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>
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>
This commit adds the age in jiffies of the current grace period along
with the duration in jiffies of the longest grace period since boot
to the rcu/rcugp debugfs file. It also adds an additional "O" state
to kthread tracing to differentiate between the kthread waiting due to
having nothing to do on the one hand and waiting due to being on the
wrong CPU on the other hand.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
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>
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>
Some recent benchmarks have indicated possible lock contention on the
leaf-level rcu_node locks. This commit therefore limits the number of
CPUs per leaf-level rcu_node structure to 16, in other words, there
can be at most 16 rcu_data structures fanning into a given rcu_node
structure. Prior to this, the limit was 32 on 32-bit systems and 64 on
64-bit systems.
Note that the fanout of non-leaf rcu_node structures is unchanged. The
organization of accesses to the rcu_node tree is such that references
to non-leaf rcu_node structures are much less frequent than to the
leaf structures.
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>
The current tracing data is not sufficient to deduce the average time
that a callback spends waiting for a grace period to end. Add three
per-CPU counters recording the number of callbacks invoked (ci), the
number of callbacks orphaned (co), and the number of callbacks adopted
(ca). Given the existing callback queue length (ql), the average wait
time in absence of CPU hotplug operations is ql/ci. The units of wait
time will be in terms of the duration over which ci was measured.
In the presence of CPU hotplug operations, there is room for argument,
but ql/(ci-co+ca) won't steer you too far wrong.
Also fixes a typo called out by Lucas De Marchi <lucas.de.marchi@gmail.com>.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Combine the duplicate definitions of ULONG_CMP_GE(), ULONG_CMP_LT(),
and rcu_preempt_depth() into include/linux/rcupdate.h.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
When using a kernel debugger, a long sojourn in the debugger can get
you lots of RCU CPU stall warnings once you resume. This might not be
helpful, especially if you are using the system console. This patch
therefore allows RCU CPU stall warnings to be suppressed, but only for
the duration of the current set of grace periods.
This differs from Jason's original patch in that it adds support for
tiny RCU and preemptible RCU, and uses a slightly different method for
suppressing the RCU CPU stall warning messages.
Signed-off-by: Jason Wessel <jason.wessel@windriver.com>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Tested-by: Jason Wessel <jason.wessel@windriver.com>
Currently, if RCU CPU stall warnings are enabled, they are enabled
immediately upon boot. They can be manually disabled via /sys (and
also re-enabled via /sys), and are automatically disabled upon panic.
However, some users need RCU CPU stalls to be disabled at boot time,
but to be enabled without rebuilding/rebooting. For example, someone
running a real-time application in production might not want the
additional latency of RCU CPU stall detection in normal operation, but
might need to enable it at any point for fault isolation purposes.
This commit therefore provides a new CONFIG_RCU_CPU_STALL_DETECTOR_RUNNABLE
kernel configuration parameter that maintains the current behavior
(enable at boot) by default, but allows a kernel to be configured
with RCU CPU stall detection built into the kernel, but disabled at
boot time.
Requested-by: Clark Williams <williams@redhat.com>
Requested-by: John Kacur <jkacur@redhat.com>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Also set the default to 60 seconds, up from the previous hard-coded timeout
of 10 seconds. This allows people who care to set short timeouts, while
avoiding people with unusual configurations (make randconfig!!!) from being
bothered with spurious CPU stall warnings.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Reviewed-by: Josh Triplett <josh@joshtriplett.org>
&percpu_data is compatible with allocated percpu data.
And we use it and remove the "->rda[NR_CPUS]" array, saving significant
storage on systems with large numbers of CPUs. This does add an additional
level of indirection and thus an additional cache line referenced, but
because ->rda is not used on the read side, this is OK.
Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Reviewed-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Reviewed-by: Josh Triplett <josh@joshtriplett.org>
