The Linux kernel coding style says that single-statement blocks should
omit curly braces unless the other leg of the "if" statement has
multiple statements, in which case the curly braces should be included.
This commit fixes RCU's violations of this rule.
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>
bigrtm: First steps towards getting RCU out of the way of
tens-of-microseconds real-time response on systems compiled
with NR_CPUS=4096. Also cleanups for and increased concurrency
of rcu_barrier() family of primitives.
doctorture: rcutorture and documentation improvements.
fixes: Miscellaneous fixes.
fnh: RCU_FAST_NO_HZ fixes and improvements.
If the nohz= boot parameter disables nohz, then RCU_FAST_NO_HZ needs to
also disable itself. This commit therefore checks for tick_nohz_enabled
being zero, disabling rcu_prepare_for_idle() if so. This commit assumes
that tick_nohz_enabled can change at runtime: If this is not the case,
then a simpler approach suffices.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Currently, if several CPUs in the same package have all lazy RCU
callbacks, their wakeups will be uncorrelated. If all the CPUs are in the
same power domain (as is often the case), this will result in unnecessary
power-ups of the package. This commit therefore uses round_jiffies()
to round the timeouts to a second boundary, increasing the odds that
they can be coalesced with each other or with other timeouts.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
An uninitialized string may be displayed at the end of the rcu_preempt
detected stall info such as
0: (1 GPs behind) idle=075/140000000000000/0 =8?^D=8?^D
^^^^^^^^^^
if CONFIG_RCU_FAST_NO_HZ is not defined.
This trivial patch clears the string in this case.
Signed-off-by: Carsten Emde <C.Emde@osadl.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
The CONFIG_TREE_PREEMPT_RCU and CONFIG_TINY_PREEMPT_RCU versions of
__rcu_read_lock() and __rcu_read_unlock() are identical, so this commit
consolidates them into kernel/rcupdate.h.
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>
The arrival of TREE_PREEMPT_RCU some years back included some ugly
code involving either #ifdef or #ifdef'ed wrapper functions to iterate
over all non-SRCU flavors of RCU. This commit therefore introduces
a for_each_rcu_flavor() iterator over the rcu_state structures for each
flavor of RCU to clean up a bit of the ugliness.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
With the advent of __this_cpu_ptr(), it is no longer necessary to pass
both the rcu_state and rcu_data structures into __rcu_process_callbacks().
This commit therefore computes the rcu_data pointer from the rcu_state
pointer within __rcu_process_callbacks() so that callers can pass in
only the pointer to the rcu_state structure. This paves the way for
linking the rcu_state structures together and iterating over them.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Reviewed-by: Josh Triplett <josh@joshtriplett.org>
This is a preparatory commit for increasing rcu_barrier()'s concurrency.
It adds a pointer in the rcu_data structure to the corresponding call_rcu()
function. This allows a pointer to the rcu_data structure to imply the
function pointer, which allows _rcu_barrier() state to be placed in the
rcu_state structure.
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>
The rcu_node tree array is sized based on compile-time constants,
including NR_CPUS. Although this approach has worked well in the past,
the recent trend by many distros to define NR_CPUS=4096 results in
excessive grace-period-initialization latencies.
This commit therefore substitutes the run-time computed nr_cpu_ids for
the compile-time NR_CPUS when building the tree. This can result in
much of the compile-time-allocated rcu_node array being unused. If
this is a major problem, you are in a specialized situation anyway,
so you can manually adjust the NR_CPUS, RCU_FANOUT, and RCU_FANOUT_LEAF
kernel config parameters.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Time to make the four-level-hierarchy setting less scary, so this
commit removes "Experimental" from the boot-time message. Leave the
message in order to get a heads-up on any possible need to expand to
a five-level hierarchy.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Although making RCU_FANOUT_LEAF a kernel configuration parameter rather
than a fixed constant makes it easier for people to decrease cache-miss
overhead for large systems, it is of little help for people who must
run a single pre-built kernel binary.
This commit therefore allows the value of RCU_FANOUT_LEAF to be
increased (but not decreased!) via a boot-time parameter named
rcutree.rcu_fanout_leaf.
Reported-by: Mike Galbraith <efault@gmx.de>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
This reverts commit 616c310e83.
(Move PREEMPT_RCU preemption to switch_to() invocation).
Testing by Sasha Levin <levinsasha928@gmail.com> showed that this
can result in deadlock due to invoking the scheduler when one of
the runqueue locks is held. Because this commit was simply a
performance optimization, revert it.
Reported-by: Sasha Levin <levinsasha928@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Tested-by: Sasha Levin <levinsasha928@gmail.com>
When a CPU is entering dyntick-idle mode, tick_nohz_stop_sched_tick()
calls rcu_needs_cpu() see if RCU needs that CPU, and, if not, computes the
next wakeup time based on the timer wheels. Only later, when actually
entering the idle loop, rcu_prepare_for_idle() will be invoked. In some
cases, rcu_prepare_for_idle() will post timers to wake the CPU back up.
But all for naught: The next wakeup time for the CPU has already been
computed, and posting a timer afterwards does not force that wakeup
time to be recomputed. This means that rcu_prepare_for_idle()'s have
no effect.
This is not a problem on a busy system because something else will wake
up the CPU soon enough. However, on lightly loaded systems, the CPU
might stay asleep for a considerable length of time. If that CPU has
a callback that the rest of the system is waiting on, the system might
run very slowly or (in theory) even hang.
This commit avoids this problem by having rcu_needs_cpu() give
tick_nohz_stop_sched_tick() an estimate of when RCU will need the CPU
to wake back up, which tick_nohz_stop_sched_tick() takes into account
when programming the CPU's wakeup time. An alternative approach is
for rcu_prepare_for_idle() to use hrtimers instead of normal timers,
but timers are much more efficient than are hrtimers for frequently
and repeatedly posting and cancelling a given timer, which is exactly
what RCU_FAST_NO_HZ does.
Reported-by: Pascal Chapperon <pascal.chapperon@wanadoo.fr>
Reported-by: Heiko Carstens <heiko.carstens@de.ibm.com>
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Tested-by: Heiko Carstens <heiko.carstens@de.ibm.com>
Tested-by: Pascal Chapperon <pascal.chapperon@wanadoo.fr>
The RCU_FAST_NO_HZ code relies on a number of per-CPU variables.
This works, but is hidden from someone scanning the data structures
in rcutree.h. This commit therefore converts these per-CPU variables
to fields in the per-CPU rcu_dynticks structures.
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Tested-by: Heiko Carstens <heiko.carstens@de.ibm.com>
Tested-by: Pascal Chapperon <pascal.chapperon@wanadoo.fr>
In the current code, a short dyntick-idle interval (where there is
at least one non-lazy callback on the CPU) and a long dyntick-idle
interval (where there are only lazy callbacks on the CPU) are traced
identically, which can be less than helpful. This commit therefore
emits different event traces in these two cases.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Tested-by: Heiko Carstens <heiko.carstens@de.ibm.com>
Tested-by: Pascal Chapperon <pascal.chapperon@wanadoo.fr>
barrier: Reduce the amount of disturbance by rcu_barrier() to the rest of
the system. This branch also includes improvements to
RCU_FAST_NO_HZ, which are included here due to conflicts.
fixes: Miscellaneous fixes.
inline: Remaining changes from an abortive attempt to inline
preemptible RCU's __rcu_read_lock(). These are (1) making
exit_rcu() avoid unnecessary work and (2) avoiding having
preemptible RCU record a blocked thread when the scheduler
declines to do a context switch.
srcu: Lai Jiangshan's algorithmic implementation of SRCU, including
call_srcu().
The current initialization of the RCU_FAST_NO_HZ per-CPU variables makes
needless and fragile assumptions about the initial value of things like
the jiffies counter. This commit therefore explicitly initializes all of
them that are better started with a non-zero value. It also adds some
comments describing the per-CPU state variables.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
The current RCU_FAST_NO_HZ assumes that timers do not migrate unless a
CPU goes offline, in which case it assumes that the CPU will have to come
out of dyntick-idle mode (cancelling the timer) in order to go offline.
This is important because when RCU_FAST_NO_HZ permits a CPU to enter
dyntick-idle mode despite having RCU callbacks pending, it posts a timer
on that CPU to force a wakeup on that CPU. This wakeup ensures that the
CPU will eventually handle the end of the grace period, including invoking
its RCU callbacks.
However, Pascal Chapperon's test setup shows that the timer handler
rcu_idle_gp_timer_func() really does get invoked in some cases. This is
problematic because this can cause the CPU that entered dyntick-idle
mode despite still having RCU callbacks pending to remain in
dyntick-idle mode indefinitely, which means that its RCU callbacks might
never be invoked. This situation can result in grace-period delays or
even system hangs, which matches Pascal's observations of slow boot-up
and shutdown (https://lkml.org/lkml/2012/4/5/142). See also the bugzilla:
https://bugzilla.redhat.com/show_bug.cgi?id=806548
This commit therefore causes the "should never be invoked" timer handler
rcu_idle_gp_timer_func() to use smp_call_function_single() to wake up
the CPU for which the timer was intended, allowing that CPU to invoke
its RCU callbacks in a timely manner.
Reported-by: Pascal Chapperon <pascal.chapperon@wanadoo.fr>
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
When running preemptible RCU, if a task exits in an RCU read-side
critical section having blocked within that same RCU read-side critical
section, the task must be removed from the list of tasks blocking a
grace period (perhaps the current grace period, perhaps the next grace
period, depending on timing). The exit() path invokes exit_rcu() to
do this cleanup.
However, the current implementation of exit_rcu() needlessly does the
cleanup even if the task did not block within the current RCU read-side
critical section, which wastes time and needlessly increases the size
of the state space. Fix this by only doing the cleanup if the current
task is actually on the list of tasks blocking some grace period.
While we are at it, consolidate the two identical exit_rcu() functions
into a single function.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Tested-by: Linus Torvalds <torvalds@linux-foundation.org>
Conflicts:
kernel/rcupdate.c
Currently, PREEMPT_RCU readers are enqueued upon entry to the scheduler.
This is inefficient because enqueuing is required only if there is a
context switch, and entry to the scheduler does not guarantee a context
switch.
The commit therefore moves the enqueuing to immediately precede the
call to switch_to() from the scheduler.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Tested-by: Linus Torvalds <torvalds@linux-foundation.org>
Timers are subject to migration, which can lead to the following
system-hang scenario when CONFIG_RCU_FAST_NO_HZ=y:
1. CPU 0 executes synchronize_rcu(), which posts an RCU callback.
2. CPU 0 then goes idle. It cannot immediately invoke the callback,
but there is nothing RCU needs from ti, so it enters dyntick-idle
mode after posting a timer.
3. The timer gets migrated to CPU 1.
4. CPU 0 never wakes up, so the synchronize_rcu() never returns, so
the system hangs.
This commit fixes this problem by using mod_timer_pinned(), as suggested
by Peter Zijlstra, to ensure that the timer is actually posted on the
running CPU.
Reported-by: Dipankar Sarma <dipankar@in.ibm.com>
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
RCU_FAST_NO_HZ uses a timer to limit the time that a CPU with callbacks
can remain in dyntick-idle mode. This timer is cancelled when the CPU
exits idle, and therefore should never fire. However, if the timer
were migrated to some other CPU for whatever reason (1) the timer could
actually fire and (2) firing on some other CPU would fail to wake up the
CPU with callbacks, possibly resulting in sluggishness or a system hang.
This commit therfore adds a WARN_ON_ONCE() to the timer handler in order
to detect this condition.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Both Steven Rostedt's new idle-capable trace macros and the RCU_NONIDLE()
macro can cause RCU to momentarily pause out of idle without the rest
of the system being involved. This can cause rcu_prepare_for_idle()
to run through its state machine too quickly, which can in turn result
in needless scheduling-clock interrupts.
This commit therefore adds code to enable rcu_prepare_for_idle() to
distinguish between an initial entry to idle on the one hand (which needs
to advance the rcu_prepare_for_idle() state machine) and an idle reentry
due to idle-capable trace macros and RCU_NONIDLE() on the other hand
(which should avoid advancing the rcu_prepare_for_idle() state machine).
Additional state is maintained to allow the timer to be correctly reposted
when returning after a momentary pause out of idle, and even more state
is maintained to detect when new non-lazy callbacks have been enqueued
(which may require re-evaluation of the approach to idleness).
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
The RCU_FAST_NO_HZ facility uses an hrtimer to wake up a CPU when
it is allowed to go into dyntick-idle mode, which is almost always
cancelled soon after. This is not what hrtimers are good at, so
this commit switches to the timer wheel.
Reported-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Traces of rcu_prep_idle events can be confusing because
rcu_cleanup_after_idle() does no tracing. This commit therefore adds
this tracing.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
This commit handles workloads that transition quickly between idle and
non-idle, and where the CPU's callbacks cannot be invoked, but where
RCU does not have anything immediate for the CPU to do. Without this
patch, the RCU_FAST_NO_HZ code can be invoked repeatedly on each entry
to idle. The commit sets the per-CPU rcu_dyntick_holdoff variable to
hold off further attempts for a tick.
Reported-by: "Abou Gazala, Neven M" <neven.m.abou.gazala@intel.com>
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
If a softirq is pending, the current CPU has RCU callbacks pending,
and RCU does not immediately need anything from this CPU, then the
current code resets the RCU_FAST_NO_HZ state machine. This means that
upon exit from the subsequent softirq handler, RCU_FAST_NO_HZ will
try really hard to force RCU into dyntick-idle mode. And if the same
conditions hold after a few tries (determined by RCU_IDLE_OPT_FLUSHES),
the same situation can repeat, possibly endlessly. This scenario is
not particularly good for battery lifetime.
This commit therefore suppresses the early exit from the RCU_FAST_NO_HZ
state machine in the case where there is a softirq pending. This change
forces the state machine to retain its memory, and to enter holdoff if
this condition persists.
Reported-by: "Abou Gazala, Neven M" <neven.m.abou.gazala@intel.com>
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
The expedited RCU primitives can be quite useful, but they have some
high costs as well. This commit updates and creates docbook comments
calling out the costs, and updates the RCU documentation as well.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Because newly offlined CPUs continue executing after completing the
CPU_DYING notifiers, they legitimately enter the scheduler and use
RCU while appearing to be offline. This calls for a more sophisticated
approach as follows:
1. RCU marks the CPU online during the CPU_UP_PREPARE phase.
2. RCU marks the CPU offline during the CPU_DEAD phase.
3. Diagnostics regarding use of read-side RCU by offline CPUs use
RCU's accounting rather than the cpu_online_map. (Note that
__call_rcu() still uses cpu_online_map to detect illegal
invocations within CPU_DYING notifiers.)
4. Offline CPUs are prevented from hanging the system by
force_quiescent_state(), which pays attention to cpu_online_map.
Some additional work (in a later commit) will be needed to
guarantee that force_quiescent_state() waits a full jiffy before
assuming that a CPU is offline, for example, when called from
idle entry. (This commit also makes the one-jiffy wait
explicit, since the old-style implicit wait can now be defeated
by RCU_FAST_NO_HZ and by rcutorture.)
This approach avoids the false positives encountered when attempting to
use more exact classification of CPU online/offline state.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
The rcu_prepare_for_idle() function is always called with interrupts
disabled, so there is no reason to disable interrupts again within
rcu_prepare_for_idle(). Therefore, this commit removes all of the
interrupt disabling, also removing a latent disabling-unbalance bug.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Now that TREE_RCU and TREE_PREEMPT_RCU no longer do anything different
for the single-CPU case, there is no need for multiple definitions of
synchronize_sched_expedited(). It is no longer in any sense a plug-in,
so move it from kernel/rcutree_plugin.h to kernel/rcutree.c.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Although it is legal to use RCU during early boot, it is anything
but legal to use RCU at runtime from an offlined CPU. After all, RCU
explicitly ignores offlined CPUs. This commit therefore adds checks
for runtime use of RCU from offlined CPUs.
These checks are not perfect, in particular, they can be subverted
through use of things like rcu_dereference_raw(). Note that it is not
possible to put checks in rcu_read_lock() and friends due to the fact
that these primitives are used in code that might be used under either
RCU or lock-based protection, which means that checking rcu_read_lock()
gets you fat piles of false positives.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
There have been situations where RCU CPU stall warnings were caused by
issues in scheduling-clock timer initialization. To make it easier to
track these down, this commit causes the RCU CPU stall-warning messages
to print out the number of scheduling-clock interrupts taken in the
current grace period for each stalled CPU.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Now that both TINY_RCU and TINY_PREEMPT_RCU have been in place for awhile,
it is time to remove UP support from TREE_RCU, which is what this commit
does.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
The recent updates to RCU_CPU_FAST_NO_HZ have an rcu_needs_cpu() that
does more than just check for callbacks, so get the name for
rcu_preempt_needs_cpu() consistent with that change, now calling it
rcu_preempt_cpu_has_callbacks().
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Currently, a given CPU is permitted to remain in dyntick-idle mode
indefinitely if it has only lazy RCU callbacks queued. This is vulnerable
to corner cases in NUMA systems, so limit the time to six seconds by
default. (Currently controlled by a cpp macro.)
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Move ->qsmaskinit and blkd_tasks[] manipulation to the CPU_DYING
notifier. This simplifies the code by eliminating a potential
deadlock and by reducing the responsibilities of force_quiescent_state().
Also rename functions to make their connection to the CPU-hotplug
stages explicit.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
When CONFIG_RCU_FAST_NO_HZ is enabled, RCU will allow a given CPU to
enter dyntick-idle mode even if it still has RCU callbacks queued.
RCU avoids system hangs in this case by scheduling a timer for several
jiffies in the future. However, if all of the callbacks on that CPU
are from kfree_rcu(), there is no reason to wake the CPU up, as it is
not a problem to defer freeing of memory.
This commit therefore tracks the number of callbacks on a given CPU
that are from kfree_rcu(), and avoids scheduling the timer if all of
a given CPU's callbacks are from kfree_rcu().
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
It is illegal to have a grace period within a same-flavor RCU read-side
critical section, so this commit adds lockdep-RCU checks to splat when
such abuse is encountered. This commit does not detect more elaborate
RCU deadlock situations. These situations might be a job for lockdep
enhancements.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Both TINY_RCU's and TREE_RCU's implementations of rcu_boost() access
the ->boost_tasks and ->exp_tasks fields without preventing concurrent
changes to these fields. This commit therefore applies ACCESS_ONCE in
order to prevent compiler mischief.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
This reverts commit 5342e269b2.
The approach taken in this patch was deemed too abusive to mutexes,
and thus too likely to result in maintenance problems in the future.
Instead, we will disallow RCU read-side critical sections that partially
overlap with interrupt-disbled code segments.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
If there are other CPUs active at a given point in time, then there is a
limit to what a given CPU can do to advance the current RCU grace period.
Beyond this limit, attempting to force the RCU grace period forward will
do nothing but consume energy burning CPU cycles.
Therefore, this commit takes an adaptive approach to RCU_FAST_NO_HZ
preparations for idle. It pushes the RCU core state machine for
two cycles unconditionally, and then it will push from zero to three
additional cycles, but only as long as the RCU core has work for this
CPU to do immediately. The rcu_pending() function is used to check
whether the RCU core has such work.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
The rcu_do_batch() function that invokes callbacks for TREE_RCU and
TREE_PREEMPT_RCU normally throttles callback invocation to avoid degrading
scheduling latency. However, as long as the CPU would otherwise be idle,
there is no downside to continuing to invoke any callbacks that have passed
through their grace periods. In fact, processing such callbacks in a
timely manner has the benefit of increasing the probability that the
CPU can enter the power-saving dyntick-idle mode.
Therefore, this commit allows callback invocation to continue beyond the
preset limit as long as the scheduler does not have some other task to
run and as long as context is that of the idle task or the relevant
RCU kthread.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
The current implementation of RCU_FAST_NO_HZ prevents CPUs from entering
dyntick-idle state if they have RCU callbacks pending. Unfortunately,
this has the side-effect of often preventing them from entering this
state, especially if at least one other CPU is not in dyntick-idle state.
However, the resulting per-tick wakeup is wasteful in many cases: if the
CPU has already fully responded to the current RCU grace period, there
will be nothing for it to do until this grace period ends, which will
frequently take several jiffies.
This commit therefore permits a CPU that has done everything that the
current grace period has asked of it (rcu_pending() == 0) even if it
still as RCU callbacks pending. However, such a CPU posts a timer to
wake it up several jiffies later (6 jiffies, based on experience with
grace-period lengths). This wakeup is required to handle situations
that can result in all CPUs being in dyntick-idle mode, thus failing
to ever complete the current grace period. If a CPU wakes up before
the timer goes off, then it cancels that timer, thus avoiding spurious
wakeups.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Re-enable interrupts across calls to quiescent-state functions and
also across force_quiescent_state() to reduce latency.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
With the new implementation of RCU_FAST_NO_HZ, it was possible to hang
RCU grace periods as follows:
o CPU 0 attempts to go idle, cycles several times through the
rcu_prepare_for_idle() loop, then goes dyntick-idle when
RCU needs nothing more from it, while still having at least
on RCU callback pending.
o CPU 1 goes idle with no callbacks.
Both CPUs can then stay in dyntick-idle mode indefinitely, preventing
the RCU grace period from ever completing, possibly hanging the system.
This commit therefore prevents CPUs that have RCU callbacks from entering
dyntick-idle mode. This approach also eliminates the need for the
end-of-grace-period IPIs used previously.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
If a CPU enters dyntick-idle mode with callbacks pending, it will need
an IPI at the end of the grace period. However, if it exits dyntick-idle
mode before the grace period ends, it will be needlessly IPIed at the
end of the grace period.
Therefore, this commit clears the per-CPU rcu_awake_at_gp_end flag
when a CPU determines that it does not need it. This in turn requires
disabling interrupts across much of rcu_prepare_for_idle() in order to
avoid having nested interrupts clearing this state out from under us.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
The earlier version would attempt to push callbacks through five times
before going into dyntick-idle mode if callbacks remained, but the CPU
had done all that it needed to do for the current RCU grace periods.
This is wasteful: In most cases, once the CPU has done all that it
needs to for the current RCU grace periods, it will make no further
progress on the callbacks no matter how many times it loops through
the RCU core processing and the idle-entry code.
This commit therefore goes to dyntick-idle mode whenever the current
CPU has done all it can for the current grace period.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
This commit adds trace_rcu_prep_idle(), which is invoked from
rcu_prepare_for_idle() and rcu_wake_cpu() to trace attempts on
the part of RCU to force CPUs into dyntick-idle mode.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Currently, RCU does not permit a CPU to enter dyntick-idle mode if that
CPU has any RCU callbacks queued. This means that workloads for which
each CPU wakes up and does some RCU updates every few ticks will never
enter dyntick-idle mode. This can result in significant unnecessary power
consumption, so this patch permits a given to enter dyntick-idle mode if
it has callbacks, but only if that same CPU has completed all current
work for the RCU core. We determine use rcu_pending() to determine
whether a given CPU has completed all current work for the RCU core.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Empty void functions do not need "return", so this commit removes it
from rcu_report_exp_rnp().
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
When setting up an expedited grace period, if there were no readers, the
task will awaken itself. This commit removes this useless self-awakening.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
When synchronize_sched_expedited() takes its second and subsequent
snapshots of sync_sched_expedited_started, it subtracts 1. This
means that the concurrent caller of synchronize_sched_expedited()
that incremented to that value sees our successful completion, it
will not be able to take advantage of it. This restriction is
pointless, given that our full expedited grace period would have
happened after the other guy started, and thus should be able to
serve as a proxy for the other guy successfully executing
try_stop_cpus().
This commit therefore removes the subtraction of 1.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Reviewed-by: Josh Triplett <josh@joshtriplett.org>
Because rcu_read_unlock_special() samples rcu_preempted_readers_exp(rnp)
after dropping rnp->lock, the following sequence of events is possible:
1. Task A exits its RCU read-side critical section, and removes
itself from the ->blkd_tasks list, releases rnp->lock, and is
then preempted. Task B remains on the ->blkd_tasks list, and
blocks the current expedited grace period.
2. Task B exits from its RCU read-side critical section and removes
itself from the ->blkd_tasks list. Because it is the last task
blocking the current expedited grace period, it ends that
expedited grace period.
3. Task A resumes, and samples rcu_preempted_readers_exp(rnp) which
of course indicates that nothing is blocking the nonexistent
expedited grace period. Task A is again preempted.
4. Some other CPU starts an expedited grace period. There are several
tasks blocking this expedited grace period queued on the
same rcu_node structure that Task A was using in step 1 above.
5. Task A examines its state and incorrectly concludes that it was
the last task blocking the expedited grace period on the current
rcu_node structure. It therefore reports completion up the
rcu_node tree.
6. The expedited grace period can then incorrectly complete before
the tasks blocked on this same rcu_node structure exit their
RCU read-side critical sections. Arbitrarily bad things happen.
This commit therefore takes a snapshot of rcu_preempted_readers_exp(rnp)
prior to dropping the lock, so that only the last task thinks that it is
the last task, thus avoiding the failure scenario laid out above.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Reviewed-by: Josh Triplett <josh@joshtriplett.org>
The purpose of rcu_needs_cpu_flush() was to iterate on pushing the
current grace period in order to help the current CPU enter dyntick-idle
mode. However, this can result in failures if the CPU starts entering
dyntick-idle mode, but then backs out. In this case, the call to
rcu_pending() from rcu_needs_cpu_flush() might end up announcing a
non-existing quiescent state.
This commit therefore removes rcu_needs_cpu_flush() in favor of letting
the dyntick-idle machinery at the end of the softirq handler push the
loop along via its call to rcu_pending().
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
RCU boost threads start life at RCU_BOOST_PRIO, while others remain
at RCU_KTHREAD_PRIO. While here, change thread names to match other
kthreads, and adjust rcu_yield() to not override the priority set by
the user. This last change sets the stage for runtime changes to
priority in the -rt tree.
Signed-off-by: Mike Galbraith <efault@gmx.de>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Create a separate lockdep class for the rt_mutex used for RCU priority
boosting and enable use of rt_mutex_lock() with irqs disabled. This
prevents RCU priority boosting from falling prey to deadlocks when
someone begins an RCU read-side critical section in preemptible state,
but releases it with an irq-disabled lock held.
Unfortunately, the scheduler's runqueue and priority-inheritance locks
still must either completely enclose or be completely enclosed by any
overlapping RCU read-side critical section.
This version removes a redundant local_irq_restore() noted by
Yong Zhang.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
It is possible for an RCU CPU stall to end just as it is detected, in
which case the current code will uselessly dump all CPU's stacks.
This commit therefore checks for this condition and refrains from
sending needless NMIs.
And yes, the stall might also end just after we checked all CPUs and
tasks, but in that case we would at least have given some clue as
to which CPU/task was at fault.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Commit 7765be (Fix RCU_BOOST race handling current->rcu_read_unlock_special)
introduced a new ->rcu_boosted field in the task structure. This is
redundant because the existing ->rcu_boost_mutex will be non-NULL at
any time that ->rcu_boosted is nonzero. Therefore, this commit removes
->rcu_boosted and tests ->rcu_boost_mutex instead.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
We only need to constrain the compiler if we are actually exiting
the top-level RCU read-side critical section. This commit therefore
moves the first barrier() cal in __rcu_read_unlock() to inside the
"if" statement, thus avoiding needless register flushes for inner
rcu_read_unlock() calls.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
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>
Add trace events to record grace-period start and end, quiescent states,
CPUs noticing grace-period start and end, grace-period initialization,
call_rcu() invocation, tasks blocking in RCU read-side critical sections,
tasks exiting those same critical sections, force_quiescent_state()
detection of dyntick-idle and offline CPUs, CPUs entering and leaving
dyntick-idle mode (except from NMIs), CPUs coming online and going
offline, and CPUs being kicked for staying in dyntick-idle mode for too
long (as in many weeks, even on 32-bit systems).
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
rcu: Add the rcu flavor to callback trace events
The earlier trace events for registering RCU callbacks and for invoking
them did not include the RCU flavor (rcu_bh, rcu_preempt, or rcu_sched).
This commit adds the RCU flavor to those trace events.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Add event-trace markers to TREE_RCU kthreads to allow including these
kthread's CPU time in the utilization calculations.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
We now have kthreads only for flavors of RCU that support boosting,
so update the now-misleading comments accordingly.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
In order to allow event tracing to distinguish between flavors of
RCU, we need those names in the relevant RCU data structures. TINY_RCU
has avoided them for memory-footprint reasons, so add them only if
CONFIG_RCU_TRACE=y.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Pull the code that waits for an RCU grace period into a single function,
which is then called by synchronize_rcu() and friends in the case of
TREE_RCU and TREE_PREEMPT_RCU, and from rcu_barrier() and friends in
the case of TINY_RCU and TINY_PREEMPT_RCU.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
There are a number of cases where the RCU can find additional work
for the per-CPU kthread within the context of that per-CPU kthread.
In such cases, the per-CPU kthread is already running, so attempting
to wake itself up does nothing except waste CPU cycles. This commit
therefore checks to see if it is in the per-CPU kthread context,
omitting the wakeup in this case.
Signed-off-by: Shaohua Li <shaohua.li@intel.com>
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Commit a26ac2455f (move TREE_RCU from softirq to kthread) added
per-CPU kthreads. However, kthread creation uses kthread_create(), which
can put the kthread's stack and task struct on the wrong NUMA node.
Therefore, use kthread_create_on_node() instead of kthread_create()
so that the stacks and task structs are placed on the correct NUMA node.
A similar change was carried out in commit 94dcf29a11 (kthread:
use kthread_create_on_node()).
Also change rcutorture's priority-boost-test kthread creation.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
CC: Tejun Heo <tj@kernel.org>
CC: Rusty Russell <rusty@rustcorp.com.au>
CC: Andrew Morton <akpm@linux-foundation.org>
CC: Andi Kleen <ak@linux.intel.com>
CC: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
The rcu_read_unlock_special() function relies on in_irq() to exclude
scheduler activity from interrupt level. This fails because exit_irq()
can invoke the scheduler after clearing the preempt_count() bits that
in_irq() uses to determine that it is at interrupt level. This situation
can result in failures as follows:
$task IRQ SoftIRQ
rcu_read_lock()
/* do stuff */
<preempt> |= UNLOCK_BLOCKED
rcu_read_unlock()
--t->rcu_read_lock_nesting
irq_enter();
/* do stuff, don't use RCU */
irq_exit();
sub_preempt_count(IRQ_EXIT_OFFSET);
invoke_softirq()
ttwu();
spin_lock_irq(&pi->lock)
rcu_read_lock();
/* do stuff */
rcu_read_unlock();
rcu_read_unlock_special()
rcu_report_exp_rnp()
ttwu()
spin_lock_irq(&pi->lock) /* deadlock */
rcu_read_unlock_special(t);
Ed can simply trigger this 'easy' because invoke_softirq() immediately
does a ttwu() of ksoftirqd/# instead of doing the in-place softirq stuff
first, but even without that the above happens.
Cure this by also excluding softirqs from the
rcu_read_unlock_special() handler and ensuring the force_irqthreads
ksoftirqd/# wakeup is done from full softirq context.
[ Alternatively, delaying the ->rcu_read_lock_nesting decrement
until after the special handling would make the thing more robust
in the face of interrupts as well. And there is a separate patch
for that. ]
Cc: Thomas Gleixner <tglx@linutronix.de>
Reported-and-tested-by: Ed Tomlinson <edt@aei.ca>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
The addition of RCU read-side critical sections within runqueue and
priority-inheritance lock critical sections introduced some deadlock
cycles, for example, involving interrupts from __rcu_read_unlock()
where the interrupt handlers call wake_up(). This situation can cause
the instance of __rcu_read_unlock() invoked from interrupt to do some
of the processing that would otherwise have been carried out by the
task-level instance of __rcu_read_unlock(). When the interrupt-level
instance of __rcu_read_unlock() is called with a scheduler lock held
from interrupt-entry/exit situations where in_irq() returns false,
deadlock can result.
This commit resolves these deadlocks by using negative values of
the per-task ->rcu_read_lock_nesting counter to indicate that an
instance of __rcu_read_unlock() is in flight, which in turn prevents
instances from interrupt handlers from doing any special processing.
This patch is inspired by Steven Rostedt's earlier patch that similarly
made __rcu_read_unlock() guard against interrupt-mediated recursion
(see https://lkml.org/lkml/2011/7/15/326), but this commit refines
Steven's approach to avoid the need for preemption disabling on the
__rcu_read_unlock() fastpath and to also avoid the need for manipulating
a separate per-CPU variable.
This patch avoids need for preempt_disable() by instead using negative
values of the per-task ->rcu_read_lock_nesting counter. Note that nested
rcu_read_lock()/rcu_read_unlock() pairs are still permitted, but they will
never see ->rcu_read_lock_nesting go to zero, and will therefore never
invoke rcu_read_unlock_special(), thus preventing them from seeing the
RCU_READ_UNLOCK_BLOCKED bit should it be set in ->rcu_read_unlock_special.
This patch also adds a check for ->rcu_read_unlock_special being negative
in rcu_check_callbacks(), thus preventing the RCU_READ_UNLOCK_NEED_QS
bit from being set should a scheduling-clock interrupt occur while
__rcu_read_unlock() is exiting from an outermost RCU read-side critical
section.
Of course, __rcu_read_unlock() can be preempted during the time that
->rcu_read_lock_nesting is negative. This could result in the setting
of the RCU_READ_UNLOCK_BLOCKED bit after __rcu_read_unlock() checks it,
and would also result it this task being queued on the corresponding
rcu_node structure's blkd_tasks list. Therefore, some later RCU read-side
critical section would enter rcu_read_unlock_special() to clean up --
which could result in deadlock if that critical section happened to be in
the scheduler where the runqueue or priority-inheritance locks were held.
This situation is dealt with by making rcu_preempt_note_context_switch()
check for negative ->rcu_read_lock_nesting, thus refraining from
queuing the task (and from setting RCU_READ_UNLOCK_BLOCKED) if we are
already exiting from the outermost RCU read-side critical section (in
other words, we really are no longer actually in that RCU read-side
critical section). In addition, rcu_preempt_note_context_switch()
invokes rcu_read_unlock_special() to carry out the cleanup in this case,
which clears out the ->rcu_read_unlock_special bits and dequeues the task
(if necessary), in turn avoiding needless delay of the current RCU grace
period and needless RCU priority boosting.
It is still illegal to call rcu_read_unlock() while holding a scheduler
lock if the prior RCU read-side critical section has ever had either
preemption or irqs enabled. However, the common use case is legal,
namely where then entire RCU read-side critical section executes with
irqs disabled, for example, when the scheduler lock is held across the
entire lifetime of the RCU read-side critical section.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Given some common flag combinations, particularly -Os, gcc will inline
rcu_read_unlock_special() despite its being in an unlikely() clause.
Use noinline to prohibit this misoptimization.
In addition, move the second barrier() in __rcu_read_unlock() so that
it is not on the common-case code path. This will allow the compiler to
generate better code for the common-case path through __rcu_read_unlock().
Suggested-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Acked-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
The RCU_BOOST commits for TREE_PREEMPT_RCU introduced an other-task
write to a new RCU_READ_UNLOCK_BOOSTED bit in the task_struct structure's
->rcu_read_unlock_special field, but, as noted by Steven Rostedt, without
correctly synchronizing all accesses to ->rcu_read_unlock_special.
This could result in bits in ->rcu_read_unlock_special being spuriously
set and cleared due to conflicting accesses, which in turn could result
in deadlocks between the rcu_node structure's ->lock and the scheduler's
rq and pi locks. These deadlocks would result from RCU incorrectly
believing that the just-ended RCU read-side critical section had been
preempted and/or boosted. If that RCU read-side critical section was
executed with either rq or pi locks held, RCU's ensuing (incorrect)
calls to the scheduler would cause the scheduler to attempt to once
again acquire the rq and pi locks, resulting in deadlock. More complex
deadlock cycles are also possible, involving multiple rq and pi locks
as well as locks from multiple rcu_node structures.
This commit fixes synchronization by creating ->rcu_boosted field in
task_struct that is accessed and modified only when holding the ->lock
in the rcu_node structure on which the task is queued (on that rcu_node
structure's ->blkd_tasks list). This results in tasks accessing only
their own current->rcu_read_unlock_special fields, making unsynchronized
access once again legal, and keeping the rcu_read_unlock() fastpath free
of atomic instructions and memory barriers.
The reason that the rcu_read_unlock() fastpath does not need to access
the new current->rcu_boosted field is that this new field cannot
be non-zero unless the RCU_READ_UNLOCK_BLOCKED bit is set in the
current->rcu_read_unlock_special field. Therefore, rcu_read_unlock()
need only test current->rcu_read_unlock_special: if that is zero, then
current->rcu_boosted must also be zero.
This bug does not affect TINY_PREEMPT_RCU because this implementation
of RCU accesses current->rcu_read_unlock_special with irqs disabled,
thus preventing races on the !SMP systems that TINY_PREEMPT_RCU runs on.
Maybe-reported-by: Dave Jones <davej@redhat.com>
Maybe-reported-by: Sergey Senozhatsky <sergey.senozhatsky@gmail.com>
Reported-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Reviewed-by: Steven Rostedt <rostedt@goodmis.org>
PREEMPT_RCU read-side critical sections blocking an expedited grace
period invoke rcu_report_exp_rnp(). When the last such critical section
has completed, rcu_report_exp_rnp() invokes the scheduler to wake up the
task that invoked synchronize_rcu_expedited() -- needlessly holding the
root rcu_node structure's lock while doing so, thus needlessly providing
a way for RCU and the scheduler to deadlock.
This commit therefore releases the root rcu_node structure's lock before
calling wake_up().
Reported-by: Ed Tomlinson <edt@aei.ca>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Under some rare but real combinations of configuration parameters, RCU
callbacks are posted during early boot that use kernel facilities that
are not yet initialized. Therefore, when these callbacks are invoked,
hard hangs and crashes ensue. This commit therefore prevents RCU
callbacks from being invoked until after the scheduler is fully up and
running, as in after multiple tasks have been spawned.
It might well turn out that a better approach is to identify the specific
RCU callbacks that are causing this problem, but that discussion will
wait until such time as someone really needs an RCU callback to be invoked
(as opposed to merely registered) during early boot.
Reported-by: julie Sullivan <kernelmail.jms@gmail.com>
Reported-by: RKK <kulkarni.ravi4@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Tested-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Tested-by: julie Sullivan <kernelmail.jms@gmail.com>
Tested-by: RKK <kulkarni.ravi4@gmail.com>
The commit "use softirq instead of kthreads except when RCU_BOOST=y"
just applied #ifdef in place. This commit is a cleanup that moves
the newly #ifdef'ed code to the header file kernel/rcutree_plugin.h.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
This patch #ifdefs RCU kthreads out of the kernel unless RCU_BOOST=y,
thus eliminating context-switch overhead if RCU priority boosting has
not been configured.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
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>
Make the functions creating the kthreads wake them up. Leverage the
fact that the per-node and boost kthreads can run anywhere, thus
dispensing with the need to wake them up once the incoming CPU has
gone fully online.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Tested-by: Daniel J Blueman <daniel.blueman@gmail.com>
Commit cc3ce5176d (rcu: Start RCU kthreads in TASK_INTERRUPTIBLE
state) fudges a sleeping task' state, resulting in the scheduler seeing
a TASK_UNINTERRUPTIBLE task going to sleep, but a TASK_INTERRUPTIBLE
task waking up. The result is unbalanced load calculation.
The problem that patch tried to address is that the RCU threads could
stay in UNINTERRUPTIBLE state for quite a while and triggering the hung
task detector due to on-demand wake-ups.
Cure the problem differently by always giving the tasks at least one
wake-up once the CPU is fully up and running, this will kick them out of
the initial UNINTERRUPTIBLE state and into the regular INTERRUPTIBLE
wait state.
[ The alternative would be teaching kthread_create() to start threads as
INTERRUPTIBLE but that needs a tad more thought. ]
Reported-by: Damien Wyart <damien.wyart@free.fr>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Paul E. McKenney <paul.mckenney@linaro.org>
Link: http://lkml.kernel.org/r/1306755291.1200.2872.camel@twins
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Upon creation, kthreads are in TASK_UNINTERRUPTIBLE state, which can
result in softlockup warnings. Because some of RCU's kthreads can
legitimately be idle indefinitely, start them in TASK_INTERRUPTIBLE
state in order to avoid those warnings.
Suggested-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Tested-by: Yinghai Lu <yinghai@kernel.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
It is not necessary to use waitqueues for the RCU kthreads because
we always know exactly which thread is to be awakened. In addition,
wake_up() only issues an actual wakeup when there is a thread waiting on
the queue, which was why there was an extra explicit wake_up_process()
to get the RCU kthreads started.
Eliminating the waitqueues (and wake_up()) in favor of wake_up_process()
eliminates the need for the initial wake_up_process() and also shrinks
the data structure size a bit. The wakeup logic is placed in a new
rcu_wait() macro.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
(Note: this was reverted, and is now being re-applied in pieces, with
this being the fifth and final piece. See below for the reason that
it is now felt to be safe to re-apply this.)
Commit d09b62d fixed grace-period synchronization, but left some smp_mb()
invocations in rcu_process_callbacks() that are no longer needed, but
sheer paranoia prevented them from being removed. This commit removes
them and provides a proof of correctness in their absence. It also adds
a memory barrier to rcu_report_qs_rsp() immediately before the update to
rsp->completed in order to handle the theoretical possibility that the
compiler or CPU might move massive quantities of code into a lock-based
critical section. This also proves that the sheer paranoia was not
entirely unjustified, at least from a theoretical point of view.
In addition, the old dyntick-idle synchronization depended on the fact
that grace periods were many milliseconds in duration, so that it could
be assumed that no dyntick-idle CPU could reorder a memory reference
across an entire grace period. Unfortunately for this design, the
addition of expedited grace periods breaks this assumption, which has
the unfortunate side-effect of requiring atomic operations in the
functions that track dyntick-idle state for RCU. (There is some hope
that the algorithms used in user-level RCU might be applied here, but
some work is required to handle the NMIs that user-space applications
can happily ignore. For the short term, better safe than sorry.)
This proof assumes that neither compiler nor CPU will allow a lock
acquisition and release to be reordered, as doing so can result in
deadlock. The proof is as follows:
1. A given CPU declares a quiescent state under the protection of
its leaf rcu_node's lock.
2. If there is more than one level of rcu_node hierarchy, the
last CPU to declare a quiescent state will also acquire the
->lock of the next rcu_node up in the hierarchy, but only
after releasing the lower level's lock. The acquisition of this
lock clearly cannot occur prior to the acquisition of the leaf
node's lock.
3. Step 2 repeats until we reach the root rcu_node structure.
Please note again that only one lock is held at a time through
this process. The acquisition of the root rcu_node's ->lock
must occur after the release of that of the leaf rcu_node.
4. At this point, we set the ->completed field in the rcu_state
structure in rcu_report_qs_rsp(). However, if the rcu_node
hierarchy contains only one rcu_node, then in theory the code
preceding the quiescent state could leak into the critical
section. We therefore precede the update of ->completed with a
memory barrier. All CPUs will therefore agree that any updates
preceding any report of a quiescent state will have happened
before the update of ->completed.
5. Regardless of whether a new grace period is needed, rcu_start_gp()
will propagate the new value of ->completed to all of the leaf
rcu_node structures, under the protection of each rcu_node's ->lock.
If a new grace period is needed immediately, this propagation
will occur in the same critical section that ->completed was
set in, but courtesy of the memory barrier in #4 above, is still
seen to follow any pre-quiescent-state activity.
6. When a given CPU invokes __rcu_process_gp_end(), it becomes
aware of the end of the old grace period and therefore makes
any RCU callbacks that were waiting on that grace period eligible
for invocation.
If this CPU is the same one that detected the end of the grace
period, and if there is but a single rcu_node in the hierarchy,
we will still be in the single critical section. In this case,
the memory barrier in step #4 guarantees that all callbacks will
be seen to execute after each CPU's quiescent state.
On the other hand, if this is a different CPU, it will acquire
the leaf rcu_node's ->lock, and will again be serialized after
each CPU's quiescent state for the old grace period.
On the strength of this proof, this commit therefore removes the memory
barriers from rcu_process_callbacks() and adds one to rcu_report_qs_rsp().
The effect is to reduce the number of memory barriers by one and to
reduce the frequency of execution from about once per scheduling tick
per CPU to once per grace period.
This was reverted do to hangs found during testing by Yinghai Lu and
Ingo Molnar. Frederic Weisbecker supplied Yinghai with tracing that
located the underlying problem, and Frederic also provided the fix.
The underlying problem was that the HARDIRQ_ENTER() macro from
lib/locking-selftest.c invoked irq_enter(), which in turn invokes
rcu_irq_enter(), but HARDIRQ_EXIT() invoked __irq_exit(), which
does not invoke rcu_irq_exit(). This situation resulted in calls
to rcu_irq_enter() that were not balanced by the required calls to
rcu_irq_exit(). Therefore, after these locking selftests completed,
RCU's dyntick-idle nesting count was a large number (for example,
72), which caused RCU to to conclude that the affected CPU was not in
dyntick-idle mode when in fact it was.
RCU would therefore incorrectly wait for this dyntick-idle CPU, resulting
in hangs.
In contrast, with Frederic's patch, which replaces the irq_enter()
in HARDIRQ_ENTER() with an __irq_enter(), these tests don't ever call
either rcu_irq_enter() or rcu_irq_exit(), which works because the CPU
running the test is already marked as not being in dyntick-idle mode.
This means that the rcu_irq_enter() and rcu_irq_exit() calls and RCU
then has no problem working out which CPUs are in dyntick-idle mode and
which are not.
The reason that the imbalance was not noticed before the barrier patch
was applied is that the old implementation of rcu_enter_nohz() ignored
the nesting depth. This could still result in delays, but much shorter
ones. Whenever there was a delay, RCU would IPI the CPU with the
unbalanced nesting level, which would eventually result in rcu_enter_nohz()
being called, which in turn would force RCU to see that the CPU was in
dyntick-idle mode.
The reason that very few people noticed the problem is that the mismatched
irq_enter() vs. __irq_exit() occured only when the kernel was built with
CONFIG_DEBUG_LOCKING_API_SELFTESTS.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Reviewed-by: Josh Triplett <josh@joshtriplett.org>
This reverts commit e59fb3120b.
This reversion was due to (extreme) boot-time slowdowns on SPARC seen by
Yinghai Lu and on x86 by Ingo
.
This is a non-trivial reversion due to intervening commits.
Conflicts:
Documentation/RCU/trace.txt
kernel/rcutree.c
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Avoid calling into the scheduler while holding core RCU locks. This
allows rcu_read_unlock() to be called while holding the runqueue locks,
but only as long as there was no chance of the RCU read-side critical
section having been preempted. (Otherwise, if RCU priority boosting
is enabled, rcu_read_unlock() might call into the scheduler in order to
unboost itself, which might allows self-deadlock on the runqueue locks
within the scheduler.)
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
The "preemptible" spelling is preferable. May as well fix it.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Reviewed-by: Josh Triplett <josh@joshtriplett.org>
Using __rcu_read_lock() in place of rcu_read_lock() leaves any debug
state as it really should be, namely with the lock still held.
Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Reviewed-by: Josh Triplett <josh@joshtriplett.org>
The rcu_initiate_boost_trace() function mis-attributed refusals to
initiate RCU priority boosting that were in fact due to its not yet
being time to boost. This patch fixes the faulty comparison.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Add tracing to help debugging situations when RCU's kthreads are not
running but are supposed to be.
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>
Includes total number of tasks boosted, number boosted on behalf of each
of normal and expedited grace periods, and statistics on attempts to
initiate boosting that failed for various reasons.
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>
The scheduler has had some heartburn in the past when too many real-time
kthreads were affinitied to the outgoing CPU. So, this commit lightens
the load by forcing the per-rcu_node and the boost kthreads off of the
outgoing CPU. Note that RCU's per-CPU kthread remains on the outgoing
CPU until the bitter end, as it must in order to preserve correctness.
Also avoid disabling hardirqs across calls to set_cpus_allowed_ptr(),
given that this function can block.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Add priority boosting for TREE_PREEMPT_RCU, similar to that for
TINY_PREEMPT_RCU. This is enabled by the default-off RCU_BOOST
kernel parameter. The priority to which to boost preempted
RCU readers is controlled by the RCU_BOOST_PRIO kernel parameter
(defaulting to real-time priority 1) and the time to wait before
boosting the readers who are blocking a given grace period is
controlled by the RCU_BOOST_DELAY kernel parameter (defaulting to
500 milliseconds).
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>
If RCU priority boosting is to be meaningful, callback invocation must
be boosted in addition to preempted RCU readers. Otherwise, in presence
of CPU real-time threads, the grace period ends, but the callbacks don't
get invoked. If the callbacks don't get invoked, the associated memory
doesn't get freed, so the system is still subject to OOM.
But it is not reasonable to priority-boost RCU_SOFTIRQ, so this commit
moves the callback invocations to a kthread, which can be boosted easily.
Also add comments and properly synchronized all accesses to
rcu_cpu_kthread_task, as suggested by Lai Jiangshan.
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>
Combine the current TREE_PREEMPT_RCU ->blocked_tasks[] lists in the
rcu_node structure into a single ->blkd_tasks list with ->gp_tasks
and ->exp_tasks tail pointers. This is in preparation for RCU priority
boosting, which will add a third dimension to the combinatorial explosion
in the ->blocked_tasks[] case, but simply a third pointer in the new
->blkd_tasks case.
Also update documentation to reflect blocked_tasks[] merge
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>
Commit d09b62d fixed grace-period synchronization, but left some smp_mb()
invocations in rcu_process_callbacks() that are no longer needed, but
sheer paranoia prevented them from being removed. This commit removes
them and provides a proof of correctness in their absence. It also adds
a memory barrier to rcu_report_qs_rsp() immediately before the update to
rsp->completed in order to handle the theoretical possibility that the
compiler or CPU might move massive quantities of code into a lock-based
critical section. This also proves that the sheer paranoia was not
entirely unjustified, at least from a theoretical point of view.
In addition, the old dyntick-idle synchronization depended on the fact
that grace periods were many milliseconds in duration, so that it could
be assumed that no dyntick-idle CPU could reorder a memory reference
across an entire grace period. Unfortunately for this design, the
addition of expedited grace periods breaks this assumption, which has
the unfortunate side-effect of requiring atomic operations in the
functions that track dyntick-idle state for RCU. (There is some hope
that the algorithms used in user-level RCU might be applied here, but
some work is required to handle the NMIs that user-space applications
can happily ignore. For the short term, better safe than sorry.)
This proof assumes that neither compiler nor CPU will allow a lock
acquisition and release to be reordered, as doing so can result in
deadlock. The proof is as follows:
1. A given CPU declares a quiescent state under the protection of
its leaf rcu_node's lock.
2. If there is more than one level of rcu_node hierarchy, the
last CPU to declare a quiescent state will also acquire the
->lock of the next rcu_node up in the hierarchy, but only
after releasing the lower level's lock. The acquisition of this
lock clearly cannot occur prior to the acquisition of the leaf
node's lock.
3. Step 2 repeats until we reach the root rcu_node structure.
Please note again that only one lock is held at a time through
this process. The acquisition of the root rcu_node's ->lock
must occur after the release of that of the leaf rcu_node.
4. At this point, we set the ->completed field in the rcu_state
structure in rcu_report_qs_rsp(). However, if the rcu_node
hierarchy contains only one rcu_node, then in theory the code
preceding the quiescent state could leak into the critical
section. We therefore precede the update of ->completed with a
memory barrier. All CPUs will therefore agree that any updates
preceding any report of a quiescent state will have happened
before the update of ->completed.
5. Regardless of whether a new grace period is needed, rcu_start_gp()
will propagate the new value of ->completed to all of the leaf
rcu_node structures, under the protection of each rcu_node's ->lock.
If a new grace period is needed immediately, this propagation
will occur in the same critical section that ->completed was
set in, but courtesy of the memory barrier in #4 above, is still
seen to follow any pre-quiescent-state activity.
6. When a given CPU invokes __rcu_process_gp_end(), it becomes
aware of the end of the old grace period and therefore makes
any RCU callbacks that were waiting on that grace period eligible
for invocation.
If this CPU is the same one that detected the end of the grace
period, and if there is but a single rcu_node in the hierarchy,
we will still be in the single critical section. In this case,
the memory barrier in step #4 guarantees that all callbacks will
be seen to execute after each CPU's quiescent state.
On the other hand, if this is a different CPU, it will acquire
the leaf rcu_node's ->lock, and will again be serialized after
each CPU's quiescent state for the old grace period.
On the strength of this proof, this commit therefore removes the memory
barriers from rcu_process_callbacks() and adds one to rcu_report_qs_rsp().
The effect is to reduce the number of memory barriers by one and to
reduce the frequency of execution from about once per scheduling tick
per CPU to once per grace period.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Reviewed-by: Josh Triplett <josh@joshtriplett.org>
The RCU CPU stall warnings can now be controlled using the
rcu_cpu_stall_suppress boot-time parameter or via the same parameter
from sysfs. There is therefore no longer any reason to have
kernel config parameters for this feature. This commit therefore
removes the RCU_CPU_STALL_DETECTOR and RCU_CPU_STALL_DETECTOR_RUNNABLE
kernel config parameters. The RCU_CPU_STALL_TIMEOUT parameter remains
to allow the timeout to be tuned and the RCU_CPU_STALL_VERBOSE parameter
remains to allow task-stall information to be suppressed if desired.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Reviewed-by: Josh Triplett <josh@joshtriplett.org>
The fix in commit #6a0cc49 requires more than three concurrent instances
of synchronize_sched_expedited() before batching is possible. This
patch uses a ticket-counter-like approach that is also not unrelated to
Lai Jiangshan's Ring RCU to allow sharing of expedited grace periods even
when there are only two concurrent instances of synchronize_sched_expedited().
This commit builds on Tejun's original posting, which may be found at
http://lkml.org/lkml/2010/11/9/204, adding memory barriers, avoiding
overflow of signed integers (other than via atomic_t), and fixing the
detection of batching.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
The new (early 2010) implementation of synchronize_sched_expedited() uses
try_stop_cpu() to force a context switch on every CPU. It also permits
concurrent calls to synchronize_sched_expedited() to share a single call
to try_stop_cpu() through use of an atomically incremented
synchronize_sched_expedited_count variable. Unfortunately, this is
subject to failure as follows:
o Task A invokes synchronize_sched_expedited(), try_stop_cpus()
succeeds, but Task A is preempted before getting to the atomic
increment of synchronize_sched_expedited_count.
o Task B also invokes synchronize_sched_expedited(), with exactly
the same outcome as Task A.
o Task C also invokes synchronize_sched_expedited(), again with
exactly the same outcome as Tasks A and B.
o Task D also invokes synchronize_sched_expedited(), but only
gets as far as acquiring the mutex within try_stop_cpus()
before being preempted, interrupted, or otherwise delayed.
o Task E also invokes synchronize_sched_expedited(), but only
gets to the snapshotting of synchronize_sched_expedited_count.
o Tasks A, B, and C all increment synchronize_sched_expedited_count.
o Task E fails to get the mutex, so checks the new value
of synchronize_sched_expedited_count. It finds that the
value has increased, so (wrongly) assumes that its work
has been done, returning despite there having been no
expedited grace period since it began.
The solution is to have the lowest-numbered CPU atomically increment
the synchronize_sched_expedited_count variable within the
synchronize_sched_expedited_cpu_stop() function, which is under
the protection of the mutex acquired by try_stop_cpus(). However, this
also requires that piggybacking tasks wait for three rather than two
instances of try_stop_cpu(), because we cannot control the order in
which the per-CPU callback function occur.
Cc: Tejun Heo <tj@kernel.org>
Cc: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Lai's RCU-callback immediate-adoption patch changes the RCU tracing
output, so update tracing.txt. Also update a few comments to clarify
the synchronization design.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
When we handle the CPU_DYING notifier, the whole system is stopped except
for the current CPU. We therefore need no synchronization with the other
CPUs. This allows us to move any orphaned RCU callbacks directly to the
list of any online CPU without needing to run them through the global
orphan lists. These global orphan lists can therefore be dispensed with.
This commit makes thes changes, though currently victimizes CPU 0 @@@.
Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Paul E. McKenney