Hi,
A user reported a kernel bug when running a particular program that did
the following:
created 32 threads
- each thread took a mutex, grabbed a global offset, added a buffer size
to that offset, released the lock
- read from the given offset in the file
- created a new thread to do the same
- exited
The result is that cfq's close cooperator logic would trigger, as the
threads were issuing I/O within the mean seek distance of one another.
This workload managed to routinely trigger a use after free bug when
walking the list of merge candidates for a particular cfqq
(cfqq->new_cfqq). The logic used for merging queues looks like this:
static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
{
int process_refs, new_process_refs;
struct cfq_queue *__cfqq;
/* Avoid a circular list and skip interim queue merges */
while ((__cfqq = new_cfqq->new_cfqq)) {
if (__cfqq == cfqq)
return;
new_cfqq = __cfqq;
}
process_refs = cfqq_process_refs(cfqq);
/*
* If the process for the cfqq has gone away, there is no
* sense in merging the queues.
*/
if (process_refs == 0)
return;
/*
* Merge in the direction of the lesser amount of work.
*/
new_process_refs = cfqq_process_refs(new_cfqq);
if (new_process_refs >= process_refs) {
cfqq->new_cfqq = new_cfqq;
atomic_add(process_refs, &new_cfqq->ref);
} else {
new_cfqq->new_cfqq = cfqq;
atomic_add(new_process_refs, &cfqq->ref);
}
}
When a merge candidate is found, we add the process references for the
queue with less references to the queue with more. The actual merging
of queues happens when a new request is issued for a given cfqq. In the
case of the test program, it only does a single pread call to read in
1MB, so the actual merge never happens.
Normally, this is fine, as when the queue exits, we simply drop the
references we took on the other cfqqs in the merge chain:
/*
* If this queue was scheduled to merge with another queue, be
* sure to drop the reference taken on that queue (and others in
* the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
*/
__cfqq = cfqq->new_cfqq;
while (__cfqq) {
if (__cfqq == cfqq) {
WARN(1, "cfqq->new_cfqq loop detected\n");
break;
}
next = __cfqq->new_cfqq;
cfq_put_queue(__cfqq);
__cfqq = next;
}
However, there is a hole in this logic. Consider the following (and
keep in mind that each I/O keeps a reference to the cfqq):
q1->new_cfqq = q2 // q2 now has 2 process references
q3->new_cfqq = q2 // q2 now has 3 process references
// the process associated with q2 exits
// q2 now has 2 process references
// queue 1 exits, drops its reference on q2
// q2 now has 1 process reference
// q3 exits, so has 0 process references, and hence drops its references
// to q2, which leaves q2 also with 0 process references
q4 comes along and wants to merge with q3
q3->new_cfqq still points at q2! We follow that link and end up at an
already freed cfqq.
So, the fix is to not follow a merge chain if the top-most queue does
not have a process reference, otherwise any queue in the chain could be
already freed. I also changed the logic to disallow merging with a
queue that does not have any process references. Previously, we did
this check for one of the merge candidates, but not the other. That
doesn't really make sense.
Without the attached patch, my system would BUG within a couple of
seconds of running the reproducer program. With the patch applied, my
system ran the program for over an hour without issues.
This addresses the following bugzilla:
https://bugzilla.kernel.org/show_bug.cgi?id=16217
Thanks a ton to Phil Carns for providing the bug report and an excellent
reproducer.
[ Note for stable: this applies to 2.6.32/33/34 ].
Signed-off-by: Jeff Moyer <jmoyer@redhat.com>
Reported-by: Phil Carns <carns@mcs.anl.gov>
Cc: stable@kernel.org
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>
Use small consequent indexes as radix tree keys instead of sparse cfqd address.
This change will reduce radix tree depth from 11 (6 for 32-bit hosts)
to 1 if host have <=64 disks under cfq control, or to 0 if there only one disk.
So, this patch save 10*560 bytes for each process (5*296 for 32-bit hosts)
For each cfqd allocate cic index from ida.
To unlink dead cic from tree without cfqd access store index into ->key.
(bit 0 -- dead mark, bits 1..30 -- index: ida produce id in range 0..2^31-1)
Signed-off-by: Konstantin Khlebnikov <khlebnikov@openvz.org>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Remove ->dead_key field from cfq_io_context to shrink its size to 128 bytes.
(64 bytes for 32-bit hosts)
Use lower bit in ->key as dead-mark, instead of moving key to separate field.
After this for dead cfq_io_context we got cic->key != cfqd automatically.
Thus, io_context's last-hit cache should work without changing.
Now to check ->key for non-dead state compare it with cfqd,
instead of checking ->key for non-null value as it was before.
Plus remove obsolete race protection in cfq_cic_lookup.
This race gone after v2.6.24-1728-g4ac845a
Signed-off-by: Konstantin Khlebnikov <khlebnikov@openvz.org>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Remove all rcu head inits. We don't care about the RCU head state before passing
it to call_rcu() anyway. Only leave the "on_stack" variants so debugobjects can
keep track of objects on stack.
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
This patch fixes few usability and configurability issues.
o All the cgroup based controller options are configurable from
"Genral Setup/Control Group Support/" menu. blkio is the only exception.
Hence make this option visible in above menu and make it configurable from
there to bring it inline with rest of the cgroup based controllers.
o Get rid of CONFIG_DEBUG_CFQ_IOSCHED.
This option currently does two things.
- Enable printing of cgroup paths in blktrace
- Enables CONFIG_DEBUG_BLK_CGROUP, which in turn displays additional stat
files in cgroup.
If we are using group scheduling, blktrace data is of not really much use
if cgroup information is not present. To get this data, currently one has to
also enable CONFIG_DEBUG_CFQ_IOSCHED, which in turn brings the overhead of
all the additional debug stat files which is not desired.
Hence, this patch moves printing of cgroup paths under
CONFIG_CFQ_GROUP_IOSCHED.
This allows us to get rid of CONFIG_DEBUG_CFQ_IOSCHED completely. Now all
the debug stat files are controlled only by CONFIG_DEBUG_BLK_CGROUP which
can be enabled through config menu.
Signed-off-by: Vivek Goyal <vgoyal@redhat.com>
Acked-by: Divyesh Shah <dpshah@google.com>
Reviewed-by: Gui Jianfeng <guijianfeng@cn.fujitsu.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Fixes compile errors in blk-cgroup code for empty_time stat and a merge fix in
CFQ. The first error was when CONFIG_DEBUG_CFQ_IOSCHED is not set.
Signed-off-by: Divyesh Shah <dpshah@google.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Changelog from v1:
o Call blkiocg_update_idle_time_stats() at cfq_rq_enqueued() instead of at
dispatch time.
Changelog from original patchset: (in response to Vivek Goyal's comments)
o group blkiocg_update_blkio_group_dequeue_stats() with other DEBUG functions
o rename blkiocg_update_set_active_queue_stats() to
blkiocg_update_avg_queue_size_stats()
o s/request/io/ in blkiocg_update_request_add_stats() and
blkiocg_update_request_remove_stats()
o Call cfq_del_timer() at request dispatch() instead of
blkiocg_update_idle_time_stats()
Signed-off-by: Divyesh Shah<dpshah@google.com>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Currently, IO Controller makes use of blkio.weight to assign weight for
all devices. Here a new user interface "blkio.weight_device" is introduced to
assign different weights for different devices. blkio.weight becomes the
default value for devices which are not configured by "blkio.weight_device"
You can use the following format to assigned specific weight for a given
device:
#echo "major:minor weight" > blkio.weight_device
major:minor represents device number.
And you can remove weight for a given device as following:
#echo "major:minor 0" > blkio.weight_device
V1->V2 changes:
- use user interface "weight_device" instead of "policy" suggested by Vivek
- rename some struct suggested by Vivek
- rebase to 2.6-block "for-linus" branch
- remove an useless list_empty check pointed out by Li Zefan
- some trivial typo fix
V2->V3 changes:
- Move policy_*_node() functions up to get rid of forward declarations
- rename related functions by adding prefix "blkio_"
Signed-off-by: Gui Jianfeng <guijianfeng@cn.fujitsu.com>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
* 'for-linus' of git://git.kernel.dk/linux-2.6-block: (34 commits)
cfq-iosched: Fix the incorrect timeslice accounting with forced_dispatch
loop: Update mtime when writing using aops
block: expose the statistics in blkio.time and blkio.sectors for the root cgroup
backing-dev: Handle class_create() failure
Block: Fix block/elevator.c elevator_get() off-by-one error
drbd: lc_element_by_index() never returns NULL
cciss: unlock on error path
cfq-iosched: Do not merge queues of BE and IDLE classes
cfq-iosched: Add additional blktrace log messages in CFQ for easier debugging
i2o: Remove the dangerous kobj_to_i2o_device macro
block: remove 16 bytes of padding from struct request on 64bits
cfq-iosched: fix a kbuild regression
block: make CONFIG_BLK_CGROUP visible
Remove GENHD_FL_DRIVERFS
block: Export max number of segments and max segment size in sysfs
block: Finalize conversion of block limits functions
block: Fix overrun in lcm() and move it to lib
vfs: improve writeback_inodes_wb()
paride: fix off-by-one test
drbd: fix al-to-on-disk-bitmap for 4k logical_block_size
...
When CFQ dispatches requests forcefully due to a barrier or changing iosched,
it runs through all cfqq's dispatching requests and then expires each queue.
However, it does not activate a cfqq before flushing its IOs resulting in
using stale values for computing slice_used.
This patch fixes it by calling activate queue before flushing reuqests from
each queue.
This is useful mostly for barrier requests because when the iosched is changing
it really doesnt matter if we have incorrect accounting since we're going to
break down all structures anyway.
We also now expire the current timeslice before moving on with the dispatch
to accurately account slice used for that cfqq.
Signed-off-by: Divyesh Shah<dpshah@google.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
1) group_wait_time - This is the amount of time the cgroup had to wait to get a
timeslice for one of its queues from when it became busy, i.e., went from 0
to 1 request queued. This is different from the io_wait_time which is the
cumulative total of the amount of time spent by each IO in that cgroup waiting
in the scheduler queue. This stat is a great way to find out any jobs in the
fleet that are being starved or waiting for longer than what is expected (due
to an IO controller bug or any other issue).
2) empty_time - This is the amount of time a cgroup spends w/o any pending
requests. This stat is useful when a job does not seem to be able to use its
assigned disk share by helping check if that is happening due to an IO
controller bug or because the job is not submitting enough IOs.
3) idle_time - This is the amount of time spent by the IO scheduler idling
for a given cgroup in anticipation of a better request than the exising ones
from other queues/cgroups.
All these stats are recorded using start and stop events. When reading these
stats, we do not add the delta between the current time and the last start time
if we're between the start and stop events. We avoid doing this to make sure
that these numbers are always monotonically increasing when read. Since we're
using sched_clock() which may use the tsc as its source, it may induce some
inconsistency (due to tsc resync across cpus) if we included the current delta.
Signed-off-by: Divyesh Shah<dpshah@google.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
These stats are useful for getting a feel for the queue depth of the cgroup,
i.e., how filled up its queues are at a given instant and over the existence of
the cgroup. This ability is useful when debugging problems in the wild as it
helps understand the application's IO pattern w/o having to read through the
userspace code (coz its tedious or just not available) or w/o the ability
to run blktrace (since you may not have root access and/or not want to disturb
performance).
Signed-off-by: Divyesh Shah<dpshah@google.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
This includes both the number of bios merged into requests belonging to this
cgroup as well as the number of requests merged together.
In the past, we've observed different merging behavior across upstream kernels,
some by design some actual bugs. This stat helps a lot in debugging such
problems when applications report decreased throughput with a new kernel
version.
This needed adding an extra elevator function to capture bios being merged as I
did not want to pollute elevator code with blkiocg knowledge and hence needed
the accounting invocation to come from CFQ.
Signed-off-by: Divyesh Shah<dpshah@google.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
that include some minor fixes and addresses all comments.
Changelog: (most based on Vivek Goyal's comments)
o renamed blkiocg_reset_write to blkiocg_reset_stats
o more clarification in the documentation on io_service_time and io_wait_time
o Initialize blkg->stats_lock
o rename io_add_stat to blkio_add_stat and declare it static
o use bool for direction and sync
o derive direction and sync info from existing rq methods
o use 12 for major:minor string length
o define io_service_time better to cover the NCQ case
o add a separate reset_stats interface
o make the indexed stats a 2d array to simplify macro and function pointer code
o blkio.time now exports in jiffies as before
o Added stats description in patch description and
Documentation/cgroup/blkio-controller.txt
o Prefix all stats functions with blkio and make them static as applicable
o replace IO_TYPE_MAX with IO_TYPE_TOTAL
o Moved #define constant to top of blk-cgroup.c
o Pass dev_t around instead of char *
o Add note to documentation file about resetting stats
o use BLK_CGROUP_MODULE in addition to BLK_CGROUP config option in #ifdef
statements
o Avoid struct request specific knowledge in blk-cgroup. blk-cgroup.h now has
rq_direction() and rq_sync() functions which are used by CFQ and when using
io-controller at a higher level, bio_* functions can be added.
Signed-off-by: Divyesh Shah<dpshah@google.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Currently, the io statistics for the root cgroup are maintained, but
they are not shown because the device information is not available at
the point that the root blkio cgroup is created. This patch updates
the device information when the statistics are updated so that the
statistics become visible.
Signed-off-by: Ricky Benitez <rickyb@google.com>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
We also add start_time_ns and io_start_time_ns fields to struct request
here to record the time when a request is created and when it is
dispatched to device. We use ns uints here as ms and jiffies are
not very useful for non-rotational media.
Signed-off-by: Divyesh Shah<dpshah@google.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
- io_service_time
- io_wait_time
- io_serviced
- io_service_bytes
These stats are accumulated per operation type helping us to distinguish between
read and write, and sync and async IO. This patch does not increment any of
these stats.
Signed-off-by: Divyesh Shah<dpshah@google.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
that info at request dispatch with other stats now. This patch removes the
existing support for accounting sectors for a blkio_group. This will be added
back differently in the next two patches.
Signed-off-by: Divyesh Shah<dpshah@google.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
percpu.h is included by sched.h and module.h and thus ends up being
included when building most .c files. percpu.h includes slab.h which
in turn includes gfp.h making everything defined by the two files
universally available and complicating inclusion dependencies.
percpu.h -> slab.h dependency is about to be removed. Prepare for
this change by updating users of gfp and slab facilities include those
headers directly instead of assuming availability. As this conversion
needs to touch large number of source files, the following script is
used as the basis of conversion.
http://userweb.kernel.org/~tj/misc/slabh-sweep.py
The script does the followings.
* Scan files for gfp and slab usages and update includes such that
only the necessary includes are there. ie. if only gfp is used,
gfp.h, if slab is used, slab.h.
* When the script inserts a new include, it looks at the include
blocks and try to put the new include such that its order conforms
to its surrounding. It's put in the include block which contains
core kernel includes, in the same order that the rest are ordered -
alphabetical, Christmas tree, rev-Xmas-tree or at the end if there
doesn't seem to be any matching order.
* If the script can't find a place to put a new include (mostly
because the file doesn't have fitting include block), it prints out
an error message indicating which .h file needs to be added to the
file.
The conversion was done in the following steps.
1. The initial automatic conversion of all .c files updated slightly
over 4000 files, deleting around 700 includes and adding ~480 gfp.h
and ~3000 slab.h inclusions. The script emitted errors for ~400
files.
2. Each error was manually checked. Some didn't need the inclusion,
some needed manual addition while adding it to implementation .h or
embedding .c file was more appropriate for others. This step added
inclusions to around 150 files.
3. The script was run again and the output was compared to the edits
from #2 to make sure no file was left behind.
4. Several build tests were done and a couple of problems were fixed.
e.g. lib/decompress_*.c used malloc/free() wrappers around slab
APIs requiring slab.h to be added manually.
5. The script was run on all .h files but without automatically
editing them as sprinkling gfp.h and slab.h inclusions around .h
files could easily lead to inclusion dependency hell. Most gfp.h
inclusion directives were ignored as stuff from gfp.h was usually
wildly available and often used in preprocessor macros. Each
slab.h inclusion directive was examined and added manually as
necessary.
6. percpu.h was updated not to include slab.h.
7. Build test were done on the following configurations and failures
were fixed. CONFIG_GCOV_KERNEL was turned off for all tests (as my
distributed build env didn't work with gcov compiles) and a few
more options had to be turned off depending on archs to make things
build (like ipr on powerpc/64 which failed due to missing writeq).
* x86 and x86_64 UP and SMP allmodconfig and a custom test config.
* powerpc and powerpc64 SMP allmodconfig
* sparc and sparc64 SMP allmodconfig
* ia64 SMP allmodconfig
* s390 SMP allmodconfig
* alpha SMP allmodconfig
* um on x86_64 SMP allmodconfig
8. percpu.h modifications were reverted so that it could be applied as
a separate patch and serve as bisection point.
Given the fact that I had only a couple of failures from tests on step
6, I'm fairly confident about the coverage of this conversion patch.
If there is a breakage, it's likely to be something in one of the arch
headers which should be easily discoverable easily on most builds of
the specific arch.
Signed-off-by: Tejun Heo <tj@kernel.org>
Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Even if they are found to be co-operating.
The prio_trees do not have any IDLE cfqqs on them. cfq_close_cooperator()
is called from cfq_select_queue() and cfq_completed_request(). The latter
ensures that the close cooperator code does not get invoked if the current
cfqq is of class IDLE but the former doesn't seem to have any such checks.
So an IDLE cfqq may get merged with a BE cfqq from the same group which
should be avoided.
Signed-off-by: Divyesh Shah<dpshah@google.com>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
These have helped us debug some issues we've noticed in earlier IO
controller versions and should be useful now as well. The extra logging
covers:
- idling behavior. Since there are so many conditions based on which we decide
to idle or not, this patch adds a log message for some conditions that we've
found useful.
- workload slices and current prio and workload type
Changelog from v1:
o moved log message from cfq_set_active_queue() to __cfq_set_active_queue()
o changed queue_count to st->count
Signed-off-by: Divyesh Shah<dpshah@google.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Alex Shi reported a kbuild regression which is about 10% performance lost.
He bisected to this commit: 3dde36ddea.
The reason is cfqq_close() can't find close cooperator. Restoring
cfq_rq_close()'s threshold to original value makes the regression go away.
Since for_preempt parameter isn't used anymore, this patch deletes it.
Reported-by: Alex Shi <alex.shi@intel.com>
Signed-off-by: Shaohua Li <shaohua.li@intel.com>
Acked-by: Corrado Zoccolo <czoccolo@gmail.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Reorder cfq_rb_root to remove 8 bytes of padding on 64 bit builds.
Consequently removing 56 bytes from cfq_group and 64 bytes from
cfq_data.
Signed-off-by: Richard Kennedy <richard@rsk.demon.co.uk>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Currently a queue can only dispatch up to 4 requests if there are other queues.
This isn't optimal, device can handle more requests, for example, AHCI can
handle 31 requests. I can understand the limit is for fairness, but we could
do a tweak: if the queue still has a lot of slice left, sounds we could
ignore the limit. Test shows this boost my workload (two thread randread of
a SSD) from 78m/s to 100m/s.
Thanks for suggestions from Corrado and Vivek for the patch.
Signed-off-by: Shaohua Li <shaohua.li@intel.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Counters for requests "in flight" and "in driver" are used asymmetrically
in cfq_may_dispatch, and have slightly different meaning.
We split the rq_in_flight counter (was sync_flight) to count both sync
and async requests, in order to use this one, which is more accurate in
some corner cases.
The rq_in_driver counter is coalesced, since individual sync/async counts
are not used any more.
Signed-off-by: Corrado Zoccolo <czoccolo@gmail.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
CFQ currently applies the same logic of detecting seeky queues and
grouping them together for rotational disks as well as SSDs.
For SSDs, the time to complete a request doesn't depend on the
request location, but only on the size.
This patch therefore changes the criterion to group queues by
request size in case of SSDs, in order to achieve better fairness.
Signed-off-by: Corrado Zoccolo <czoccolo@gmail.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Current seeky detection is based on average seek lenght.
This is suboptimal, since the average will not distinguish between:
* a process doing medium sized seeks
* a process doing some sequential requests interleaved with larger seeks
and even a medium seek can take lot of time, if the requested sector
happens to be behind the disk head in the rotation (50% probability).
Therefore, we change the seeky queue detection to work as follows:
* each request can be classified as sequential if it is very close to
the current head position, i.e. it is likely in the disk cache (disks
usually read more data than requested, and put it in cache for
subsequent reads). Otherwise, the request is classified as seeky.
* an history window of the last 32 requests is kept, storing the
classification result.
* A queue is marked as seeky if more than 1/8 of the last 32 requests
were seeky.
This patch fixes a regression reported by Yanmin, on mmap 64k random
reads.
Reported-by: Yanmin Zhang <yanmin_zhang@linux.intel.com>
Signed-off-by: Corrado Zoccolo <czoccolo@gmail.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
There's no need to take css reference here, for the caller
has already called rcu_read_lock() to prevent cgroup from
being removed.
Signed-off-by: Gui Jianfeng <guijianfeng@cn.fujitsu.com>
Reviewed-by: Li Zefan <lizf@cn.fujitsu.com>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
This removes 8 bytes of padding from struct cfq_queue on 64 bit builds,
shrinking it's size to 256 bytes, so fitting into 1 fewer cachelines and
allowing 1 more object/slab in it's kmem_cache.
Signed-off-by: Richard Kennedy <richard@rsk.demon.co.uk>
Reviewed-by: Jeff Moyer <jmoyer@redhat.com>
----
patch against 2.6.33-rc8
tested on x86_64 AMDX2
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Currently we split seeky coop queues after 1s, which is too big. Below patch
marks seeky coop queue split_coop flag after one slice. After that, if new
requests come in, the queues will be splitted. Patch is suggested by Corrado.
Signed-off-by: Shaohua Li <shaohua.li@intel.com>
Reviewed-by: Corrado Zoccolo <czoccolo@gmail.com>
Acked-by: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Few weeks back, Shaohua Li had posted similar patch. I am reposting it
with more test results.
This patch does two things.
- Do not idle on async queues.
- It also changes the write queue depth CFQ drives (cfq_may_dispatch()).
Currently, we seem to driving queue depth of 1 always for WRITES. This is
true even if there is only one write queue in the system and all the logic
of infinite queue depth in case of single busy queue as well as slowly
increasing queue depth based on last delayed sync request does not seem to
be kicking in at all.
This patch will allow deeper WRITE queue depths (subjected to the other
WRITE queue depth contstraints like cfq_quantum and last delayed sync
request).
Shaohua Li had reported getting more out of his SSD. For me, I have got
one Lun exported from an HP EVA and when pure buffered writes are on, I
can get more out of the system. Following are test results of pure
buffered writes (with end_fsync=1) with vanilla and patched kernel. These
results are average of 3 sets of run with increasing number of threads.
AVERAGE[bufwfs][vanilla]
-------
job Set NR ReadBW(KB/s) MaxClat(us) WriteBW(KB/s) MaxClat(us)
--- --- -- ------------ ----------- ------------- -----------
bufwfs 3 1 0 0 95349 474141
bufwfs 3 2 0 0 100282 806926
bufwfs 3 4 0 0 109989 2.7301e+06
bufwfs 3 8 0 0 116642 3762231
bufwfs 3 16 0 0 118230 6902970
AVERAGE[bufwfs] [patched kernel]
-------
bufwfs 3 1 0 0 270722 404352
bufwfs 3 2 0 0 206770 1.06552e+06
bufwfs 3 4 0 0 195277 1.62283e+06
bufwfs 3 8 0 0 260960 2.62979e+06
bufwfs 3 16 0 0 299260 1.70731e+06
I also ran buffered writes along with some sequential reads and some
buffered reads going on in the system on a SATA disk because the potential
risk could be that we should not be driving queue depth higher in presence
of sync IO going to keep the max clat low.
With some random and sequential reads going on in the system on one SATA
disk I did not see any significant increase in max clat. So it looks like
other WRITE queue depth control logic is doing its job. Here are the
results.
AVERAGE[brr, bsr, bufw together] [vanilla]
-------
job Set NR ReadBW(KB/s) MaxClat(us) WriteBW(KB/s) MaxClat(us)
--- --- -- ------------ ----------- ------------- -----------
brr 3 1 850 546345 0 0
bsr 3 1 14650 729543 0 0
bufw 3 1 0 0 23908 8274517
brr 3 2 981.333 579395 0 0
bsr 3 2 14149.7 1175689 0 0
bufw 3 2 0 0 21921 1.28108e+07
brr 3 4 898.333 1.75527e+06 0 0
bsr 3 4 12230.7 1.40072e+06 0 0
bufw 3 4 0 0 19722.3 2.4901e+07
brr 3 8 900 3160594 0 0
bsr 3 8 9282.33 1.91314e+06 0 0
bufw 3 8 0 0 18789.3 23890622
AVERAGE[brr, bsr, bufw mixed] [patched kernel]
-------
job Set NR ReadBW(KB/s) MaxClat(us) WriteBW(KB/s) MaxClat(us)
--- --- -- ------------ ----------- ------------- -----------
brr 3 1 837 417973 0 0
bsr 3 1 14357.7 591275 0 0
bufw 3 1 0 0 24869.7 8910662
brr 3 2 1038.33 543434 0 0
bsr 3 2 13351.3 1205858 0 0
bufw 3 2 0 0 18626.3 13280370
brr 3 4 913 1.86861e+06 0 0
bsr 3 4 12652.3 1430974 0 0
bufw 3 4 0 0 15343.3 2.81305e+07
brr 3 8 890 2.92695e+06 0 0
bsr 3 8 9635.33 1.90244e+06 0 0
bufw 3 8 0 0 17200.3 24424392
So looks like it might make sense to include this patch.
Thanks
Vivek
Signed-off-by: Vivek Goyal <vgoyal@redhat.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
In cfq_should_preempt(), we currently allow some cases where a non-RT request
can preempt an ongoing RT cfqq timeslice. This should not happen.
Examples include:
o A sync_noidle wl type non-RT request pre-empting a sync_noidle wl type cfqq
on which we are idling.
o Once we have per-cgroup async queues, a non-RT sync request pre-empting a RT
async cfqq.
Signed-off-by: Divyesh Shah<dpshah@google.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
seek_mean could be very big sometimes, using it as close criteria is meaningless
as this doen't improve any performance. So if it's big, let's fallback to
default value.
Reviewed-by: Corrado Zoccolo <czoccolo@gmail.com>
Signed-off-by: Shaohua Li<shaohua.li@intel.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
o CFQ now internally divides cfq queues in therr workload categories. sync-idle,
sync-noidle and async. Which workload to run depends primarily on rb_key
offset across three service trees. Which is a combination of mulitiple things
including what time queue got queued on the service tree.
There is one exception though. That is if we switched the prio class, say
we served some RT tasks and again started serving BE class, then with-in
BE class we always started with sync-noidle workload irrespective of rb_key
offset in service trees.
This can provide better latencies for sync-noidle workload in the presence
of RT tasks.
o This patch gets rid of that exception and which workload to run with-in
class always depends on lowest rb_key across service trees. The reason
being that now we have multiple BE class groups and if we always switch
to sync-noidle workload with-in group, we can potentially starve a sync-idle
workload with-in group. Same is true for async workload which will be in
root group. Also the workload-switching with-in group will become very
unpredictable as it now depends whether some RT workload was running in
the system or not.
Signed-off-by: Vivek Goyal <vgoyal@redhat.com>
Reviewed-by: Gui Jianfeng <guijianfeng@cn.fujitsu.com>
Acked-by: Corrado Zoccolo <czoccolo@gmail.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
o Currently code does not seem to be using cfqd->nr_groups. Get rid of it.
Signed-off-by: Vivek Goyal <vgoyal@redhat.com>
Reviewed-by: Gui Jianfeng <guijianfeng@cn.fujitsu.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
o allow_merge() already checks if submitting task is pointing to same cfqq
as rq has been queued in. If everything is fine, we should not be having
a task in one cgroup and having a pointer to cfqq in other cgroup.
Well I guess in some situations it can happen and that is, when a random
IO queue has been moved into root cgroup for group_isolation=0. In
this case, tasks's cgroup/group is different from where actually cfqq is,
but this is intentional and in this case merging should be allowed.
The second situation is where due to close cooperator patches, multiple
processes can be sharing a cfqq. If everything implemented right, we should
not end up in a situation where tasks from different processes in different
groups are sharing the same cfqq as we allow merging of cooperating queues
only if they are in same group.
Signed-off-by: Vivek Goyal <vgoyal@redhat.com>
Reviewed-by: Gui Jianfeng <guijianfeng@cn.fujitsu.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
When a group is resumed, if it doesn't have workload slice left,
we should set workload_expires as expired. Otherwise, we might
start from where we left in previous group by error.
Thanks the idea from Corrado.
Signed-off-by: Gui Jianfeng <guijianfeng@cn.fujitsu.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
I think my previous patch introduced a bug which can lead to CFQ hitting
BUG_ON().
The offending commit in for-2.6.33 branch is.
commit 7667aa0630
Author: Vivek Goyal <vgoyal@redhat.com>
Date: Tue Dec 8 17:52:58 2009 -0500
cfq-iosched: Take care of corner cases of group losing share due to deletion
While doing some stress testing on my box, I enountered following.
login: [ 3165.148841] BUG: scheduling while
atomic: swapper/0/0x10000100
[ 3165.149821] Modules linked in: cfq_iosched dm_multipath qla2xxx igb
scsi_transport_fc dm_snapshot [last unloaded: scsi_wait_scan]
[ 3165.149821] Pid: 0, comm: swapper Not tainted
2.6.32-block-for-33-merged-new #3
[ 3165.149821] Call Trace:
[ 3165.149821] <IRQ> [<ffffffff8103fab8>] __schedule_bug+0x5c/0x60
[ 3165.149821] [<ffffffff8103afd7>] ? __wake_up+0x44/0x4d
[ 3165.149821] [<ffffffff8153a979>] schedule+0xe3/0x7bc
[ 3165.149821] [<ffffffff8103a796>] ? cpumask_next+0x1d/0x1f
[ 3165.149821] [<ffffffffa000b21d>] ? cfq_dispatch_requests+0x6ba/0x93e
[cfq_iosched]
[ 3165.149821] [<ffffffff810422d8>] __cond_resched+0x2a/0x35
[ 3165.149821] [<ffffffffa000b21d>] ? cfq_dispatch_requests+0x6ba/0x93e
[cfq_iosched]
[ 3165.149821] [<ffffffff8153b1ee>] _cond_resched+0x2c/0x37
[ 3165.149821] [<ffffffff8100e2db>] is_valid_bugaddr+0x16/0x2f
[ 3165.149821] [<ffffffff811e4161>] report_bug+0x18/0xac
[ 3165.149821] [<ffffffff8100f1fc>] die+0x39/0x63
[ 3165.149821] [<ffffffff8153cde1>] do_trap+0x11a/0x129
[ 3165.149821] [<ffffffff8100d470>] do_invalid_op+0x96/0x9f
[ 3165.149821] [<ffffffffa000b21d>] ? cfq_dispatch_requests+0x6ba/0x93e
[cfq_iosched]
[ 3165.149821] [<ffffffff81034b4d>] ? enqueue_task+0x5c/0x67
[ 3165.149821] [<ffffffff8103ae83>] ? task_rq_unlock+0x11/0x13
[ 3165.149821] [<ffffffff81041aae>] ? try_to_wake_up+0x292/0x2a4
[ 3165.149821] [<ffffffff8100c935>] invalid_op+0x15/0x20
[ 3165.149821] [<ffffffffa000b21d>] ? cfq_dispatch_requests+0x6ba/0x93e
[cfq_iosched]
[ 3165.149821] [<ffffffff810df5a6>] ? virt_to_head_page+0xe/0x2f
[ 3165.149821] [<ffffffff811d8c2a>] blk_peek_request+0x191/0x1a7
[ 3165.149821] [<ffffffff811e5b8d>] ? kobject_get+0x1a/0x21
[ 3165.149821] [<ffffffff812c8d4c>] scsi_request_fn+0x82/0x3df
[ 3165.149821] [<ffffffff8110b2de>] ? bio_fs_destructor+0x15/0x17
[ 3165.149821] [<ffffffff810df5a6>] ? virt_to_head_page+0xe/0x2f
[ 3165.149821] [<ffffffff811d931f>] __blk_run_queue+0x42/0x71
[ 3165.149821] [<ffffffff811d9403>] blk_run_queue+0x26/0x3a
[ 3165.149821] [<ffffffff812c8761>] scsi_run_queue+0x2de/0x375
[ 3165.149821] [<ffffffff812b60ac>] ? put_device+0x17/0x19
[ 3165.149821] [<ffffffff812c92d7>] scsi_next_command+0x3b/0x4b
[ 3165.149821] [<ffffffff812c9b9f>] scsi_io_completion+0x1c9/0x3f5
[ 3165.149821] [<ffffffff812c3c36>] scsi_finish_command+0xb5/0xbe
I think I have hit following BUG_ON() in cfq_dispatch_request().
BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
Please find attached the patch to fix it. I have done some stress testing
with it and have not seen it happening again.
o We should wait on a queue even after slice expiry only if it is empty. If
queue is not empty then continue to expire it.
o If we decide to keep the queue then make cfqq=NULL. Otherwise select_queue()
will return a valid cfqq and cfq_dispatch_request() can hit following
BUG_ON().
BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list))
Reviewed-by: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Vivek Goyal <vgoyal@redhat.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Remove wait_request flag when idle time is being deleted, otherwise
it'll hit this path every time when a request is enqueued.
Signed-off-by: Gui Jianfeng <guijianfeng@cn.fujitsu.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Added a comment to explain the initialization of last_delayed_sync.
Signed-off-by: Corrado Zoccolo <czoccolo@gmail.com>
Acked-by: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
If there is a sequential reader running in a group, we wait for next request
to come in that group after slice expiry and once new request is in, we expire
the queue. Otherwise we delete the group from service tree and group looses
its fair share.
So far I was marking a queue as wait_busy if it had consumed its slice and
it was last queue in the group. But this condition did not cover following
two cases.
1.If a request completed and slice has not expired yet. Next request comes
in and is dispatched to disk. Now select_queue() hits and slice has expired.
This group will be deleted. Because request is still in the disk, this queue
will never get a chance to wait_busy.
2.If request completed and slice has not expired yet. Before next request
comes in (delay due to think time), select_queue() hits and expires the
queue hence group. This queue never got a chance to wait busy.
Gui was hitting the boundary condition 1 and not getting fairness numbers
proportional to weight.
This patch puts the checks for above two conditions and improves the fairness
numbers for sequential workload on rotational media. Check in select_queue()
takes care of case 1 and additional check in should_wait_busy() takes care
of case 2.
Reported-by: Gui Jianfeng <guijianfeng@cn.fujitsu.com>
Signed-off-by: Vivek Goyal <vgoyal@redhat.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Vivek Goyal