PREEMPT_RT, spinlock_t is converted into sleeplock. This means that invoking the page allocator from atomic contexts results in "BUG: scheduling while atomic". c) Please note that call_rcu() is already invoked from raw atomic context, so it is only reasonable to expaect that kfree_rcu() and kvfree_rcu() will also be called from atomic raw context. This commit therefore defers page allocation to a clean context using the combination of an hrtimer and a workqueue. The hrtimer stage is required in order to avoid deadlocks with the scheduler. This deferred allocation is required only when kvfree_rcu()'s per-CPU page cache is empty. Link: https://lore.kernel.org/lkml/20200630164543.4mdcf6zb4zfclhln@linutronix.de/ Fixes: 3042f83f19be ("rcu: Support reclaim for head-less object") Reported-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-10-30 00:50:04 +08:00
bool pushed;
int i;
nr_pages = atomic_read(&krcp->backoff_page_cache_fill) ?
1 : rcu_min_cached_objs;
for (i = READ_ONCE(krcp->nr_bkv_objs); i < nr_pages; i++) {
rcu/tree: Defer kvfree_rcu() allocation to a clean context The current memmory-allocation interface causes the following difficulties for kvfree_rcu(): a) If built with CONFIG_PROVE_RAW_LOCK_NESTING, the lockdep will complain about violation of the nesting rules, as in "BUG: Invalid wait context". This Kconfig option checks for proper raw_spinlock vs. spinlock nesting, in particular, it is not legal to acquire a spinlock_t while holding a raw_spinlock_t. This is a problem because kfree_rcu() uses raw_spinlock_t whereas the "page allocator" internally deals with spinlock_t to access to its zones. The code also can be broken from higher level of view: <snip> raw_spin_lock(&some_lock); kfree_rcu(some_pointer, some_field_offset); <snip> b) If built with CONFIG_PREEMPT_RT, spinlock_t is converted into sleeplock. This means that invoking the page allocator from atomic contexts results in "BUG: scheduling while atomic". c) Please note that call_rcu() is already invoked from raw atomic context, so it is only reasonable to expaect that kfree_rcu() and kvfree_rcu() will also be called from atomic raw context. This commit therefore defers page allocation to a clean context using the combination of an hrtimer and a workqueue. The hrtimer stage is required in order to avoid deadlocks with the scheduler. This deferred allocation is required only when kvfree_rcu()'s per-CPU page cache is empty. Link: https://lore.kernel.org/lkml/20200630164543.4mdcf6zb4zfclhln@linutronix.de/ Fixes: 3042f83f19be ("rcu: Support reclaim for head-less object") Reported-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-10-30 00:50:04 +08:00
bnode = (struct kvfree_rcu_bulk_data *)
kvfree_rcu: Use same set of GFP flags as does single-argument Running an rcuscale stress-suite can lead to "Out of memory" of a system. This can happen under high memory pressure with a small amount of physical memory. For example, a KVM test configuration with 64 CPUs and 512 megabytes can result in OOM when running rcuscale with below parameters: ../kvm.sh --torture rcuscale --allcpus --duration 10 --kconfig CONFIG_NR_CPUS=64 \ --bootargs "rcuscale.kfree_rcu_test=1 rcuscale.kfree_nthreads=16 rcuscale.holdoff=20 \ rcuscale.kfree_loops=10000 torture.disable_onoff_at_boot" --trust-make <snip> [ 12.054448] kworker/1:1H invoked oom-killer: gfp_mask=0x2cc0(GFP_KERNEL|__GFP_NOWARN), order=0, oom_score_adj=0 [ 12.055303] CPU: 1 PID: 377 Comm: kworker/1:1H Not tainted 5.11.0-rc3+ #510 [ 12.055416] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.12.0-1 04/01/2014 [ 12.056485] Workqueue: events_highpri fill_page_cache_func [ 12.056485] Call Trace: [ 12.056485] dump_stack+0x57/0x6a [ 12.056485] dump_header+0x4c/0x30a [ 12.056485] ? del_timer_sync+0x20/0x30 [ 12.056485] out_of_memory.cold.47+0xa/0x7e [ 12.056485] __alloc_pages_slowpath.constprop.123+0x82f/0xc00 [ 12.056485] __alloc_pages_nodemask+0x289/0x2c0 [ 12.056485] __get_free_pages+0x8/0x30 [ 12.056485] fill_page_cache_func+0x39/0xb0 [ 12.056485] process_one_work+0x1ed/0x3b0 [ 12.056485] ? process_one_work+0x3b0/0x3b0 [ 12.060485] worker_thread+0x28/0x3c0 [ 12.060485] ? process_one_work+0x3b0/0x3b0 [ 12.060485] kthread+0x138/0x160 [ 12.060485] ? kthread_park+0x80/0x80 [ 12.060485] ret_from_fork+0x22/0x30 [ 12.062156] Mem-Info: [ 12.062350] active_anon:0 inactive_anon:0 isolated_anon:0 [ 12.062350] active_file:0 inactive_file:0 isolated_file:0 [ 12.062350] unevictable:0 dirty:0 writeback:0 [ 12.062350] slab_reclaimable:2797 slab_unreclaimable:80920 [ 12.062350] mapped:1 shmem:2 pagetables:8 bounce:0 [ 12.062350] free:10488 free_pcp:1227 free_cma:0 ... [ 12.101610] Out of memory and no killable processes... [ 12.102042] Kernel panic - not syncing: System is deadlocked on memory [ 12.102583] CPU: 1 PID: 377 Comm: kworker/1:1H Not tainted 5.11.0-rc3+ #510 [ 12.102600] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.12.0-1 04/01/2014 <snip> Because kvfree_rcu() has a fallback path, memory allocation failure is not the end of the world. Furthermore, the added overhead of aggressive GFP settings must be balanced against the overhead of the fallback path, which is a cache miss for double-argument kvfree_rcu() and a call to synchronize_rcu() for single-argument kvfree_rcu(). The current choice of GFP_KERNEL|__GFP_NOWARN can result in longer latencies than a call to synchronize_rcu(), so less-tenacious GFP flags would be helpful. Here is the tradeoff that must be balanced: a) Minimize use of the fallback path, b) Avoid pushing the system into OOM, c) Bound allocation latency to that of synchronize_rcu(), and d) Leave the emergency reserves to use cases lacking fallbacks. This commit therefore changes GFP flags from GFP_KERNEL|__GFP_NOWARN to GFP_KERNEL|__GFP_NORETRY|__GFP_NOMEMALLOC|__GFP_NOWARN. This combination leaves the emergency reserves alone and can initiate reclaim, but will not invoke the OOM killer. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2021-01-30 04:05:05 +08:00
__get_free_page(GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
rcu/tree: Defer kvfree_rcu() allocation to a clean context The current memmory-allocation interface causes the following difficulties for kvfree_rcu(): a) If built with CONFIG_PROVE_RAW_LOCK_NESTING, the lockdep will complain about violation of the nesting rules, as in "BUG: Invalid wait context". This Kconfig option checks for proper raw_spinlock vs. spinlock nesting, in particular, it is not legal to acquire a spinlock_t while holding a raw_spinlock_t. This is a problem because kfree_rcu() uses raw_spinlock_t whereas the "page allocator" internally deals with spinlock_t to access to its zones. The code also can be broken from higher level of view: <snip> raw_spin_lock(&some_lock); kfree_rcu(some_pointer, some_field_offset); <snip> b) If built with CONFIG_PREEMPT_RT, spinlock_t is converted into sleeplock. This means that invoking the page allocator from atomic contexts results in "BUG: scheduling while atomic". c) Please note that call_rcu() is already invoked from raw atomic context, so it is only reasonable to expaect that kfree_rcu() and kvfree_rcu() will also be called from atomic raw context. This commit therefore defers page allocation to a clean context using the combination of an hrtimer and a workqueue. The hrtimer stage is required in order to avoid deadlocks with the scheduler. This deferred allocation is required only when kvfree_rcu()'s per-CPU page cache is empty. Link: https://lore.kernel.org/lkml/20200630164543.4mdcf6zb4zfclhln@linutronix.de/ Fixes: 3042f83f19be ("rcu: Support reclaim for head-less object") Reported-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-10-30 00:50:04 +08:00
rcu: Back off upon fill_page_cache_func() allocation failure The fill_page_cache_func() function allocates couple of pages to store kvfree_rcu_bulk_data structures. This is a lightweight (GFP_NORETRY) allocation which can fail under memory pressure. The function will, however keep retrying even when the previous attempt has failed. This retrying is in theory correct, but in practice the allocation is invoked from workqueue context, which means that if the memory reclaim gets stuck, these retries can hog the worker for quite some time. Although the workqueues subsystem automatically adjusts concurrency, such adjustment is not guaranteed to happen until the worker context sleeps. And the fill_page_cache_func() function's retry loop is not guaranteed to sleep (see the should_reclaim_retry() function). And we have seen this function cause workqueue lockups: kernel: BUG: workqueue lockup - pool cpus=93 node=1 flags=0x1 nice=0 stuck for 32s! [...] kernel: pool 74: cpus=37 node=0 flags=0x1 nice=0 hung=32s workers=2 manager: 2146 kernel: pwq 498: cpus=249 node=1 flags=0x1 nice=0 active=4/256 refcnt=5 kernel: in-flight: 1917:fill_page_cache_func kernel: pending: dbs_work_handler, free_work, kfree_rcu_monitor Originally, we thought that the root cause of this lockup was several retries with direct reclaim, but this is not yet confirmed. Furthermore, we have seen similar lockups without any heavy memory pressure. This suggests that there are other factors contributing to these lockups. However, it is not really clear that endless retries are desireable. So let's make the fill_page_cache_func() function back off after allocation failure. Cc: Uladzislau Rezki (Sony) <urezki@gmail.com> Cc: "Paul E. McKenney" <paulmck@kernel.org> Cc: Frederic Weisbecker <frederic@kernel.org> Cc: Neeraj Upadhyay <quic_neeraju@quicinc.com> Cc: Josh Triplett <josh@joshtriplett.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Joel Fernandes <joel@joelfernandes.org> Signed-off-by: Michal Hocko <mhocko@suse.com> Reviewed-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2022-06-22 19:47:11 +08:00
if (!bnode)
break;
rcu/tree: Defer kvfree_rcu() allocation to a clean context The current memmory-allocation interface causes the following difficulties for kvfree_rcu(): a) If built with CONFIG_PROVE_RAW_LOCK_NESTING, the lockdep will complain about violation of the nesting rules, as in "BUG: Invalid wait context". This Kconfig option checks for proper raw_spinlock vs. spinlock nesting, in particular, it is not legal to acquire a spinlock_t while holding a raw_spinlock_t. This is a problem because kfree_rcu() uses raw_spinlock_t whereas the "page allocator" internally deals with spinlock_t to access to its zones. The code also can be broken from higher level of view: <snip> raw_spin_lock(&some_lock); kfree_rcu(some_pointer, some_field_offset); <snip> b) If built with CONFIG_PREEMPT_RT, spinlock_t is converted into sleeplock. This means that invoking the page allocator from atomic contexts results in "BUG: scheduling while atomic". c) Please note that call_rcu() is already invoked from raw atomic context, so it is only reasonable to expaect that kfree_rcu() and kvfree_rcu() will also be called from atomic raw context. This commit therefore defers page allocation to a clean context using the combination of an hrtimer and a workqueue. The hrtimer stage is required in order to avoid deadlocks with the scheduler. This deferred allocation is required only when kvfree_rcu()'s per-CPU page cache is empty. Link: https://lore.kernel.org/lkml/20200630164543.4mdcf6zb4zfclhln@linutronix.de/ Fixes: 3042f83f19be ("rcu: Support reclaim for head-less object") Reported-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-10-30 00:50:04 +08:00
rcu: Back off upon fill_page_cache_func() allocation failure The fill_page_cache_func() function allocates couple of pages to store kvfree_rcu_bulk_data structures. This is a lightweight (GFP_NORETRY) allocation which can fail under memory pressure. The function will, however keep retrying even when the previous attempt has failed. This retrying is in theory correct, but in practice the allocation is invoked from workqueue context, which means that if the memory reclaim gets stuck, these retries can hog the worker for quite some time. Although the workqueues subsystem automatically adjusts concurrency, such adjustment is not guaranteed to happen until the worker context sleeps. And the fill_page_cache_func() function's retry loop is not guaranteed to sleep (see the should_reclaim_retry() function). And we have seen this function cause workqueue lockups: kernel: BUG: workqueue lockup - pool cpus=93 node=1 flags=0x1 nice=0 stuck for 32s! [...] kernel: pool 74: cpus=37 node=0 flags=0x1 nice=0 hung=32s workers=2 manager: 2146 kernel: pwq 498: cpus=249 node=1 flags=0x1 nice=0 active=4/256 refcnt=5 kernel: in-flight: 1917:fill_page_cache_func kernel: pending: dbs_work_handler, free_work, kfree_rcu_monitor Originally, we thought that the root cause of this lockup was several retries with direct reclaim, but this is not yet confirmed. Furthermore, we have seen similar lockups without any heavy memory pressure. This suggests that there are other factors contributing to these lockups. However, it is not really clear that endless retries are desireable. So let's make the fill_page_cache_func() function back off after allocation failure. Cc: Uladzislau Rezki (Sony) <urezki@gmail.com> Cc: "Paul E. McKenney" <paulmck@kernel.org> Cc: Frederic Weisbecker <frederic@kernel.org> Cc: Neeraj Upadhyay <quic_neeraju@quicinc.com> Cc: Josh Triplett <josh@joshtriplett.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Joel Fernandes <joel@joelfernandes.org> Signed-off-by: Michal Hocko <mhocko@suse.com> Reviewed-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2022-06-22 19:47:11 +08:00
raw_spin_lock_irqsave(&krcp->lock, flags);
pushed = put_cached_bnode(krcp, bnode);
raw_spin_unlock_irqrestore(&krcp->lock, flags);
if (!pushed) {
free_page((unsigned long) bnode);
break;
rcu/tree: Defer kvfree_rcu() allocation to a clean context The current memmory-allocation interface causes the following difficulties for kvfree_rcu(): a) If built with CONFIG_PROVE_RAW_LOCK_NESTING, the lockdep will complain about violation of the nesting rules, as in "BUG: Invalid wait context". This Kconfig option checks for proper raw_spinlock vs. spinlock nesting, in particular, it is not legal to acquire a spinlock_t while holding a raw_spinlock_t. This is a problem because kfree_rcu() uses raw_spinlock_t whereas the "page allocator" internally deals with spinlock_t to access to its zones. The code also can be broken from higher level of view: <snip> raw_spin_lock(&some_lock); kfree_rcu(some_pointer, some_field_offset); <snip> b) If built with CONFIG_PREEMPT_RT, spinlock_t is converted into sleeplock. This means that invoking the page allocator from atomic contexts results in "BUG: scheduling while atomic". c) Please note that call_rcu() is already invoked from raw atomic context, so it is only reasonable to expaect that kfree_rcu() and kvfree_rcu() will also be called from atomic raw context. This commit therefore defers page allocation to a clean context using the combination of an hrtimer and a workqueue. The hrtimer stage is required in order to avoid deadlocks with the scheduler. This deferred allocation is required only when kvfree_rcu()'s per-CPU page cache is empty. Link: https://lore.kernel.org/lkml/20200630164543.4mdcf6zb4zfclhln@linutronix.de/ Fixes: 3042f83f19be ("rcu: Support reclaim for head-less object") Reported-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-10-30 00:50:04 +08:00
}
}
atomic_set(&krcp->work_in_progress, 0);
atomic_set(&krcp->backoff_page_cache_fill, 0);
rcu/tree: Defer kvfree_rcu() allocation to a clean context The current memmory-allocation interface causes the following difficulties for kvfree_rcu(): a) If built with CONFIG_PROVE_RAW_LOCK_NESTING, the lockdep will complain about violation of the nesting rules, as in "BUG: Invalid wait context". This Kconfig option checks for proper raw_spinlock vs. spinlock nesting, in particular, it is not legal to acquire a spinlock_t while holding a raw_spinlock_t. This is a problem because kfree_rcu() uses raw_spinlock_t whereas the "page allocator" internally deals with spinlock_t to access to its zones. The code also can be broken from higher level of view: <snip> raw_spin_lock(&some_lock); kfree_rcu(some_pointer, some_field_offset); <snip> b) If built with CONFIG_PREEMPT_RT, spinlock_t is converted into sleeplock. This means that invoking the page allocator from atomic contexts results in "BUG: scheduling while atomic". c) Please note that call_rcu() is already invoked from raw atomic context, so it is only reasonable to expaect that kfree_rcu() and kvfree_rcu() will also be called from atomic raw context. This commit therefore defers page allocation to a clean context using the combination of an hrtimer and a workqueue. The hrtimer stage is required in order to avoid deadlocks with the scheduler. This deferred allocation is required only when kvfree_rcu()'s per-CPU page cache is empty. Link: https://lore.kernel.org/lkml/20200630164543.4mdcf6zb4zfclhln@linutronix.de/ Fixes: 3042f83f19be ("rcu: Support reclaim for head-less object") Reported-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-10-30 00:50:04 +08:00
}
static void
run_page_cache_worker(struct kfree_rcu_cpu *krcp)
{
// If cache disabled, bail out.
if (!rcu_min_cached_objs)
return;
rcu/tree: Defer kvfree_rcu() allocation to a clean context The current memmory-allocation interface causes the following difficulties for kvfree_rcu(): a) If built with CONFIG_PROVE_RAW_LOCK_NESTING, the lockdep will complain about violation of the nesting rules, as in "BUG: Invalid wait context". This Kconfig option checks for proper raw_spinlock vs. spinlock nesting, in particular, it is not legal to acquire a spinlock_t while holding a raw_spinlock_t. This is a problem because kfree_rcu() uses raw_spinlock_t whereas the "page allocator" internally deals with spinlock_t to access to its zones. The code also can be broken from higher level of view: <snip> raw_spin_lock(&some_lock); kfree_rcu(some_pointer, some_field_offset); <snip> b) If built with CONFIG_PREEMPT_RT, spinlock_t is converted into sleeplock. This means that invoking the page allocator from atomic contexts results in "BUG: scheduling while atomic". c) Please note that call_rcu() is already invoked from raw atomic context, so it is only reasonable to expaect that kfree_rcu() and kvfree_rcu() will also be called from atomic raw context. This commit therefore defers page allocation to a clean context using the combination of an hrtimer and a workqueue. The hrtimer stage is required in order to avoid deadlocks with the scheduler. This deferred allocation is required only when kvfree_rcu()'s per-CPU page cache is empty. Link: https://lore.kernel.org/lkml/20200630164543.4mdcf6zb4zfclhln@linutronix.de/ Fixes: 3042f83f19be ("rcu: Support reclaim for head-less object") Reported-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-10-30 00:50:04 +08:00
if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING &&
!atomic_xchg(&krcp->work_in_progress, 1)) {
if (atomic_read(&krcp->backoff_page_cache_fill)) {
queue_delayed_work(system_wq,
&krcp->page_cache_work,
msecs_to_jiffies(rcu_delay_page_cache_fill_msec));
} else {
hrtimer_init(&krcp->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
krcp->hrtimer.function = schedule_page_work_fn;
hrtimer_start(&krcp->hrtimer, 0, HRTIMER_MODE_REL);
}
rcu/tree: Defer kvfree_rcu() allocation to a clean context The current memmory-allocation interface causes the following difficulties for kvfree_rcu(): a) If built with CONFIG_PROVE_RAW_LOCK_NESTING, the lockdep will complain about violation of the nesting rules, as in "BUG: Invalid wait context". This Kconfig option checks for proper raw_spinlock vs. spinlock nesting, in particular, it is not legal to acquire a spinlock_t while holding a raw_spinlock_t. This is a problem because kfree_rcu() uses raw_spinlock_t whereas the "page allocator" internally deals with spinlock_t to access to its zones. The code also can be broken from higher level of view: <snip> raw_spin_lock(&some_lock); kfree_rcu(some_pointer, some_field_offset); <snip> b) If built with CONFIG_PREEMPT_RT, spinlock_t is converted into sleeplock. This means that invoking the page allocator from atomic contexts results in "BUG: scheduling while atomic". c) Please note that call_rcu() is already invoked from raw atomic context, so it is only reasonable to expaect that kfree_rcu() and kvfree_rcu() will also be called from atomic raw context. This commit therefore defers page allocation to a clean context using the combination of an hrtimer and a workqueue. The hrtimer stage is required in order to avoid deadlocks with the scheduler. This deferred allocation is required only when kvfree_rcu()'s per-CPU page cache is empty. Link: https://lore.kernel.org/lkml/20200630164543.4mdcf6zb4zfclhln@linutronix.de/ Fixes: 3042f83f19be ("rcu: Support reclaim for head-less object") Reported-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-10-30 00:50:04 +08:00
}
}
// Record ptr in a page managed by krcp, with the pre-krc_this_cpu_lock()
// state specified by flags. If can_alloc is true, the caller must
// be schedulable and not be holding any locks or mutexes that might be
// acquired by the memory allocator or anything that it might invoke.
// Returns true if ptr was successfully recorded, else the caller must
// use a fallback.
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
static inline bool
add_ptr_to_bulk_krc_lock(struct kfree_rcu_cpu **krcp,
unsigned long *flags, void *ptr, bool can_alloc)
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
{
struct kvfree_rcu_bulk_data *bnode;
int idx;
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
*krcp = krc_this_cpu_lock(flags);
if (unlikely(!(*krcp)->initialized))
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
return false;
idx = !!is_vmalloc_addr(ptr);
bnode = list_first_entry_or_null(&(*krcp)->bulk_head[idx],
struct kvfree_rcu_bulk_data, list);
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
/* Check if a new block is required. */
if (!bnode || bnode->nr_records == KVFREE_BULK_MAX_ENTR) {
bnode = get_cached_bnode(*krcp);
if (!bnode && can_alloc) {
krc_this_cpu_unlock(*krcp, *flags);
// __GFP_NORETRY - allows a light-weight direct reclaim
// what is OK from minimizing of fallback hitting point of
// view. Apart of that it forbids any OOM invoking what is
// also beneficial since we are about to release memory soon.
//
// __GFP_NOMEMALLOC - prevents from consuming of all the
// memory reserves. Please note we have a fallback path.
//
// __GFP_NOWARN - it is supposed that an allocation can
// be failed under low memory or high memory pressure
// scenarios.
bnode = (struct kvfree_rcu_bulk_data *)
__get_free_page(GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
raw_spin_lock_irqsave(&(*krcp)->lock, *flags);
}
rcu/tree: Defer kvfree_rcu() allocation to a clean context The current memmory-allocation interface causes the following difficulties for kvfree_rcu(): a) If built with CONFIG_PROVE_RAW_LOCK_NESTING, the lockdep will complain about violation of the nesting rules, as in "BUG: Invalid wait context". This Kconfig option checks for proper raw_spinlock vs. spinlock nesting, in particular, it is not legal to acquire a spinlock_t while holding a raw_spinlock_t. This is a problem because kfree_rcu() uses raw_spinlock_t whereas the "page allocator" internally deals with spinlock_t to access to its zones. The code also can be broken from higher level of view: <snip> raw_spin_lock(&some_lock); kfree_rcu(some_pointer, some_field_offset); <snip> b) If built with CONFIG_PREEMPT_RT, spinlock_t is converted into sleeplock. This means that invoking the page allocator from atomic contexts results in "BUG: scheduling while atomic". c) Please note that call_rcu() is already invoked from raw atomic context, so it is only reasonable to expaect that kfree_rcu() and kvfree_rcu() will also be called from atomic raw context. This commit therefore defers page allocation to a clean context using the combination of an hrtimer and a workqueue. The hrtimer stage is required in order to avoid deadlocks with the scheduler. This deferred allocation is required only when kvfree_rcu()'s per-CPU page cache is empty. Link: https://lore.kernel.org/lkml/20200630164543.4mdcf6zb4zfclhln@linutronix.de/ Fixes: 3042f83f19be ("rcu: Support reclaim for head-less object") Reported-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-10-30 00:50:04 +08:00
if (!bnode)
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
return false;
// Initialize the new block and attach it.
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
bnode->nr_records = 0;
list_add(&bnode->list, &(*krcp)->bulk_head[idx]);
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
}
rcu/kvfree: Use a polled API to speedup a reclaim process Currently all objects placed into a batch wait for a full grace period to elapse after that batch is ready to send to RCU. However, this can unnecessarily delay freeing of the first objects that were added to the batch. After all, several RCU grace periods might have elapsed since those objects were added, and if so, there is no point in further deferring their freeing. This commit therefore adds per-page grace-period snapshots which are obtained from get_state_synchronize_rcu(). When the batch is ready to be passed to call_rcu(), each page's snapshot is checked by passing it to poll_state_synchronize_rcu(). If a given page's RCU grace period has already elapsed, its objects are freed immediately by kvfree_rcu_bulk(). Otherwise, these objects are freed after a call to synchronize_rcu(). This approach requires that the pages be traversed in reverse order, that is, the oldest ones first. Test example: kvm.sh --memory 10G --torture rcuscale --allcpus --duration 1 \ --kconfig CONFIG_NR_CPUS=64 \ --kconfig CONFIG_RCU_NOCB_CPU=y \ --kconfig CONFIG_RCU_NOCB_CPU_DEFAULT_ALL=y \ --kconfig CONFIG_RCU_LAZY=n \ --bootargs "rcuscale.kfree_rcu_test=1 rcuscale.kfree_nthreads=16 \ rcuscale.holdoff=20 rcuscale.kfree_loops=10000 \ torture.disable_onoff_at_boot" --trust-make Before this commit: Total time taken by all kfree'ers: 8535693700 ns, loops: 10000, batches: 1188, memory footprint: 2248MB Total time taken by all kfree'ers: 8466933582 ns, loops: 10000, batches: 1157, memory footprint: 2820MB Total time taken by all kfree'ers: 5375602446 ns, loops: 10000, batches: 1130, memory footprint: 6502MB Total time taken by all kfree'ers: 7523283832 ns, loops: 10000, batches: 1006, memory footprint: 3343MB Total time taken by all kfree'ers: 6459171956 ns, loops: 10000, batches: 1150, memory footprint: 6549MB After this commit: Total time taken by all kfree'ers: 8560060176 ns, loops: 10000, batches: 1787, memory footprint: 61MB Total time taken by all kfree'ers: 8573885501 ns, loops: 10000, batches: 1777, memory footprint: 93MB Total time taken by all kfree'ers: 8320000202 ns, loops: 10000, batches: 1727, memory footprint: 66MB Total time taken by all kfree'ers: 8552718794 ns, loops: 10000, batches: 1790, memory footprint: 75MB Total time taken by all kfree'ers: 8601368792 ns, loops: 10000, batches: 1724, memory footprint: 62MB The reduction in memory footprint is well in excess of an order of magnitude. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2022-11-29 23:58:22 +08:00
// Finally insert and update the GP for this page.
bnode->records[bnode->nr_records++] = ptr;
get_state_synchronize_rcu_full(&bnode->gp_snap);
atomic_inc(&(*krcp)->bulk_count[idx]);
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
return true;
}
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
/*
* Queue a request for lazy invocation of the appropriate free routine
* after a grace period. Please note that three paths are maintained,
* two for the common case using arrays of pointers and a third one that
* is used only when the main paths cannot be used, for example, due to
* memory pressure.
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
*
* Each kvfree_call_rcu() request is added to a batch. The batch will be drained
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
* every KFREE_DRAIN_JIFFIES number of jiffies. All the objects in the batch will
* be free'd in workqueue context. This allows us to: batch requests together to
* reduce the number of grace periods during heavy kfree_rcu()/kvfree_rcu() load.
*/
void kvfree_call_rcu(struct rcu_head *head, void *ptr)
{
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
unsigned long flags;
struct kfree_rcu_cpu *krcp;
bool success;
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
/*
* Please note there is a limitation for the head-less
* variant, that is why there is a clear rule for such
* objects: it can be used from might_sleep() context
* only. For other places please embed an rcu_head to
* your data.
*/
if (!head)
might_sleep();
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
// Queue the object but don't yet schedule the batch.
if (debug_rcu_head_queue(ptr)) {
// Probable double kfree_rcu(), just leak.
WARN_ONCE(1, "%s(): Double-freed call. rcu_head %p\n",
__func__, head);
// Mark as success and leave.
return;
}
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
rcu: Avoid alloc_pages() when recording stack The default kasan_record_aux_stack() calls stack_depot_save() with GFP_NOWAIT, which in turn can then call alloc_pages(GFP_NOWAIT, ...). In general, however, it is not even possible to use either GFP_ATOMIC nor GFP_NOWAIT in certain non-preemptive contexts/RT kernel including raw_spin_locks (see gfp.h and ab00db216c9c7). Fix it by instructing stackdepot to not expand stack storage via alloc_pages() in case it runs out by using kasan_record_aux_stack_noalloc(). Jianwei Hu reported: BUG: sleeping function called from invalid context at kernel/locking/rtmutex.c:969 in_atomic(): 0, irqs_disabled(): 1, non_block: 0, pid: 15319, name: python3 INFO: lockdep is turned off. irq event stamp: 0 hardirqs last enabled at (0): [<0000000000000000>] 0x0 hardirqs last disabled at (0): [<ffffffff856c8b13>] copy_process+0xaf3/0x2590 softirqs last enabled at (0): [<ffffffff856c8b13>] copy_process+0xaf3/0x2590 softirqs last disabled at (0): [<0000000000000000>] 0x0 CPU: 6 PID: 15319 Comm: python3 Tainted: G W O 5.15-rc7-preempt-rt #1 Hardware name: Supermicro SYS-E300-9A-8C/A2SDi-8C-HLN4F, BIOS 1.1b 12/17/2018 Call Trace: show_stack+0x52/0x58 dump_stack+0xa1/0xd6 ___might_sleep.cold+0x11c/0x12d rt_spin_lock+0x3f/0xc0 rmqueue+0x100/0x1460 rmqueue+0x100/0x1460 mark_usage+0x1a0/0x1a0 ftrace_graph_ret_addr+0x2a/0xb0 rmqueue_pcplist.constprop.0+0x6a0/0x6a0 __kasan_check_read+0x11/0x20 __zone_watermark_ok+0x114/0x270 get_page_from_freelist+0x148/0x630 is_module_text_address+0x32/0xa0 __alloc_pages_nodemask+0x2f6/0x790 __alloc_pages_slowpath.constprop.0+0x12d0/0x12d0 create_prof_cpu_mask+0x30/0x30 alloc_pages_current+0xb1/0x150 stack_depot_save+0x39f/0x490 kasan_save_stack+0x42/0x50 kasan_save_stack+0x23/0x50 kasan_record_aux_stack+0xa9/0xc0 __call_rcu+0xff/0x9c0 call_rcu+0xe/0x10 put_object+0x53/0x70 __delete_object+0x7b/0x90 kmemleak_free+0x46/0x70 slab_free_freelist_hook+0xb4/0x160 kfree+0xe5/0x420 kfree_const+0x17/0x30 kobject_cleanup+0xaa/0x230 kobject_put+0x76/0x90 netdev_queue_update_kobjects+0x17d/0x1f0 ... ... ksys_write+0xd9/0x180 __x64_sys_write+0x42/0x50 do_syscall_64+0x38/0x50 entry_SYSCALL_64_after_hwframe+0x44/0xa9 Links: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/include/linux/kasan.h?id=7cb3007ce2da27ec02a1a3211941e7fe6875b642 Fixes: 84109ab58590 ("rcu: Record kvfree_call_rcu() call stack for KASAN") Fixes: 26e760c9a7c8 ("rcu: kasan: record and print call_rcu() call stack") Reported-by: Jianwei Hu <jianwei.hu@windriver.com> Reviewed-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Acked-by: Marco Elver <elver@google.com> Tested-by: Juri Lelli <juri.lelli@redhat.com> Signed-off-by: Jun Miao <jun.miao@intel.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2021-11-16 07:23:02 +08:00
kasan_record_aux_stack_noalloc(ptr);
success = add_ptr_to_bulk_krc_lock(&krcp, &flags, ptr, !head);
if (!success) {
rcu/tree: Defer kvfree_rcu() allocation to a clean context The current memmory-allocation interface causes the following difficulties for kvfree_rcu(): a) If built with CONFIG_PROVE_RAW_LOCK_NESTING, the lockdep will complain about violation of the nesting rules, as in "BUG: Invalid wait context". This Kconfig option checks for proper raw_spinlock vs. spinlock nesting, in particular, it is not legal to acquire a spinlock_t while holding a raw_spinlock_t. This is a problem because kfree_rcu() uses raw_spinlock_t whereas the "page allocator" internally deals with spinlock_t to access to its zones. The code also can be broken from higher level of view: <snip> raw_spin_lock(&some_lock); kfree_rcu(some_pointer, some_field_offset); <snip> b) If built with CONFIG_PREEMPT_RT, spinlock_t is converted into sleeplock. This means that invoking the page allocator from atomic contexts results in "BUG: scheduling while atomic". c) Please note that call_rcu() is already invoked from raw atomic context, so it is only reasonable to expaect that kfree_rcu() and kvfree_rcu() will also be called from atomic raw context. This commit therefore defers page allocation to a clean context using the combination of an hrtimer and a workqueue. The hrtimer stage is required in order to avoid deadlocks with the scheduler. This deferred allocation is required only when kvfree_rcu()'s per-CPU page cache is empty. Link: https://lore.kernel.org/lkml/20200630164543.4mdcf6zb4zfclhln@linutronix.de/ Fixes: 3042f83f19be ("rcu: Support reclaim for head-less object") Reported-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-10-30 00:50:04 +08:00
run_page_cache_worker(krcp);
if (head == NULL)
// Inline if kvfree_rcu(one_arg) call.
goto unlock_return;
head->func = ptr;
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
head->next = krcp->head;
WRITE_ONCE(krcp->head, head);
atomic_inc(&krcp->head_count);
// Take a snapshot for this krcp.
krcp->head_gp_snap = get_state_synchronize_rcu();
success = true;
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
}
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
/*
* The kvfree_rcu() caller considers the pointer freed at this point
* and likely removes any references to it. Since the actual slab
* freeing (and kmemleak_free()) is deferred, tell kmemleak to ignore
* this object (no scanning or false positives reporting).
*/
kmemleak_ignore(ptr);
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
// Set timer to drain after KFREE_DRAIN_JIFFIES.
if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING)
rcu/kvfree: Update KFREE_DRAIN_JIFFIES interval Currently the monitor work is scheduled with a fixed interval of HZ/20, which is roughly 50 milliseconds. The drawback of this approach is low utilization of the 512 page slots in scenarios with infrequence kvfree_rcu() calls. For example on an Android system: <snip> kworker/3:3-507 [003] .... 470.286305: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000d0f0dde5 nr_records=6 kworker/6:1-76 [006] .... 470.416613: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000ea0d6556 nr_records=1 kworker/6:1-76 [006] .... 470.416625: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000003e025849 nr_records=9 kworker/3:3-507 [003] .... 471.390000: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000815a8713 nr_records=48 kworker/1:1-73 [001] .... 471.725785: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000fda9bf20 nr_records=3 kworker/1:1-73 [001] .... 471.725833: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000a425b67b nr_records=76 kworker/0:4-1411 [000] .... 472.085673: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000007996be9d nr_records=1 kworker/0:4-1411 [000] .... 472.085728: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000d0f0dde5 nr_records=5 kworker/6:1-76 [006] .... 472.260340: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000065630ee4 nr_records=102 <snip> In many cases, out of 512 slots, fewer than 10 were actually used. In order to improve batching and make utilization more efficient this commit sets a drain interval to a fixed 5-seconds interval. Floods are detected when a page fills quickly, and in that case, the reclaim work is re-scheduled for the next scheduling-clock tick (jiffy). After this change: <snip> kworker/7:1-371 [007] .... 5630.725708: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000005ab0ffb3 nr_records=121 kworker/7:1-371 [007] .... 5630.989702: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000060c84761 nr_records=47 kworker/7:1-371 [007] .... 5630.989714: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000000babf308 nr_records=510 kworker/7:1-371 [007] .... 5631.553790: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000bb7bd0ef nr_records=169 kworker/7:1-371 [007] .... 5631.553808: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000044c78753 nr_records=510 kworker/5:6-9428 [005] .... 5631.746102: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000d98519aa nr_records=123 kworker/4:7-9434 [004] .... 5632.001758: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000526c9d44 nr_records=322 kworker/4:7-9434 [004] .... 5632.002073: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000002c6a8afa nr_records=185 kworker/7:1-371 [007] .... 5632.277515: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000007f4a962f nr_records=510 <snip> Here, all but one of the cases, more than one hundreds slots were used, representing an order-of-magnitude improvement. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2022-07-01 00:33:35 +08:00
schedule_delayed_monitor_work(krcp);
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
unlock_return:
krc_this_cpu_unlock(krcp, flags);
/*
* Inline kvfree() after synchronize_rcu(). We can do
* it from might_sleep() context only, so the current
* CPU can pass the QS state.
*/
if (!success) {
debug_rcu_head_unqueue((struct rcu_head *) ptr);
synchronize_rcu();
kvfree(ptr);
}
}
EXPORT_SYMBOL_GPL(kvfree_call_rcu);
static unsigned long
kfree_rcu_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
{
int cpu;
unsigned long count = 0;
/* Snapshot count of all CPUs */
for_each_possible_cpu(cpu) {
struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
count += krc_count(krcp);
count += READ_ONCE(krcp->nr_bkv_objs);
atomic_set(&krcp->backoff_page_cache_fill, 1);
}
return count == 0 ? SHRINK_EMPTY : count;
}
static unsigned long
kfree_rcu_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
{
int cpu, freed = 0;
for_each_possible_cpu(cpu) {
int count;
struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
count = krc_count(krcp);
count += drain_page_cache(krcp);
kfree_rcu_monitor(&krcp->monitor_work.work);
sc->nr_to_scan -= count;
freed += count;
if (sc->nr_to_scan <= 0)
break;
}
return freed == 0 ? SHRINK_STOP : freed;
}
static struct shrinker kfree_rcu_shrinker = {
.count_objects = kfree_rcu_shrink_count,
.scan_objects = kfree_rcu_shrink_scan,
.batch = 0,
.seeks = DEFAULT_SEEKS,
};
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
void __init kfree_rcu_scheduler_running(void)
{
int cpu;
for_each_possible_cpu(cpu) {
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
if (need_offload_krc(krcp))
rcu/kvfree: Update KFREE_DRAIN_JIFFIES interval Currently the monitor work is scheduled with a fixed interval of HZ/20, which is roughly 50 milliseconds. The drawback of this approach is low utilization of the 512 page slots in scenarios with infrequence kvfree_rcu() calls. For example on an Android system: <snip> kworker/3:3-507 [003] .... 470.286305: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000d0f0dde5 nr_records=6 kworker/6:1-76 [006] .... 470.416613: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000ea0d6556 nr_records=1 kworker/6:1-76 [006] .... 470.416625: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000003e025849 nr_records=9 kworker/3:3-507 [003] .... 471.390000: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000815a8713 nr_records=48 kworker/1:1-73 [001] .... 471.725785: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000fda9bf20 nr_records=3 kworker/1:1-73 [001] .... 471.725833: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000a425b67b nr_records=76 kworker/0:4-1411 [000] .... 472.085673: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000007996be9d nr_records=1 kworker/0:4-1411 [000] .... 472.085728: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000d0f0dde5 nr_records=5 kworker/6:1-76 [006] .... 472.260340: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000065630ee4 nr_records=102 <snip> In many cases, out of 512 slots, fewer than 10 were actually used. In order to improve batching and make utilization more efficient this commit sets a drain interval to a fixed 5-seconds interval. Floods are detected when a page fills quickly, and in that case, the reclaim work is re-scheduled for the next scheduling-clock tick (jiffy). After this change: <snip> kworker/7:1-371 [007] .... 5630.725708: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000005ab0ffb3 nr_records=121 kworker/7:1-371 [007] .... 5630.989702: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000060c84761 nr_records=47 kworker/7:1-371 [007] .... 5630.989714: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000000babf308 nr_records=510 kworker/7:1-371 [007] .... 5631.553790: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000bb7bd0ef nr_records=169 kworker/7:1-371 [007] .... 5631.553808: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000044c78753 nr_records=510 kworker/5:6-9428 [005] .... 5631.746102: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000d98519aa nr_records=123 kworker/4:7-9434 [004] .... 5632.001758: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000526c9d44 nr_records=322 kworker/4:7-9434 [004] .... 5632.002073: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000002c6a8afa nr_records=185 kworker/7:1-371 [007] .... 5632.277515: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000007f4a962f nr_records=510 <snip> Here, all but one of the cases, more than one hundreds slots were used, representing an order-of-magnitude improvement. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2022-07-01 00:33:35 +08:00
schedule_delayed_monitor_work(krcp);
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
}
}
/*
* During early boot, any blocking grace-period wait automatically
* implies a grace period.
*
* Later on, this could in theory be the case for kernels built with
* CONFIG_SMP=y && CONFIG_PREEMPTION=y running on a single CPU, but this
* is not a common case. Furthermore, this optimization would cause
* the rcu_gp_oldstate structure to expand by 50%, so this potential
* grace-period optimization is ignored once the scheduler is running.
*/
static int rcu_blocking_is_gp(void)
{
if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE) {
might_sleep();
return false;
}
return true;
}
/**
* synchronize_rcu - wait until a grace period has elapsed.
*
* Control will return to the caller some time after a full grace
* period has elapsed, in other words after all currently executing RCU
* read-side critical sections have completed. Note, however, that
* upon return from synchronize_rcu(), the caller might well be executing
* concurrently with new RCU read-side critical sections that began while
* synchronize_rcu() was waiting.
*
* RCU read-side critical sections are delimited by rcu_read_lock()
* and rcu_read_unlock(), and may be nested. In addition, but only in
* v5.0 and later, regions of code across which interrupts, preemption,
* or softirqs have been disabled also serve as RCU read-side critical
* sections. This includes hardware interrupt handlers, softirq handlers,
* and NMI handlers.
*
* Note that this guarantee implies further memory-ordering guarantees.
* On systems with more than one CPU, when synchronize_rcu() returns,
* each CPU is guaranteed to have executed a full memory barrier since
* the end of its last RCU read-side critical section whose beginning
* preceded the call to synchronize_rcu(). In addition, each CPU having
* an RCU read-side critical section that extends beyond the return from
* synchronize_rcu() is guaranteed to have executed a full memory barrier
* after the beginning of synchronize_rcu() and before the beginning of
* that RCU read-side critical section. Note that these guarantees include
* CPUs that are offline, idle, or executing in user mode, as well as CPUs
* that are executing in the kernel.
*
* Furthermore, if CPU A invoked synchronize_rcu(), which returned
* to its caller on CPU B, then both CPU A and CPU B are guaranteed
* to have executed a full memory barrier during the execution of
* synchronize_rcu() -- even if CPU A and CPU B are the same CPU (but
* again only if the system has more than one CPU).
*
* Implementation of these memory-ordering guarantees is described here:
* Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst.
*/
void synchronize_rcu(void)
{
unsigned long flags;
struct rcu_node *rnp;
RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
lock_is_held(&rcu_lock_map) ||
lock_is_held(&rcu_sched_lock_map),
"Illegal synchronize_rcu() in RCU read-side critical section");
if (!rcu_blocking_is_gp()) {
if (rcu_gp_is_expedited())
synchronize_rcu_expedited();
else
wait_rcu_gp(call_rcu_hurry);
return;
}
// Context allows vacuous grace periods.
// Note well that this code runs with !PREEMPT && !SMP.
// In addition, all code that advances grace periods runs at
// process level. Therefore, this normal GP overlaps with other
// normal GPs only by being fully nested within them, which allows
// reuse of ->gp_seq_polled_snap.
rcu_poll_gp_seq_start_unlocked(&rcu_state.gp_seq_polled_snap);
rcu_poll_gp_seq_end_unlocked(&rcu_state.gp_seq_polled_snap);
// Update the normal grace-period counters to record
// this grace period, but only those used by the boot CPU.
// The rcu_scheduler_starting() will take care of the rest of
// these counters.
local_irq_save(flags);
WARN_ON_ONCE(num_online_cpus() > 1);
rcu_state.gp_seq += (1 << RCU_SEQ_CTR_SHIFT);
for (rnp = this_cpu_ptr(&rcu_data)->mynode; rnp; rnp = rnp->parent)
rnp->gp_seq_needed = rnp->gp_seq = rcu_state.gp_seq;
local_irq_restore(flags);
}
EXPORT_SYMBOL_GPL(synchronize_rcu);
/**
* get_completed_synchronize_rcu_full - Return a full pre-completed polled state cookie
* @rgosp: Place to put state cookie
*
* Stores into @rgosp a value that will always be treated by functions
* like poll_state_synchronize_rcu_full() as a cookie whose grace period
* has already completed.
*/
void get_completed_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
{
rgosp->rgos_norm = RCU_GET_STATE_COMPLETED;
rgosp->rgos_exp = RCU_GET_STATE_COMPLETED;
}
EXPORT_SYMBOL_GPL(get_completed_synchronize_rcu_full);
/**
* get_state_synchronize_rcu - Snapshot current RCU state
*
* Returns a cookie that is used by a later call to cond_synchronize_rcu()
* or poll_state_synchronize_rcu() to determine whether or not a full
* grace period has elapsed in the meantime.
*/
unsigned long get_state_synchronize_rcu(void)
{
/*
* Any prior manipulation of RCU-protected data must happen
* before the load from ->gp_seq.
*/
smp_mb(); /* ^^^ */
return rcu_seq_snap(&rcu_state.gp_seq_polled);
}
EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
/**
* get_state_synchronize_rcu_full - Snapshot RCU state, both normal and expedited
* @rgosp: location to place combined normal/expedited grace-period state
*
* Places the normal and expedited grace-period states in @rgosp. This
* state value can be passed to a later call to cond_synchronize_rcu_full()
* or poll_state_synchronize_rcu_full() to determine whether or not a
* grace period (whether normal or expedited) has elapsed in the meantime.
* The rcu_gp_oldstate structure takes up twice the memory of an unsigned
* long, but is guaranteed to see all grace periods. In contrast, the
* combined state occupies less memory, but can sometimes fail to take
* grace periods into account.
*
* This does not guarantee that the needed grace period will actually
* start.
*/
void get_state_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
{
struct rcu_node *rnp = rcu_get_root();
/*
* Any prior manipulation of RCU-protected data must happen
* before the loads from ->gp_seq and ->expedited_sequence.
*/
smp_mb(); /* ^^^ */
rgosp->rgos_norm = rcu_seq_snap(&rnp->gp_seq);
rgosp->rgos_exp = rcu_seq_snap(&rcu_state.expedited_sequence);
}
EXPORT_SYMBOL_GPL(get_state_synchronize_rcu_full);
/*
* Helper function for start_poll_synchronize_rcu() and
* start_poll_synchronize_rcu_full().
*/
static void start_poll_synchronize_rcu_common(void)
{
unsigned long flags;
bool needwake;
struct rcu_data *rdp;
struct rcu_node *rnp;
lockdep_assert_irqs_enabled();
local_irq_save(flags);
rdp = this_cpu_ptr(&rcu_data);
rnp = rdp->mynode;
raw_spin_lock_rcu_node(rnp); // irqs already disabled.
// Note it is possible for a grace period to have elapsed between
// the above call to get_state_synchronize_rcu() and the below call
// to rcu_seq_snap. This is OK, the worst that happens is that we
// get a grace period that no one needed. These accesses are ordered
// by smp_mb(), and we are accessing them in the opposite order
// from which they are updated at grace-period start, as required.
needwake = rcu_start_this_gp(rnp, rdp, rcu_seq_snap(&rcu_state.gp_seq));
raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
if (needwake)
rcu_gp_kthread_wake();
}
/**
* start_poll_synchronize_rcu - Snapshot and start RCU grace period
*
* Returns a cookie that is used by a later call to cond_synchronize_rcu()
* or poll_state_synchronize_rcu() to determine whether or not a full
* grace period has elapsed in the meantime. If the needed grace period
* is not already slated to start, notifies RCU core of the need for that
* grace period.
*
* Interrupts must be enabled for the case where it is necessary to awaken
* the grace-period kthread.
*/
unsigned long start_poll_synchronize_rcu(void)
{
unsigned long gp_seq = get_state_synchronize_rcu();
start_poll_synchronize_rcu_common();
return gp_seq;
}
EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu);
/**
* start_poll_synchronize_rcu_full - Take a full snapshot and start RCU grace period
* @rgosp: value from get_state_synchronize_rcu_full() or start_poll_synchronize_rcu_full()
*
* Places the normal and expedited grace-period states in *@rgos. This
* state value can be passed to a later call to cond_synchronize_rcu_full()
* or poll_state_synchronize_rcu_full() to determine whether or not a
* grace period (whether normal or expedited) has elapsed in the meantime.
* If the needed grace period is not already slated to start, notifies
* RCU core of the need for that grace period.
*
* Interrupts must be enabled for the case where it is necessary to awaken
* the grace-period kthread.
*/
void start_poll_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
{
get_state_synchronize_rcu_full(rgosp);
start_poll_synchronize_rcu_common();
}
EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu_full);
/**
* poll_state_synchronize_rcu - Has the specified RCU grace period completed?
* @oldstate: value from get_state_synchronize_rcu() or start_poll_synchronize_rcu()
*
* If a full RCU grace period has elapsed since the earlier call from
* which @oldstate was obtained, return @true, otherwise return @false.
* If @false is returned, it is the caller's responsibility to invoke this
* function later on until it does return @true. Alternatively, the caller
* can explicitly wait for a grace period, for example, by passing @oldstate
* to either cond_synchronize_rcu() or cond_synchronize_rcu_expedited()
* on the one hand or by directly invoking either synchronize_rcu() or
* synchronize_rcu_expedited() on the other.
*
* Yes, this function does not take counter wrap into account.
* But counter wrap is harmless. If the counter wraps, we have waited for
* more than a billion grace periods (and way more on a 64-bit system!).
* Those needing to keep old state values for very long time periods
* (many hours even on 32-bit systems) should check them occasionally and
* either refresh them or set a flag indicating that the grace period has
* completed. Alternatively, they can use get_completed_synchronize_rcu()
* to get a guaranteed-completed grace-period state.
*
* In addition, because oldstate compresses the grace-period state for
* both normal and expedited grace periods into a single unsigned long,
* it can miss a grace period when synchronize_rcu() runs concurrently
* with synchronize_rcu_expedited(). If this is unacceptable, please
* instead use the _full() variant of these polling APIs.
*
* This function provides the same memory-ordering guarantees that
* would be provided by a synchronize_rcu() that was invoked at the call
* to the function that provided @oldstate, and that returned at the end
* of this function.
*/
bool poll_state_synchronize_rcu(unsigned long oldstate)
{
2022-04-14 06:17:25 +08:00
if (oldstate == RCU_GET_STATE_COMPLETED ||
rcu_seq_done_exact(&rcu_state.gp_seq_polled, oldstate)) {
smp_mb(); /* Ensure GP ends before subsequent accesses. */
return true;
}
return false;
}
EXPORT_SYMBOL_GPL(poll_state_synchronize_rcu);
/**
* poll_state_synchronize_rcu_full - Has the specified RCU grace period completed?
* @rgosp: value from get_state_synchronize_rcu_full() or start_poll_synchronize_rcu_full()
*
* If a full RCU grace period has elapsed since the earlier call from
* which *rgosp was obtained, return @true, otherwise return @false.
* If @false is returned, it is the caller's responsibility to invoke this
* function later on until it does return @true. Alternatively, the caller
* can explicitly wait for a grace period, for example, by passing @rgosp
* to cond_synchronize_rcu() or by directly invoking synchronize_rcu().
*
* Yes, this function does not take counter wrap into account.
* But counter wrap is harmless. If the counter wraps, we have waited
* for more than a billion grace periods (and way more on a 64-bit
* system!). Those needing to keep rcu_gp_oldstate values for very
* long time periods (many hours even on 32-bit systems) should check
* them occasionally and either refresh them or set a flag indicating
* that the grace period has completed. Alternatively, they can use
* get_completed_synchronize_rcu_full() to get a guaranteed-completed
* grace-period state.
*
* This function provides the same memory-ordering guarantees that would
* be provided by a synchronize_rcu() that was invoked at the call to
* the function that provided @rgosp, and that returned at the end of this
* function. And this guarantee requires that the root rcu_node structure's
* ->gp_seq field be checked instead of that of the rcu_state structure.
* The problem is that the just-ending grace-period's callbacks can be
* invoked between the time that the root rcu_node structure's ->gp_seq
* field is updated and the time that the rcu_state structure's ->gp_seq
* field is updated. Therefore, if a single synchronize_rcu() is to
* cause a subsequent poll_state_synchronize_rcu_full() to return @true,
* then the root rcu_node structure is the one that needs to be polled.
*/
bool poll_state_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
{
struct rcu_node *rnp = rcu_get_root();
smp_mb(); // Order against root rcu_node structure grace-period cleanup.
if (rgosp->rgos_norm == RCU_GET_STATE_COMPLETED ||
rcu_seq_done_exact(&rnp->gp_seq, rgosp->rgos_norm) ||
rgosp->rgos_exp == RCU_GET_STATE_COMPLETED ||
rcu_seq_done_exact(&rcu_state.expedited_sequence, rgosp->rgos_exp)) {
smp_mb(); /* Ensure GP ends before subsequent accesses. */
return true;
}
return false;
}
EXPORT_SYMBOL_GPL(poll_state_synchronize_rcu_full);
/**
* cond_synchronize_rcu - Conditionally wait for an RCU grace period
* @oldstate: value from get_state_synchronize_rcu(), start_poll_synchronize_rcu(), or start_poll_synchronize_rcu_expedited()
*
* If a full RCU grace period has elapsed since the earlier call to
* get_state_synchronize_rcu() or start_poll_synchronize_rcu(), just return.
* Otherwise, invoke synchronize_rcu() to wait for a full grace period.
*
* Yes, this function does not take counter wrap into account.
* But counter wrap is harmless. If the counter wraps, we have waited for
* more than 2 billion grace periods (and way more on a 64-bit system!),
* so waiting for a couple of additional grace periods should be just fine.
*
* This function provides the same memory-ordering guarantees that
* would be provided by a synchronize_rcu() that was invoked at the call
* to the function that provided @oldstate and that returned at the end
* of this function.
*/
void cond_synchronize_rcu(unsigned long oldstate)
{
if (!poll_state_synchronize_rcu(oldstate))
synchronize_rcu();
}
EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
/**
* cond_synchronize_rcu_full - Conditionally wait for an RCU grace period
* @rgosp: value from get_state_synchronize_rcu_full(), start_poll_synchronize_rcu_full(), or start_poll_synchronize_rcu_expedited_full()
*
* If a full RCU grace period has elapsed since the call to
* get_state_synchronize_rcu_full(), start_poll_synchronize_rcu_full(),
* or start_poll_synchronize_rcu_expedited_full() from which @rgosp was
* obtained, just return. Otherwise, invoke synchronize_rcu() to wait
* for a full grace period.
*
* Yes, this function does not take counter wrap into account.
* But counter wrap is harmless. If the counter wraps, we have waited for
* more than 2 billion grace periods (and way more on a 64-bit system!),
* so waiting for a couple of additional grace periods should be just fine.
*
* This function provides the same memory-ordering guarantees that
* would be provided by a synchronize_rcu() that was invoked at the call
* to the function that provided @rgosp and that returned at the end of
* this function.
*/
void cond_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
{
if (!poll_state_synchronize_rcu_full(rgosp))
synchronize_rcu();
}
EXPORT_SYMBOL_GPL(cond_synchronize_rcu_full);
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/*
* Check to see if there is any immediate RCU-related work to be done by
* the current CPU, returning 1 if so and zero otherwise. The checks are
* in order of increasing expense: checks that can be carried out against
* CPU-local state are performed first. However, we must check for CPU
* stalls first, else we might not get a chance.
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
*/
static int rcu_pending(int user)
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
{
bool gp_in_progress;
struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
struct rcu_node *rnp = rdp->mynode;
lockdep_assert_irqs_disabled();
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/* Check for CPU stalls, if enabled. */
check_cpu_stall(rdp);
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/* Does this CPU need a deferred NOCB wakeup? */
if (rcu_nocb_need_deferred_wakeup(rdp, RCU_NOCB_WAKE))
return 1;
/* Is this a nohz_full CPU in userspace or idle? (Ignore RCU if so.) */
if ((user || rcu_is_cpu_rrupt_from_idle()) && rcu_nohz_full_cpu())
return 0;
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/* Is the RCU core waiting for a quiescent state from this CPU? */
gp_in_progress = rcu_gp_in_progress();
if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm && gp_in_progress)
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
return 1;
/* Does this CPU have callbacks ready to invoke? */
if (!rcu_rdp_is_offloaded(rdp) &&
rcu/tree: nocb: Avoid raising softirq for offloaded ready-to-execute CBs Testing showed that rcu_pending() can return 1 when offloaded callbacks are ready to execute. This invokes RCU core processing, for example, by raising RCU_SOFTIRQ, eventually resulting in a call to rcu_core(). However, rcu_core() explicitly avoids in any way manipulating offloaded callbacks, which are instead handled by the rcuog and rcuoc kthreads, which work independently of rcu_core(). One exception to this independence is that rcu_core() invokes do_nocb_deferred_wakeup(), however, rcu_pending() also checks rcu_nocb_need_deferred_wakeup() in order to correctly handle this case, invoking rcu_core() when needed. This commit therefore avoids needlessly invoking RCU core processing by checking rcu_segcblist_ready_cbs() only on non-offloaded CPUs. This reduces overhead, for example, by reducing softirq activity. This change passed 30 minute tests of TREE01 through TREE09 each. On TREE08, there is at most 150us from the time that rcu_pending() chose not to invoke RCU core processing to the time when the ready callbacks were invoked by the rcuoc kthread. This provides further evidence that there is no need to invoke rcu_core() for offloaded callbacks that are ready to invoke. Cc: Neeraj Upadhyay <neeraju@codeaurora.org> Reviewed-by: Frederic Weisbecker <frederic@kernel.org> Reviewed-by: Neeraj Upadhyay <neeraju@codeaurora.org> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-10-08 04:50:36 +08:00
rcu_segcblist_ready_cbs(&rdp->cblist))
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
return 1;
/* Has RCU gone idle with this CPU needing another grace period? */
if (!gp_in_progress && rcu_segcblist_is_enabled(&rdp->cblist) &&
!rcu_rdp_is_offloaded(rdp) &&
!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
return 1;
/* Have RCU grace period completed or started? */
if (rcu_seq_current(&rnp->gp_seq) != rdp->gp_seq ||
unlikely(READ_ONCE(rdp->gpwrap))) /* outside lock */
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
return 1;
/* nothing to do */
return 0;
}
/*
* Helper function for rcu_barrier() tracing. If tracing is disabled,
* the compiler is expected to optimize this away.
*/
static void rcu_barrier_trace(const char *s, int cpu, unsigned long done)
{
trace_rcu_barrier(rcu_state.name, s, cpu,
atomic_read(&rcu_state.barrier_cpu_count), done);
}
/*
* RCU callback function for rcu_barrier(). If we are last, wake
* up the task executing rcu_barrier().
rcu: Fix rcu_barrier_callback() race condition The rcu_barrier_callback() function does an atomic_dec_and_test(), and if it is the last CPU to check in, does the required wakeup. Either way, it does an event trace. Unfortunately, this is susceptible to the following sequence of events: o CPU 0 invokes rcu_barrier_callback(), but atomic_dec_and_test() says that it is not last. But at this point, CPU 0 is delayed, perhaps due to an NMI, SMI, or vCPU preemption. o CPU 1 invokes rcu_barrier_callback(), and atomic_dec_and_test() says that it is last. So CPU 1 traces completion and does the needed wakeup. o The awakened rcu_barrier() function does cleanup and releases rcu_state.barrier_mutex. o Another CPU now acquires rcu_state.barrier_mutex and starts another round of rcu_barrier() processing, including updating rcu_state.barrier_sequence. o CPU 0 gets its act back together and does its tracing. Except that rcu_state.barrier_sequence has already been updated, so its tracing is incorrect and probably quite confusing. (Wait! Why did this CPU check in twice for one rcu_barrier() invocation???) This commit therefore causes rcu_barrier_callback() to take a snapshot of the value of rcu_state.barrier_sequence before invoking atomic_dec_and_test(), thus guaranteeing that the event-trace output is sensible, even if the timing of the event-trace output might still be confusing. (Wait! Why did the old rcu_barrier() complete before all of its CPUs checked in???) But being that this is RCU, only so much confusion can reasonably be eliminated. This data race was reported by KCSAN. Not appropriate for backporting due to failure being unlikely and due to the mild consequences of the failure, namely a confusing event trace. Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-21 07:43:45 +08:00
*
* Note that the value of rcu_state.barrier_sequence must be captured
* before the atomic_dec_and_test(). Otherwise, if this CPU is not last,
* other CPUs might count the value down to zero before this CPU gets
* around to invoking rcu_barrier_trace(), which might result in bogus
* data from the next instance of rcu_barrier().
*/
static void rcu_barrier_callback(struct rcu_head *rhp)
{
rcu: Fix rcu_barrier_callback() race condition The rcu_barrier_callback() function does an atomic_dec_and_test(), and if it is the last CPU to check in, does the required wakeup. Either way, it does an event trace. Unfortunately, this is susceptible to the following sequence of events: o CPU 0 invokes rcu_barrier_callback(), but atomic_dec_and_test() says that it is not last. But at this point, CPU 0 is delayed, perhaps due to an NMI, SMI, or vCPU preemption. o CPU 1 invokes rcu_barrier_callback(), and atomic_dec_and_test() says that it is last. So CPU 1 traces completion and does the needed wakeup. o The awakened rcu_barrier() function does cleanup and releases rcu_state.barrier_mutex. o Another CPU now acquires rcu_state.barrier_mutex and starts another round of rcu_barrier() processing, including updating rcu_state.barrier_sequence. o CPU 0 gets its act back together and does its tracing. Except that rcu_state.barrier_sequence has already been updated, so its tracing is incorrect and probably quite confusing. (Wait! Why did this CPU check in twice for one rcu_barrier() invocation???) This commit therefore causes rcu_barrier_callback() to take a snapshot of the value of rcu_state.barrier_sequence before invoking atomic_dec_and_test(), thus guaranteeing that the event-trace output is sensible, even if the timing of the event-trace output might still be confusing. (Wait! Why did the old rcu_barrier() complete before all of its CPUs checked in???) But being that this is RCU, only so much confusion can reasonably be eliminated. This data race was reported by KCSAN. Not appropriate for backporting due to failure being unlikely and due to the mild consequences of the failure, namely a confusing event trace. Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-21 07:43:45 +08:00
unsigned long __maybe_unused s = rcu_state.barrier_sequence;
if (atomic_dec_and_test(&rcu_state.barrier_cpu_count)) {
rcu: Fix rcu_barrier_callback() race condition The rcu_barrier_callback() function does an atomic_dec_and_test(), and if it is the last CPU to check in, does the required wakeup. Either way, it does an event trace. Unfortunately, this is susceptible to the following sequence of events: o CPU 0 invokes rcu_barrier_callback(), but atomic_dec_and_test() says that it is not last. But at this point, CPU 0 is delayed, perhaps due to an NMI, SMI, or vCPU preemption. o CPU 1 invokes rcu_barrier_callback(), and atomic_dec_and_test() says that it is last. So CPU 1 traces completion and does the needed wakeup. o The awakened rcu_barrier() function does cleanup and releases rcu_state.barrier_mutex. o Another CPU now acquires rcu_state.barrier_mutex and starts another round of rcu_barrier() processing, including updating rcu_state.barrier_sequence. o CPU 0 gets its act back together and does its tracing. Except that rcu_state.barrier_sequence has already been updated, so its tracing is incorrect and probably quite confusing. (Wait! Why did this CPU check in twice for one rcu_barrier() invocation???) This commit therefore causes rcu_barrier_callback() to take a snapshot of the value of rcu_state.barrier_sequence before invoking atomic_dec_and_test(), thus guaranteeing that the event-trace output is sensible, even if the timing of the event-trace output might still be confusing. (Wait! Why did the old rcu_barrier() complete before all of its CPUs checked in???) But being that this is RCU, only so much confusion can reasonably be eliminated. This data race was reported by KCSAN. Not appropriate for backporting due to failure being unlikely and due to the mild consequences of the failure, namely a confusing event trace. Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-21 07:43:45 +08:00
rcu_barrier_trace(TPS("LastCB"), -1, s);
complete(&rcu_state.barrier_completion);
} else {
rcu: Fix rcu_barrier_callback() race condition The rcu_barrier_callback() function does an atomic_dec_and_test(), and if it is the last CPU to check in, does the required wakeup. Either way, it does an event trace. Unfortunately, this is susceptible to the following sequence of events: o CPU 0 invokes rcu_barrier_callback(), but atomic_dec_and_test() says that it is not last. But at this point, CPU 0 is delayed, perhaps due to an NMI, SMI, or vCPU preemption. o CPU 1 invokes rcu_barrier_callback(), and atomic_dec_and_test() says that it is last. So CPU 1 traces completion and does the needed wakeup. o The awakened rcu_barrier() function does cleanup and releases rcu_state.barrier_mutex. o Another CPU now acquires rcu_state.barrier_mutex and starts another round of rcu_barrier() processing, including updating rcu_state.barrier_sequence. o CPU 0 gets its act back together and does its tracing. Except that rcu_state.barrier_sequence has already been updated, so its tracing is incorrect and probably quite confusing. (Wait! Why did this CPU check in twice for one rcu_barrier() invocation???) This commit therefore causes rcu_barrier_callback() to take a snapshot of the value of rcu_state.barrier_sequence before invoking atomic_dec_and_test(), thus guaranteeing that the event-trace output is sensible, even if the timing of the event-trace output might still be confusing. (Wait! Why did the old rcu_barrier() complete before all of its CPUs checked in???) But being that this is RCU, only so much confusion can reasonably be eliminated. This data race was reported by KCSAN. Not appropriate for backporting due to failure being unlikely and due to the mild consequences of the failure, namely a confusing event trace. Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-21 07:43:45 +08:00
rcu_barrier_trace(TPS("CB"), -1, s);
}
}
/*
* If needed, entrain an rcu_barrier() callback on rdp->cblist.
*/
static void rcu_barrier_entrain(struct rcu_data *rdp)
{
unsigned long gseq = READ_ONCE(rcu_state.barrier_sequence);
unsigned long lseq = READ_ONCE(rdp->barrier_seq_snap);
bool wake_nocb = false;
bool was_alldone = false;
rcu: Make rcu_barrier() no longer block CPU-hotplug operations This commit removes the cpus_read_lock() and cpus_read_unlock() calls from rcu_barrier(), thus allowing CPUs to come and go during the course of rcu_barrier() execution. Posting of the ->barrier_head callbacks does synchronize with portions of RCU's CPU-hotplug notifiers, but these locks are held for short time periods on both sides. Thus, full CPU-hotplug operations could both start and finish during the execution of a given rcu_barrier() invocation. Additional synchronization is provided by a global ->barrier_lock. Since the ->barrier_lock is only used during rcu_barrier() execution and during onlining/offlining a CPU, the contention for this lock should be low. It might be tempting to make use of a per-CPU lock just on general principles, but straightforward attempts to do this have the problems shown below. Initial state: 3 CPUs present, CPU 0 and CPU1 do not have any callback and CPU2 has callbacks. 1. CPU0 calls rcu_barrier(). 2. CPU1 starts offlining for CPU2. CPU1 calls rcutree_migrate_callbacks(). rcu_barrier_entrain() is called from rcutree_migrate_callbacks(), with CPU2's rdp->barrier_lock. It does not entrain ->barrier_head for CPU2, as rcu_barrier() on CPU0 hasn't started the barrier sequence (by calling rcu_seq_start(&rcu_state.barrier_sequence)) yet. 3. CPU0 starts new barrier sequence. It iterates over CPU0 and CPU1, after acquiring their per-cpu ->barrier_lock and finds 0 segcblist length. It updates ->barrier_seq_snap for CPU0 and CPU1 and continues loop iteration to CPU2. for_each_possible_cpu(cpu) { raw_spin_lock_irqsave(&rdp->barrier_lock, flags); if (!rcu_segcblist_n_cbs(&rdp->cblist)) { WRITE_ONCE(rdp->barrier_seq_snap, gseq); raw_spin_unlock_irqrestore(&rdp->barrier_lock, flags); rcu_barrier_trace(TPS("NQ"), cpu, rcu_state.barrier_sequence); continue; } 4. rcutree_migrate_callbacks() completes execution on CPU1. Segcblist len for CPU2 becomes 0. 5. The loop iteration on CPU0, checks rcu_segcblist_n_cbs(&rdp->cblist) for CPU2 and completes the loop iteration after setting ->barrier_seq_snap. 6. As there isn't any ->barrier_head callback entrained; at this point, rcu_barrier() in CPU0 returns. 7. The callbacks, which migrated from CPU2 to CPU1, execute. Straightforward per-CPU locking is also subject to the following race condition noted by Boqun Feng: 1. CPU0 calls rcu_barrier(), starting a new barrier sequence by invoking rcu_seq_start() and init_completion(), but does not yet initialize rcu_state.barrier_cpu_count. 2. CPU1 starts offlining for CPU2, calling rcutree_migrate_callbacks(), which in turn calls rcu_barrier_entrain() holding CPU2's. rdp->barrier_lock. It then entrains ->barrier_head for CPU2 and atomically increments rcu_state.barrier_cpu_count, which is unfortunately not yet initialized to the value 2. 3. The just-entrained RCU callback is invoked. It atomically decrements rcu_state.barrier_cpu_count and sees that it is now zero. This callback therefore invokes complete(). 4. CPU0 continues executing rcu_barrier(), but is not blocked by its call to wait_for_completion(). This results in rcu_barrier() returning before all pre-existing callbacks have been invoked, which is a bug. Therefore, synchronization is provided by rcu_state.barrier_lock, which is also held across the initialization sequence, especially the rcu_seq_start() and the atomic_set() that sets rcu_state.barrier_cpu_count to the value 2. In addition, this lock is held when entraining the rcu_barrier() callback, when deciding whether or not a CPU has callbacks that rcu_barrier() must wait on, when setting the ->qsmaskinitnext for incoming CPUs, and when migrating callbacks from a CPU that is going offline. Reviewed-by: Frederic Weisbecker <frederic@kernel.org> Co-developed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com> Signed-off-by: Neeraj Upadhyay <quic_neeraju@quicinc.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2021-12-15 05:35:17 +08:00
lockdep_assert_held(&rcu_state.barrier_lock);
if (rcu_seq_state(lseq) || !rcu_seq_state(gseq) || rcu_seq_ctr(lseq) != rcu_seq_ctr(gseq))
return;
rcu_barrier_trace(TPS("IRQ"), -1, rcu_state.barrier_sequence);
rdp->barrier_head.func = rcu_barrier_callback;
debug_rcu_head_queue(&rdp->barrier_head);
rcu_nocb_lock(rdp);
/*
* Flush bypass and wakeup rcuog if we add callbacks to an empty regular
* queue. This way we don't wait for bypass timer that can reach seconds
* if it's fully lazy.
*/
was_alldone = rcu_rdp_is_offloaded(rdp) && !rcu_segcblist_pend_cbs(&rdp->cblist);
WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies, false));
wake_nocb = was_alldone && rcu_segcblist_pend_cbs(&rdp->cblist);
if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head)) {
atomic_inc(&rcu_state.barrier_cpu_count);
} else {
debug_rcu_head_unqueue(&rdp->barrier_head);
rcu_barrier_trace(TPS("IRQNQ"), -1, rcu_state.barrier_sequence);
}
rcu_nocb_unlock(rdp);
if (wake_nocb)
wake_nocb_gp(rdp, false);
smp_store_release(&rdp->barrier_seq_snap, gseq);
}
/*
* Called with preemption disabled, and from cross-cpu IRQ context.
*/
static void rcu_barrier_handler(void *cpu_in)
{
uintptr_t cpu = (uintptr_t)cpu_in;
struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
lockdep_assert_irqs_disabled();
WARN_ON_ONCE(cpu != rdp->cpu);
WARN_ON_ONCE(cpu != smp_processor_id());
rcu: Make rcu_barrier() no longer block CPU-hotplug operations This commit removes the cpus_read_lock() and cpus_read_unlock() calls from rcu_barrier(), thus allowing CPUs to come and go during the course of rcu_barrier() execution. Posting of the ->barrier_head callbacks does synchronize with portions of RCU's CPU-hotplug notifiers, but these locks are held for short time periods on both sides. Thus, full CPU-hotplug operations could both start and finish during the execution of a given rcu_barrier() invocation. Additional synchronization is provided by a global ->barrier_lock. Since the ->barrier_lock is only used during rcu_barrier() execution and during onlining/offlining a CPU, the contention for this lock should be low. It might be tempting to make use of a per-CPU lock just on general principles, but straightforward attempts to do this have the problems shown below. Initial state: 3 CPUs present, CPU 0 and CPU1 do not have any callback and CPU2 has callbacks. 1. CPU0 calls rcu_barrier(). 2. CPU1 starts offlining for CPU2. CPU1 calls rcutree_migrate_callbacks(). rcu_barrier_entrain() is called from rcutree_migrate_callbacks(), with CPU2's rdp->barrier_lock. It does not entrain ->barrier_head for CPU2, as rcu_barrier() on CPU0 hasn't started the barrier sequence (by calling rcu_seq_start(&rcu_state.barrier_sequence)) yet. 3. CPU0 starts new barrier sequence. It iterates over CPU0 and CPU1, after acquiring their per-cpu ->barrier_lock and finds 0 segcblist length. It updates ->barrier_seq_snap for CPU0 and CPU1 and continues loop iteration to CPU2. for_each_possible_cpu(cpu) { raw_spin_lock_irqsave(&rdp->barrier_lock, flags); if (!rcu_segcblist_n_cbs(&rdp->cblist)) { WRITE_ONCE(rdp->barrier_seq_snap, gseq); raw_spin_unlock_irqrestore(&rdp->barrier_lock, flags); rcu_barrier_trace(TPS("NQ"), cpu, rcu_state.barrier_sequence); continue; } 4. rcutree_migrate_callbacks() completes execution on CPU1. Segcblist len for CPU2 becomes 0. 5. The loop iteration on CPU0, checks rcu_segcblist_n_cbs(&rdp->cblist) for CPU2 and completes the loop iteration after setting ->barrier_seq_snap. 6. As there isn't any ->barrier_head callback entrained; at this point, rcu_barrier() in CPU0 returns. 7. The callbacks, which migrated from CPU2 to CPU1, execute. Straightforward per-CPU locking is also subject to the following race condition noted by Boqun Feng: 1. CPU0 calls rcu_barrier(), starting a new barrier sequence by invoking rcu_seq_start() and init_completion(), but does not yet initialize rcu_state.barrier_cpu_count. 2. CPU1 starts offlining for CPU2, calling rcutree_migrate_callbacks(), which in turn calls rcu_barrier_entrain() holding CPU2's. rdp->barrier_lock. It then entrains ->barrier_head for CPU2 and atomically increments rcu_state.barrier_cpu_count, which is unfortunately not yet initialized to the value 2. 3. The just-entrained RCU callback is invoked. It atomically decrements rcu_state.barrier_cpu_count and sees that it is now zero. This callback therefore invokes complete(). 4. CPU0 continues executing rcu_barrier(), but is not blocked by its call to wait_for_completion(). This results in rcu_barrier() returning before all pre-existing callbacks have been invoked, which is a bug. Therefore, synchronization is provided by rcu_state.barrier_lock, which is also held across the initialization sequence, especially the rcu_seq_start() and the atomic_set() that sets rcu_state.barrier_cpu_count to the value 2. In addition, this lock is held when entraining the rcu_barrier() callback, when deciding whether or not a CPU has callbacks that rcu_barrier() must wait on, when setting the ->qsmaskinitnext for incoming CPUs, and when migrating callbacks from a CPU that is going offline. Reviewed-by: Frederic Weisbecker <frederic@kernel.org> Co-developed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com> Signed-off-by: Neeraj Upadhyay <quic_neeraju@quicinc.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2021-12-15 05:35:17 +08:00
raw_spin_lock(&rcu_state.barrier_lock);
rcu_barrier_entrain(rdp);
rcu: Make rcu_barrier() no longer block CPU-hotplug operations This commit removes the cpus_read_lock() and cpus_read_unlock() calls from rcu_barrier(), thus allowing CPUs to come and go during the course of rcu_barrier() execution. Posting of the ->barrier_head callbacks does synchronize with portions of RCU's CPU-hotplug notifiers, but these locks are held for short time periods on both sides. Thus, full CPU-hotplug operations could both start and finish during the execution of a given rcu_barrier() invocation. Additional synchronization is provided by a global ->barrier_lock. Since the ->barrier_lock is only used during rcu_barrier() execution and during onlining/offlining a CPU, the contention for this lock should be low. It might be tempting to make use of a per-CPU lock just on general principles, but straightforward attempts to do this have the problems shown below. Initial state: 3 CPUs present, CPU 0 and CPU1 do not have any callback and CPU2 has callbacks. 1. CPU0 calls rcu_barrier(). 2. CPU1 starts offlining for CPU2. CPU1 calls rcutree_migrate_callbacks(). rcu_barrier_entrain() is called from rcutree_migrate_callbacks(), with CPU2's rdp->barrier_lock. It does not entrain ->barrier_head for CPU2, as rcu_barrier() on CPU0 hasn't started the barrier sequence (by calling rcu_seq_start(&rcu_state.barrier_sequence)) yet. 3. CPU0 starts new barrier sequence. It iterates over CPU0 and CPU1, after acquiring their per-cpu ->barrier_lock and finds 0 segcblist length. It updates ->barrier_seq_snap for CPU0 and CPU1 and continues loop iteration to CPU2. for_each_possible_cpu(cpu) { raw_spin_lock_irqsave(&rdp->barrier_lock, flags); if (!rcu_segcblist_n_cbs(&rdp->cblist)) { WRITE_ONCE(rdp->barrier_seq_snap, gseq); raw_spin_unlock_irqrestore(&rdp->barrier_lock, flags); rcu_barrier_trace(TPS("NQ"), cpu, rcu_state.barrier_sequence); continue; } 4. rcutree_migrate_callbacks() completes execution on CPU1. Segcblist len for CPU2 becomes 0. 5. The loop iteration on CPU0, checks rcu_segcblist_n_cbs(&rdp->cblist) for CPU2 and completes the loop iteration after setting ->barrier_seq_snap. 6. As there isn't any ->barrier_head callback entrained; at this point, rcu_barrier() in CPU0 returns. 7. The callbacks, which migrated from CPU2 to CPU1, execute. Straightforward per-CPU locking is also subject to the following race condition noted by Boqun Feng: 1. CPU0 calls rcu_barrier(), starting a new barrier sequence by invoking rcu_seq_start() and init_completion(), but does not yet initialize rcu_state.barrier_cpu_count. 2. CPU1 starts offlining for CPU2, calling rcutree_migrate_callbacks(), which in turn calls rcu_barrier_entrain() holding CPU2's. rdp->barrier_lock. It then entrains ->barrier_head for CPU2 and atomically increments rcu_state.barrier_cpu_count, which is unfortunately not yet initialized to the value 2. 3. The just-entrained RCU callback is invoked. It atomically decrements rcu_state.barrier_cpu_count and sees that it is now zero. This callback therefore invokes complete(). 4. CPU0 continues executing rcu_barrier(), but is not blocked by its call to wait_for_completion(). This results in rcu_barrier() returning before all pre-existing callbacks have been invoked, which is a bug. Therefore, synchronization is provided by rcu_state.barrier_lock, which is also held across the initialization sequence, especially the rcu_seq_start() and the atomic_set() that sets rcu_state.barrier_cpu_count to the value 2. In addition, this lock is held when entraining the rcu_barrier() callback, when deciding whether or not a CPU has callbacks that rcu_barrier() must wait on, when setting the ->qsmaskinitnext for incoming CPUs, and when migrating callbacks from a CPU that is going offline. Reviewed-by: Frederic Weisbecker <frederic@kernel.org> Co-developed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com> Signed-off-by: Neeraj Upadhyay <quic_neeraju@quicinc.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2021-12-15 05:35:17 +08:00
raw_spin_unlock(&rcu_state.barrier_lock);
}
/**
* rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
*
* Note that this primitive does not necessarily wait for an RCU grace period
* to complete. For example, if there are no RCU callbacks queued anywhere
* in the system, then rcu_barrier() is within its rights to return
* immediately, without waiting for anything, much less an RCU grace period.
*/
void rcu_barrier(void)
{
uintptr_t cpu;
unsigned long flags;
unsigned long gseq;
struct rcu_data *rdp;
unsigned long s = rcu_seq_snap(&rcu_state.barrier_sequence);
rcu_barrier_trace(TPS("Begin"), -1, s);
/* Take mutex to serialize concurrent rcu_barrier() requests. */
mutex_lock(&rcu_state.barrier_mutex);
/* Did someone else do our work for us? */
if (rcu_seq_done(&rcu_state.barrier_sequence, s)) {
rcu_barrier_trace(TPS("EarlyExit"), -1, rcu_state.barrier_sequence);
smp_mb(); /* caller's subsequent code after above check. */
mutex_unlock(&rcu_state.barrier_mutex);
return;
}
/* Mark the start of the barrier operation. */
rcu: Make rcu_barrier() no longer block CPU-hotplug operations This commit removes the cpus_read_lock() and cpus_read_unlock() calls from rcu_barrier(), thus allowing CPUs to come and go during the course of rcu_barrier() execution. Posting of the ->barrier_head callbacks does synchronize with portions of RCU's CPU-hotplug notifiers, but these locks are held for short time periods on both sides. Thus, full CPU-hotplug operations could both start and finish during the execution of a given rcu_barrier() invocation. Additional synchronization is provided by a global ->barrier_lock. Since the ->barrier_lock is only used during rcu_barrier() execution and during onlining/offlining a CPU, the contention for this lock should be low. It might be tempting to make use of a per-CPU lock just on general principles, but straightforward attempts to do this have the problems shown below. Initial state: 3 CPUs present, CPU 0 and CPU1 do not have any callback and CPU2 has callbacks. 1. CPU0 calls rcu_barrier(). 2. CPU1 starts offlining for CPU2. CPU1 calls rcutree_migrate_callbacks(). rcu_barrier_entrain() is called from rcutree_migrate_callbacks(), with CPU2's rdp->barrier_lock. It does not entrain ->barrier_head for CPU2, as rcu_barrier() on CPU0 hasn't started the barrier sequence (by calling rcu_seq_start(&rcu_state.barrier_sequence)) yet. 3. CPU0 starts new barrier sequence. It iterates over CPU0 and CPU1, after acquiring their per-cpu ->barrier_lock and finds 0 segcblist length. It updates ->barrier_seq_snap for CPU0 and CPU1 and continues loop iteration to CPU2. for_each_possible_cpu(cpu) { raw_spin_lock_irqsave(&rdp->barrier_lock, flags); if (!rcu_segcblist_n_cbs(&rdp->cblist)) { WRITE_ONCE(rdp->barrier_seq_snap, gseq); raw_spin_unlock_irqrestore(&rdp->barrier_lock, flags); rcu_barrier_trace(TPS("NQ"), cpu, rcu_state.barrier_sequence); continue; } 4. rcutree_migrate_callbacks() completes execution on CPU1. Segcblist len for CPU2 becomes 0. 5. The loop iteration on CPU0, checks rcu_segcblist_n_cbs(&rdp->cblist) for CPU2 and completes the loop iteration after setting ->barrier_seq_snap. 6. As there isn't any ->barrier_head callback entrained; at this point, rcu_barrier() in CPU0 returns. 7. The callbacks, which migrated from CPU2 to CPU1, execute. Straightforward per-CPU locking is also subject to the following race condition noted by Boqun Feng: 1. CPU0 calls rcu_barrier(), starting a new barrier sequence by invoking rcu_seq_start() and init_completion(), but does not yet initialize rcu_state.barrier_cpu_count. 2. CPU1 starts offlining for CPU2, calling rcutree_migrate_callbacks(), which in turn calls rcu_barrier_entrain() holding CPU2's. rdp->barrier_lock. It then entrains ->barrier_head for CPU2 and atomically increments rcu_state.barrier_cpu_count, which is unfortunately not yet initialized to the value 2. 3. The just-entrained RCU callback is invoked. It atomically decrements rcu_state.barrier_cpu_count and sees that it is now zero. This callback therefore invokes complete(). 4. CPU0 continues executing rcu_barrier(), but is not blocked by its call to wait_for_completion(). This results in rcu_barrier() returning before all pre-existing callbacks have been invoked, which is a bug. Therefore, synchronization is provided by rcu_state.barrier_lock, which is also held across the initialization sequence, especially the rcu_seq_start() and the atomic_set() that sets rcu_state.barrier_cpu_count to the value 2. In addition, this lock is held when entraining the rcu_barrier() callback, when deciding whether or not a CPU has callbacks that rcu_barrier() must wait on, when setting the ->qsmaskinitnext for incoming CPUs, and when migrating callbacks from a CPU that is going offline. Reviewed-by: Frederic Weisbecker <frederic@kernel.org> Co-developed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com> Signed-off-by: Neeraj Upadhyay <quic_neeraju@quicinc.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2021-12-15 05:35:17 +08:00
raw_spin_lock_irqsave(&rcu_state.barrier_lock, flags);
rcu_seq_start(&rcu_state.barrier_sequence);
gseq = rcu_state.barrier_sequence;
rcu_barrier_trace(TPS("Inc1"), -1, rcu_state.barrier_sequence);
/*
* Initialize the count to two rather than to zero in order
* to avoid a too-soon return to zero in case of an immediate
* invocation of the just-enqueued callback (or preemption of
* this task). Exclude CPU-hotplug operations to ensure that no
* offline non-offloaded CPU has callbacks queued.
*/
init_completion(&rcu_state.barrier_completion);
atomic_set(&rcu_state.barrier_cpu_count, 2);
rcu: Make rcu_barrier() no longer block CPU-hotplug operations This commit removes the cpus_read_lock() and cpus_read_unlock() calls from rcu_barrier(), thus allowing CPUs to come and go during the course of rcu_barrier() execution. Posting of the ->barrier_head callbacks does synchronize with portions of RCU's CPU-hotplug notifiers, but these locks are held for short time periods on both sides. Thus, full CPU-hotplug operations could both start and finish during the execution of a given rcu_barrier() invocation. Additional synchronization is provided by a global ->barrier_lock. Since the ->barrier_lock is only used during rcu_barrier() execution and during onlining/offlining a CPU, the contention for this lock should be low. It might be tempting to make use of a per-CPU lock just on general principles, but straightforward attempts to do this have the problems shown below. Initial state: 3 CPUs present, CPU 0 and CPU1 do not have any callback and CPU2 has callbacks. 1. CPU0 calls rcu_barrier(). 2. CPU1 starts offlining for CPU2. CPU1 calls rcutree_migrate_callbacks(). rcu_barrier_entrain() is called from rcutree_migrate_callbacks(), with CPU2's rdp->barrier_lock. It does not entrain ->barrier_head for CPU2, as rcu_barrier() on CPU0 hasn't started the barrier sequence (by calling rcu_seq_start(&rcu_state.barrier_sequence)) yet. 3. CPU0 starts new barrier sequence. It iterates over CPU0 and CPU1, after acquiring their per-cpu ->barrier_lock and finds 0 segcblist length. It updates ->barrier_seq_snap for CPU0 and CPU1 and continues loop iteration to CPU2. for_each_possible_cpu(cpu) { raw_spin_lock_irqsave(&rdp->barrier_lock, flags); if (!rcu_segcblist_n_cbs(&rdp->cblist)) { WRITE_ONCE(rdp->barrier_seq_snap, gseq); raw_spin_unlock_irqrestore(&rdp->barrier_lock, flags); rcu_barrier_trace(TPS("NQ"), cpu, rcu_state.barrier_sequence); continue; } 4. rcutree_migrate_callbacks() completes execution on CPU1. Segcblist len for CPU2 becomes 0. 5. The loop iteration on CPU0, checks rcu_segcblist_n_cbs(&rdp->cblist) for CPU2 and completes the loop iteration after setting ->barrier_seq_snap. 6. As there isn't any ->barrier_head callback entrained; at this point, rcu_barrier() in CPU0 returns. 7. The callbacks, which migrated from CPU2 to CPU1, execute. Straightforward per-CPU locking is also subject to the following race condition noted by Boqun Feng: 1. CPU0 calls rcu_barrier(), starting a new barrier sequence by invoking rcu_seq_start() and init_completion(), but does not yet initialize rcu_state.barrier_cpu_count. 2. CPU1 starts offlining for CPU2, calling rcutree_migrate_callbacks(), which in turn calls rcu_barrier_entrain() holding CPU2's. rdp->barrier_lock. It then entrains ->barrier_head for CPU2 and atomically increments rcu_state.barrier_cpu_count, which is unfortunately not yet initialized to the value 2. 3. The just-entrained RCU callback is invoked. It atomically decrements rcu_state.barrier_cpu_count and sees that it is now zero. This callback therefore invokes complete(). 4. CPU0 continues executing rcu_barrier(), but is not blocked by its call to wait_for_completion(). This results in rcu_barrier() returning before all pre-existing callbacks have been invoked, which is a bug. Therefore, synchronization is provided by rcu_state.barrier_lock, which is also held across the initialization sequence, especially the rcu_seq_start() and the atomic_set() that sets rcu_state.barrier_cpu_count to the value 2. In addition, this lock is held when entraining the rcu_barrier() callback, when deciding whether or not a CPU has callbacks that rcu_barrier() must wait on, when setting the ->qsmaskinitnext for incoming CPUs, and when migrating callbacks from a CPU that is going offline. Reviewed-by: Frederic Weisbecker <frederic@kernel.org> Co-developed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com> Signed-off-by: Neeraj Upadhyay <quic_neeraju@quicinc.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2021-12-15 05:35:17 +08:00
raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags);
/*
* Force each CPU with callbacks to register a new callback.
* When that callback is invoked, we will know that all of the
* corresponding CPU's preceding callbacks have been invoked.
*/
for_each_possible_cpu(cpu) {
rdp = per_cpu_ptr(&rcu_data, cpu);
retry:
if (smp_load_acquire(&rdp->barrier_seq_snap) == gseq)
continue;
rcu: Make rcu_barrier() no longer block CPU-hotplug operations This commit removes the cpus_read_lock() and cpus_read_unlock() calls from rcu_barrier(), thus allowing CPUs to come and go during the course of rcu_barrier() execution. Posting of the ->barrier_head callbacks does synchronize with portions of RCU's CPU-hotplug notifiers, but these locks are held for short time periods on both sides. Thus, full CPU-hotplug operations could both start and finish during the execution of a given rcu_barrier() invocation. Additional synchronization is provided by a global ->barrier_lock. Since the ->barrier_lock is only used during rcu_barrier() execution and during onlining/offlining a CPU, the contention for this lock should be low. It might be tempting to make use of a per-CPU lock just on general principles, but straightforward attempts to do this have the problems shown below. Initial state: 3 CPUs present, CPU 0 and CPU1 do not have any callback and CPU2 has callbacks. 1. CPU0 calls rcu_barrier(). 2. CPU1 starts offlining for CPU2. CPU1 calls rcutree_migrate_callbacks(). rcu_barrier_entrain() is called from rcutree_migrate_callbacks(), with CPU2's rdp->barrier_lock. It does not entrain ->barrier_head for CPU2, as rcu_barrier() on CPU0 hasn't started the barrier sequence (by calling rcu_seq_start(&rcu_state.barrier_sequence)) yet. 3. CPU0 starts new barrier sequence. It iterates over CPU0 and CPU1, after acquiring their per-cpu ->barrier_lock and finds 0 segcblist length. It updates ->barrier_seq_snap for CPU0 and CPU1 and continues loop iteration to CPU2. for_each_possible_cpu(cpu) { raw_spin_lock_irqsave(&rdp->barrier_lock, flags); if (!rcu_segcblist_n_cbs(&rdp->cblist)) { WRITE_ONCE(rdp->barrier_seq_snap, gseq); raw_spin_unlock_irqrestore(&rdp->barrier_lock, flags); rcu_barrier_trace(TPS("NQ"), cpu, rcu_state.barrier_sequence); continue; } 4. rcutree_migrate_callbacks() completes execution on CPU1. Segcblist len for CPU2 becomes 0. 5. The loop iteration on CPU0, checks rcu_segcblist_n_cbs(&rdp->cblist) for CPU2 and completes the loop iteration after setting ->barrier_seq_snap. 6. As there isn't any ->barrier_head callback entrained; at this point, rcu_barrier() in CPU0 returns. 7. The callbacks, which migrated from CPU2 to CPU1, execute. Straightforward per-CPU locking is also subject to the following race condition noted by Boqun Feng: 1. CPU0 calls rcu_barrier(), starting a new barrier sequence by invoking rcu_seq_start() and init_completion(), but does not yet initialize rcu_state.barrier_cpu_count. 2. CPU1 starts offlining for CPU2, calling rcutree_migrate_callbacks(), which in turn calls rcu_barrier_entrain() holding CPU2's. rdp->barrier_lock. It then entrains ->barrier_head for CPU2 and atomically increments rcu_state.barrier_cpu_count, which is unfortunately not yet initialized to the value 2. 3. The just-entrained RCU callback is invoked. It atomically decrements rcu_state.barrier_cpu_count and sees that it is now zero. This callback therefore invokes complete(). 4. CPU0 continues executing rcu_barrier(), but is not blocked by its call to wait_for_completion(). This results in rcu_barrier() returning before all pre-existing callbacks have been invoked, which is a bug. Therefore, synchronization is provided by rcu_state.barrier_lock, which is also held across the initialization sequence, especially the rcu_seq_start() and the atomic_set() that sets rcu_state.barrier_cpu_count to the value 2. In addition, this lock is held when entraining the rcu_barrier() callback, when deciding whether or not a CPU has callbacks that rcu_barrier() must wait on, when setting the ->qsmaskinitnext for incoming CPUs, and when migrating callbacks from a CPU that is going offline. Reviewed-by: Frederic Weisbecker <frederic@kernel.org> Co-developed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com> Signed-off-by: Neeraj Upadhyay <quic_neeraju@quicinc.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2021-12-15 05:35:17 +08:00
raw_spin_lock_irqsave(&rcu_state.barrier_lock, flags);
if (!rcu_segcblist_n_cbs(&rdp->cblist)) {
WRITE_ONCE(rdp->barrier_seq_snap, gseq);
rcu: Make rcu_barrier() no longer block CPU-hotplug operations This commit removes the cpus_read_lock() and cpus_read_unlock() calls from rcu_barrier(), thus allowing CPUs to come and go during the course of rcu_barrier() execution. Posting of the ->barrier_head callbacks does synchronize with portions of RCU's CPU-hotplug notifiers, but these locks are held for short time periods on both sides. Thus, full CPU-hotplug operations could both start and finish during the execution of a given rcu_barrier() invocation. Additional synchronization is provided by a global ->barrier_lock. Since the ->barrier_lock is only used during rcu_barrier() execution and during onlining/offlining a CPU, the contention for this lock should be low. It might be tempting to make use of a per-CPU lock just on general principles, but straightforward attempts to do this have the problems shown below. Initial state: 3 CPUs present, CPU 0 and CPU1 do not have any callback and CPU2 has callbacks. 1. CPU0 calls rcu_barrier(). 2. CPU1 starts offlining for CPU2. CPU1 calls rcutree_migrate_callbacks(). rcu_barrier_entrain() is called from rcutree_migrate_callbacks(), with CPU2's rdp->barrier_lock. It does not entrain ->barrier_head for CPU2, as rcu_barrier() on CPU0 hasn't started the barrier sequence (by calling rcu_seq_start(&rcu_state.barrier_sequence)) yet. 3. CPU0 starts new barrier sequence. It iterates over CPU0 and CPU1, after acquiring their per-cpu ->barrier_lock and finds 0 segcblist length. It updates ->barrier_seq_snap for CPU0 and CPU1 and continues loop iteration to CPU2. for_each_possible_cpu(cpu) { raw_spin_lock_irqsave(&rdp->barrier_lock, flags); if (!rcu_segcblist_n_cbs(&rdp->cblist)) { WRITE_ONCE(rdp->barrier_seq_snap, gseq); raw_spin_unlock_irqrestore(&rdp->barrier_lock, flags); rcu_barrier_trace(TPS("NQ"), cpu, rcu_state.barrier_sequence); continue; } 4. rcutree_migrate_callbacks() completes execution on CPU1. Segcblist len for CPU2 becomes 0. 5. The loop iteration on CPU0, checks rcu_segcblist_n_cbs(&rdp->cblist) for CPU2 and completes the loop iteration after setting ->barrier_seq_snap. 6. As there isn't any ->barrier_head callback entrained; at this point, rcu_barrier() in CPU0 returns. 7. The callbacks, which migrated from CPU2 to CPU1, execute. Straightforward per-CPU locking is also subject to the following race condition noted by Boqun Feng: 1. CPU0 calls rcu_barrier(), starting a new barrier sequence by invoking rcu_seq_start() and init_completion(), but does not yet initialize rcu_state.barrier_cpu_count. 2. CPU1 starts offlining for CPU2, calling rcutree_migrate_callbacks(), which in turn calls rcu_barrier_entrain() holding CPU2's. rdp->barrier_lock. It then entrains ->barrier_head for CPU2 and atomically increments rcu_state.barrier_cpu_count, which is unfortunately not yet initialized to the value 2. 3. The just-entrained RCU callback is invoked. It atomically decrements rcu_state.barrier_cpu_count and sees that it is now zero. This callback therefore invokes complete(). 4. CPU0 continues executing rcu_barrier(), but is not blocked by its call to wait_for_completion(). This results in rcu_barrier() returning before all pre-existing callbacks have been invoked, which is a bug. Therefore, synchronization is provided by rcu_state.barrier_lock, which is also held across the initialization sequence, especially the rcu_seq_start() and the atomic_set() that sets rcu_state.barrier_cpu_count to the value 2. In addition, this lock is held when entraining the rcu_barrier() callback, when deciding whether or not a CPU has callbacks that rcu_barrier() must wait on, when setting the ->qsmaskinitnext for incoming CPUs, and when migrating callbacks from a CPU that is going offline. Reviewed-by: Frederic Weisbecker <frederic@kernel.org> Co-developed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com> Signed-off-by: Neeraj Upadhyay <quic_neeraju@quicinc.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2021-12-15 05:35:17 +08:00
raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags);
rcu_barrier_trace(TPS("NQ"), cpu, rcu_state.barrier_sequence);
continue;
}
if (!rcu_rdp_cpu_online(rdp)) {
rcu_barrier_entrain(rdp);
WARN_ON_ONCE(READ_ONCE(rdp->barrier_seq_snap) != gseq);
rcu: Make rcu_barrier() no longer block CPU-hotplug operations This commit removes the cpus_read_lock() and cpus_read_unlock() calls from rcu_barrier(), thus allowing CPUs to come and go during the course of rcu_barrier() execution. Posting of the ->barrier_head callbacks does synchronize with portions of RCU's CPU-hotplug notifiers, but these locks are held for short time periods on both sides. Thus, full CPU-hotplug operations could both start and finish during the execution of a given rcu_barrier() invocation. Additional synchronization is provided by a global ->barrier_lock. Since the ->barrier_lock is only used during rcu_barrier() execution and during onlining/offlining a CPU, the contention for this lock should be low. It might be tempting to make use of a per-CPU lock just on general principles, but straightforward attempts to do this have the problems shown below. Initial state: 3 CPUs present, CPU 0 and CPU1 do not have any callback and CPU2 has callbacks. 1. CPU0 calls rcu_barrier(). 2. CPU1 starts offlining for CPU2. CPU1 calls rcutree_migrate_callbacks(). rcu_barrier_entrain() is called from rcutree_migrate_callbacks(), with CPU2's rdp->barrier_lock. It does not entrain ->barrier_head for CPU2, as rcu_barrier() on CPU0 hasn't started the barrier sequence (by calling rcu_seq_start(&rcu_state.barrier_sequence)) yet. 3. CPU0 starts new barrier sequence. It iterates over CPU0 and CPU1, after acquiring their per-cpu ->barrier_lock and finds 0 segcblist length. It updates ->barrier_seq_snap for CPU0 and CPU1 and continues loop iteration to CPU2. for_each_possible_cpu(cpu) { raw_spin_lock_irqsave(&rdp->barrier_lock, flags); if (!rcu_segcblist_n_cbs(&rdp->cblist)) { WRITE_ONCE(rdp->barrier_seq_snap, gseq); raw_spin_unlock_irqrestore(&rdp->barrier_lock, flags); rcu_barrier_trace(TPS("NQ"), cpu, rcu_state.barrier_sequence); continue; } 4. rcutree_migrate_callbacks() completes execution on CPU1. Segcblist len for CPU2 becomes 0. 5. The loop iteration on CPU0, checks rcu_segcblist_n_cbs(&rdp->cblist) for CPU2 and completes the loop iteration after setting ->barrier_seq_snap. 6. As there isn't any ->barrier_head callback entrained; at this point, rcu_barrier() in CPU0 returns. 7. The callbacks, which migrated from CPU2 to CPU1, execute. Straightforward per-CPU locking is also subject to the following race condition noted by Boqun Feng: 1. CPU0 calls rcu_barrier(), starting a new barrier sequence by invoking rcu_seq_start() and init_completion(), but does not yet initialize rcu_state.barrier_cpu_count. 2. CPU1 starts offlining for CPU2, calling rcutree_migrate_callbacks(), which in turn calls rcu_barrier_entrain() holding CPU2's. rdp->barrier_lock. It then entrains ->barrier_head for CPU2 and atomically increments rcu_state.barrier_cpu_count, which is unfortunately not yet initialized to the value 2. 3. The just-entrained RCU callback is invoked. It atomically decrements rcu_state.barrier_cpu_count and sees that it is now zero. This callback therefore invokes complete(). 4. CPU0 continues executing rcu_barrier(), but is not blocked by its call to wait_for_completion(). This results in rcu_barrier() returning before all pre-existing callbacks have been invoked, which is a bug. Therefore, synchronization is provided by rcu_state.barrier_lock, which is also held across the initialization sequence, especially the rcu_seq_start() and the atomic_set() that sets rcu_state.barrier_cpu_count to the value 2. In addition, this lock is held when entraining the rcu_barrier() callback, when deciding whether or not a CPU has callbacks that rcu_barrier() must wait on, when setting the ->qsmaskinitnext for incoming CPUs, and when migrating callbacks from a CPU that is going offline. Reviewed-by: Frederic Weisbecker <frederic@kernel.org> Co-developed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com> Signed-off-by: Neeraj Upadhyay <quic_neeraju@quicinc.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2021-12-15 05:35:17 +08:00
raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags);
rcu_barrier_trace(TPS("OfflineNoCBQ"), cpu, rcu_state.barrier_sequence);
continue;
}
rcu: Make rcu_barrier() no longer block CPU-hotplug operations This commit removes the cpus_read_lock() and cpus_read_unlock() calls from rcu_barrier(), thus allowing CPUs to come and go during the course of rcu_barrier() execution. Posting of the ->barrier_head callbacks does synchronize with portions of RCU's CPU-hotplug notifiers, but these locks are held for short time periods on both sides. Thus, full CPU-hotplug operations could both start and finish during the execution of a given rcu_barrier() invocation. Additional synchronization is provided by a global ->barrier_lock. Since the ->barrier_lock is only used during rcu_barrier() execution and during onlining/offlining a CPU, the contention for this lock should be low. It might be tempting to make use of a per-CPU lock just on general principles, but straightforward attempts to do this have the problems shown below. Initial state: 3 CPUs present, CPU 0 and CPU1 do not have any callback and CPU2 has callbacks. 1. CPU0 calls rcu_barrier(). 2. CPU1 starts offlining for CPU2. CPU1 calls rcutree_migrate_callbacks(). rcu_barrier_entrain() is called from rcutree_migrate_callbacks(), with CPU2's rdp->barrier_lock. It does not entrain ->barrier_head for CPU2, as rcu_barrier() on CPU0 hasn't started the barrier sequence (by calling rcu_seq_start(&rcu_state.barrier_sequence)) yet. 3. CPU0 starts new barrier sequence. It iterates over CPU0 and CPU1, after acquiring their per-cpu ->barrier_lock and finds 0 segcblist length. It updates ->barrier_seq_snap for CPU0 and CPU1 and continues loop iteration to CPU2. for_each_possible_cpu(cpu) { raw_spin_lock_irqsave(&rdp->barrier_lock, flags); if (!rcu_segcblist_n_cbs(&rdp->cblist)) { WRITE_ONCE(rdp->barrier_seq_snap, gseq); raw_spin_unlock_irqrestore(&rdp->barrier_lock, flags); rcu_barrier_trace(TPS("NQ"), cpu, rcu_state.barrier_sequence); continue; } 4. rcutree_migrate_callbacks() completes execution on CPU1. Segcblist len for CPU2 becomes 0. 5. The loop iteration on CPU0, checks rcu_segcblist_n_cbs(&rdp->cblist) for CPU2 and completes the loop iteration after setting ->barrier_seq_snap. 6. As there isn't any ->barrier_head callback entrained; at this point, rcu_barrier() in CPU0 returns. 7. The callbacks, which migrated from CPU2 to CPU1, execute. Straightforward per-CPU locking is also subject to the following race condition noted by Boqun Feng: 1. CPU0 calls rcu_barrier(), starting a new barrier sequence by invoking rcu_seq_start() and init_completion(), but does not yet initialize rcu_state.barrier_cpu_count. 2. CPU1 starts offlining for CPU2, calling rcutree_migrate_callbacks(), which in turn calls rcu_barrier_entrain() holding CPU2's. rdp->barrier_lock. It then entrains ->barrier_head for CPU2 and atomically increments rcu_state.barrier_cpu_count, which is unfortunately not yet initialized to the value 2. 3. The just-entrained RCU callback is invoked. It atomically decrements rcu_state.barrier_cpu_count and sees that it is now zero. This callback therefore invokes complete(). 4. CPU0 continues executing rcu_barrier(), but is not blocked by its call to wait_for_completion(). This results in rcu_barrier() returning before all pre-existing callbacks have been invoked, which is a bug. Therefore, synchronization is provided by rcu_state.barrier_lock, which is also held across the initialization sequence, especially the rcu_seq_start() and the atomic_set() that sets rcu_state.barrier_cpu_count to the value 2. In addition, this lock is held when entraining the rcu_barrier() callback, when deciding whether or not a CPU has callbacks that rcu_barrier() must wait on, when setting the ->qsmaskinitnext for incoming CPUs, and when migrating callbacks from a CPU that is going offline. Reviewed-by: Frederic Weisbecker <frederic@kernel.org> Co-developed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com> Signed-off-by: Neeraj Upadhyay <quic_neeraju@quicinc.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2021-12-15 05:35:17 +08:00
raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags);
if (smp_call_function_single(cpu, rcu_barrier_handler, (void *)cpu, 1)) {
schedule_timeout_uninterruptible(1);
goto retry;
}
WARN_ON_ONCE(READ_ONCE(rdp->barrier_seq_snap) != gseq);
rcu_barrier_trace(TPS("OnlineQ"), cpu, rcu_state.barrier_sequence);
}
/*
* Now that we have an rcu_barrier_callback() callback on each
* CPU, and thus each counted, remove the initial count.
*/
if (atomic_sub_and_test(2, &rcu_state.barrier_cpu_count))
complete(&rcu_state.barrier_completion);
/* Wait for all rcu_barrier_callback() callbacks to be invoked. */
wait_for_completion(&rcu_state.barrier_completion);
/* Mark the end of the barrier operation. */
rcu_barrier_trace(TPS("Inc2"), -1, rcu_state.barrier_sequence);
rcu_seq_end(&rcu_state.barrier_sequence);
gseq = rcu_state.barrier_sequence;
for_each_possible_cpu(cpu) {
rdp = per_cpu_ptr(&rcu_data, cpu);
WRITE_ONCE(rdp->barrier_seq_snap, gseq);
}
/* Other rcu_barrier() invocations can now safely proceed. */
mutex_unlock(&rcu_state.barrier_mutex);
}
EXPORT_SYMBOL_GPL(rcu_barrier);
/*
* Compute the mask of online CPUs for the specified rcu_node structure.
* This will not be stable unless the rcu_node structure's ->lock is
* held, but the bit corresponding to the current CPU will be stable
* in most contexts.
*/
static unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
{
return READ_ONCE(rnp->qsmaskinitnext);
}
/*
* Is the CPU corresponding to the specified rcu_data structure online
* from RCU's perspective? This perspective is given by that structure's
* ->qsmaskinitnext field rather than by the global cpu_online_mask.
*/
static bool rcu_rdp_cpu_online(struct rcu_data *rdp)
{
return !!(rdp->grpmask & rcu_rnp_online_cpus(rdp->mynode));
}
bool rcu_cpu_online(int cpu)
{
struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
return rcu_rdp_cpu_online(rdp);
}
#if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
/*
* Is the current CPU online as far as RCU is concerned?
*
* Disable preemption to avoid false positives that could otherwise
* happen due to the current CPU number being sampled, this task being
* preempted, its old CPU being taken offline, resuming on some other CPU,
* then determining that its old CPU is now offline.
*
* Disable checking if in an NMI handler because we cannot safely
* report errors from NMI handlers anyway. In addition, it is OK to use
* RCU on an offline processor during initial boot, hence the check for
* rcu_scheduler_fully_active.
*/
bool rcu_lockdep_current_cpu_online(void)
{
struct rcu_data *rdp;
bool ret = false;
if (in_nmi() || !rcu_scheduler_fully_active)
return true;
preempt_disable_notrace();
rdp = this_cpu_ptr(&rcu_data);
/*
* Strictly, we care here about the case where the current CPU is
* in rcu_cpu_starting() and thus has an excuse for rdp->grpmask
* not being up to date. So arch_spin_is_locked() might have a
* false positive if it's held by some *other* CPU, but that's
* OK because that just means a false *negative* on the warning.
*/
if (rcu_rdp_cpu_online(rdp) || arch_spin_is_locked(&rcu_state.ofl_lock))
ret = true;
preempt_enable_notrace();
return ret;
}
EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
#endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
// Has rcu_init() been invoked? This is used (for example) to determine
// whether spinlocks may be acquired safely.
static bool rcu_init_invoked(void)
{
return !!rcu_state.n_online_cpus;
}
/*
* Near the end of the offline process. Trace the fact that this CPU
* is going offline.
*/
int rcutree_dying_cpu(unsigned int cpu)
{
bool blkd;
struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
struct rcu_node *rnp = rdp->mynode;
if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
return 0;
blkd = !!(READ_ONCE(rnp->qsmask) & rdp->grpmask);
trace_rcu_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq),
blkd ? TPS("cpuofl-bgp") : TPS("cpuofl"));
return 0;
}
/*
* All CPUs for the specified rcu_node structure have gone offline,
* and all tasks that were preempted within an RCU read-side critical
* section while running on one of those CPUs have since exited their RCU
* read-side critical section. Some other CPU is reporting this fact with
* the specified rcu_node structure's ->lock held and interrupts disabled.
* This function therefore goes up the tree of rcu_node structures,
* clearing the corresponding bits in the ->qsmaskinit fields. Note that
* the leaf rcu_node structure's ->qsmaskinit field has already been
* updated.
*
* This function does check that the specified rcu_node structure has
* all CPUs offline and no blocked tasks, so it is OK to invoke it
* prematurely. That said, invoking it after the fact will cost you
* a needless lock acquisition. So once it has done its work, don't
* invoke it again.
*/
static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
{
long mask;
struct rcu_node *rnp = rnp_leaf;
raw_lockdep_assert_held_rcu_node(rnp_leaf);
if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
WARN_ON_ONCE(rnp_leaf->qsmaskinit) ||
WARN_ON_ONCE(rcu_preempt_has_tasks(rnp_leaf)))
return;
for (;;) {
mask = rnp->grpmask;
rnp = rnp->parent;
if (!rnp)
break;
raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
rnp->qsmaskinit &= ~mask;
/* Between grace periods, so better already be zero! */
WARN_ON_ONCE(rnp->qsmask);
if (rnp->qsmaskinit) {
raw_spin_unlock_rcu_node(rnp);
/* irqs remain disabled. */
return;
}
raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
}
}
/*
* The CPU has been completely removed, and some other CPU is reporting
* this fact from process context. Do the remainder of the cleanup.
* There can only be one CPU hotplug operation at a time, so no need for
* explicit locking.
*/
int rcutree_dead_cpu(unsigned int cpu)
{
if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
return 0;
WRITE_ONCE(rcu_state.n_online_cpus, rcu_state.n_online_cpus - 1);
// Stop-machine done, so allow nohz_full to disable tick.
tick_dep_clear(TICK_DEP_BIT_RCU);
return 0;
}
rcu: Process offlining and onlining only at grace-period start Races between CPU hotplug and grace periods can be difficult to resolve, so the ->onoff_mutex is used to exclude the two events. Unfortunately, this means that it is impossible for an outgoing CPU to perform the last bits of its offlining from its last pass through the idle loop, because sleeplocks cannot be acquired in that context. This commit avoids these problems by buffering online and offline events in a new ->qsmaskinitnext field in the leaf rcu_node structures. When a grace period starts, the events accumulated in this mask are applied to the ->qsmaskinit field, and, if needed, up the rcu_node tree. The special case of all CPUs corresponding to a given leaf rcu_node structure being offline while there are still elements in that structure's ->blkd_tasks list is handled using a new ->wait_blkd_tasks field. In this case, propagating the offline bits up the tree is deferred until the beginning of the grace period after all of the tasks have exited their RCU read-side critical sections and removed themselves from the list, at which point the ->wait_blkd_tasks flag is cleared. If one of that leaf rcu_node structure's CPUs comes back online before the list empties, then the ->wait_blkd_tasks flag is simply cleared. This of course means that RCU's notion of which CPUs are offline can be out of date. This is OK because RCU need only wait on CPUs that were online at the time that the grace period started. In addition, RCU's force-quiescent-state actions will handle the case where a CPU goes offline after the grace period starts. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2015-01-24 13:52:37 +08:00
/*
* Propagate ->qsinitmask bits up the rcu_node tree to account for the
* first CPU in a given leaf rcu_node structure coming online. The caller
* must hold the corresponding leaf rcu_node ->lock with interrupts
rcu: Process offlining and onlining only at grace-period start Races between CPU hotplug and grace periods can be difficult to resolve, so the ->onoff_mutex is used to exclude the two events. Unfortunately, this means that it is impossible for an outgoing CPU to perform the last bits of its offlining from its last pass through the idle loop, because sleeplocks cannot be acquired in that context. This commit avoids these problems by buffering online and offline events in a new ->qsmaskinitnext field in the leaf rcu_node structures. When a grace period starts, the events accumulated in this mask are applied to the ->qsmaskinit field, and, if needed, up the rcu_node tree. The special case of all CPUs corresponding to a given leaf rcu_node structure being offline while there are still elements in that structure's ->blkd_tasks list is handled using a new ->wait_blkd_tasks field. In this case, propagating the offline bits up the tree is deferred until the beginning of the grace period after all of the tasks have exited their RCU read-side critical sections and removed themselves from the list, at which point the ->wait_blkd_tasks flag is cleared. If one of that leaf rcu_node structure's CPUs comes back online before the list empties, then the ->wait_blkd_tasks flag is simply cleared. This of course means that RCU's notion of which CPUs are offline can be out of date. This is OK because RCU need only wait on CPUs that were online at the time that the grace period started. In addition, RCU's force-quiescent-state actions will handle the case where a CPU goes offline after the grace period starts. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2015-01-24 13:52:37 +08:00
* disabled.
*/
static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
{
long mask;
long oldmask;
rcu: Process offlining and onlining only at grace-period start Races between CPU hotplug and grace periods can be difficult to resolve, so the ->onoff_mutex is used to exclude the two events. Unfortunately, this means that it is impossible for an outgoing CPU to perform the last bits of its offlining from its last pass through the idle loop, because sleeplocks cannot be acquired in that context. This commit avoids these problems by buffering online and offline events in a new ->qsmaskinitnext field in the leaf rcu_node structures. When a grace period starts, the events accumulated in this mask are applied to the ->qsmaskinit field, and, if needed, up the rcu_node tree. The special case of all CPUs corresponding to a given leaf rcu_node structure being offline while there are still elements in that structure's ->blkd_tasks list is handled using a new ->wait_blkd_tasks field. In this case, propagating the offline bits up the tree is deferred until the beginning of the grace period after all of the tasks have exited their RCU read-side critical sections and removed themselves from the list, at which point the ->wait_blkd_tasks flag is cleared. If one of that leaf rcu_node structure's CPUs comes back online before the list empties, then the ->wait_blkd_tasks flag is simply cleared. This of course means that RCU's notion of which CPUs are offline can be out of date. This is OK because RCU need only wait on CPUs that were online at the time that the grace period started. In addition, RCU's force-quiescent-state actions will handle the case where a CPU goes offline after the grace period starts. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2015-01-24 13:52:37 +08:00
struct rcu_node *rnp = rnp_leaf;
raw_lockdep_assert_held_rcu_node(rnp_leaf);
rcu: Clean up handling of tasks blocked across full-rcu_node offline Commit 0aa04b055e71 ("rcu: Process offlining and onlining only at grace-period start") deferred handling of CPU-hotplug events until the start of the next grace period, but consider the following sequence of events: 1. A task is preempted within an RCU-preempt read-side critical section. 2. The CPU that this task was running on goes offline, along with all other CPUs sharing the corresponding leaf rcu_node structure. 3. The task resumes execution. 4. One of those CPUs comes back online before a new grace period starts. In step 2, the code in the next rcu_gp_init() invocation will (correctly) defer removing the leaf rcu_node structure from the upper-level bitmasks, and will (correctly) set that structure's ->wait_blkd_tasks field. During the ensuing interval, RCU will (correctly) track the tasks preempted on that structure because they must block any subsequent grace period. In step 3, the code in rcu_read_unlock_special() will (correctly) remove the task from the leaf rcu_node structure. From this point forward, RCU need not pay attention to this structure, at least not until one of the corresponding CPUs comes back online. In step 4, the code in the next rcu_gp_init() invocation will (incorrectly) invoke rcu_init_new_rnp(). This is incorrect because the corresponding rcu_cleanup_dead_rnp() was never invoked. This is nevertheless harmless because the upper-level bits are still set. So, no harm, no foul, right? At least, all is well until a little further into rcu_gp_init() invocation, which will notice that there are no longer any tasks blocked on the leaf rcu_node structure, conclude that there is no longer anything left over from step 2's offline operation, and will therefore invoke rcu_cleanup_dead_rnp(). But this invocation of rcu_cleanup_dead_rnp() is for the beginning of the earlier offline interval, and the previous invocation of rcu_init_new_rnp() is for the end of that same interval. That is right, they are invoked out of order. That cannot be good, can it? It turns out that this is not a (correctness!) problem because rcu_cleanup_dead_rnp() checks to see if any of the corresponding CPUs are online, and refuses to do anything if so. In other words, in the case where rcu_init_new_rnp() and rcu_cleanup_dead_rnp() execute out of order, they both have no effect. But this is at best an accident waiting to happen. This commit therefore adds logic to rcu_gp_init() so that rcu_init_new_rnp() and rcu_cleanup_dead_rnp() are always invoked in order, and so that neither are invoked at all in cases where RCU had to pay attention to the leaf rcu_node structure during the entire time that all corresponding CPUs were offline. And, while in the area, this commit reduces confusion by using formal parameters rather than local variables that just happen to have the same value at that particular point in the code. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2018-05-03 03:49:21 +08:00
WARN_ON_ONCE(rnp->wait_blkd_tasks);
rcu: Process offlining and onlining only at grace-period start Races between CPU hotplug and grace periods can be difficult to resolve, so the ->onoff_mutex is used to exclude the two events. Unfortunately, this means that it is impossible for an outgoing CPU to perform the last bits of its offlining from its last pass through the idle loop, because sleeplocks cannot be acquired in that context. This commit avoids these problems by buffering online and offline events in a new ->qsmaskinitnext field in the leaf rcu_node structures. When a grace period starts, the events accumulated in this mask are applied to the ->qsmaskinit field, and, if needed, up the rcu_node tree. The special case of all CPUs corresponding to a given leaf rcu_node structure being offline while there are still elements in that structure's ->blkd_tasks list is handled using a new ->wait_blkd_tasks field. In this case, propagating the offline bits up the tree is deferred until the beginning of the grace period after all of the tasks have exited their RCU read-side critical sections and removed themselves from the list, at which point the ->wait_blkd_tasks flag is cleared. If one of that leaf rcu_node structure's CPUs comes back online before the list empties, then the ->wait_blkd_tasks flag is simply cleared. This of course means that RCU's notion of which CPUs are offline can be out of date. This is OK because RCU need only wait on CPUs that were online at the time that the grace period started. In addition, RCU's force-quiescent-state actions will handle the case where a CPU goes offline after the grace period starts. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2015-01-24 13:52:37 +08:00
for (;;) {
mask = rnp->grpmask;
rnp = rnp->parent;
if (rnp == NULL)
return;
raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
oldmask = rnp->qsmaskinit;
rcu: Process offlining and onlining only at grace-period start Races between CPU hotplug and grace periods can be difficult to resolve, so the ->onoff_mutex is used to exclude the two events. Unfortunately, this means that it is impossible for an outgoing CPU to perform the last bits of its offlining from its last pass through the idle loop, because sleeplocks cannot be acquired in that context. This commit avoids these problems by buffering online and offline events in a new ->qsmaskinitnext field in the leaf rcu_node structures. When a grace period starts, the events accumulated in this mask are applied to the ->qsmaskinit field, and, if needed, up the rcu_node tree. The special case of all CPUs corresponding to a given leaf rcu_node structure being offline while there are still elements in that structure's ->blkd_tasks list is handled using a new ->wait_blkd_tasks field. In this case, propagating the offline bits up the tree is deferred until the beginning of the grace period after all of the tasks have exited their RCU read-side critical sections and removed themselves from the list, at which point the ->wait_blkd_tasks flag is cleared. If one of that leaf rcu_node structure's CPUs comes back online before the list empties, then the ->wait_blkd_tasks flag is simply cleared. This of course means that RCU's notion of which CPUs are offline can be out of date. This is OK because RCU need only wait on CPUs that were online at the time that the grace period started. In addition, RCU's force-quiescent-state actions will handle the case where a CPU goes offline after the grace period starts. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2015-01-24 13:52:37 +08:00
rnp->qsmaskinit |= mask;
raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
if (oldmask)
return;
rcu: Process offlining and onlining only at grace-period start Races between CPU hotplug and grace periods can be difficult to resolve, so the ->onoff_mutex is used to exclude the two events. Unfortunately, this means that it is impossible for an outgoing CPU to perform the last bits of its offlining from its last pass through the idle loop, because sleeplocks cannot be acquired in that context. This commit avoids these problems by buffering online and offline events in a new ->qsmaskinitnext field in the leaf rcu_node structures. When a grace period starts, the events accumulated in this mask are applied to the ->qsmaskinit field, and, if needed, up the rcu_node tree. The special case of all CPUs corresponding to a given leaf rcu_node structure being offline while there are still elements in that structure's ->blkd_tasks list is handled using a new ->wait_blkd_tasks field. In this case, propagating the offline bits up the tree is deferred until the beginning of the grace period after all of the tasks have exited their RCU read-side critical sections and removed themselves from the list, at which point the ->wait_blkd_tasks flag is cleared. If one of that leaf rcu_node structure's CPUs comes back online before the list empties, then the ->wait_blkd_tasks flag is simply cleared. This of course means that RCU's notion of which CPUs are offline can be out of date. This is OK because RCU need only wait on CPUs that were online at the time that the grace period started. In addition, RCU's force-quiescent-state actions will handle the case where a CPU goes offline after the grace period starts. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2015-01-24 13:52:37 +08:00
}
}
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/*
* Do boot-time initialization of a CPU's per-CPU RCU data.
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
*/
static void __init
rcu_boot_init_percpu_data(int cpu)
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
{
struct context_tracking *ct = this_cpu_ptr(&context_tracking);
struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
/* Set up local state, ensuring consistent view of global state. */
rcu: Correctly handle sparse possible cpus In many cases in the RCU tree code, we iterate over the set of cpus for a leaf node described by rcu_node::grplo and rcu_node::grphi, checking per-cpu data for each cpu in this range. However, if the set of possible cpus is sparse, some cpus described in this range are not possible, and thus no per-cpu region will have been allocated (or initialised) for them by the generic percpu code. Erroneous accesses to a per-cpu area for these !possible cpus may fault or may hit other data depending on the addressed generated when the erroneous per cpu offset is applied. In practice, both cases have been observed on arm64 hardware (the former being silent, but detectable with additional patches). To avoid issues resulting from this, we must iterate over the set of *possible* cpus for a given leaf node. This patch add a new helper, for_each_leaf_node_possible_cpu, to enable this. As iteration is often intertwined with rcu_node local bitmask manipulation, a new leaf_node_cpu_bit helper is added to make this simpler and more consistent. The RCU tree code is made to use both of these where appropriate. Without this patch, running reboot at a shell can result in an oops like: [ 3369.075979] Unable to handle kernel paging request at virtual address ffffff8008b21b4c [ 3369.083881] pgd = ffffffc3ecdda000 [ 3369.087270] [ffffff8008b21b4c] *pgd=00000083eca48003, *pud=00000083eca48003, *pmd=0000000000000000 [ 3369.096222] Internal error: Oops: 96000007 [#1] PREEMPT SMP [ 3369.101781] Modules linked in: [ 3369.104825] CPU: 2 PID: 1817 Comm: NetworkManager Tainted: G W 4.6.0+ #3 [ 3369.121239] task: ffffffc0fa13e000 ti: ffffffc3eb940000 task.ti: ffffffc3eb940000 [ 3369.128708] PC is at sync_rcu_exp_select_cpus+0x188/0x510 [ 3369.134094] LR is at sync_rcu_exp_select_cpus+0x104/0x510 [ 3369.139479] pc : [<ffffff80081109a8>] lr : [<ffffff8008110924>] pstate: 200001c5 [ 3369.146860] sp : ffffffc3eb9435a0 [ 3369.150162] x29: ffffffc3eb9435a0 x28: ffffff8008be4f88 [ 3369.155465] x27: ffffff8008b66c80 x26: ffffffc3eceb2600 [ 3369.160767] x25: 0000000000000001 x24: ffffff8008be4f88 [ 3369.166070] x23: ffffff8008b51c3c x22: ffffff8008b66c80 [ 3369.171371] x21: 0000000000000001 x20: ffffff8008b21b40 [ 3369.176673] x19: ffffff8008b66c80 x18: 0000000000000000 [ 3369.181975] x17: 0000007fa951a010 x16: ffffff80086a30f0 [ 3369.187278] x15: 0000007fa9505590 x14: 0000000000000000 [ 3369.192580] x13: ffffff8008b51000 x12: ffffffc3eb940000 [ 3369.197882] x11: 0000000000000006 x10: ffffff8008b51b78 [ 3369.203184] x9 : 0000000000000001 x8 : ffffff8008be4000 [ 3369.208486] x7 : ffffff8008b21b40 x6 : 0000000000001003 [ 3369.213788] x5 : 0000000000000000 x4 : ffffff8008b27280 [ 3369.219090] x3 : ffffff8008b21b4c x2 : 0000000000000001 [ 3369.224406] x1 : 0000000000000001 x0 : 0000000000000140 ... [ 3369.972257] [<ffffff80081109a8>] sync_rcu_exp_select_cpus+0x188/0x510 [ 3369.978685] [<ffffff80081128b4>] synchronize_rcu_expedited+0x64/0xa8 [ 3369.985026] [<ffffff80086b987c>] synchronize_net+0x24/0x30 [ 3369.990499] [<ffffff80086ddb54>] dev_deactivate_many+0x28c/0x298 [ 3369.996493] [<ffffff80086b6bb8>] __dev_close_many+0x60/0xd0 [ 3370.002052] [<ffffff80086b6d48>] __dev_close+0x28/0x40 [ 3370.007178] [<ffffff80086bf62c>] __dev_change_flags+0x8c/0x158 [ 3370.012999] [<ffffff80086bf718>] dev_change_flags+0x20/0x60 [ 3370.018558] [<ffffff80086cf7f0>] do_setlink+0x288/0x918 [ 3370.023771] [<ffffff80086d0798>] rtnl_newlink+0x398/0x6a8 [ 3370.029158] [<ffffff80086cee84>] rtnetlink_rcv_msg+0xe4/0x220 [ 3370.034891] [<ffffff80086e274c>] netlink_rcv_skb+0xc4/0xf8 [ 3370.040364] [<ffffff80086ced8c>] rtnetlink_rcv+0x2c/0x40 [ 3370.045663] [<ffffff80086e1fe8>] netlink_unicast+0x160/0x238 [ 3370.051309] [<ffffff80086e24b8>] netlink_sendmsg+0x2f0/0x358 [ 3370.056956] [<ffffff80086a0070>] sock_sendmsg+0x18/0x30 [ 3370.062168] [<ffffff80086a21cc>] ___sys_sendmsg+0x26c/0x280 [ 3370.067728] [<ffffff80086a30ac>] __sys_sendmsg+0x44/0x88 [ 3370.073027] [<ffffff80086a3100>] SyS_sendmsg+0x10/0x20 [ 3370.078153] [<ffffff8008085e70>] el0_svc_naked+0x24/0x28 Signed-off-by: Mark Rutland <mark.rutland@arm.com> Reported-by: Dennis Chen <dennis.chen@arm.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Josh Triplett <josh@joshtriplett.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Cc: Steve Capper <steve.capper@arm.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Will Deacon <will.deacon@arm.com> Cc: linux-kernel@vger.kernel.org Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2016-06-03 22:20:04 +08:00
rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
INIT_WORK(&rdp->strict_work, strict_work_handler);
WARN_ON_ONCE(ct->dynticks_nesting != 1);
WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(cpu)));
rdp->barrier_seq_snap = rcu_state.barrier_sequence;
rdp->rcu_ofl_gp_seq = rcu_state.gp_seq;
rdp->rcu_ofl_gp_flags = RCU_GP_CLEANED;
rdp->rcu_onl_gp_seq = rcu_state.gp_seq;
rdp->rcu_onl_gp_flags = RCU_GP_CLEANED;
rdp->last_sched_clock = jiffies;
rdp->cpu = cpu;
rcu_boot_init_nocb_percpu_data(rdp);
}
/*
* Invoked early in the CPU-online process, when pretty much all services
* are available. The incoming CPU is not present.
*
* Initializes a CPU's per-CPU RCU data. Note that only one online or
* offline event can be happening at a given time. Note also that we can
* accept some slop in the rsp->gp_seq access due to the fact that this
* CPU cannot possibly have any non-offloaded RCU callbacks in flight yet.
* And any offloaded callbacks are being numbered elsewhere.
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
*/
int rcutree_prepare_cpu(unsigned int cpu)
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
{
unsigned long flags;
struct context_tracking *ct = per_cpu_ptr(&context_tracking, cpu);
struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
struct rcu_node *rnp = rcu_get_root();
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/* Set up local state, ensuring consistent view of global state. */
raw_spin_lock_irqsave_rcu_node(rnp, flags);
rdp->qlen_last_fqs_check = 0;
rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs);
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
rdp->blimit = blimit;
ct->dynticks_nesting = 1; /* CPU not up, no tearing. */
raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
rcu/nocb: Only (re-)initialize segcblist when needed on CPU up At the start of a CPU-hotplug operation, the incoming CPU's callback list can be in a number of states: 1. Disabled and empty. This is the case when the boot CPU has not invoked call_rcu(), when a non-boot CPU first comes online, and when a non-offloaded CPU comes back online. In this case, it is both necessary and permissible to initialize ->cblist. Because either the CPU is currently running with interrupts disabled (boot CPU) or is not yet running at all (other CPUs), it is not necessary to acquire ->nocb_lock. In this case, initialization is required. 2. Disabled and non-empty. This cannot occur, because early boot call_rcu() invocations enable the callback list before enqueuing their callback. 3. Enabled, whether empty or not. In this case, the callback list has already been initialized. This case occurs when the boot CPU has executed an early boot call_rcu() and also when an offloaded CPU comes back online. In both cases, there is no need to initialize the callback list: In the boot-CPU case, the CPU has not (yet) gone offline, and in the offloaded case, the rcuo kthreads are taking care of business. Because it is not necessary to initialize the callback list, it is also not necessary to acquire ->nocb_lock. Therefore, checking if the segcblist is enabled suffices. This commit therefore initializes the callback list at rcutree_prepare_cpu() time only if that list is disabled. Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Cc: Josh Triplett <josh@joshtriplett.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Joel Fernandes <joel@joelfernandes.org> Cc: Neeraj Upadhyay <neeraju@codeaurora.org> Cc: Boqun Feng <boqun.feng@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2021-01-29 01:12:10 +08:00
rcu/nocb: Always init segcblist on CPU up How the rdp->cblist enabled state is treated at CPU-hotplug time depends on whether or not that ->cblist is offloaded. 1) Not offloaded: The ->cblist is disabled when the CPU goes down. All its callbacks are migrated and none can to enqueued until after some later CPU-hotplug operation brings the CPU back up. 2) Offloaded: The ->cblist is not disabled on CPU down because the CB/GP kthreads must finish invoking the remaining callbacks. There is thus no need to re-enable it on CPU up. Since the ->cblist offloaded state is set in stone at boot, it cannot change between CPU down and CPU up. So 1) and 2) are symmetrical. However, given runtime toggling of the offloaded state, there are two additional asymmetrical scenarios: 3) The ->cblist is not offloaded when the CPU goes down. The ->cblist is later toggled to offloaded and then the CPU comes back up. 4) The ->cblist is offloaded when the CPU goes down. The ->cblist is later toggled to no longer be offloaded and then the CPU comes back up. Scenario 4) is currently handled correctly. The ->cblist remains enabled on CPU down and gets re-initialized on CPU up. The toggling operation will wait until ->cblist is empty, so ->cblist will remain empty until CPU-up time. The scenario 3) would run into trouble though, as the rdp is disabled on CPU down and not re-initialized/re-enabled on CPU up. Except that in this case, ->cblist is guaranteed to be empty because all its callbacks were migrated away at CPU-down time. And the CPU-up code already initializes and enables any empty ->cblist structures in order to handle the possibility of early-boot invocations of call_rcu() in the case where such invocations don't occur. So all that need be done is to adjust the locking. Cc: Josh Triplett <josh@joshtriplett.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Joel Fernandes <joel@joelfernandes.org> Cc: Neeraj Upadhyay <neeraju@codeaurora.org> Cc: Thomas Gleixner <tglx@linutronix.de> Inspired-by: Paul E. McKenney <paulmck@kernel.org> Tested-by: Boqun Feng <boqun.feng@gmail.com> Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-11-13 20:13:18 +08:00
/*
rcu/nocb: Only (re-)initialize segcblist when needed on CPU up At the start of a CPU-hotplug operation, the incoming CPU's callback list can be in a number of states: 1. Disabled and empty. This is the case when the boot CPU has not invoked call_rcu(), when a non-boot CPU first comes online, and when a non-offloaded CPU comes back online. In this case, it is both necessary and permissible to initialize ->cblist. Because either the CPU is currently running with interrupts disabled (boot CPU) or is not yet running at all (other CPUs), it is not necessary to acquire ->nocb_lock. In this case, initialization is required. 2. Disabled and non-empty. This cannot occur, because early boot call_rcu() invocations enable the callback list before enqueuing their callback. 3. Enabled, whether empty or not. In this case, the callback list has already been initialized. This case occurs when the boot CPU has executed an early boot call_rcu() and also when an offloaded CPU comes back online. In both cases, there is no need to initialize the callback list: In the boot-CPU case, the CPU has not (yet) gone offline, and in the offloaded case, the rcuo kthreads are taking care of business. Because it is not necessary to initialize the callback list, it is also not necessary to acquire ->nocb_lock. Therefore, checking if the segcblist is enabled suffices. This commit therefore initializes the callback list at rcutree_prepare_cpu() time only if that list is disabled. Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Cc: Josh Triplett <josh@joshtriplett.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Joel Fernandes <joel@joelfernandes.org> Cc: Neeraj Upadhyay <neeraju@codeaurora.org> Cc: Boqun Feng <boqun.feng@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2021-01-29 01:12:10 +08:00
* Only non-NOCB CPUs that didn't have early-boot callbacks need to be
* (re-)initialized.
rcu/nocb: Always init segcblist on CPU up How the rdp->cblist enabled state is treated at CPU-hotplug time depends on whether or not that ->cblist is offloaded. 1) Not offloaded: The ->cblist is disabled when the CPU goes down. All its callbacks are migrated and none can to enqueued until after some later CPU-hotplug operation brings the CPU back up. 2) Offloaded: The ->cblist is not disabled on CPU down because the CB/GP kthreads must finish invoking the remaining callbacks. There is thus no need to re-enable it on CPU up. Since the ->cblist offloaded state is set in stone at boot, it cannot change between CPU down and CPU up. So 1) and 2) are symmetrical. However, given runtime toggling of the offloaded state, there are two additional asymmetrical scenarios: 3) The ->cblist is not offloaded when the CPU goes down. The ->cblist is later toggled to offloaded and then the CPU comes back up. 4) The ->cblist is offloaded when the CPU goes down. The ->cblist is later toggled to no longer be offloaded and then the CPU comes back up. Scenario 4) is currently handled correctly. The ->cblist remains enabled on CPU down and gets re-initialized on CPU up. The toggling operation will wait until ->cblist is empty, so ->cblist will remain empty until CPU-up time. The scenario 3) would run into trouble though, as the rdp is disabled on CPU down and not re-initialized/re-enabled on CPU up. Except that in this case, ->cblist is guaranteed to be empty because all its callbacks were migrated away at CPU-down time. And the CPU-up code already initializes and enables any empty ->cblist structures in order to handle the possibility of early-boot invocations of call_rcu() in the case where such invocations don't occur. So all that need be done is to adjust the locking. Cc: Josh Triplett <josh@joshtriplett.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Joel Fernandes <joel@joelfernandes.org> Cc: Neeraj Upadhyay <neeraju@codeaurora.org> Cc: Thomas Gleixner <tglx@linutronix.de> Inspired-by: Paul E. McKenney <paulmck@kernel.org> Tested-by: Boqun Feng <boqun.feng@gmail.com> Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-11-13 20:13:18 +08:00
*/
rcu/nocb: Only (re-)initialize segcblist when needed on CPU up At the start of a CPU-hotplug operation, the incoming CPU's callback list can be in a number of states: 1. Disabled and empty. This is the case when the boot CPU has not invoked call_rcu(), when a non-boot CPU first comes online, and when a non-offloaded CPU comes back online. In this case, it is both necessary and permissible to initialize ->cblist. Because either the CPU is currently running with interrupts disabled (boot CPU) or is not yet running at all (other CPUs), it is not necessary to acquire ->nocb_lock. In this case, initialization is required. 2. Disabled and non-empty. This cannot occur, because early boot call_rcu() invocations enable the callback list before enqueuing their callback. 3. Enabled, whether empty or not. In this case, the callback list has already been initialized. This case occurs when the boot CPU has executed an early boot call_rcu() and also when an offloaded CPU comes back online. In both cases, there is no need to initialize the callback list: In the boot-CPU case, the CPU has not (yet) gone offline, and in the offloaded case, the rcuo kthreads are taking care of business. Because it is not necessary to initialize the callback list, it is also not necessary to acquire ->nocb_lock. Therefore, checking if the segcblist is enabled suffices. This commit therefore initializes the callback list at rcutree_prepare_cpu() time only if that list is disabled. Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Cc: Josh Triplett <josh@joshtriplett.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Joel Fernandes <joel@joelfernandes.org> Cc: Neeraj Upadhyay <neeraju@codeaurora.org> Cc: Boqun Feng <boqun.feng@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2021-01-29 01:12:10 +08:00
if (!rcu_segcblist_is_enabled(&rdp->cblist))
rcu/nocb: Always init segcblist on CPU up How the rdp->cblist enabled state is treated at CPU-hotplug time depends on whether or not that ->cblist is offloaded. 1) Not offloaded: The ->cblist is disabled when the CPU goes down. All its callbacks are migrated and none can to enqueued until after some later CPU-hotplug operation brings the CPU back up. 2) Offloaded: The ->cblist is not disabled on CPU down because the CB/GP kthreads must finish invoking the remaining callbacks. There is thus no need to re-enable it on CPU up. Since the ->cblist offloaded state is set in stone at boot, it cannot change between CPU down and CPU up. So 1) and 2) are symmetrical. However, given runtime toggling of the offloaded state, there are two additional asymmetrical scenarios: 3) The ->cblist is not offloaded when the CPU goes down. The ->cblist is later toggled to offloaded and then the CPU comes back up. 4) The ->cblist is offloaded when the CPU goes down. The ->cblist is later toggled to no longer be offloaded and then the CPU comes back up. Scenario 4) is currently handled correctly. The ->cblist remains enabled on CPU down and gets re-initialized on CPU up. The toggling operation will wait until ->cblist is empty, so ->cblist will remain empty until CPU-up time. The scenario 3) would run into trouble though, as the rdp is disabled on CPU down and not re-initialized/re-enabled on CPU up. Except that in this case, ->cblist is guaranteed to be empty because all its callbacks were migrated away at CPU-down time. And the CPU-up code already initializes and enables any empty ->cblist structures in order to handle the possibility of early-boot invocations of call_rcu() in the case where such invocations don't occur. So all that need be done is to adjust the locking. Cc: Josh Triplett <josh@joshtriplett.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Joel Fernandes <joel@joelfernandes.org> Cc: Neeraj Upadhyay <neeraju@codeaurora.org> Cc: Thomas Gleixner <tglx@linutronix.de> Inspired-by: Paul E. McKenney <paulmck@kernel.org> Tested-by: Boqun Feng <boqun.feng@gmail.com> Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-11-13 20:13:18 +08:00
rcu_segcblist_init(&rdp->cblist); /* Re-enable callbacks. */
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
rcu: Process offlining and onlining only at grace-period start Races between CPU hotplug and grace periods can be difficult to resolve, so the ->onoff_mutex is used to exclude the two events. Unfortunately, this means that it is impossible for an outgoing CPU to perform the last bits of its offlining from its last pass through the idle loop, because sleeplocks cannot be acquired in that context. This commit avoids these problems by buffering online and offline events in a new ->qsmaskinitnext field in the leaf rcu_node structures. When a grace period starts, the events accumulated in this mask are applied to the ->qsmaskinit field, and, if needed, up the rcu_node tree. The special case of all CPUs corresponding to a given leaf rcu_node structure being offline while there are still elements in that structure's ->blkd_tasks list is handled using a new ->wait_blkd_tasks field. In this case, propagating the offline bits up the tree is deferred until the beginning of the grace period after all of the tasks have exited their RCU read-side critical sections and removed themselves from the list, at which point the ->wait_blkd_tasks flag is cleared. If one of that leaf rcu_node structure's CPUs comes back online before the list empties, then the ->wait_blkd_tasks flag is simply cleared. This of course means that RCU's notion of which CPUs are offline can be out of date. This is OK because RCU need only wait on CPUs that were online at the time that the grace period started. In addition, RCU's force-quiescent-state actions will handle the case where a CPU goes offline after the grace period starts. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2015-01-24 13:52:37 +08:00
/*
* Add CPU to leaf rcu_node pending-online bitmask. Any needed
* propagation up the rcu_node tree will happen at the beginning
* of the next grace period.
*/
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
rnp = rdp->mynode;
raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
rdp->gp_seq = READ_ONCE(rnp->gp_seq);
rdp->gp_seq_needed = rdp->gp_seq;
rdp->cpu_no_qs.b.norm = true;
rdp->core_needs_qs = false;
rdp->rcu_iw_pending = false;
rdp->rcu_iw = IRQ_WORK_INIT_HARD(rcu_iw_handler);
rdp->rcu_iw_gp_seq = rdp->gp_seq - 1;
trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuonl"));
raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
rcu: Make RCU priority boosting work on single-CPU rcu_node structures When any CPU comes online, it checks to see if an RCU-boost kthread has already been created for that CPU's leaf rcu_node structure, and if not, it creates one. Unfortunately, it also verifies that this leaf rcu_node structure actually has at least one online CPU, and if not, it declines to create the kthread. Although this behavior makes sense during early boot, especially on systems that claim far more CPUs than they actually have, it makes no sense for the first CPU to come online for a given rcu_node structure. There is no point in checking because we know there is a CPU on its way in. The problem is that timing differences can cause this incoming CPU to not yet be reflected in the various bit masks even at rcutree_online_cpu() time, and there is no chance at rcutree_prepare_cpu() time. Plus it would be better to create the RCU-boost kthread at rcutree_prepare_cpu() to handle the case where the CPU is involved in an RCU priority inversion very shortly after it comes online. This commit therefore moves the checking to rcu_prepare_kthreads(), which is called only at early boot, when the check is appropriate. In addition, it makes rcutree_prepare_cpu() invoke rcu_spawn_one_boost_kthread(), which no longer does any checking for online CPUs. With this change, RCU priority boosting tests now pass for short rcutorture runs, even with single-CPU leaf rcu_node structures. Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Cc: Scott Wood <swood@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2021-04-06 11:42:09 +08:00
rcu_spawn_one_boost_kthread(rnp);
rcu_spawn_cpu_nocb_kthread(cpu);
rcu: Fix single-CPU check in rcu_blocking_is_gp() Currently, for CONFIG_PREEMPTION=n kernels, rcu_blocking_is_gp() uses num_online_cpus() to determine whether there is only one CPU online. When there is only a single CPU online, the simple fact that synchronize_rcu() could be legally called implies that a full grace period has elapsed. Therefore, in the single-CPU case, synchronize_rcu() simply returns immediately. Unfortunately, num_online_cpus() is unreliable while a CPU-hotplug operation is transitioning to or from single-CPU operation because: 1. num_online_cpus() uses atomic_read(&__num_online_cpus) to locklessly sample the number of online CPUs. The hotplug locks are not held, which means that an incoming CPU can concurrently update this count. This in turn means that an RCU read-side critical section on the incoming CPU might observe updates prior to the grace period, but also that this critical section might extend beyond the end of the optimized synchronize_rcu(). This breaks RCU's fundamental guarantee. 2. In addition, num_online_cpus() does no ordering, thus providing another way that RCU's fundamental guarantee can be broken by the current code. 3. The most probable failure mode happens on outgoing CPUs. The outgoing CPU updates the count of online CPUs in the CPUHP_TEARDOWN_CPU stop-machine handler, which is fine in and of itself due to preemption being disabled at the call to num_online_cpus(). Unfortunately, after that stop-machine handler returns, the CPU takes one last trip through the scheduler (which has RCU readers) and, after the resulting context switch, one final dive into the idle loop. During this time, RCU needs to keep track of two CPUs, but num_online_cpus() will say that there is only one, which in turn means that the surviving CPU will incorrectly ignore the outgoing CPU's RCU read-side critical sections. This problem is illustrated by the following litmus test in which P0() corresponds to synchronize_rcu() and P1() corresponds to the incoming CPU. The herd7 tool confirms that the "exists" clause can be satisfied, thus demonstrating that this breakage can happen according to the Linux kernel memory model. { int x = 0; atomic_t numonline = ATOMIC_INIT(1); } P0(int *x, atomic_t *numonline) { int r0; WRITE_ONCE(*x, 1); r0 = atomic_read(numonline); if (r0 == 1) { smp_mb(); } else { synchronize_rcu(); } WRITE_ONCE(*x, 2); } P1(int *x, atomic_t *numonline) { int r0; int r1; atomic_inc(numonline); smp_mb(); rcu_read_lock(); r0 = READ_ONCE(*x); smp_rmb(); r1 = READ_ONCE(*x); rcu_read_unlock(); } locations [x;numonline;] exists (1:r0=0 /\ 1:r1=2) It is important to note that these problems arise only when the system is transitioning to or from single-CPU operation. One solution would be to hold the CPU-hotplug locks while sampling num_online_cpus(), which was in fact the intent of the (redundant) preempt_disable() and preempt_enable() surrounding this call to num_online_cpus(). Actually blocking CPU hotplug would not only result in excessive overhead, but would also unnecessarily impede CPU-hotplug operations. This commit therefore follows long-standing RCU tradition by maintaining a separate RCU-specific set of CPU-hotplug books. This separate set of books is implemented by a new ->n_online_cpus field in the rcu_state structure that maintains RCU's count of the online CPUs. This count is incremented early in the CPU-online process, so that the critical transition away from single-CPU operation will occur when there is only a single CPU. Similarly for the critical transition to single-CPU operation, the counter is decremented late in the CPU-offline process, again while there is only a single CPU. Because there is only ever a single CPU when the ->n_online_cpus field undergoes the critical 1->2 and 2->1 transitions, full memory ordering and mutual exclusion is provided implicitly and, better yet, for free. In the case where the CPU is coming online, nothing will happen until the current CPU helps it come online. Therefore, the new CPU will see all accesses prior to the optimized grace period, which means that RCU does not need to further delay this new CPU. In the case where the CPU is going offline, the outgoing CPU is totally out of the picture before the optimized grace period starts, which means that this outgoing CPU cannot see any of the accesses following that grace period. Again, RCU needs no further interaction with the outgoing CPU. This does mean that synchronize_rcu() will unnecessarily do a few grace periods the hard way just before the second CPU comes online and just after the second-to-last CPU goes offline, but it is not worth optimizing this uncommon case. Reviewed-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Neeraj Upadhyay <neeraju@codeaurora.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-09-23 15:29:33 +08:00
WRITE_ONCE(rcu_state.n_online_cpus, rcu_state.n_online_cpus + 1);
return 0;
}
/*
* Update RCU priority boot kthread affinity for CPU-hotplug changes.
*/
static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
{
struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
}
rcu-tasks: Stop rcu_tasks_invoke_cbs() from using never-onlined CPUs The rcu_tasks_invoke_cbs() function relies on queue_work_on() to silently fall back to WORK_CPU_UNBOUND when the specified CPU is offline. However, the queue_work_on() function's silent fallback mechanism relies on that CPU having been online at some time in the past. When queue_work_on() is passed a CPU that has never been online, workqueue lockups ensue, which can be bad for your kernel's general health and well-being. This commit therefore checks whether a given CPU has ever been online, and, if not substitutes WORK_CPU_UNBOUND in the subsequent call to queue_work_on(). Why not simply omit the queue_work_on() call entirely? Because this function is flooding callback-invocation notifications to all CPUs, and must deal with possibilities that include a sparse cpu_possible_mask. This commit also moves the setting of the rcu_data structure's ->beenonline field to rcu_cpu_starting(), which executes on the incoming CPU before that CPU has ever enabled interrupts. This ensures that the required workqueues are present. In addition, because the incoming CPU has not yet enabled its interrupts, there cannot yet have been any softirq handlers running on this CPU, which means that the WARN_ON_ONCE(!rdp->beenonline) within the RCU_SOFTIRQ handler cannot have triggered yet. Fixes: d363f833c6d88 ("rcu-tasks: Use workqueues for multiple rcu_tasks_invoke_cbs() invocations") Reported-by: Tejun Heo <tj@kernel.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2023-04-27 02:11:29 +08:00
/*
* Has the specified (known valid) CPU ever been fully online?
*/
bool rcu_cpu_beenfullyonline(int cpu)
{
struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
return smp_load_acquire(&rdp->beenonline);
}
/*
* Near the end of the CPU-online process. Pretty much all services
* enabled, and the CPU is now very much alive.
*/
int rcutree_online_cpu(unsigned int cpu)
{
unsigned long flags;
struct rcu_data *rdp;
struct rcu_node *rnp;
rdp = per_cpu_ptr(&rcu_data, cpu);
rnp = rdp->mynode;
raw_spin_lock_irqsave_rcu_node(rnp, flags);
rnp->ffmask |= rdp->grpmask;
raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
return 0; /* Too early in boot for scheduler work. */
sync_sched_exp_online_cleanup(cpu);
rcutree_affinity_setting(cpu, -1);
// Stop-machine done, so allow nohz_full to disable tick.
tick_dep_clear(TICK_DEP_BIT_RCU);
return 0;
}
/*
* Near the beginning of the process. The CPU is still very much alive
* with pretty much all services enabled.
*/
int rcutree_offline_cpu(unsigned int cpu)
{
unsigned long flags;
struct rcu_data *rdp;
struct rcu_node *rnp;
rdp = per_cpu_ptr(&rcu_data, cpu);
rnp = rdp->mynode;
raw_spin_lock_irqsave_rcu_node(rnp, flags);
rnp->ffmask &= ~rdp->grpmask;
raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
rcutree_affinity_setting(cpu, cpu);
// nohz_full CPUs need the tick for stop-machine to work quickly
tick_dep_set(TICK_DEP_BIT_RCU);
return 0;
}
rcu: Provide exact CPU-online tracking for RCU Up to now, RCU has assumed that the CPU-online process makes it from CPU_UP_PREPARE to set_cpu_online() within one jiffy. Given the recent rise of virtualized environments, this assumption is very clearly obsolete. Failing to meet this deadline can result in RCU paying attention to an incoming CPU for one jiffy, then ignoring it until the grace period following the one in which that CPU sets itself online. This situation might prove to be fatally disappointing to any RCU read-side critical sections that had the misfortune to execute during the time in which RCU was ignoring the slow-to-come-online CPU. This commit therefore updates RCU's internal CPU state-tracking information at notify_cpu_starting() time, thus providing RCU with an exact transition of the CPU's state from offline to online. Note that this means that incoming CPUs must not use RCU read-side critical section (other than those of SRCU) until notify_cpu_starting() time. Note also that the CPU_STARTING notifiers -are- allowed to use RCU read-side critical sections. (Of course, CPU-hotplug notifiers are rapidly becoming obsolete, so you need to act fast!) If a given architecture or CPU family needs to use RCU read-side critical sections earlier, the call to rcu_cpu_starting() from notify_cpu_starting() will need to be architecture-specific, with architectures that need early use being required to hand-place the call to rcu_cpu_starting() at some point preceding the call to notify_cpu_starting(). Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2016-07-01 04:58:26 +08:00
/*
* Mark the specified CPU as being online so that subsequent grace periods
* (both expedited and normal) will wait on it. Note that this means that
* incoming CPUs are not allowed to use RCU read-side critical sections
* until this function is called. Failing to observe this restriction
* will result in lockdep splats.
*
* Note that this function is special in that it is invoked directly
* from the incoming CPU rather than from the cpuhp_step mechanism.
* This is because this function must be invoked at a precise location.
* This incoming CPU must not have enabled interrupts yet.
rcu: Provide exact CPU-online tracking for RCU Up to now, RCU has assumed that the CPU-online process makes it from CPU_UP_PREPARE to set_cpu_online() within one jiffy. Given the recent rise of virtualized environments, this assumption is very clearly obsolete. Failing to meet this deadline can result in RCU paying attention to an incoming CPU for one jiffy, then ignoring it until the grace period following the one in which that CPU sets itself online. This situation might prove to be fatally disappointing to any RCU read-side critical sections that had the misfortune to execute during the time in which RCU was ignoring the slow-to-come-online CPU. This commit therefore updates RCU's internal CPU state-tracking information at notify_cpu_starting() time, thus providing RCU with an exact transition of the CPU's state from offline to online. Note that this means that incoming CPUs must not use RCU read-side critical section (other than those of SRCU) until notify_cpu_starting() time. Note also that the CPU_STARTING notifiers -are- allowed to use RCU read-side critical sections. (Of course, CPU-hotplug notifiers are rapidly becoming obsolete, so you need to act fast!) If a given architecture or CPU family needs to use RCU read-side critical sections earlier, the call to rcu_cpu_starting() from notify_cpu_starting() will need to be architecture-specific, with architectures that need early use being required to hand-place the call to rcu_cpu_starting() at some point preceding the call to notify_cpu_starting(). Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2016-07-01 04:58:26 +08:00
*/
void rcu_cpu_starting(unsigned int cpu)
{
unsigned long mask;
struct rcu_data *rdp;
struct rcu_node *rnp;
bool newcpu;
rcu: Provide exact CPU-online tracking for RCU Up to now, RCU has assumed that the CPU-online process makes it from CPU_UP_PREPARE to set_cpu_online() within one jiffy. Given the recent rise of virtualized environments, this assumption is very clearly obsolete. Failing to meet this deadline can result in RCU paying attention to an incoming CPU for one jiffy, then ignoring it until the grace period following the one in which that CPU sets itself online. This situation might prove to be fatally disappointing to any RCU read-side critical sections that had the misfortune to execute during the time in which RCU was ignoring the slow-to-come-online CPU. This commit therefore updates RCU's internal CPU state-tracking information at notify_cpu_starting() time, thus providing RCU with an exact transition of the CPU's state from offline to online. Note that this means that incoming CPUs must not use RCU read-side critical section (other than those of SRCU) until notify_cpu_starting() time. Note also that the CPU_STARTING notifiers -are- allowed to use RCU read-side critical sections. (Of course, CPU-hotplug notifiers are rapidly becoming obsolete, so you need to act fast!) If a given architecture or CPU family needs to use RCU read-side critical sections earlier, the call to rcu_cpu_starting() from notify_cpu_starting() will need to be architecture-specific, with architectures that need early use being required to hand-place the call to rcu_cpu_starting() at some point preceding the call to notify_cpu_starting(). Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2016-07-01 04:58:26 +08:00
lockdep_assert_irqs_disabled();
rdp = per_cpu_ptr(&rcu_data, cpu);
if (rdp->cpu_started)
return;
rdp->cpu_started = true;
rnp = rdp->mynode;
mask = rdp->grpmask;
arch_spin_lock(&rcu_state.ofl_lock);
rcu: Move rcu_dynticks_eqs_online() to rcu_cpu_starting() The purpose of rcu_dynticks_eqs_online() is to adjust the ->dynticks counter of an incoming CPU when required. It is currently invoked from rcutree_prepare_cpu(), which runs before the incoming CPU is running, and thus on some other CPU. This makes the per-CPU accesses in rcu_dynticks_eqs_online() iffy at best, and it all "works" only because the running CPU cannot possibly be in dyntick-idle mode, which means that rcu_dynticks_eqs_online() never has any effect. It is currently OK for rcu_dynticks_eqs_online() to have no effect, but only because the CPU-offline process just happens to leave ->dynticks in the correct state. After all, if ->dynticks were in the wrong state on a just-onlined CPU, rcutorture would complain bitterly the next time that CPU went idle, at least in kernels built with CONFIG_RCU_EQS_DEBUG=y, for example, those built by rcutorture scenario TREE04. One could argue that this means that rcu_dynticks_eqs_online() is unnecessary, however, removing it would make the CPU-online process vulnerable to slight changes in the CPU-offline process. One could also ask why it is safe to move the rcu_dynticks_eqs_online() call so late in the CPU-online process. Indeed, there was a time when it would not have been safe, which does much to explain its current location. However, the marking of a CPU as online from an RCU perspective has long since moved from rcutree_prepare_cpu() to rcu_cpu_starting(), and all that is required is that ->dynticks be set correctly by the time that the CPU is marked as online from an RCU perspective. After all, the RCU grace-period kthread does not check to see if offline CPUs are also idle. (In case you were curious, this is one reason why there is quiescent-state reporting as part of the offlining process.) This commit therefore moves the call to rcu_dynticks_eqs_online() from rcutree_prepare_cpu() to rcu_cpu_starting(), this latter being guaranteed to be running on the incoming CPU. The call to this function must of course be placed before this rcu_cpu_starting() announces this CPU's presence to RCU. Reported-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2021-07-29 03:38:42 +08:00
rcu_dynticks_eqs_online();
rcu: Make rcu_barrier() no longer block CPU-hotplug operations This commit removes the cpus_read_lock() and cpus_read_unlock() calls from rcu_barrier(), thus allowing CPUs to come and go during the course of rcu_barrier() execution. Posting of the ->barrier_head callbacks does synchronize with portions of RCU's CPU-hotplug notifiers, but these locks are held for short time periods on both sides. Thus, full CPU-hotplug operations could both start and finish during the execution of a given rcu_barrier() invocation. Additional synchronization is provided by a global ->barrier_lock. Since the ->barrier_lock is only used during rcu_barrier() execution and during onlining/offlining a CPU, the contention for this lock should be low. It might be tempting to make use of a per-CPU lock just on general principles, but straightforward attempts to do this have the problems shown below. Initial state: 3 CPUs present, CPU 0 and CPU1 do not have any callback and CPU2 has callbacks. 1. CPU0 calls rcu_barrier(). 2. CPU1 starts offlining for CPU2. CPU1 calls rcutree_migrate_callbacks(). rcu_barrier_entrain() is called from rcutree_migrate_callbacks(), with CPU2's rdp->barrier_lock. It does not entrain ->barrier_head for CPU2, as rcu_barrier() on CPU0 hasn't started the barrier sequence (by calling rcu_seq_start(&rcu_state.barrier_sequence)) yet. 3. CPU0 starts new barrier sequence. It iterates over CPU0 and CPU1, after acquiring their per-cpu ->barrier_lock and finds 0 segcblist length. It updates ->barrier_seq_snap for CPU0 and CPU1 and continues loop iteration to CPU2. for_each_possible_cpu(cpu) { raw_spin_lock_irqsave(&rdp->barrier_lock, flags); if (!rcu_segcblist_n_cbs(&rdp->cblist)) { WRITE_ONCE(rdp->barrier_seq_snap, gseq); raw_spin_unlock_irqrestore(&rdp->barrier_lock, flags); rcu_barrier_trace(TPS("NQ"), cpu, rcu_state.barrier_sequence); continue; } 4. rcutree_migrate_callbacks() completes execution on CPU1. Segcblist len for CPU2 becomes 0. 5. The loop iteration on CPU0, checks rcu_segcblist_n_cbs(&rdp->cblist) for CPU2 and completes the loop iteration after setting ->barrier_seq_snap. 6. As there isn't any ->barrier_head callback entrained; at this point, rcu_barrier() in CPU0 returns. 7. The callbacks, which migrated from CPU2 to CPU1, execute. Straightforward per-CPU locking is also subject to the following race condition noted by Boqun Feng: 1. CPU0 calls rcu_barrier(), starting a new barrier sequence by invoking rcu_seq_start() and init_completion(), but does not yet initialize rcu_state.barrier_cpu_count. 2. CPU1 starts offlining for CPU2, calling rcutree_migrate_callbacks(), which in turn calls rcu_barrier_entrain() holding CPU2's. rdp->barrier_lock. It then entrains ->barrier_head for CPU2 and atomically increments rcu_state.barrier_cpu_count, which is unfortunately not yet initialized to the value 2. 3. The just-entrained RCU callback is invoked. It atomically decrements rcu_state.barrier_cpu_count and sees that it is now zero. This callback therefore invokes complete(). 4. CPU0 continues executing rcu_barrier(), but is not blocked by its call to wait_for_completion(). This results in rcu_barrier() returning before all pre-existing callbacks have been invoked, which is a bug. Therefore, synchronization is provided by rcu_state.barrier_lock, which is also held across the initialization sequence, especially the rcu_seq_start() and the atomic_set() that sets rcu_state.barrier_cpu_count to the value 2. In addition, this lock is held when entraining the rcu_barrier() callback, when deciding whether or not a CPU has callbacks that rcu_barrier() must wait on, when setting the ->qsmaskinitnext for incoming CPUs, and when migrating callbacks from a CPU that is going offline. Reviewed-by: Frederic Weisbecker <frederic@kernel.org> Co-developed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com> Signed-off-by: Neeraj Upadhyay <quic_neeraju@quicinc.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2021-12-15 05:35:17 +08:00
raw_spin_lock(&rcu_state.barrier_lock);
raw_spin_lock_rcu_node(rnp);
WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext | mask);
rcu: Make rcu_barrier() no longer block CPU-hotplug operations This commit removes the cpus_read_lock() and cpus_read_unlock() calls from rcu_barrier(), thus allowing CPUs to come and go during the course of rcu_barrier() execution. Posting of the ->barrier_head callbacks does synchronize with portions of RCU's CPU-hotplug notifiers, but these locks are held for short time periods on both sides. Thus, full CPU-hotplug operations could both start and finish during the execution of a given rcu_barrier() invocation. Additional synchronization is provided by a global ->barrier_lock. Since the ->barrier_lock is only used during rcu_barrier() execution and during onlining/offlining a CPU, the contention for this lock should be low. It might be tempting to make use of a per-CPU lock just on general principles, but straightforward attempts to do this have the problems shown below. Initial state: 3 CPUs present, CPU 0 and CPU1 do not have any callback and CPU2 has callbacks. 1. CPU0 calls rcu_barrier(). 2. CPU1 starts offlining for CPU2. CPU1 calls rcutree_migrate_callbacks(). rcu_barrier_entrain() is called from rcutree_migrate_callbacks(), with CPU2's rdp->barrier_lock. It does not entrain ->barrier_head for CPU2, as rcu_barrier() on CPU0 hasn't started the barrier sequence (by calling rcu_seq_start(&rcu_state.barrier_sequence)) yet. 3. CPU0 starts new barrier sequence. It iterates over CPU0 and CPU1, after acquiring their per-cpu ->barrier_lock and finds 0 segcblist length. It updates ->barrier_seq_snap for CPU0 and CPU1 and continues loop iteration to CPU2. for_each_possible_cpu(cpu) { raw_spin_lock_irqsave(&rdp->barrier_lock, flags); if (!rcu_segcblist_n_cbs(&rdp->cblist)) { WRITE_ONCE(rdp->barrier_seq_snap, gseq); raw_spin_unlock_irqrestore(&rdp->barrier_lock, flags); rcu_barrier_trace(TPS("NQ"), cpu, rcu_state.barrier_sequence); continue; } 4. rcutree_migrate_callbacks() completes execution on CPU1. Segcblist len for CPU2 becomes 0. 5. The loop iteration on CPU0, checks rcu_segcblist_n_cbs(&rdp->cblist) for CPU2 and completes the loop iteration after setting ->barrier_seq_snap. 6. As there isn't any ->barrier_head callback entrained; at this point, rcu_barrier() in CPU0 returns. 7. The callbacks, which migrated from CPU2 to CPU1, execute. Straightforward per-CPU locking is also subject to the following race condition noted by Boqun Feng: 1. CPU0 calls rcu_barrier(), starting a new barrier sequence by invoking rcu_seq_start() and init_completion(), but does not yet initialize rcu_state.barrier_cpu_count. 2. CPU1 starts offlining for CPU2, calling rcutree_migrate_callbacks(), which in turn calls rcu_barrier_entrain() holding CPU2's. rdp->barrier_lock. It then entrains ->barrier_head for CPU2 and atomically increments rcu_state.barrier_cpu_count, which is unfortunately not yet initialized to the value 2. 3. The just-entrained RCU callback is invoked. It atomically decrements rcu_state.barrier_cpu_count and sees that it is now zero. This callback therefore invokes complete(). 4. CPU0 continues executing rcu_barrier(), but is not blocked by its call to wait_for_completion(). This results in rcu_barrier() returning before all pre-existing callbacks have been invoked, which is a bug. Therefore, synchronization is provided by rcu_state.barrier_lock, which is also held across the initialization sequence, especially the rcu_seq_start() and the atomic_set() that sets rcu_state.barrier_cpu_count to the value 2. In addition, this lock is held when entraining the rcu_barrier() callback, when deciding whether or not a CPU has callbacks that rcu_barrier() must wait on, when setting the ->qsmaskinitnext for incoming CPUs, and when migrating callbacks from a CPU that is going offline. Reviewed-by: Frederic Weisbecker <frederic@kernel.org> Co-developed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com> Signed-off-by: Neeraj Upadhyay <quic_neeraju@quicinc.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2021-12-15 05:35:17 +08:00
raw_spin_unlock(&rcu_state.barrier_lock);
newcpu = !(rnp->expmaskinitnext & mask);
rnp->expmaskinitnext |= mask;
/* Allow lockless access for expedited grace periods. */
smp_store_release(&rcu_state.ncpus, rcu_state.ncpus + newcpu); /* ^^^ */
ASSERT_EXCLUSIVE_WRITER(rcu_state.ncpus);
rcu_gpnum_ovf(rnp, rdp); /* Offline-induced counter wrap? */
rdp->rcu_onl_gp_seq = READ_ONCE(rcu_state.gp_seq);
rdp->rcu_onl_gp_flags = READ_ONCE(rcu_state.gp_flags);
/* An incoming CPU should never be blocking a grace period. */
if (WARN_ON_ONCE(rnp->qsmask & mask)) { /* RCU waiting on incoming CPU? */
/* rcu_report_qs_rnp() *really* wants some flags to restore */
unsigned long flags;
local_irq_save(flags);
rcu_disable_urgency_upon_qs(rdp);
/* Report QS -after- changing ->qsmaskinitnext! */
rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
} else {
raw_spin_unlock_rcu_node(rnp);
rcu: Provide exact CPU-online tracking for RCU Up to now, RCU has assumed that the CPU-online process makes it from CPU_UP_PREPARE to set_cpu_online() within one jiffy. Given the recent rise of virtualized environments, this assumption is very clearly obsolete. Failing to meet this deadline can result in RCU paying attention to an incoming CPU for one jiffy, then ignoring it until the grace period following the one in which that CPU sets itself online. This situation might prove to be fatally disappointing to any RCU read-side critical sections that had the misfortune to execute during the time in which RCU was ignoring the slow-to-come-online CPU. This commit therefore updates RCU's internal CPU state-tracking information at notify_cpu_starting() time, thus providing RCU with an exact transition of the CPU's state from offline to online. Note that this means that incoming CPUs must not use RCU read-side critical section (other than those of SRCU) until notify_cpu_starting() time. Note also that the CPU_STARTING notifiers -are- allowed to use RCU read-side critical sections. (Of course, CPU-hotplug notifiers are rapidly becoming obsolete, so you need to act fast!) If a given architecture or CPU family needs to use RCU read-side critical sections earlier, the call to rcu_cpu_starting() from notify_cpu_starting() will need to be architecture-specific, with architectures that need early use being required to hand-place the call to rcu_cpu_starting() at some point preceding the call to notify_cpu_starting(). Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2016-07-01 04:58:26 +08:00
}
arch_spin_unlock(&rcu_state.ofl_lock);
rcu-tasks: Stop rcu_tasks_invoke_cbs() from using never-onlined CPUs The rcu_tasks_invoke_cbs() function relies on queue_work_on() to silently fall back to WORK_CPU_UNBOUND when the specified CPU is offline. However, the queue_work_on() function's silent fallback mechanism relies on that CPU having been online at some time in the past. When queue_work_on() is passed a CPU that has never been online, workqueue lockups ensue, which can be bad for your kernel's general health and well-being. This commit therefore checks whether a given CPU has ever been online, and, if not substitutes WORK_CPU_UNBOUND in the subsequent call to queue_work_on(). Why not simply omit the queue_work_on() call entirely? Because this function is flooding callback-invocation notifications to all CPUs, and must deal with possibilities that include a sparse cpu_possible_mask. This commit also moves the setting of the rcu_data structure's ->beenonline field to rcu_cpu_starting(), which executes on the incoming CPU before that CPU has ever enabled interrupts. This ensures that the required workqueues are present. In addition, because the incoming CPU has not yet enabled its interrupts, there cannot yet have been any softirq handlers running on this CPU, which means that the WARN_ON_ONCE(!rdp->beenonline) within the RCU_SOFTIRQ handler cannot have triggered yet. Fixes: d363f833c6d88 ("rcu-tasks: Use workqueues for multiple rcu_tasks_invoke_cbs() invocations") Reported-by: Tejun Heo <tj@kernel.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2023-04-27 02:11:29 +08:00
smp_store_release(&rdp->beenonline, true);
rcu: Make expedited GPs correctly handle hardware CPU insertion The update of the ->expmaskinitnext and of ->ncpus are unsynchronized, with the value of ->ncpus being incremented long before the corresponding ->expmaskinitnext mask is updated. If an RCU expedited grace period sees ->ncpus change, it will update the ->expmaskinit masks from the new ->expmaskinitnext masks. But it is possible that ->ncpus has already been updated, but the ->expmaskinitnext masks still have their old values. For the current expedited grace period, no harm done. The CPU could not have been online before the grace period started, so there is no need to wait for its non-existent pre-existing readers. But the next RCU expedited grace period is in a world of hurt. The value of ->ncpus has already been updated, so this grace period will assume that the ->expmaskinitnext masks have not changed. But they have, and they won't be taken into account until the next never-been-online CPU comes online. This means that RCU will be ignoring some CPUs that it should be paying attention to. The solution is to update ->ncpus and ->expmaskinitnext while holding the ->lock for the rcu_node structure containing the ->expmaskinitnext mask. Because smp_store_release() is now used to update ->ncpus and smp_load_acquire() is now used to locklessly read it, if the expedited grace period sees ->ncpus change, then the updating CPU has to already be holding the corresponding ->lock. Therefore, when the expedited grace period later acquires that ->lock, it is guaranteed to see the new value of ->expmaskinitnext. On the other hand, if the expedited grace period loads ->ncpus just before an update, earlier full memory barriers guarantee that the incoming CPU isn't far enough along to be running any RCU readers. This commit therefore makes the required change. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2017-06-09 07:55:40 +08:00
smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
rcu: Provide exact CPU-online tracking for RCU Up to now, RCU has assumed that the CPU-online process makes it from CPU_UP_PREPARE to set_cpu_online() within one jiffy. Given the recent rise of virtualized environments, this assumption is very clearly obsolete. Failing to meet this deadline can result in RCU paying attention to an incoming CPU for one jiffy, then ignoring it until the grace period following the one in which that CPU sets itself online. This situation might prove to be fatally disappointing to any RCU read-side critical sections that had the misfortune to execute during the time in which RCU was ignoring the slow-to-come-online CPU. This commit therefore updates RCU's internal CPU state-tracking information at notify_cpu_starting() time, thus providing RCU with an exact transition of the CPU's state from offline to online. Note that this means that incoming CPUs must not use RCU read-side critical section (other than those of SRCU) until notify_cpu_starting() time. Note also that the CPU_STARTING notifiers -are- allowed to use RCU read-side critical sections. (Of course, CPU-hotplug notifiers are rapidly becoming obsolete, so you need to act fast!) If a given architecture or CPU family needs to use RCU read-side critical sections earlier, the call to rcu_cpu_starting() from notify_cpu_starting() will need to be architecture-specific, with architectures that need early use being required to hand-place the call to rcu_cpu_starting() at some point preceding the call to notify_cpu_starting(). Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2016-07-01 04:58:26 +08:00
}
/*
* The outgoing function has no further need of RCU, so remove it from
* the rcu_node tree's ->qsmaskinitnext bit masks.
*
* Note that this function is special in that it is invoked directly
* from the outgoing CPU rather than from the cpuhp_step mechanism.
* This is because this function must be invoked at a precise location.
*/
void rcu_report_dead(unsigned int cpu)
{
unsigned long flags, seq_flags;
unsigned long mask;
struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
// Do any dangling deferred wakeups.
do_nocb_deferred_wakeup(rdp);
rcu_preempt_deferred_qs(current);
/* Remove outgoing CPU from mask in the leaf rcu_node structure. */
mask = rdp->grpmask;
local_irq_save(seq_flags);
arch_spin_lock(&rcu_state.ofl_lock);
raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
rdp->rcu_ofl_gp_seq = READ_ONCE(rcu_state.gp_seq);
rdp->rcu_ofl_gp_flags = READ_ONCE(rcu_state.gp_flags);
rcu: Suppress false-positive preempted-task splats Consider the following sequence of events in a PREEMPT=y kernel: 1. All CPUs corresponding to a given rcu_node structure go offline. A new grace period starts just after the CPU-hotplug code path does its synchronize_rcu() for the last CPU, so at least this CPU is present in that structure's ->qsmask. 2. Before the grace period ends, a CPU comes back online, and not just any CPU, but the one corresponding to a non-zero bit in the leaf rcu_node structure's ->qsmask. 3. A task running on the newly onlined CPU is preempted while in an RCU read-side critical section. Because this CPU's ->qsmask bit is net, not only does this task queue itself on the leaf rcu_node structure's ->blkd_tasks list, it also sets that structure's ->gp_tasks pointer to reference it. 4. The grace period started in #1 above comes to an end. This results in rcu_gp_cleanup() being invoked, which, among other things, checks to make sure that there are no tasks blocking the just-ended grace period, that is, that all ->gp_tasks pointers are NULL. The ->gp_tasks pointer corresponding to the task preempted in #3 above is non-NULL, which results in a splat. This splat is a false positive. The task's RCU read-side critical section cannot have begun before the just-ended grace period because this would mean either: (1) The CPU came online before the grace period started, which cannot have happened because the grace period started before that CPU was all the way offline, or (2) The task started its RCU read-side critical section on some other CPU, but then it would have had to have been preempted before migrating to this CPU, which would mean that it would have instead queued itself on that other CPU's rcu_node structure. This commit eliminates this false positive by adding code to the end of rcu_cleanup_dying_idle_cpu() that reports a quiescent state to RCU, which has the side-effect of clearing that CPU's ->qsmask bit, preventing the above scenario. This approach has the added benefit of more promptly reporting quiescent states corresponding to offline CPUs. Note well that the call to rcu_report_qs_rnp() reporting the quiescent state must come -before- the clearing of this CPU's bit in the leaf rcu_node structure's ->qsmaskinitnext field. Otherwise, lockdep-RCU will complain bitterly about quiescent states coming from an offline CPU. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2018-05-03 11:04:12 +08:00
if (rnp->qsmask & mask) { /* RCU waiting on outgoing CPU? */
/* Report quiescent state -before- changing ->qsmaskinitnext! */
rcu_disable_urgency_upon_qs(rdp);
rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
rcu: Suppress false-positive preempted-task splats Consider the following sequence of events in a PREEMPT=y kernel: 1. All CPUs corresponding to a given rcu_node structure go offline. A new grace period starts just after the CPU-hotplug code path does its synchronize_rcu() for the last CPU, so at least this CPU is present in that structure's ->qsmask. 2. Before the grace period ends, a CPU comes back online, and not just any CPU, but the one corresponding to a non-zero bit in the leaf rcu_node structure's ->qsmask. 3. A task running on the newly onlined CPU is preempted while in an RCU read-side critical section. Because this CPU's ->qsmask bit is net, not only does this task queue itself on the leaf rcu_node structure's ->blkd_tasks list, it also sets that structure's ->gp_tasks pointer to reference it. 4. The grace period started in #1 above comes to an end. This results in rcu_gp_cleanup() being invoked, which, among other things, checks to make sure that there are no tasks blocking the just-ended grace period, that is, that all ->gp_tasks pointers are NULL. The ->gp_tasks pointer corresponding to the task preempted in #3 above is non-NULL, which results in a splat. This splat is a false positive. The task's RCU read-side critical section cannot have begun before the just-ended grace period because this would mean either: (1) The CPU came online before the grace period started, which cannot have happened because the grace period started before that CPU was all the way offline, or (2) The task started its RCU read-side critical section on some other CPU, but then it would have had to have been preempted before migrating to this CPU, which would mean that it would have instead queued itself on that other CPU's rcu_node structure. This commit eliminates this false positive by adding code to the end of rcu_cleanup_dying_idle_cpu() that reports a quiescent state to RCU, which has the side-effect of clearing that CPU's ->qsmask bit, preventing the above scenario. This approach has the added benefit of more promptly reporting quiescent states corresponding to offline CPUs. Note well that the call to rcu_report_qs_rnp() reporting the quiescent state must come -before- the clearing of this CPU's bit in the leaf rcu_node structure's ->qsmaskinitnext field. Otherwise, lockdep-RCU will complain bitterly about quiescent states coming from an offline CPU. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2018-05-03 11:04:12 +08:00
raw_spin_lock_irqsave_rcu_node(rnp, flags);
}
WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext & ~mask);
Merge branch 'smp-hotplug-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip Pull cpu hotplug updates from Thomas Gleixner: "This is the first part of the ongoing cpu hotplug rework: - Initial implementation of the state machine - Runs all online and prepare down callbacks on the plugged cpu and not on some random processor - Replaces busy loop waiting with completions - Adds tracepoints so the states can be followed" More detailed commentary on this work from an earlier email: "What's wrong with the current cpu hotplug infrastructure? - Asymmetry The hotplug notifier mechanism is asymmetric versus the bringup and teardown. This is mostly caused by the notifier mechanism. - Largely undocumented dependencies While some notifiers use explicitely defined notifier priorities, we have quite some notifiers which use numerical priorities to express dependencies without any documentation why. - Control processor driven Most of the bringup/teardown of a cpu is driven by a control processor. While it is understandable, that preperatory steps, like idle thread creation, memory allocation for and initialization of essential facilities needs to be done before a cpu can boot, there is no reason why everything else must run on a control processor. Before this patch series, bringup looks like this: Control CPU Booting CPU do preparatory steps kick cpu into life do low level init sync with booting cpu sync with control cpu bring the rest up - All or nothing approach There is no way to do partial bringups. That's something which is really desired because we waste e.g. at boot substantial amount of time just busy waiting that the cpu comes to life. That's stupid as we could very well do preparatory steps and the initial IPI for other cpus and then go back and do the necessary low level synchronization with the freshly booted cpu. - Minimal debuggability Due to the notifier based design, it's impossible to switch between two stages of the bringup/teardown back and forth in order to test the correctness. So in many hotplug notifiers the cancel mechanisms are either not existant or completely untested. - Notifier [un]registering is tedious To [un]register notifiers we need to protect against hotplug at every callsite. There is no mechanism that bringup/teardown callbacks are issued on the online cpus, so every caller needs to do it itself. That also includes error rollback. What's the new design? The base of the new design is a symmetric state machine, where both the control processor and the booting/dying cpu execute a well defined set of states. Each state is symmetric in the end, except for some well defined exceptions, and the bringup/teardown can be stopped and reversed at almost all states. So the bringup of a cpu will look like this in the future: Control CPU Booting CPU do preparatory steps kick cpu into life do low level init sync with booting cpu sync with control cpu bring itself up The synchronization step does not require the control cpu to wait. That mechanism can be done asynchronously via a worker or some other mechanism. The teardown can be made very similar, so that the dying cpu cleans up and brings itself down. Cleanups which need to be done after the cpu is gone, can be scheduled asynchronously as well. There is a long way to this, as we need to refactor the notion when a cpu is available. Today we set the cpu online right after it comes out of the low level bringup, which is not really correct. The proper mechanism is to set it to available, i.e. cpu local threads, like softirqd, hotplug thread etc. can be scheduled on that cpu, and once it finished all booting steps, it's set to online, so general workloads can be scheduled on it. The reverse happens on teardown. First thing to do is to forbid scheduling of general workloads, then teardown all the per cpu resources and finally shut it off completely. This patch series implements the basic infrastructure for this at the core level. This includes the following: - Basic state machine implementation with well defined states, so ordering and prioritization can be expressed. - Interfaces to [un]register state callbacks This invokes the bringup/teardown callback on all online cpus with the proper protection in place and [un]installs the callbacks in the state machine array. For callbacks which have no particular ordering requirement we have a dynamic state space, so that drivers don't have to register an explicit hotplug state. If a callback fails, the code automatically does a rollback to the previous state. - Sysfs interface to drive the state machine to a particular step. This is only partially functional today. Full functionality and therefor testability will be achieved once we converted all existing hotplug notifiers over to the new scheme. - Run all CPU_ONLINE/DOWN_PREPARE notifiers on the booting/dying processor: Control CPU Booting CPU do preparatory steps kick cpu into life do low level init sync with booting cpu sync with control cpu wait for boot bring itself up Signal completion to control cpu In a previous step of this work we've done a full tree mechanical conversion of all hotplug notifiers to the new scheme. The balance is a net removal of about 4000 lines of code. This is not included in this series, as we decided to take a different approach. Instead of mechanically converting everything over, we will do a proper overhaul of the usage sites one by one so they nicely fit into the symmetric callback scheme. I decided to do that after I looked at the ugliness of some of the converted sites and figured out that their hotplug mechanism is completely buggered anyway. So there is no point to do a mechanical conversion first as we need to go through the usage sites one by one again in order to achieve a full symmetric and testable behaviour" * 'smp-hotplug-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (23 commits) cpu/hotplug: Document states better cpu/hotplug: Fix smpboot thread ordering cpu/hotplug: Remove redundant state check cpu/hotplug: Plug death reporting race rcu: Make CPU_DYING_IDLE an explicit call cpu/hotplug: Make wait for dead cpu completion based cpu/hotplug: Let upcoming cpu bring itself fully up arch/hotplug: Call into idle with a proper state cpu/hotplug: Move online calls to hotplugged cpu cpu/hotplug: Create hotplug threads cpu/hotplug: Split out the state walk into functions cpu/hotplug: Unpark smpboot threads from the state machine cpu/hotplug: Move scheduler cpu_online notifier to hotplug core cpu/hotplug: Implement setup/removal interface cpu/hotplug: Make target state writeable cpu/hotplug: Add sysfs state interface cpu/hotplug: Hand in target state to _cpu_up/down cpu/hotplug: Convert the hotplugged cpu work to a state machine cpu/hotplug: Convert to a state machine for the control processor cpu/hotplug: Add tracepoints ...
2016-03-16 04:50:29 +08:00
raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
arch_spin_unlock(&rcu_state.ofl_lock);
local_irq_restore(seq_flags);
rdp->cpu_started = false;
}
rcu: Migrate callbacks earlier in the CPU-offline timeline RCU callbacks must be migrated away from an outgoing CPU, and this is done near the end of the CPU-hotplug operation, after the outgoing CPU is long gone. Unfortunately, this means that other CPU-hotplug callbacks can execute while the outgoing CPU's callbacks are still immobilized on the long-gone CPU's callback lists. If any of these CPU-hotplug callbacks must wait, either directly or indirectly, for the invocation of any of the immobilized RCU callbacks, the system will hang. This commit avoids such hangs by migrating the callbacks away from the outgoing CPU immediately upon its departure, shortly after the return from __cpu_die() in takedown_cpu(). Thus, RCU is able to advance these callbacks and invoke them, which allows all the after-the-fact CPU-hotplug callbacks to wait on these RCU callbacks without risk of a hang. While in the neighborhood, this commit also moves rcu_send_cbs_to_orphanage() and rcu_adopt_orphan_cbs() under a pre-existing #ifdef to avoid including dead code on the one hand and to avoid define-without-use warnings on the other hand. Reported-by: Jeffrey Hugo <jhugo@codeaurora.org> Link: http://lkml.kernel.org/r/db9c91f6-1b17-6136-84f0-03c3c2581ab4@codeaurora.org Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Cc: Ingo Molnar <mingo@kernel.org> Cc: Anna-Maria Gleixner <anna-maria@linutronix.de> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Richard Weinberger <richard@nod.at>
2017-06-21 03:11:34 +08:00
#ifdef CONFIG_HOTPLUG_CPU
/*
* The outgoing CPU has just passed through the dying-idle state, and we
* are being invoked from the CPU that was IPIed to continue the offline
* operation. Migrate the outgoing CPU's callbacks to the current CPU.
*/
void rcutree_migrate_callbacks(int cpu)
rcu: Migrate callbacks earlier in the CPU-offline timeline RCU callbacks must be migrated away from an outgoing CPU, and this is done near the end of the CPU-hotplug operation, after the outgoing CPU is long gone. Unfortunately, this means that other CPU-hotplug callbacks can execute while the outgoing CPU's callbacks are still immobilized on the long-gone CPU's callback lists. If any of these CPU-hotplug callbacks must wait, either directly or indirectly, for the invocation of any of the immobilized RCU callbacks, the system will hang. This commit avoids such hangs by migrating the callbacks away from the outgoing CPU immediately upon its departure, shortly after the return from __cpu_die() in takedown_cpu(). Thus, RCU is able to advance these callbacks and invoke them, which allows all the after-the-fact CPU-hotplug callbacks to wait on these RCU callbacks without risk of a hang. While in the neighborhood, this commit also moves rcu_send_cbs_to_orphanage() and rcu_adopt_orphan_cbs() under a pre-existing #ifdef to avoid including dead code on the one hand and to avoid define-without-use warnings on the other hand. Reported-by: Jeffrey Hugo <jhugo@codeaurora.org> Link: http://lkml.kernel.org/r/db9c91f6-1b17-6136-84f0-03c3c2581ab4@codeaurora.org Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Cc: Ingo Molnar <mingo@kernel.org> Cc: Anna-Maria Gleixner <anna-maria@linutronix.de> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Richard Weinberger <richard@nod.at>
2017-06-21 03:11:34 +08:00
{
unsigned long flags;
struct rcu_data *my_rdp;
struct rcu_node *my_rnp;
struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
bool needwake;
rcu: Migrate callbacks earlier in the CPU-offline timeline RCU callbacks must be migrated away from an outgoing CPU, and this is done near the end of the CPU-hotplug operation, after the outgoing CPU is long gone. Unfortunately, this means that other CPU-hotplug callbacks can execute while the outgoing CPU's callbacks are still immobilized on the long-gone CPU's callback lists. If any of these CPU-hotplug callbacks must wait, either directly or indirectly, for the invocation of any of the immobilized RCU callbacks, the system will hang. This commit avoids such hangs by migrating the callbacks away from the outgoing CPU immediately upon its departure, shortly after the return from __cpu_die() in takedown_cpu(). Thus, RCU is able to advance these callbacks and invoke them, which allows all the after-the-fact CPU-hotplug callbacks to wait on these RCU callbacks without risk of a hang. While in the neighborhood, this commit also moves rcu_send_cbs_to_orphanage() and rcu_adopt_orphan_cbs() under a pre-existing #ifdef to avoid including dead code on the one hand and to avoid define-without-use warnings on the other hand. Reported-by: Jeffrey Hugo <jhugo@codeaurora.org> Link: http://lkml.kernel.org/r/db9c91f6-1b17-6136-84f0-03c3c2581ab4@codeaurora.org Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Cc: Ingo Molnar <mingo@kernel.org> Cc: Anna-Maria Gleixner <anna-maria@linutronix.de> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Richard Weinberger <richard@nod.at>
2017-06-21 03:11:34 +08:00
if (rcu_rdp_is_offloaded(rdp) ||
rcu_segcblist_empty(&rdp->cblist))
return; /* No callbacks to migrate. */
rcu: Make rcu_barrier() no longer block CPU-hotplug operations This commit removes the cpus_read_lock() and cpus_read_unlock() calls from rcu_barrier(), thus allowing CPUs to come and go during the course of rcu_barrier() execution. Posting of the ->barrier_head callbacks does synchronize with portions of RCU's CPU-hotplug notifiers, but these locks are held for short time periods on both sides. Thus, full CPU-hotplug operations could both start and finish during the execution of a given rcu_barrier() invocation. Additional synchronization is provided by a global ->barrier_lock. Since the ->barrier_lock is only used during rcu_barrier() execution and during onlining/offlining a CPU, the contention for this lock should be low. It might be tempting to make use of a per-CPU lock just on general principles, but straightforward attempts to do this have the problems shown below. Initial state: 3 CPUs present, CPU 0 and CPU1 do not have any callback and CPU2 has callbacks. 1. CPU0 calls rcu_barrier(). 2. CPU1 starts offlining for CPU2. CPU1 calls rcutree_migrate_callbacks(). rcu_barrier_entrain() is called from rcutree_migrate_callbacks(), with CPU2's rdp->barrier_lock. It does not entrain ->barrier_head for CPU2, as rcu_barrier() on CPU0 hasn't started the barrier sequence (by calling rcu_seq_start(&rcu_state.barrier_sequence)) yet. 3. CPU0 starts new barrier sequence. It iterates over CPU0 and CPU1, after acquiring their per-cpu ->barrier_lock and finds 0 segcblist length. It updates ->barrier_seq_snap for CPU0 and CPU1 and continues loop iteration to CPU2. for_each_possible_cpu(cpu) { raw_spin_lock_irqsave(&rdp->barrier_lock, flags); if (!rcu_segcblist_n_cbs(&rdp->cblist)) { WRITE_ONCE(rdp->barrier_seq_snap, gseq); raw_spin_unlock_irqrestore(&rdp->barrier_lock, flags); rcu_barrier_trace(TPS("NQ"), cpu, rcu_state.barrier_sequence); continue; } 4. rcutree_migrate_callbacks() completes execution on CPU1. Segcblist len for CPU2 becomes 0. 5. The loop iteration on CPU0, checks rcu_segcblist_n_cbs(&rdp->cblist) for CPU2 and completes the loop iteration after setting ->barrier_seq_snap. 6. As there isn't any ->barrier_head callback entrained; at this point, rcu_barrier() in CPU0 returns. 7. The callbacks, which migrated from CPU2 to CPU1, execute. Straightforward per-CPU locking is also subject to the following race condition noted by Boqun Feng: 1. CPU0 calls rcu_barrier(), starting a new barrier sequence by invoking rcu_seq_start() and init_completion(), but does not yet initialize rcu_state.barrier_cpu_count. 2. CPU1 starts offlining for CPU2, calling rcutree_migrate_callbacks(), which in turn calls rcu_barrier_entrain() holding CPU2's. rdp->barrier_lock. It then entrains ->barrier_head for CPU2 and atomically increments rcu_state.barrier_cpu_count, which is unfortunately not yet initialized to the value 2. 3. The just-entrained RCU callback is invoked. It atomically decrements rcu_state.barrier_cpu_count and sees that it is now zero. This callback therefore invokes complete(). 4. CPU0 continues executing rcu_barrier(), but is not blocked by its call to wait_for_completion(). This results in rcu_barrier() returning before all pre-existing callbacks have been invoked, which is a bug. Therefore, synchronization is provided by rcu_state.barrier_lock, which is also held across the initialization sequence, especially the rcu_seq_start() and the atomic_set() that sets rcu_state.barrier_cpu_count to the value 2. In addition, this lock is held when entraining the rcu_barrier() callback, when deciding whether or not a CPU has callbacks that rcu_barrier() must wait on, when setting the ->qsmaskinitnext for incoming CPUs, and when migrating callbacks from a CPU that is going offline. Reviewed-by: Frederic Weisbecker <frederic@kernel.org> Co-developed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com> Signed-off-by: Neeraj Upadhyay <quic_neeraju@quicinc.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2021-12-15 05:35:17 +08:00
raw_spin_lock_irqsave(&rcu_state.barrier_lock, flags);
WARN_ON_ONCE(rcu_rdp_cpu_online(rdp));
rcu_barrier_entrain(rdp);
my_rdp = this_cpu_ptr(&rcu_data);
my_rnp = my_rdp->mynode;
rcu_nocb_lock(my_rdp); /* irqs already disabled. */
WARN_ON_ONCE(!rcu_nocb_flush_bypass(my_rdp, NULL, jiffies, false));
raw_spin_lock_rcu_node(my_rnp); /* irqs already disabled. */
/* Leverage recent GPs and set GP for new callbacks. */
needwake = rcu_advance_cbs(my_rnp, rdp) ||
rcu_advance_cbs(my_rnp, my_rdp);
rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist);
rcu: Make rcu_barrier() no longer block CPU-hotplug operations This commit removes the cpus_read_lock() and cpus_read_unlock() calls from rcu_barrier(), thus allowing CPUs to come and go during the course of rcu_barrier() execution. Posting of the ->barrier_head callbacks does synchronize with portions of RCU's CPU-hotplug notifiers, but these locks are held for short time periods on both sides. Thus, full CPU-hotplug operations could both start and finish during the execution of a given rcu_barrier() invocation. Additional synchronization is provided by a global ->barrier_lock. Since the ->barrier_lock is only used during rcu_barrier() execution and during onlining/offlining a CPU, the contention for this lock should be low. It might be tempting to make use of a per-CPU lock just on general principles, but straightforward attempts to do this have the problems shown below. Initial state: 3 CPUs present, CPU 0 and CPU1 do not have any callback and CPU2 has callbacks. 1. CPU0 calls rcu_barrier(). 2. CPU1 starts offlining for CPU2. CPU1 calls rcutree_migrate_callbacks(). rcu_barrier_entrain() is called from rcutree_migrate_callbacks(), with CPU2's rdp->barrier_lock. It does not entrain ->barrier_head for CPU2, as rcu_barrier() on CPU0 hasn't started the barrier sequence (by calling rcu_seq_start(&rcu_state.barrier_sequence)) yet. 3. CPU0 starts new barrier sequence. It iterates over CPU0 and CPU1, after acquiring their per-cpu ->barrier_lock and finds 0 segcblist length. It updates ->barrier_seq_snap for CPU0 and CPU1 and continues loop iteration to CPU2. for_each_possible_cpu(cpu) { raw_spin_lock_irqsave(&rdp->barrier_lock, flags); if (!rcu_segcblist_n_cbs(&rdp->cblist)) { WRITE_ONCE(rdp->barrier_seq_snap, gseq); raw_spin_unlock_irqrestore(&rdp->barrier_lock, flags); rcu_barrier_trace(TPS("NQ"), cpu, rcu_state.barrier_sequence); continue; } 4. rcutree_migrate_callbacks() completes execution on CPU1. Segcblist len for CPU2 becomes 0. 5. The loop iteration on CPU0, checks rcu_segcblist_n_cbs(&rdp->cblist) for CPU2 and completes the loop iteration after setting ->barrier_seq_snap. 6. As there isn't any ->barrier_head callback entrained; at this point, rcu_barrier() in CPU0 returns. 7. The callbacks, which migrated from CPU2 to CPU1, execute. Straightforward per-CPU locking is also subject to the following race condition noted by Boqun Feng: 1. CPU0 calls rcu_barrier(), starting a new barrier sequence by invoking rcu_seq_start() and init_completion(), but does not yet initialize rcu_state.barrier_cpu_count. 2. CPU1 starts offlining for CPU2, calling rcutree_migrate_callbacks(), which in turn calls rcu_barrier_entrain() holding CPU2's. rdp->barrier_lock. It then entrains ->barrier_head for CPU2 and atomically increments rcu_state.barrier_cpu_count, which is unfortunately not yet initialized to the value 2. 3. The just-entrained RCU callback is invoked. It atomically decrements rcu_state.barrier_cpu_count and sees that it is now zero. This callback therefore invokes complete(). 4. CPU0 continues executing rcu_barrier(), but is not blocked by its call to wait_for_completion(). This results in rcu_barrier() returning before all pre-existing callbacks have been invoked, which is a bug. Therefore, synchronization is provided by rcu_state.barrier_lock, which is also held across the initialization sequence, especially the rcu_seq_start() and the atomic_set() that sets rcu_state.barrier_cpu_count to the value 2. In addition, this lock is held when entraining the rcu_barrier() callback, when deciding whether or not a CPU has callbacks that rcu_barrier() must wait on, when setting the ->qsmaskinitnext for incoming CPUs, and when migrating callbacks from a CPU that is going offline. Reviewed-by: Frederic Weisbecker <frederic@kernel.org> Co-developed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com> Signed-off-by: Neeraj Upadhyay <quic_neeraju@quicinc.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2021-12-15 05:35:17 +08:00
raw_spin_unlock(&rcu_state.barrier_lock); /* irqs remain disabled. */
needwake = needwake || rcu_advance_cbs(my_rnp, my_rdp);
rcu_segcblist_disable(&rdp->cblist);
WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) != !rcu_segcblist_n_cbs(&my_rdp->cblist));
check_cb_ovld_locked(my_rdp, my_rnp);
if (rcu_rdp_is_offloaded(my_rdp)) {
raw_spin_unlock_rcu_node(my_rnp); /* irqs remain disabled. */
__call_rcu_nocb_wake(my_rdp, true, flags);
} else {
rcu_nocb_unlock(my_rdp); /* irqs remain disabled. */
raw_spin_unlock_irqrestore_rcu_node(my_rnp, flags);
}
if (needwake)
rcu_gp_kthread_wake();
lockdep_assert_irqs_enabled();
rcu: Migrate callbacks earlier in the CPU-offline timeline RCU callbacks must be migrated away from an outgoing CPU, and this is done near the end of the CPU-hotplug operation, after the outgoing CPU is long gone. Unfortunately, this means that other CPU-hotplug callbacks can execute while the outgoing CPU's callbacks are still immobilized on the long-gone CPU's callback lists. If any of these CPU-hotplug callbacks must wait, either directly or indirectly, for the invocation of any of the immobilized RCU callbacks, the system will hang. This commit avoids such hangs by migrating the callbacks away from the outgoing CPU immediately upon its departure, shortly after the return from __cpu_die() in takedown_cpu(). Thus, RCU is able to advance these callbacks and invoke them, which allows all the after-the-fact CPU-hotplug callbacks to wait on these RCU callbacks without risk of a hang. While in the neighborhood, this commit also moves rcu_send_cbs_to_orphanage() and rcu_adopt_orphan_cbs() under a pre-existing #ifdef to avoid including dead code on the one hand and to avoid define-without-use warnings on the other hand. Reported-by: Jeffrey Hugo <jhugo@codeaurora.org> Link: http://lkml.kernel.org/r/db9c91f6-1b17-6136-84f0-03c3c2581ab4@codeaurora.org Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Cc: Ingo Molnar <mingo@kernel.org> Cc: Anna-Maria Gleixner <anna-maria@linutronix.de> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Richard Weinberger <richard@nod.at>
2017-06-21 03:11:34 +08:00
WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
!rcu_segcblist_empty(&rdp->cblist),
"rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
cpu, rcu_segcblist_n_cbs(&rdp->cblist),
rcu_segcblist_first_cb(&rdp->cblist));
}
#endif
/*
* On non-huge systems, use expedited RCU grace periods to make suspend
* and hibernation run faster.
*/
static int rcu_pm_notify(struct notifier_block *self,
unsigned long action, void *hcpu)
{
switch (action) {
case PM_HIBERNATION_PREPARE:
case PM_SUSPEND_PREPARE:
rcu_async_hurry();
rcu_expedite_gp();
break;
case PM_POST_HIBERNATION:
case PM_POST_SUSPEND:
rcu_unexpedite_gp();
rcu_async_relax();
break;
default:
break;
}
return NOTIFY_OK;
}
rcu: Move expedited grace period (GP) work to RT kthread_worker Enabling CONFIG_RCU_BOOST did not reduce RCU expedited grace-period latency because its workqueues run at SCHED_OTHER, and thus can be delayed by normal processes. This commit avoids these delays by moving the expedited GP work items to a real-time-priority kthread_worker. This option is controlled by CONFIG_RCU_EXP_KTHREAD and disabled by default on PREEMPT_RT=y kernels which disable expedited grace periods after boot by unconditionally setting rcupdate.rcu_normal_after_boot=1. The results were evaluated on arm64 Android devices (6GB ram) running 5.10 kernel, and capturing trace data in critical user-level code. The table below shows the resulting order-of-magnitude improvements in synchronize_rcu_expedited() latency: ------------------------------------------------------------------------ | | workqueues | kthread_worker | Diff | ------------------------------------------------------------------------ | Count | 725 | 688 | | ------------------------------------------------------------------------ | Min Duration (ns) | 326 | 447 | 37.12% | ------------------------------------------------------------------------ | Q1 (ns) | 39,428 | 38,971 | -1.16% | ------------------------------------------------------------------------ | Q2 - Median (ns) | 98,225 | 69,743 | -29.00% | ------------------------------------------------------------------------ | Q3 (ns) | 342,122 | 126,638 | -62.98% | ------------------------------------------------------------------------ | Max Duration (ns) | 372,766,967 | 2,329,671 | -99.38% | ------------------------------------------------------------------------ | Avg Duration (ns) | 2,746,353 | 151,242 | -94.49% | ------------------------------------------------------------------------ | Standard Deviation (ns) | 19,327,765 | 294,408 | | ------------------------------------------------------------------------ The below table show the range of maximums/minimums for synchronize_rcu_expedited() latency from all experiments: ------------------------------------------------------------------------ | | workqueues | kthread_worker | Diff | ------------------------------------------------------------------------ | Total No. of Experiments | 25 | 23 | | ------------------------------------------------------------------------ | Largest Maximum (ns) | 372,766,967 | 2,329,671 | -99.38% | ------------------------------------------------------------------------ | Smallest Maximum (ns) | 38,819 | 86,954 | 124.00% | ------------------------------------------------------------------------ | Range of Maximums (ns) | 372,728,148 | 2,242,717 | | ------------------------------------------------------------------------ | Largest Minimum (ns) | 88,623 | 27,588 | -68.87% | ------------------------------------------------------------------------ | Smallest Minimum (ns) | 326 | 447 | 37.12% | ------------------------------------------------------------------------ | Range of Minimums (ns) | 88,297 | 27,141 | | ------------------------------------------------------------------------ Cc: "Paul E. McKenney" <paulmck@kernel.org> Cc: Tejun Heo <tj@kernel.org> Reported-by: Tim Murray <timmurray@google.com> Reported-by: Wei Wang <wvw@google.com> Tested-by: Kyle Lin <kylelin@google.com> Tested-by: Chunwei Lu <chunweilu@google.com> Tested-by: Lulu Wang <luluw@google.com> Signed-off-by: Kalesh Singh <kaleshsingh@google.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2022-04-09 08:35:27 +08:00
#ifdef CONFIG_RCU_EXP_KTHREAD
struct kthread_worker *rcu_exp_gp_kworker;
struct kthread_worker *rcu_exp_par_gp_kworker;
static void __init rcu_start_exp_gp_kworkers(void)
{
const char *par_gp_kworker_name = "rcu_exp_par_gp_kthread_worker";
const char *gp_kworker_name = "rcu_exp_gp_kthread_worker";
struct sched_param param = { .sched_priority = kthread_prio };
rcu_exp_gp_kworker = kthread_create_worker(0, gp_kworker_name);
if (IS_ERR_OR_NULL(rcu_exp_gp_kworker)) {
pr_err("Failed to create %s!\n", gp_kworker_name);
rcu_exp_gp_kworker = NULL;
rcu: Move expedited grace period (GP) work to RT kthread_worker Enabling CONFIG_RCU_BOOST did not reduce RCU expedited grace-period latency because its workqueues run at SCHED_OTHER, and thus can be delayed by normal processes. This commit avoids these delays by moving the expedited GP work items to a real-time-priority kthread_worker. This option is controlled by CONFIG_RCU_EXP_KTHREAD and disabled by default on PREEMPT_RT=y kernels which disable expedited grace periods after boot by unconditionally setting rcupdate.rcu_normal_after_boot=1. The results were evaluated on arm64 Android devices (6GB ram) running 5.10 kernel, and capturing trace data in critical user-level code. The table below shows the resulting order-of-magnitude improvements in synchronize_rcu_expedited() latency: ------------------------------------------------------------------------ | | workqueues | kthread_worker | Diff | ------------------------------------------------------------------------ | Count | 725 | 688 | | ------------------------------------------------------------------------ | Min Duration (ns) | 326 | 447 | 37.12% | ------------------------------------------------------------------------ | Q1 (ns) | 39,428 | 38,971 | -1.16% | ------------------------------------------------------------------------ | Q2 - Median (ns) | 98,225 | 69,743 | -29.00% | ------------------------------------------------------------------------ | Q3 (ns) | 342,122 | 126,638 | -62.98% | ------------------------------------------------------------------------ | Max Duration (ns) | 372,766,967 | 2,329,671 | -99.38% | ------------------------------------------------------------------------ | Avg Duration (ns) | 2,746,353 | 151,242 | -94.49% | ------------------------------------------------------------------------ | Standard Deviation (ns) | 19,327,765 | 294,408 | | ------------------------------------------------------------------------ The below table show the range of maximums/minimums for synchronize_rcu_expedited() latency from all experiments: ------------------------------------------------------------------------ | | workqueues | kthread_worker | Diff | ------------------------------------------------------------------------ | Total No. of Experiments | 25 | 23 | | ------------------------------------------------------------------------ | Largest Maximum (ns) | 372,766,967 | 2,329,671 | -99.38% | ------------------------------------------------------------------------ | Smallest Maximum (ns) | 38,819 | 86,954 | 124.00% | ------------------------------------------------------------------------ | Range of Maximums (ns) | 372,728,148 | 2,242,717 | | ------------------------------------------------------------------------ | Largest Minimum (ns) | 88,623 | 27,588 | -68.87% | ------------------------------------------------------------------------ | Smallest Minimum (ns) | 326 | 447 | 37.12% | ------------------------------------------------------------------------ | Range of Minimums (ns) | 88,297 | 27,141 | | ------------------------------------------------------------------------ Cc: "Paul E. McKenney" <paulmck@kernel.org> Cc: Tejun Heo <tj@kernel.org> Reported-by: Tim Murray <timmurray@google.com> Reported-by: Wei Wang <wvw@google.com> Tested-by: Kyle Lin <kylelin@google.com> Tested-by: Chunwei Lu <chunweilu@google.com> Tested-by: Lulu Wang <luluw@google.com> Signed-off-by: Kalesh Singh <kaleshsingh@google.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2022-04-09 08:35:27 +08:00
return;
}
rcu_exp_par_gp_kworker = kthread_create_worker(0, par_gp_kworker_name);
if (IS_ERR_OR_NULL(rcu_exp_par_gp_kworker)) {
pr_err("Failed to create %s!\n", par_gp_kworker_name);
rcu_exp_par_gp_kworker = NULL;
rcu: Move expedited grace period (GP) work to RT kthread_worker Enabling CONFIG_RCU_BOOST did not reduce RCU expedited grace-period latency because its workqueues run at SCHED_OTHER, and thus can be delayed by normal processes. This commit avoids these delays by moving the expedited GP work items to a real-time-priority kthread_worker. This option is controlled by CONFIG_RCU_EXP_KTHREAD and disabled by default on PREEMPT_RT=y kernels which disable expedited grace periods after boot by unconditionally setting rcupdate.rcu_normal_after_boot=1. The results were evaluated on arm64 Android devices (6GB ram) running 5.10 kernel, and capturing trace data in critical user-level code. The table below shows the resulting order-of-magnitude improvements in synchronize_rcu_expedited() latency: ------------------------------------------------------------------------ | | workqueues | kthread_worker | Diff | ------------------------------------------------------------------------ | Count | 725 | 688 | | ------------------------------------------------------------------------ | Min Duration (ns) | 326 | 447 | 37.12% | ------------------------------------------------------------------------ | Q1 (ns) | 39,428 | 38,971 | -1.16% | ------------------------------------------------------------------------ | Q2 - Median (ns) | 98,225 | 69,743 | -29.00% | ------------------------------------------------------------------------ | Q3 (ns) | 342,122 | 126,638 | -62.98% | ------------------------------------------------------------------------ | Max Duration (ns) | 372,766,967 | 2,329,671 | -99.38% | ------------------------------------------------------------------------ | Avg Duration (ns) | 2,746,353 | 151,242 | -94.49% | ------------------------------------------------------------------------ | Standard Deviation (ns) | 19,327,765 | 294,408 | | ------------------------------------------------------------------------ The below table show the range of maximums/minimums for synchronize_rcu_expedited() latency from all experiments: ------------------------------------------------------------------------ | | workqueues | kthread_worker | Diff | ------------------------------------------------------------------------ | Total No. of Experiments | 25 | 23 | | ------------------------------------------------------------------------ | Largest Maximum (ns) | 372,766,967 | 2,329,671 | -99.38% | ------------------------------------------------------------------------ | Smallest Maximum (ns) | 38,819 | 86,954 | 124.00% | ------------------------------------------------------------------------ | Range of Maximums (ns) | 372,728,148 | 2,242,717 | | ------------------------------------------------------------------------ | Largest Minimum (ns) | 88,623 | 27,588 | -68.87% | ------------------------------------------------------------------------ | Smallest Minimum (ns) | 326 | 447 | 37.12% | ------------------------------------------------------------------------ | Range of Minimums (ns) | 88,297 | 27,141 | | ------------------------------------------------------------------------ Cc: "Paul E. McKenney" <paulmck@kernel.org> Cc: Tejun Heo <tj@kernel.org> Reported-by: Tim Murray <timmurray@google.com> Reported-by: Wei Wang <wvw@google.com> Tested-by: Kyle Lin <kylelin@google.com> Tested-by: Chunwei Lu <chunweilu@google.com> Tested-by: Lulu Wang <luluw@google.com> Signed-off-by: Kalesh Singh <kaleshsingh@google.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2022-04-09 08:35:27 +08:00
kthread_destroy_worker(rcu_exp_gp_kworker);
rcu_exp_gp_kworker = NULL;
rcu: Move expedited grace period (GP) work to RT kthread_worker Enabling CONFIG_RCU_BOOST did not reduce RCU expedited grace-period latency because its workqueues run at SCHED_OTHER, and thus can be delayed by normal processes. This commit avoids these delays by moving the expedited GP work items to a real-time-priority kthread_worker. This option is controlled by CONFIG_RCU_EXP_KTHREAD and disabled by default on PREEMPT_RT=y kernels which disable expedited grace periods after boot by unconditionally setting rcupdate.rcu_normal_after_boot=1. The results were evaluated on arm64 Android devices (6GB ram) running 5.10 kernel, and capturing trace data in critical user-level code. The table below shows the resulting order-of-magnitude improvements in synchronize_rcu_expedited() latency: ------------------------------------------------------------------------ | | workqueues | kthread_worker | Diff | ------------------------------------------------------------------------ | Count | 725 | 688 | | ------------------------------------------------------------------------ | Min Duration (ns) | 326 | 447 | 37.12% | ------------------------------------------------------------------------ | Q1 (ns) | 39,428 | 38,971 | -1.16% | ------------------------------------------------------------------------ | Q2 - Median (ns) | 98,225 | 69,743 | -29.00% | ------------------------------------------------------------------------ | Q3 (ns) | 342,122 | 126,638 | -62.98% | ------------------------------------------------------------------------ | Max Duration (ns) | 372,766,967 | 2,329,671 | -99.38% | ------------------------------------------------------------------------ | Avg Duration (ns) | 2,746,353 | 151,242 | -94.49% | ------------------------------------------------------------------------ | Standard Deviation (ns) | 19,327,765 | 294,408 | | ------------------------------------------------------------------------ The below table show the range of maximums/minimums for synchronize_rcu_expedited() latency from all experiments: ------------------------------------------------------------------------ | | workqueues | kthread_worker | Diff | ------------------------------------------------------------------------ | Total No. of Experiments | 25 | 23 | | ------------------------------------------------------------------------ | Largest Maximum (ns) | 372,766,967 | 2,329,671 | -99.38% | ------------------------------------------------------------------------ | Smallest Maximum (ns) | 38,819 | 86,954 | 124.00% | ------------------------------------------------------------------------ | Range of Maximums (ns) | 372,728,148 | 2,242,717 | | ------------------------------------------------------------------------ | Largest Minimum (ns) | 88,623 | 27,588 | -68.87% | ------------------------------------------------------------------------ | Smallest Minimum (ns) | 326 | 447 | 37.12% | ------------------------------------------------------------------------ | Range of Minimums (ns) | 88,297 | 27,141 | | ------------------------------------------------------------------------ Cc: "Paul E. McKenney" <paulmck@kernel.org> Cc: Tejun Heo <tj@kernel.org> Reported-by: Tim Murray <timmurray@google.com> Reported-by: Wei Wang <wvw@google.com> Tested-by: Kyle Lin <kylelin@google.com> Tested-by: Chunwei Lu <chunweilu@google.com> Tested-by: Lulu Wang <luluw@google.com> Signed-off-by: Kalesh Singh <kaleshsingh@google.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2022-04-09 08:35:27 +08:00
return;
}
sched_setscheduler_nocheck(rcu_exp_gp_kworker->task, SCHED_FIFO, &param);
sched_setscheduler_nocheck(rcu_exp_par_gp_kworker->task, SCHED_FIFO,
&param);
}
static inline void rcu_alloc_par_gp_wq(void)
{
}
#else /* !CONFIG_RCU_EXP_KTHREAD */
struct workqueue_struct *rcu_par_gp_wq;
static void __init rcu_start_exp_gp_kworkers(void)
{
}
static inline void rcu_alloc_par_gp_wq(void)
{
rcu_par_gp_wq = alloc_workqueue("rcu_par_gp", WQ_MEM_RECLAIM, 0);
WARN_ON(!rcu_par_gp_wq);
}
#endif /* CONFIG_RCU_EXP_KTHREAD */
/*
* Spawn the kthreads that handle RCU's grace periods.
*/
static int __init rcu_spawn_gp_kthread(void)
{
unsigned long flags;
struct rcu_node *rnp;
struct sched_param sp;
struct task_struct *t;
struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
rcu_scheduler_fully_active = 1;
t = kthread_create(rcu_gp_kthread, NULL, "%s", rcu_state.name);
if (WARN_ONCE(IS_ERR(t), "%s: Could not start grace-period kthread, OOM is now expected behavior\n", __func__))
return 0;
if (kthread_prio) {
sp.sched_priority = kthread_prio;
sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
}
rnp = rcu_get_root();
raw_spin_lock_irqsave_rcu_node(rnp, flags);
WRITE_ONCE(rcu_state.gp_activity, jiffies);
WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
// Reset .gp_activity and .gp_req_activity before setting .gp_kthread.
smp_store_release(&rcu_state.gp_kthread, t); /* ^^^ */
raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
wake_up_process(t);
/* This is a pre-SMP initcall, we expect a single CPU */
WARN_ON(num_online_cpus() > 1);
/*
* Those kthreads couldn't be created on rcu_init() -> rcutree_prepare_cpu()
* due to rcu_scheduler_fully_active.
*/
rcu_spawn_cpu_nocb_kthread(smp_processor_id());
rcu_spawn_one_boost_kthread(rdp->mynode);
rcu_spawn_core_kthreads();
rcu: Move expedited grace period (GP) work to RT kthread_worker Enabling CONFIG_RCU_BOOST did not reduce RCU expedited grace-period latency because its workqueues run at SCHED_OTHER, and thus can be delayed by normal processes. This commit avoids these delays by moving the expedited GP work items to a real-time-priority kthread_worker. This option is controlled by CONFIG_RCU_EXP_KTHREAD and disabled by default on PREEMPT_RT=y kernels which disable expedited grace periods after boot by unconditionally setting rcupdate.rcu_normal_after_boot=1. The results were evaluated on arm64 Android devices (6GB ram) running 5.10 kernel, and capturing trace data in critical user-level code. The table below shows the resulting order-of-magnitude improvements in synchronize_rcu_expedited() latency: ------------------------------------------------------------------------ | | workqueues | kthread_worker | Diff | ------------------------------------------------------------------------ | Count | 725 | 688 | | ------------------------------------------------------------------------ | Min Duration (ns) | 326 | 447 | 37.12% | ------------------------------------------------------------------------ | Q1 (ns) | 39,428 | 38,971 | -1.16% | ------------------------------------------------------------------------ | Q2 - Median (ns) | 98,225 | 69,743 | -29.00% | ------------------------------------------------------------------------ | Q3 (ns) | 342,122 | 126,638 | -62.98% | ------------------------------------------------------------------------ | Max Duration (ns) | 372,766,967 | 2,329,671 | -99.38% | ------------------------------------------------------------------------ | Avg Duration (ns) | 2,746,353 | 151,242 | -94.49% | ------------------------------------------------------------------------ | Standard Deviation (ns) | 19,327,765 | 294,408 | | ------------------------------------------------------------------------ The below table show the range of maximums/minimums for synchronize_rcu_expedited() latency from all experiments: ------------------------------------------------------------------------ | | workqueues | kthread_worker | Diff | ------------------------------------------------------------------------ | Total No. of Experiments | 25 | 23 | | ------------------------------------------------------------------------ | Largest Maximum (ns) | 372,766,967 | 2,329,671 | -99.38% | ------------------------------------------------------------------------ | Smallest Maximum (ns) | 38,819 | 86,954 | 124.00% | ------------------------------------------------------------------------ | Range of Maximums (ns) | 372,728,148 | 2,242,717 | | ------------------------------------------------------------------------ | Largest Minimum (ns) | 88,623 | 27,588 | -68.87% | ------------------------------------------------------------------------ | Smallest Minimum (ns) | 326 | 447 | 37.12% | ------------------------------------------------------------------------ | Range of Minimums (ns) | 88,297 | 27,141 | | ------------------------------------------------------------------------ Cc: "Paul E. McKenney" <paulmck@kernel.org> Cc: Tejun Heo <tj@kernel.org> Reported-by: Tim Murray <timmurray@google.com> Reported-by: Wei Wang <wvw@google.com> Tested-by: Kyle Lin <kylelin@google.com> Tested-by: Chunwei Lu <chunweilu@google.com> Tested-by: Lulu Wang <luluw@google.com> Signed-off-by: Kalesh Singh <kaleshsingh@google.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2022-04-09 08:35:27 +08:00
/* Create kthread worker for expedited GPs */
rcu_start_exp_gp_kworkers();
return 0;
}
early_initcall(rcu_spawn_gp_kthread);
/*
rcu: Narrow early boot window of illegal synchronous grace periods The current preemptible RCU implementation goes through three phases during bootup. In the first phase, there is only one CPU that is running with preemption disabled, so that a no-op is a synchronous grace period. In the second mid-boot phase, the scheduler is running, but RCU has not yet gotten its kthreads spawned (and, for expedited grace periods, workqueues are not yet running. During this time, any attempt to do a synchronous grace period will hang the system (or complain bitterly, depending). In the third and final phase, RCU is fully operational and everything works normally. This has been OK for some time, but there has recently been some synchronous grace periods showing up during the second mid-boot phase. This code worked "by accident" for awhile, but started failing as soon as expedited RCU grace periods switched over to workqueues in commit 8b355e3bc140 ("rcu: Drive expedited grace periods from workqueue"). Note that the code was buggy even before this commit, as it was subject to failure on real-time systems that forced all expedited grace periods to run as normal grace periods (for example, using the rcu_normal ksysfs parameter). The callchain from the failure case is as follows: early_amd_iommu_init() |-> acpi_put_table(ivrs_base); |-> acpi_tb_put_table(table_desc); |-> acpi_tb_invalidate_table(table_desc); |-> acpi_tb_release_table(...) |-> acpi_os_unmap_memory |-> acpi_os_unmap_iomem |-> acpi_os_map_cleanup |-> synchronize_rcu_expedited The kernel showing this callchain was built with CONFIG_PREEMPT_RCU=y, which caused the code to try using workqueues before they were initialized, which did not go well. This commit therefore reworks RCU to permit synchronous grace periods to proceed during this mid-boot phase. This commit is therefore a fix to a regression introduced in v4.9, and is therefore being put forward post-merge-window in v4.10. This commit sets a flag from the existing rcu_scheduler_starting() function which causes all synchronous grace periods to take the expedited path. The expedited path now checks this flag, using the requesting task to drive the expedited grace period forward during the mid-boot phase. Finally, this flag is updated by a core_initcall() function named rcu_exp_runtime_mode(), which causes the runtime codepaths to be used. Note that this arrangement assumes that tasks are not sent POSIX signals (or anything similar) from the time that the first task is spawned through core_initcall() time. Fixes: 8b355e3bc140 ("rcu: Drive expedited grace periods from workqueue") Reported-by: "Zheng, Lv" <lv.zheng@intel.com> Reported-by: Borislav Petkov <bp@alien8.de> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Tested-by: Stan Kain <stan.kain@gmail.com> Tested-by: Ivan <waffolz@hotmail.com> Tested-by: Emanuel Castelo <emanuel.castelo@gmail.com> Tested-by: Bruno Pesavento <bpesavento@infinito.it> Tested-by: Borislav Petkov <bp@suse.de> Tested-by: Frederic Bezies <fredbezies@gmail.com> Cc: <stable@vger.kernel.org> # 4.9.0-
2017-01-10 18:28:26 +08:00
* This function is invoked towards the end of the scheduler's
* initialization process. Before this is called, the idle task might
* contain synchronous grace-period primitives (during which time, this idle
* task is booting the system, and such primitives are no-ops). After this
* function is called, any synchronous grace-period primitives are run as
* expedited, with the requesting task driving the grace period forward.
* A later core_initcall() rcu_set_runtime_mode() will switch to full
rcu: Narrow early boot window of illegal synchronous grace periods The current preemptible RCU implementation goes through three phases during bootup. In the first phase, there is only one CPU that is running with preemption disabled, so that a no-op is a synchronous grace period. In the second mid-boot phase, the scheduler is running, but RCU has not yet gotten its kthreads spawned (and, for expedited grace periods, workqueues are not yet running. During this time, any attempt to do a synchronous grace period will hang the system (or complain bitterly, depending). In the third and final phase, RCU is fully operational and everything works normally. This has been OK for some time, but there has recently been some synchronous grace periods showing up during the second mid-boot phase. This code worked "by accident" for awhile, but started failing as soon as expedited RCU grace periods switched over to workqueues in commit 8b355e3bc140 ("rcu: Drive expedited grace periods from workqueue"). Note that the code was buggy even before this commit, as it was subject to failure on real-time systems that forced all expedited grace periods to run as normal grace periods (for example, using the rcu_normal ksysfs parameter). The callchain from the failure case is as follows: early_amd_iommu_init() |-> acpi_put_table(ivrs_base); |-> acpi_tb_put_table(table_desc); |-> acpi_tb_invalidate_table(table_desc); |-> acpi_tb_release_table(...) |-> acpi_os_unmap_memory |-> acpi_os_unmap_iomem |-> acpi_os_map_cleanup |-> synchronize_rcu_expedited The kernel showing this callchain was built with CONFIG_PREEMPT_RCU=y, which caused the code to try using workqueues before they were initialized, which did not go well. This commit therefore reworks RCU to permit synchronous grace periods to proceed during this mid-boot phase. This commit is therefore a fix to a regression introduced in v4.9, and is therefore being put forward post-merge-window in v4.10. This commit sets a flag from the existing rcu_scheduler_starting() function which causes all synchronous grace periods to take the expedited path. The expedited path now checks this flag, using the requesting task to drive the expedited grace period forward during the mid-boot phase. Finally, this flag is updated by a core_initcall() function named rcu_exp_runtime_mode(), which causes the runtime codepaths to be used. Note that this arrangement assumes that tasks are not sent POSIX signals (or anything similar) from the time that the first task is spawned through core_initcall() time. Fixes: 8b355e3bc140 ("rcu: Drive expedited grace periods from workqueue") Reported-by: "Zheng, Lv" <lv.zheng@intel.com> Reported-by: Borislav Petkov <bp@alien8.de> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Tested-by: Stan Kain <stan.kain@gmail.com> Tested-by: Ivan <waffolz@hotmail.com> Tested-by: Emanuel Castelo <emanuel.castelo@gmail.com> Tested-by: Bruno Pesavento <bpesavento@infinito.it> Tested-by: Borislav Petkov <bp@suse.de> Tested-by: Frederic Bezies <fredbezies@gmail.com> Cc: <stable@vger.kernel.org> # 4.9.0-
2017-01-10 18:28:26 +08:00
* runtime RCU functionality.
*/
void rcu_scheduler_starting(void)
{
unsigned long flags;
struct rcu_node *rnp;
WARN_ON(num_online_cpus() != 1);
WARN_ON(nr_context_switches() > 0);
rcu: Narrow early boot window of illegal synchronous grace periods The current preemptible RCU implementation goes through three phases during bootup. In the first phase, there is only one CPU that is running with preemption disabled, so that a no-op is a synchronous grace period. In the second mid-boot phase, the scheduler is running, but RCU has not yet gotten its kthreads spawned (and, for expedited grace periods, workqueues are not yet running. During this time, any attempt to do a synchronous grace period will hang the system (or complain bitterly, depending). In the third and final phase, RCU is fully operational and everything works normally. This has been OK for some time, but there has recently been some synchronous grace periods showing up during the second mid-boot phase. This code worked "by accident" for awhile, but started failing as soon as expedited RCU grace periods switched over to workqueues in commit 8b355e3bc140 ("rcu: Drive expedited grace periods from workqueue"). Note that the code was buggy even before this commit, as it was subject to failure on real-time systems that forced all expedited grace periods to run as normal grace periods (for example, using the rcu_normal ksysfs parameter). The callchain from the failure case is as follows: early_amd_iommu_init() |-> acpi_put_table(ivrs_base); |-> acpi_tb_put_table(table_desc); |-> acpi_tb_invalidate_table(table_desc); |-> acpi_tb_release_table(...) |-> acpi_os_unmap_memory |-> acpi_os_unmap_iomem |-> acpi_os_map_cleanup |-> synchronize_rcu_expedited The kernel showing this callchain was built with CONFIG_PREEMPT_RCU=y, which caused the code to try using workqueues before they were initialized, which did not go well. This commit therefore reworks RCU to permit synchronous grace periods to proceed during this mid-boot phase. This commit is therefore a fix to a regression introduced in v4.9, and is therefore being put forward post-merge-window in v4.10. This commit sets a flag from the existing rcu_scheduler_starting() function which causes all synchronous grace periods to take the expedited path. The expedited path now checks this flag, using the requesting task to drive the expedited grace period forward during the mid-boot phase. Finally, this flag is updated by a core_initcall() function named rcu_exp_runtime_mode(), which causes the runtime codepaths to be used. Note that this arrangement assumes that tasks are not sent POSIX signals (or anything similar) from the time that the first task is spawned through core_initcall() time. Fixes: 8b355e3bc140 ("rcu: Drive expedited grace periods from workqueue") Reported-by: "Zheng, Lv" <lv.zheng@intel.com> Reported-by: Borislav Petkov <bp@alien8.de> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Tested-by: Stan Kain <stan.kain@gmail.com> Tested-by: Ivan <waffolz@hotmail.com> Tested-by: Emanuel Castelo <emanuel.castelo@gmail.com> Tested-by: Bruno Pesavento <bpesavento@infinito.it> Tested-by: Borislav Petkov <bp@suse.de> Tested-by: Frederic Bezies <fredbezies@gmail.com> Cc: <stable@vger.kernel.org> # 4.9.0-
2017-01-10 18:28:26 +08:00
rcu_test_sync_prims();
// Fix up the ->gp_seq counters.
local_irq_save(flags);
rcu_for_each_node_breadth_first(rnp)
rnp->gp_seq_needed = rnp->gp_seq = rcu_state.gp_seq;
local_irq_restore(flags);
// Switch out of early boot mode.
rcu: Narrow early boot window of illegal synchronous grace periods The current preemptible RCU implementation goes through three phases during bootup. In the first phase, there is only one CPU that is running with preemption disabled, so that a no-op is a synchronous grace period. In the second mid-boot phase, the scheduler is running, but RCU has not yet gotten its kthreads spawned (and, for expedited grace periods, workqueues are not yet running. During this time, any attempt to do a synchronous grace period will hang the system (or complain bitterly, depending). In the third and final phase, RCU is fully operational and everything works normally. This has been OK for some time, but there has recently been some synchronous grace periods showing up during the second mid-boot phase. This code worked "by accident" for awhile, but started failing as soon as expedited RCU grace periods switched over to workqueues in commit 8b355e3bc140 ("rcu: Drive expedited grace periods from workqueue"). Note that the code was buggy even before this commit, as it was subject to failure on real-time systems that forced all expedited grace periods to run as normal grace periods (for example, using the rcu_normal ksysfs parameter). The callchain from the failure case is as follows: early_amd_iommu_init() |-> acpi_put_table(ivrs_base); |-> acpi_tb_put_table(table_desc); |-> acpi_tb_invalidate_table(table_desc); |-> acpi_tb_release_table(...) |-> acpi_os_unmap_memory |-> acpi_os_unmap_iomem |-> acpi_os_map_cleanup |-> synchronize_rcu_expedited The kernel showing this callchain was built with CONFIG_PREEMPT_RCU=y, which caused the code to try using workqueues before they were initialized, which did not go well. This commit therefore reworks RCU to permit synchronous grace periods to proceed during this mid-boot phase. This commit is therefore a fix to a regression introduced in v4.9, and is therefore being put forward post-merge-window in v4.10. This commit sets a flag from the existing rcu_scheduler_starting() function which causes all synchronous grace periods to take the expedited path. The expedited path now checks this flag, using the requesting task to drive the expedited grace period forward during the mid-boot phase. Finally, this flag is updated by a core_initcall() function named rcu_exp_runtime_mode(), which causes the runtime codepaths to be used. Note that this arrangement assumes that tasks are not sent POSIX signals (or anything similar) from the time that the first task is spawned through core_initcall() time. Fixes: 8b355e3bc140 ("rcu: Drive expedited grace periods from workqueue") Reported-by: "Zheng, Lv" <lv.zheng@intel.com> Reported-by: Borislav Petkov <bp@alien8.de> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Tested-by: Stan Kain <stan.kain@gmail.com> Tested-by: Ivan <waffolz@hotmail.com> Tested-by: Emanuel Castelo <emanuel.castelo@gmail.com> Tested-by: Bruno Pesavento <bpesavento@infinito.it> Tested-by: Borislav Petkov <bp@suse.de> Tested-by: Frederic Bezies <fredbezies@gmail.com> Cc: <stable@vger.kernel.org> # 4.9.0-
2017-01-10 18:28:26 +08:00
rcu_scheduler_active = RCU_SCHEDULER_INIT;
rcu_test_sync_prims();
}
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/*
* Helper function for rcu_init() that initializes the rcu_state structure.
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
*/
static void __init rcu_init_one(void)
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
{
static const char * const buf[] = RCU_NODE_NAME_INIT;
static const char * const fqs[] = RCU_FQS_NAME_INIT;
static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
int cpustride = 1;
int i;
int j;
struct rcu_node *rnp;
BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
/* Silence gcc 4.8 false positive about array index out of range. */
if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
panic("rcu_init_one: rcu_num_lvls out of range");
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/* Initialize the level-tracking arrays. */
for (i = 1; i < rcu_num_lvls; i++)
rcu_state.level[i] =
rcu_state.level[i - 1] + num_rcu_lvl[i - 1];
rcu_init_levelspread(levelspread, num_rcu_lvl);
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
/* Initialize the elements themselves, starting from the leaves. */
for (i = rcu_num_lvls - 1; i >= 0; i--) {
cpustride *= levelspread[i];
rnp = rcu_state.level[i];
for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
&rcu_node_class[i], buf[i]);
raw_spin_lock_init(&rnp->fqslock);
lockdep_set_class_and_name(&rnp->fqslock,
&rcu_fqs_class[i], fqs[i]);
rnp->gp_seq = rcu_state.gp_seq;
rnp->gp_seq_needed = rcu_state.gp_seq;
rnp->completedqs = rcu_state.gp_seq;
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
rnp->qsmask = 0;
rnp->qsmaskinit = 0;
rnp->grplo = j * cpustride;
rnp->grphi = (j + 1) * cpustride - 1;
if (rnp->grphi >= nr_cpu_ids)
rnp->grphi = nr_cpu_ids - 1;
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
if (i == 0) {
rnp->grpnum = 0;
rnp->grpmask = 0;
rnp->parent = NULL;
} else {
rnp->grpnum = j % levelspread[i - 1];
rnp->grpmask = BIT(rnp->grpnum);
rnp->parent = rcu_state.level[i - 1] +
j / levelspread[i - 1];
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
}
rnp->level = i;
INIT_LIST_HEAD(&rnp->blkd_tasks);
rcu_init_one_nocb(rnp);
init_waitqueue_head(&rnp->exp_wq[0]);
init_waitqueue_head(&rnp->exp_wq[1]);
init_waitqueue_head(&rnp->exp_wq[2]);
init_waitqueue_head(&rnp->exp_wq[3]);
spin_lock_init(&rnp->exp_lock);
mutex_init(&rnp->boost_kthread_mutex);
raw_spin_lock_init(&rnp->exp_poll_lock);
rnp->exp_seq_poll_rq = RCU_GET_STATE_COMPLETED;
INIT_WORK(&rnp->exp_poll_wq, sync_rcu_do_polled_gp);
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
}
}
init_swait_queue_head(&rcu_state.gp_wq);
init_swait_queue_head(&rcu_state.expedited_wq);
rnp = rcu_first_leaf_node();
for_each_possible_cpu(i) {
while (i > rnp->grphi)
rnp++;
per_cpu_ptr(&rcu_data, i)->mynode = rnp;
rcu_boot_init_percpu_data(i);
}
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
}
/*
* Force priority from the kernel command-line into range.
*/
static void __init sanitize_kthread_prio(void)
{
int kthread_prio_in = kthread_prio;
if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 2
&& IS_BUILTIN(CONFIG_RCU_TORTURE_TEST))
kthread_prio = 2;
else if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
kthread_prio = 1;
else if (kthread_prio < 0)
kthread_prio = 0;
else if (kthread_prio > 99)
kthread_prio = 99;
if (kthread_prio != kthread_prio_in)
pr_alert("%s: Limited prio to %d from %d\n",
__func__, kthread_prio, kthread_prio_in);
}
/*
* Compute the rcu_node tree geometry from kernel parameters. This cannot
* replace the definitions in tree.h because those are needed to size
* the ->node array in the rcu_state structure.
*/
srcu: Fix broken node geometry after early ssp init An srcu_struct structure that is initialized before rcu_init_geometry() will have its srcu_node hierarchy based on CONFIG_NR_CPUS. Once rcu_init_geometry() is called, this hierarchy is compressed as needed for the actual maximum number of CPUs for this system. Later on, that srcu_struct structure is confused, sometimes referring to its initial CONFIG_NR_CPUS-based hierarchy, and sometimes instead to the new num_possible_cpus() hierarchy. For example, each of its ->mynode fields continues to reference the original leaf rcu_node structures, some of which might no longer exist. On the other hand, srcu_for_each_node_breadth_first() traverses to the new node hierarchy. There are at least two bad possible outcomes to this: 1) a) A callback enqueued early on an srcu_data structure (call it *sdp) is recorded pending on sdp->mynode->srcu_data_have_cbs in srcu_funnel_gp_start() with sdp->mynode pointing to a deep leaf (say 3 levels). b) The grace period ends after rcu_init_geometry() shrinks the nodes level to a single one. srcu_gp_end() walks through the new srcu_node hierarchy without ever reaching the old leaves so the callback is never executed. This is easily reproduced on an 8 CPUs machine with CONFIG_NR_CPUS >= 32 and "rcupdate.rcu_self_test=1". The srcu_barrier() after early tests verification never completes and the boot hangs: [ 5413.141029] INFO: task swapper/0:1 blocked for more than 4915 seconds. [ 5413.147564] Not tainted 5.12.0-rc4+ #28 [ 5413.151927] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. [ 5413.159753] task:swapper/0 state:D stack: 0 pid: 1 ppid: 0 flags:0x00004000 [ 5413.168099] Call Trace: [ 5413.170555] __schedule+0x36c/0x930 [ 5413.174057] ? wait_for_completion+0x88/0x110 [ 5413.178423] schedule+0x46/0xf0 [ 5413.181575] schedule_timeout+0x284/0x380 [ 5413.185591] ? wait_for_completion+0x88/0x110 [ 5413.189957] ? mark_held_locks+0x61/0x80 [ 5413.193882] ? mark_held_locks+0x61/0x80 [ 5413.197809] ? _raw_spin_unlock_irq+0x24/0x50 [ 5413.202173] ? wait_for_completion+0x88/0x110 [ 5413.206535] wait_for_completion+0xb4/0x110 [ 5413.210724] ? srcu_torture_stats_print+0x110/0x110 [ 5413.215610] srcu_barrier+0x187/0x200 [ 5413.219277] ? rcu_tasks_verify_self_tests+0x50/0x50 [ 5413.224244] ? rdinit_setup+0x2b/0x2b [ 5413.227907] rcu_verify_early_boot_tests+0x2d/0x40 [ 5413.232700] do_one_initcall+0x63/0x310 [ 5413.236541] ? rdinit_setup+0x2b/0x2b [ 5413.240207] ? rcu_read_lock_sched_held+0x52/0x80 [ 5413.244912] kernel_init_freeable+0x253/0x28f [ 5413.249273] ? rest_init+0x250/0x250 [ 5413.252846] kernel_init+0xa/0x110 [ 5413.256257] ret_from_fork+0x22/0x30 2) An srcu_struct structure that is initialized before rcu_init_geometry() and used afterward will always have stale rdp->mynode references, resulting in callbacks to be missed in srcu_gp_end(), just like in the previous scenario. This commit therefore causes init_srcu_struct_nodes to initialize the geometry, if needed. This ensures that the srcu_node hierarchy is properly built and distributed from the get-go. Suggested-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Cc: Boqun Feng <boqun.feng@gmail.com> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Neeraj Upadhyay <neeraju@codeaurora.org> Cc: Josh Triplett <josh@joshtriplett.org> Cc: Joel Fernandes <joel@joelfernandes.org> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2021-04-17 21:16:49 +08:00
void rcu_init_geometry(void)
{
ulong d;
int i;
srcu: Fix broken node geometry after early ssp init An srcu_struct structure that is initialized before rcu_init_geometry() will have its srcu_node hierarchy based on CONFIG_NR_CPUS. Once rcu_init_geometry() is called, this hierarchy is compressed as needed for the actual maximum number of CPUs for this system. Later on, that srcu_struct structure is confused, sometimes referring to its initial CONFIG_NR_CPUS-based hierarchy, and sometimes instead to the new num_possible_cpus() hierarchy. For example, each of its ->mynode fields continues to reference the original leaf rcu_node structures, some of which might no longer exist. On the other hand, srcu_for_each_node_breadth_first() traverses to the new node hierarchy. There are at least two bad possible outcomes to this: 1) a) A callback enqueued early on an srcu_data structure (call it *sdp) is recorded pending on sdp->mynode->srcu_data_have_cbs in srcu_funnel_gp_start() with sdp->mynode pointing to a deep leaf (say 3 levels). b) The grace period ends after rcu_init_geometry() shrinks the nodes level to a single one. srcu_gp_end() walks through the new srcu_node hierarchy without ever reaching the old leaves so the callback is never executed. This is easily reproduced on an 8 CPUs machine with CONFIG_NR_CPUS >= 32 and "rcupdate.rcu_self_test=1". The srcu_barrier() after early tests verification never completes and the boot hangs: [ 5413.141029] INFO: task swapper/0:1 blocked for more than 4915 seconds. [ 5413.147564] Not tainted 5.12.0-rc4+ #28 [ 5413.151927] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. [ 5413.159753] task:swapper/0 state:D stack: 0 pid: 1 ppid: 0 flags:0x00004000 [ 5413.168099] Call Trace: [ 5413.170555] __schedule+0x36c/0x930 [ 5413.174057] ? wait_for_completion+0x88/0x110 [ 5413.178423] schedule+0x46/0xf0 [ 5413.181575] schedule_timeout+0x284/0x380 [ 5413.185591] ? wait_for_completion+0x88/0x110 [ 5413.189957] ? mark_held_locks+0x61/0x80 [ 5413.193882] ? mark_held_locks+0x61/0x80 [ 5413.197809] ? _raw_spin_unlock_irq+0x24/0x50 [ 5413.202173] ? wait_for_completion+0x88/0x110 [ 5413.206535] wait_for_completion+0xb4/0x110 [ 5413.210724] ? srcu_torture_stats_print+0x110/0x110 [ 5413.215610] srcu_barrier+0x187/0x200 [ 5413.219277] ? rcu_tasks_verify_self_tests+0x50/0x50 [ 5413.224244] ? rdinit_setup+0x2b/0x2b [ 5413.227907] rcu_verify_early_boot_tests+0x2d/0x40 [ 5413.232700] do_one_initcall+0x63/0x310 [ 5413.236541] ? rdinit_setup+0x2b/0x2b [ 5413.240207] ? rcu_read_lock_sched_held+0x52/0x80 [ 5413.244912] kernel_init_freeable+0x253/0x28f [ 5413.249273] ? rest_init+0x250/0x250 [ 5413.252846] kernel_init+0xa/0x110 [ 5413.256257] ret_from_fork+0x22/0x30 2) An srcu_struct structure that is initialized before rcu_init_geometry() and used afterward will always have stale rdp->mynode references, resulting in callbacks to be missed in srcu_gp_end(), just like in the previous scenario. This commit therefore causes init_srcu_struct_nodes to initialize the geometry, if needed. This ensures that the srcu_node hierarchy is properly built and distributed from the get-go. Suggested-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Cc: Boqun Feng <boqun.feng@gmail.com> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Neeraj Upadhyay <neeraju@codeaurora.org> Cc: Josh Triplett <josh@joshtriplett.org> Cc: Joel Fernandes <joel@joelfernandes.org> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2021-04-17 21:16:49 +08:00
static unsigned long old_nr_cpu_ids;
int rcu_capacity[RCU_NUM_LVLS];
srcu: Fix broken node geometry after early ssp init An srcu_struct structure that is initialized before rcu_init_geometry() will have its srcu_node hierarchy based on CONFIG_NR_CPUS. Once rcu_init_geometry() is called, this hierarchy is compressed as needed for the actual maximum number of CPUs for this system. Later on, that srcu_struct structure is confused, sometimes referring to its initial CONFIG_NR_CPUS-based hierarchy, and sometimes instead to the new num_possible_cpus() hierarchy. For example, each of its ->mynode fields continues to reference the original leaf rcu_node structures, some of which might no longer exist. On the other hand, srcu_for_each_node_breadth_first() traverses to the new node hierarchy. There are at least two bad possible outcomes to this: 1) a) A callback enqueued early on an srcu_data structure (call it *sdp) is recorded pending on sdp->mynode->srcu_data_have_cbs in srcu_funnel_gp_start() with sdp->mynode pointing to a deep leaf (say 3 levels). b) The grace period ends after rcu_init_geometry() shrinks the nodes level to a single one. srcu_gp_end() walks through the new srcu_node hierarchy without ever reaching the old leaves so the callback is never executed. This is easily reproduced on an 8 CPUs machine with CONFIG_NR_CPUS >= 32 and "rcupdate.rcu_self_test=1". The srcu_barrier() after early tests verification never completes and the boot hangs: [ 5413.141029] INFO: task swapper/0:1 blocked for more than 4915 seconds. [ 5413.147564] Not tainted 5.12.0-rc4+ #28 [ 5413.151927] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. [ 5413.159753] task:swapper/0 state:D stack: 0 pid: 1 ppid: 0 flags:0x00004000 [ 5413.168099] Call Trace: [ 5413.170555] __schedule+0x36c/0x930 [ 5413.174057] ? wait_for_completion+0x88/0x110 [ 5413.178423] schedule+0x46/0xf0 [ 5413.181575] schedule_timeout+0x284/0x380 [ 5413.185591] ? wait_for_completion+0x88/0x110 [ 5413.189957] ? mark_held_locks+0x61/0x80 [ 5413.193882] ? mark_held_locks+0x61/0x80 [ 5413.197809] ? _raw_spin_unlock_irq+0x24/0x50 [ 5413.202173] ? wait_for_completion+0x88/0x110 [ 5413.206535] wait_for_completion+0xb4/0x110 [ 5413.210724] ? srcu_torture_stats_print+0x110/0x110 [ 5413.215610] srcu_barrier+0x187/0x200 [ 5413.219277] ? rcu_tasks_verify_self_tests+0x50/0x50 [ 5413.224244] ? rdinit_setup+0x2b/0x2b [ 5413.227907] rcu_verify_early_boot_tests+0x2d/0x40 [ 5413.232700] do_one_initcall+0x63/0x310 [ 5413.236541] ? rdinit_setup+0x2b/0x2b [ 5413.240207] ? rcu_read_lock_sched_held+0x52/0x80 [ 5413.244912] kernel_init_freeable+0x253/0x28f [ 5413.249273] ? rest_init+0x250/0x250 [ 5413.252846] kernel_init+0xa/0x110 [ 5413.256257] ret_from_fork+0x22/0x30 2) An srcu_struct structure that is initialized before rcu_init_geometry() and used afterward will always have stale rdp->mynode references, resulting in callbacks to be missed in srcu_gp_end(), just like in the previous scenario. This commit therefore causes init_srcu_struct_nodes to initialize the geometry, if needed. This ensures that the srcu_node hierarchy is properly built and distributed from the get-go. Suggested-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Cc: Boqun Feng <boqun.feng@gmail.com> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Neeraj Upadhyay <neeraju@codeaurora.org> Cc: Josh Triplett <josh@joshtriplett.org> Cc: Joel Fernandes <joel@joelfernandes.org> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2021-04-17 21:16:49 +08:00
static bool initialized;
if (initialized) {
/*
* Warn if setup_nr_cpu_ids() had not yet been invoked,
* unless nr_cpus_ids == NR_CPUS, in which case who cares?
*/
WARN_ON_ONCE(old_nr_cpu_ids != nr_cpu_ids);
return;
}
old_nr_cpu_ids = nr_cpu_ids;
initialized = true;
/*
* Initialize any unspecified boot parameters.
* The default values of jiffies_till_first_fqs and
* jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
* value, which is a function of HZ, then adding one for each
* RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
*/
d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
if (jiffies_till_first_fqs == ULONG_MAX)
jiffies_till_first_fqs = d;
if (jiffies_till_next_fqs == ULONG_MAX)
jiffies_till_next_fqs = d;
adjust_jiffies_till_sched_qs();
/* If the compile-time values are accurate, just leave. */
if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
nr_cpu_ids == NR_CPUS)
return;
pr_info("Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n",
rcu_fanout_leaf, nr_cpu_ids);
/*
* The boot-time rcu_fanout_leaf parameter must be at least two
* and cannot exceed the number of bits in the rcu_node masks.
* Complain and fall back to the compile-time values if this
* limit is exceeded.
*/
if (rcu_fanout_leaf < 2 ||
rcu_fanout_leaf > sizeof(unsigned long) * 8) {
rcu_fanout_leaf = RCU_FANOUT_LEAF;
WARN_ON(1);
return;
}
/*
* Compute number of nodes that can be handled an rcu_node tree
* with the given number of levels.
*/
rcu_capacity[0] = rcu_fanout_leaf;
for (i = 1; i < RCU_NUM_LVLS; i++)
rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
/*
* The tree must be able to accommodate the configured number of CPUs.
* If this limit is exceeded, fall back to the compile-time values.
*/
if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
rcu_fanout_leaf = RCU_FANOUT_LEAF;
WARN_ON(1);
return;
}
/* Calculate the number of levels in the tree. */
for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
}
rcu_num_lvls = i + 1;
/* Calculate the number of rcu_nodes at each level of the tree. */
for (i = 0; i < rcu_num_lvls; i++) {
int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
}
/* Calculate the total number of rcu_node structures. */
rcu_num_nodes = 0;
for (i = 0; i < rcu_num_lvls; i++)
rcu_num_nodes += num_rcu_lvl[i];
}
/*
* Dump out the structure of the rcu_node combining tree associated
* with the rcu_state structure.
*/
static void __init rcu_dump_rcu_node_tree(void)
{
int level = 0;
struct rcu_node *rnp;
pr_info("rcu_node tree layout dump\n");
pr_info(" ");
rcu_for_each_node_breadth_first(rnp) {
if (rnp->level != level) {
pr_cont("\n");
pr_info(" ");
level = rnp->level;
}
pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
}
pr_cont("\n");
}
struct workqueue_struct *rcu_gp_wq;
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
static void __init kfree_rcu_batch_init(void)
{
int cpu;
int i, j;
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
/* Clamp it to [0:100] seconds interval. */
if (rcu_delay_page_cache_fill_msec < 0 ||
rcu_delay_page_cache_fill_msec > 100 * MSEC_PER_SEC) {
rcu_delay_page_cache_fill_msec =
clamp(rcu_delay_page_cache_fill_msec, 0,
(int) (100 * MSEC_PER_SEC));
pr_info("Adjusting rcutree.rcu_delay_page_cache_fill_msec to %d ms.\n",
rcu_delay_page_cache_fill_msec);
}
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
for_each_possible_cpu(cpu) {
struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
for (i = 0; i < KFREE_N_BATCHES; i++) {
INIT_RCU_WORK(&krcp->krw_arr[i].rcu_work, kfree_rcu_work);
krcp->krw_arr[i].krcp = krcp;
for (j = 0; j < FREE_N_CHANNELS; j++)
INIT_LIST_HEAD(&krcp->krw_arr[i].bulk_head_free[j]);
rcu: Support kfree_bulk() interface in kfree_rcu() The kfree_rcu() logic can be improved further by using kfree_bulk() interface along with "basic batching support" introduced earlier. The are at least two advantages of using "bulk" interface: - in case of large number of kfree_rcu() requests kfree_bulk() reduces the per-object overhead caused by calling kfree() per-object. - reduces the number of cache-misses due to "pointer chasing" between objects which can be far spread between each other. This approach defines a new kfree_rcu_bulk_data structure that stores pointers in an array with a specific size. Number of entries in that array depends on PAGE_SIZE making kfree_rcu_bulk_data structure to be exactly one page. Since it deals with "block-chain" technique there is an extra need in dynamic allocation when a new block is required. Memory is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that allows to skip direct reclaim under low memory condition to prevent stalling and fails silently under high memory pressure. The "emergency path" gets maintained when a system is run out of memory. In that case objects are linked into regular list. The "rcuperf" was run to analyze this change in terms of memory consumption and kfree_bulk() throughput. 1) Testing on the Intel(R) Xeon(R) W-2135 CPU @ 3.70GHz, 12xCPUs with following parameters: kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 kfree_vary_obj_size=1 dev.2020.01.10a branch Default / CONFIG_SLAB 53607352517 ns, loops: 200000, batches: 1885, memory footprint: 1248MB 53529637912 ns, loops: 200000, batches: 1921, memory footprint: 1193MB 53570175705 ns, loops: 200000, batches: 1929, memory footprint: 1250MB Patch / CONFIG_SLAB 23981587315 ns, loops: 200000, batches: 810, memory footprint: 1219MB 23879375281 ns, loops: 200000, batches: 822, memory footprint: 1190MB 24086841707 ns, loops: 200000, batches: 794, memory footprint: 1380MB Default / CONFIG_SLUB 51291025022 ns, loops: 200000, batches: 1713, memory footprint: 741MB 51278911477 ns, loops: 200000, batches: 1671, memory footprint: 719MB 51256183045 ns, loops: 200000, batches: 1719, memory footprint: 647MB Patch / CONFIG_SLUB 50709919132 ns, loops: 200000, batches: 1618, memory footprint: 456MB 50736297452 ns, loops: 200000, batches: 1633, memory footprint: 507MB 50660403893 ns, loops: 200000, batches: 1628, memory footprint: 429MB in case of CONFIG_SLAB there is double increase in performance and slightly higher memory usage. As for CONFIG_SLUB, the performance figures are better together with lower memory usage. 2) Testing on the HiKey-960, arm64, 8xCPUs with below parameters: CONFIG_SLAB=y kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB rcuperf shows approximately ~12% better throughput in case of using "bulk" interface. The "drain logic" or its RCU callback does the work faster that leads to better throughput. Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-01-20 22:42:25 +08:00
}
for (i = 0; i < FREE_N_CHANNELS; i++)
INIT_LIST_HEAD(&krcp->bulk_head[i]);
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
INIT_DELAYED_WORK(&krcp->monitor_work, kfree_rcu_monitor);
INIT_DELAYED_WORK(&krcp->page_cache_work, fill_page_cache_func);
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
krcp->initialized = true;
}
mm: shrinkers: provide shrinkers with names Currently shrinkers are anonymous objects. For debugging purposes they can be identified by count/scan function names, but it's not always useful: e.g. for superblock's shrinkers it's nice to have at least an idea of to which superblock the shrinker belongs. This commit adds names to shrinkers. register_shrinker() and prealloc_shrinker() functions are extended to take a format and arguments to master a name. In some cases it's not possible to determine a good name at the time when a shrinker is allocated. For such cases shrinker_debugfs_rename() is provided. The expected format is: <subsystem>-<shrinker_type>[:<instance>]-<id> For some shrinkers an instance can be encoded as (MAJOR:MINOR) pair. After this change the shrinker debugfs directory looks like: $ cd /sys/kernel/debug/shrinker/ $ ls dquota-cache-16 sb-devpts-28 sb-proc-47 sb-tmpfs-42 mm-shadow-18 sb-devtmpfs-5 sb-proc-48 sb-tmpfs-43 mm-zspool:zram0-34 sb-hugetlbfs-17 sb-pstore-31 sb-tmpfs-44 rcu-kfree-0 sb-hugetlbfs-33 sb-rootfs-2 sb-tmpfs-49 sb-aio-20 sb-iomem-12 sb-securityfs-6 sb-tracefs-13 sb-anon_inodefs-15 sb-mqueue-21 sb-selinuxfs-22 sb-xfs:vda1-36 sb-bdev-3 sb-nsfs-4 sb-sockfs-8 sb-zsmalloc-19 sb-bpf-32 sb-pipefs-14 sb-sysfs-26 thp-deferred_split-10 sb-btrfs:vda2-24 sb-proc-25 sb-tmpfs-1 thp-zero-9 sb-cgroup2-30 sb-proc-39 sb-tmpfs-27 xfs-buf:vda1-37 sb-configfs-23 sb-proc-41 sb-tmpfs-29 xfs-inodegc:vda1-38 sb-dax-11 sb-proc-45 sb-tmpfs-35 sb-debugfs-7 sb-proc-46 sb-tmpfs-40 [roman.gushchin@linux.dev: fix build warnings] Link: https://lkml.kernel.org/r/Yr+ZTnLb9lJk6fJO@castle Reported-by: kernel test robot <lkp@intel.com> Link: https://lkml.kernel.org/r/20220601032227.4076670-4-roman.gushchin@linux.dev Signed-off-by: Roman Gushchin <roman.gushchin@linux.dev> Cc: Christophe JAILLET <christophe.jaillet@wanadoo.fr> Cc: Dave Chinner <dchinner@redhat.com> Cc: Hillf Danton <hdanton@sina.com> Cc: Kent Overstreet <kent.overstreet@gmail.com> Cc: Muchun Song <songmuchun@bytedance.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-06-01 11:22:24 +08:00
if (register_shrinker(&kfree_rcu_shrinker, "rcu-kfree"))
pr_err("Failed to register kfree_rcu() shrinker!\n");
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
}
void __init rcu_init(void)
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
{
int cpu = smp_processor_id();
rcu_early_boot_tests();
rcu: Add basic support for kfree_rcu() batching Recently a discussion about stability and performance of a system involving a high rate of kfree_rcu() calls surfaced on the list [1] which led to another discussion how to prepare for this situation. This patch adds basic batching support for kfree_rcu(). It is "basic" because we do none of the slab management, dynamic allocation, code moving or any of the other things, some of which previous attempts did [2]. These fancier improvements can be follow-up patches and there are different ideas being discussed in those regards. This is an effort to start simple, and build up from there. In the future, an extension to use kfree_bulk and possibly per-slab batching could be done to further improve performance due to cache-locality and slab-specific bulk free optimizations. By using an array of pointers, the worker thread processing the work would need to read lesser data since it does not need to deal with large rcu_head(s) any longer. Torture tests follow in the next patch and show improvements of around 5x reduction in number of grace periods on a 16 CPU system. More details and test data are in that patch. There is an implication with rcu_barrier() with this patch. Since the kfree_rcu() calls can be batched, and may not be handed yet to the RCU machinery in fact, the monitor may not have even run yet to do the queue_rcu_work(), there seems no easy way of implementing rcu_barrier() to wait for those kfree_rcu()s that are already made. So this means a kfree_rcu() followed by an rcu_barrier() does not imply that memory will be freed once rcu_barrier() returns. Another implication is higher active memory usage (although not run-away..) until the kfree_rcu() flooding ends, in comparison to without batching. More details about this are in the second patch which adds an rcuperf test. Finally, in the near future we will get rid of kfree_rcu() special casing within RCU such as in rcu_do_batch and switch everything to just batching. Currently we don't do that since timer subsystem is not yet up and we cannot schedule the kfree_rcu() monitor as the timer subsystem's lock are not initialized. That would also mean getting rid of kfree_call_rcu_nobatch() entirely. [1] http://lore.kernel.org/lkml/20190723035725-mutt-send-email-mst@kernel.org [2] https://lkml.org/lkml/2017/12/19/824 Cc: kernel-team@android.com Cc: kernel-team@lge.com Co-developed-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Byungchul Park <byungchul.park@lge.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> [ paulmck: Applied 0day and Paul Walmsley feedback on ->monitor_todo. ] [ paulmck: Make it work during early boot. ] [ paulmck: Add a crude early boot self-test. ] [ paulmck: Style adjustments and experimental docbook structure header. ] Link: https://lore.kernel.org/lkml/alpine.DEB.2.21.9999.1908161931110.32497@viisi.sifive.com/T/#me9956f66cb611b95d26ae92700e1d901f46e8c59 Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2019-08-06 06:22:27 +08:00
kfree_rcu_batch_init();
rcu: Merge preemptable-RCU functionality into hierarchical RCU Create a kernel/rcutree_plugin.h file that contains definitions for preemptable RCU (or, under the #else branch of the #ifdef, empty definitions for the classic non-preemptable semantics). These definitions fit into plugins defined in kernel/rcutree.c for this purpose. This variant of preemptable RCU uses a new algorithm whose read-side expense is roughly that of classic hierarchical RCU under CONFIG_PREEMPT. This new algorithm's update-side expense is similar to that of classic hierarchical RCU, and, in absence of read-side preemption or blocking, is exactly that of classic hierarchical RCU. Perhaps more important, this new algorithm has a much simpler implementation, saving well over 1,000 lines of code compared to mainline's implementation of preemptable RCU, which will hopefully be retired in favor of this new algorithm. The simplifications are obtained by maintaining per-task nesting state for running tasks, and using a simple lock-protected algorithm to handle accounting when tasks block within RCU read-side critical sections, making use of lessons learned while creating numerous user-level RCU implementations over the past 18 months. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: laijs@cn.fujitsu.com Cc: dipankar@in.ibm.com Cc: akpm@linux-foundation.org Cc: mathieu.desnoyers@polymtl.ca Cc: josht@linux.vnet.ibm.com Cc: dvhltc@us.ibm.com Cc: niv@us.ibm.com Cc: peterz@infradead.org Cc: rostedt@goodmis.org LKML-Reference: <12509746134003-git-send-email-> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-08-23 04:56:52 +08:00
rcu_bootup_announce();
sanitize_kthread_prio();
rcu_init_geometry();
rcu_init_one();
if (dump_tree)
rcu_dump_rcu_node_tree();
if (use_softirq)
open_softirq(RCU_SOFTIRQ, rcu_core_si);
/*
* We don't need protection against CPU-hotplug here because
* this is called early in boot, before either interrupts
* or the scheduler are operational.
*/
pm_notifier(rcu_pm_notify, 0);
WARN_ON(num_online_cpus() > 1); // Only one CPU this early in boot.
rcutree_prepare_cpu(cpu);
rcu_cpu_starting(cpu);
rcutree_online_cpu(cpu);
/* Create workqueue for Tree SRCU and for expedited GPs. */
rcu_gp_wq = alloc_workqueue("rcu_gp", WQ_MEM_RECLAIM, 0);
WARN_ON(!rcu_gp_wq);
rcu: Move expedited grace period (GP) work to RT kthread_worker Enabling CONFIG_RCU_BOOST did not reduce RCU expedited grace-period latency because its workqueues run at SCHED_OTHER, and thus can be delayed by normal processes. This commit avoids these delays by moving the expedited GP work items to a real-time-priority kthread_worker. This option is controlled by CONFIG_RCU_EXP_KTHREAD and disabled by default on PREEMPT_RT=y kernels which disable expedited grace periods after boot by unconditionally setting rcupdate.rcu_normal_after_boot=1. The results were evaluated on arm64 Android devices (6GB ram) running 5.10 kernel, and capturing trace data in critical user-level code. The table below shows the resulting order-of-magnitude improvements in synchronize_rcu_expedited() latency: ------------------------------------------------------------------------ | | workqueues | kthread_worker | Diff | ------------------------------------------------------------------------ | Count | 725 | 688 | | ------------------------------------------------------------------------ | Min Duration (ns) | 326 | 447 | 37.12% | ------------------------------------------------------------------------ | Q1 (ns) | 39,428 | 38,971 | -1.16% | ------------------------------------------------------------------------ | Q2 - Median (ns) | 98,225 | 69,743 | -29.00% | ------------------------------------------------------------------------ | Q3 (ns) | 342,122 | 126,638 | -62.98% | ------------------------------------------------------------------------ | Max Duration (ns) | 372,766,967 | 2,329,671 | -99.38% | ------------------------------------------------------------------------ | Avg Duration (ns) | 2,746,353 | 151,242 | -94.49% | ------------------------------------------------------------------------ | Standard Deviation (ns) | 19,327,765 | 294,408 | | ------------------------------------------------------------------------ The below table show the range of maximums/minimums for synchronize_rcu_expedited() latency from all experiments: ------------------------------------------------------------------------ | | workqueues | kthread_worker | Diff | ------------------------------------------------------------------------ | Total No. of Experiments | 25 | 23 | | ------------------------------------------------------------------------ | Largest Maximum (ns) | 372,766,967 | 2,329,671 | -99.38% | ------------------------------------------------------------------------ | Smallest Maximum (ns) | 38,819 | 86,954 | 124.00% | ------------------------------------------------------------------------ | Range of Maximums (ns) | 372,728,148 | 2,242,717 | | ------------------------------------------------------------------------ | Largest Minimum (ns) | 88,623 | 27,588 | -68.87% | ------------------------------------------------------------------------ | Smallest Minimum (ns) | 326 | 447 | 37.12% | ------------------------------------------------------------------------ | Range of Minimums (ns) | 88,297 | 27,141 | | ------------------------------------------------------------------------ Cc: "Paul E. McKenney" <paulmck@kernel.org> Cc: Tejun Heo <tj@kernel.org> Reported-by: Tim Murray <timmurray@google.com> Reported-by: Wei Wang <wvw@google.com> Tested-by: Kyle Lin <kylelin@google.com> Tested-by: Chunwei Lu <chunweilu@google.com> Tested-by: Lulu Wang <luluw@google.com> Signed-off-by: Kalesh Singh <kaleshsingh@google.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2022-04-09 08:35:27 +08:00
rcu_alloc_par_gp_wq();
/* Fill in default value for rcutree.qovld boot parameter. */
/* -After- the rcu_node ->lock fields are initialized! */
if (qovld < 0)
qovld_calc = DEFAULT_RCU_QOVLD_MULT * qhimark;
else
qovld_calc = qovld;
// Kick-start in case any polled grace periods started early.
(void)start_poll_synchronize_rcu_expedited();
rcu_test_sync_prims();
"Tree RCU": scalable classic RCU implementation This patch fixes a long-standing performance bug in classic RCU that results in massive internal-to-RCU lock contention on systems with more than a few hundred CPUs. Although this patch creates a separate flavor of RCU for ease of review and patch maintenance, it is intended to replace classic RCU. This patch still handles stress better than does mainline, so I am still calling it ready for inclusion. This patch is against the -tip tree. Nevertheless, experience on an actual 1000+ CPU machine would still be most welcome. Most of the changes noted below were found while creating an rcutiny (which should permit ejecting the current rcuclassic) and while doing detailed line-by-line documentation. Updates from v9 (http://lkml.org/lkml/2008/12/2/334): o Fixes from remainder of line-by-line code walkthrough, including comment spelling, initialization, undesirable narrowing due to type conversion, removing redundant memory barriers, removing redundant local-variable initialization, and removing redundant local variables. I do not believe that any of these fixes address the CPU-hotplug issues that Andi Kleen was seeing, but please do give it a whirl in case the machine is smarter than I am. A writeup from the walkthrough may be found at the following URL, in case you are suffering from terminal insomnia or masochism: http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf o Made rcutree tracing use seq_file, as suggested some time ago by Lai Jiangshan. o Added a .csv variant of the rcudata debugfs trace file, to allow people having thousands of CPUs to drop the data into a spreadsheet. Tested with oocalc and gnumeric. Updated documentation to suit. Updates from v8 (http://lkml.org/lkml/2008/11/15/139): o Fix a theoretical race between grace-period initialization and force_quiescent_state() that could occur if more than three jiffies were required to carry out the grace-period initialization. Which it might, if you had enough CPUs. o Apply Ingo's printk-standardization patch. o Substitute local variables for repeated accesses to global variables. o Fix comment misspellings and redundant (but harmless) increments of ->n_rcu_pending (this latter after having explicitly added it). o Apply checkpatch fixes. Updates from v7 (http://lkml.org/lkml/2008/10/10/291): o Fixed a number of problems noted by Gautham Shenoy, including the cpu-stall-detection bug that he was having difficulty convincing me was real. ;-) o Changed cpu-stall detection to wait for ten seconds rather than three in order to reduce false positive, as suggested by Ingo Molnar. o Produced a design document (http://lwn.net/Articles/305782/). The act of writing this document uncovered a number of both theoretical and "here and now" bugs as noted below. o Fix dynticks_nesting accounting confusion, simplify WARN_ON() condition, fix kerneldoc comments, and add memory barriers in dynticks interface functions. o Add more data to tracing. o Remove unused "rcu_barrier" field from rcu_data structure. o Count calls to rcu_pending() from scheduling-clock interrupt to use as a surrogate timebase should jiffies stop counting. o Fix a theoretical race between force_quiescent_state() and grace-period initialization. Yes, initialization does have to go on for some jiffies for this race to occur, but given enough CPUs... Updates from v6 (http://lkml.org/lkml/2008/9/23/448): o Fix a number of checkpatch.pl complaints. o Apply review comments from Ingo Molnar and Lai Jiangshan on the stall-detection code. o Fix several bugs in !CONFIG_SMP builds. o Fix a misspelled config-parameter name so that RCU now announces at boot time if stall detection is configured. o Run tests on numerous combinations of configurations parameters, which after the fixes above, now build and run correctly. Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line): o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a changeset some time ago, and finally got around to retesting this option). o Fix some tracing bugs in rcupreempt that caused incorrect totals to be printed. o I now test with a more brutal random-selection online/offline script (attached). Probably more brutal than it needs to be on the people reading it as well, but so it goes. o A number of optimizations and usability improvements: o Make rcu_pending() ignore the grace-period timeout when there is no grace period in progress. o Make force_quiescent_state() avoid going for a global lock in the case where there is no grace period in progress. o Rearrange struct fields to improve struct layout. o Make call_rcu() initiate a grace period if RCU was idle, rather than waiting for the next scheduling clock interrupt. o Invoke rcu_irq_enter() and rcu_irq_exit() only when idle, as suggested by Andi Kleen. I still don't completely trust this change, and might back it out. o Make CONFIG_RCU_TRACE be the single config variable manipulated for all forms of RCU, instead of the prior confusion. o Document tracing files and formats for both rcupreempt and rcutree. Updates from v4 for those missing v5 given its bad subject line: o Separated dynticks interface so that NMIs and irqs call separate functions, greatly simplifying it. In particular, this code no longer requires a proof of correctness. ;-) o Separated dynticks state out into its own per-CPU structure, avoiding the duplicated accounting. o The case where a dynticks-idle CPU runs an irq handler that invokes call_rcu() is now correctly handled, forcing that CPU out of dynticks-idle mode. o Review comments have been applied (thank you all!!!). For but one example, fixed the dynticks-ordering issue that Manfred pointed out, saving me much debugging. ;-) o Adjusted rcuclassic and rcupreempt to handle dynticks changes. Attached is an updated patch to Classic RCU that applies a hierarchy, greatly reducing the contention on the top-level lock for large machines. This passes 10-hour concurrent rcutorture and online-offline testing on 128-CPU ppc64 without dynticks enabled, and exposes some timekeeping bugs in presence of dynticks (exciting working on a system where "sleep 1" hangs until interrupted...), which were fixed in the 2.6.27 kernel. It is getting more reliable than mainline by some measures, so the next version will be against -tip for inclusion. See also Manfred Spraul's recent patches (or his earlier work from 2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2). We will converge onto a common patch in the fullness of time, but are currently exploring different regions of the design space. That said, I have already gratefully stolen quite a few of Manfred's ideas. This patch provides CONFIG_RCU_FANOUT, which controls the bushiness of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on 64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT, there is no hierarchy. By default, the RCU initialization code will adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable this balancing, allowing the hierarchy to be exactly aligned to the underlying hardware. Up to two levels of hierarchy are permitted (in addition to the root node), allowing up to 16,384 CPUs on 32-bit systems and up to 262,144 CPUs on 64-bit systems. I just know that I am going to regret saying this, but this seems more than sufficient for the foreseeable future. (Some architectures might wish to set CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs. If this becomes a real problem, additional levels can be added, but I doubt that it will make a significant difference on real hardware.) In the common case, a given CPU will manipulate its private rcu_data structure and the rcu_node structure that it shares with its immediate neighbors. This can reduce both lock and memory contention by multiple orders of magnitude, which should eliminate the need for the strange manipulations that are reported to be required when running Linux on very large systems. Some shortcomings: o More bugs will probably surface as a result of an ongoing line-by-line code inspection. Patches will be provided as required. o There are probably hangs, rcutorture failures, &c. Seems quite stable on a 128-CPU machine, but that is kind of small compared to 4096 CPUs. However, seems to do better than mainline. Patches will be provided as required. o The memory footprint of this version is several KB larger than rcuclassic. A separate UP-only rcutiny patch will be provided, which will reduce the memory footprint significantly, even compared to the old rcuclassic. One such patch passes light testing, and has a memory footprint smaller even than rcuclassic. Initial reaction from various embedded guys was "it is not worth it", so am putting it aside. Credits: o Manfred Spraul for ideas, review comments, and bugs spotted, as well as some good friendly competition. ;-) o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers, Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton for reviews and comments. o Thomas Gleixner for much-needed help with some timer issues (see patches below). o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos, Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines alive despite my heavy abuse^Wtesting. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-19 04:55:32 +08:00
}
#include "tree_stall.h"
#include "tree_exp.h"
#include "tree_nocb.h"
#include "tree_plugin.h"