The ->percpu_enqueue_shift field is used to map from the running CPU
number to the index of the corresponding callback list. This mapping
can change at runtime in response to varying callback load, resulting
in varying levels of contention on the callback-list locks.
Unfortunately, the initial value of this field is correct only if the
system happens to have a power-of-two number of CPUs, otherwise the
callbacks from the high-numbered CPUs can be placed into the callback list
indexed by 1 (rather than 0), and those index-1 callbacks will be ignored.
This can result in soft lockups and hangs.
This commit therefore corrects this mapping, adding one to this shift
count as needed for systems having odd numbers of CPUs.
Fixes: 7a30871b6a ("rcu-tasks: Introduce ->percpu_enqueue_shift for dynamic queue selection")
Reported-by: Andrii Nakryiko <andrii.nakryiko@gmail.com>
Cc: Reported-by: Martin Lau <kafai@fb.com>
Cc: Neeraj Upadhyay <neeraj.iitr10@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Merge misc updates from Andrew Morton:
"146 patches.
Subsystems affected by this patch series: kthread, ia64, scripts,
ntfs, squashfs, ocfs2, vfs, and mm (slab-generic, slab, kmemleak,
dax, kasan, debug, pagecache, gup, shmem, frontswap, memremap,
memcg, selftests, pagemap, dma, vmalloc, memory-failure, hugetlb,
userfaultfd, vmscan, mempolicy, oom-kill, hugetlbfs, migration, thp,
ksm, page-poison, percpu, rmap, zswap, zram, cleanups, hmm, and
damon)"
* emailed patches from Andrew Morton <akpm@linux-foundation.org>: (146 commits)
mm/damon: hide kernel pointer from tracepoint event
mm/damon/vaddr: hide kernel pointer from damon_va_three_regions() failure log
mm/damon/vaddr: use pr_debug() for damon_va_three_regions() failure logging
mm/damon/dbgfs: remove an unnecessary variable
mm/damon: move the implementation of damon_insert_region to damon.h
mm/damon: add access checking for hugetlb pages
Docs/admin-guide/mm/damon/usage: update for schemes statistics
mm/damon/dbgfs: support all DAMOS stats
Docs/admin-guide/mm/damon/reclaim: document statistics parameters
mm/damon/reclaim: provide reclamation statistics
mm/damon/schemes: account how many times quota limit has exceeded
mm/damon/schemes: account scheme actions that successfully applied
mm/damon: remove a mistakenly added comment for a future feature
Docs/admin-guide/mm/damon/usage: update for kdamond_pid and (mk|rm)_contexts
Docs/admin-guide/mm/damon/usage: mention tracepoint at the beginning
Docs/admin-guide/mm/damon/usage: remove redundant information
Docs/admin-guide/mm/damon/usage: update for scheme quotas and watermarks
mm/damon: convert macro functions to static inline functions
mm/damon: modify damon_rand() macro to static inline function
mm/damon: move damon_rand() definition into damon.h
...
The rcu_spawn_one_nocb_kthread() function is called only from
rcu_spawn_cpu_nocb_kthread(). Therefore, inline the former into
the latter, saving a few lines of code.
Reviewed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Cc: Boqun Feng <boqun.feng@gmail.com>
Cc: Uladzislau Rezki <urezki@gmail.com>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Joel Fernandes <joel@joelfernandes.org>
Tested-by: Juri Lelli <juri.lelli@redhat.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Allow the rcu_nocbs kernel parameter to be specified just by itself,
without specifying any CPUs. This allows systems administrators to use
"rcu_nocbs" to specify that none of the CPUs are to be offloaded at boot
time, but than any of them may be offloaded at runtime via cpusets.
In contrast, if the "rcu_nocbs" or "nohz_full" kernel parameters are not
specified at all, then not only are none of the CPUs offloaded at boot,
none of them can be offloaded at runtime, either.
While in the area, modernize the description of the "rcuo" kthreads'
naming scheme.
Reviewed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Cc: Boqun Feng <boqun.feng@gmail.com>
Cc: Uladzislau Rezki <urezki@gmail.com>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Joel Fernandes <joel@joelfernandes.org>
Tested-by: Juri Lelli <juri.lelli@redhat.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
In order to be able to (de-)offload any CPU using cpusets in the future,
create the NOCB data structures for all possible CPUs. For now this is
done only as long as the "rcu_nocbs=" or "nohz_full=" kernel parameters
are passed to avoid the unnecessary overhead for most users.
Note that the rcuog and rcuoc kthreads are not created until at least
one of the corresponding CPUs comes online. This approach avoids the
creation of excess kthreads when firmware lies about the number of CPUs
present on the system.
Reviewed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Cc: Boqun Feng <boqun.feng@gmail.com>
Cc: Uladzislau Rezki <urezki@gmail.com>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Joel Fernandes <joel@joelfernandes.org>
Tested-by: Juri Lelli <juri.lelli@redhat.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Currently cpumask_available() is used to prevent from unwanted NOCB
initialization. However if neither "rcu_nocbs=" nor "nohz_full="
parameters are passed to a kernel built with CONFIG_CPUMASK_OFFSTACK=n,
the initialization path is still taken, running through all sorts of
needless operations and iterations on an empty cpumask.
Fix this by relying on a real initialization state instead. This also
optimizes kthread creation, preventing needless iteration over all online
CPUs when the kernel is booted without any offloaded CPUs.
Reviewed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Cc: Boqun Feng <boqun.feng@gmail.com>
Cc: Uladzislau Rezki <urezki@gmail.com>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Joel Fernandes <joel@joelfernandes.org>
Tested-by: Juri Lelli <juri.lelli@redhat.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
In order to be able to toggle the offloaded state from cpusets, a nocb
kthread will need to be created for all possible CPUs whenever either
of the "rcu_nocbs=" or "nohz_full=" parameters are specified.
Therefore, the nocb_cb_wait() kthread must be prepared to start running
on a de-offloaded rdp. To accomplish this, simply move the sleeping
condition to the beginning of the nocb_cb_wait() function, which prevents
this kthread from attempting to invoke callbacks before the corresponding
CPU is offloaded.
Reviewed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Cc: Boqun Feng <boqun.feng@gmail.com>
Cc: Uladzislau Rezki <urezki@gmail.com>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Joel Fernandes <joel@joelfernandes.org>
Tested-by: Juri Lelli <juri.lelli@redhat.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
The nocb_gp_wait() function iterates over all CPUs in its group,
including even those CPUs that have been de-offloaded. This is of
course suboptimal, especially if none of the CPUs within the group are
currently offloaded. This will become even more of a problem once a
nocb kthread is created for all possible CPUs.
Therefore use a standard double linked list to link all the offloaded
rcu_data structures and safely add or delete these structure as we
offload or de-offload them, respectively.
Reviewed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Cc: Boqun Feng <boqun.feng@gmail.com>
Cc: Uladzislau Rezki <urezki@gmail.com>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Joel Fernandes <joel@joelfernandes.org>
Tested-by: Juri Lelli <juri.lelli@redhat.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
By default, when lock contention is encountered, the RCU Tasks flavors
of RCU switch to using per-CPU queueing. However, if the callback
flood ends, per-CPU queueing continues to be used, which introduces
significant additional overhead, especially for callback invocation,
which fans out a series of workqueue handlers.
This commit therefore switches back to single-queue operation if at the
beginning of a grace period there are very few callbacks. The definition
of "very few" is set by the rcupdate.rcu_task_collapse_lim module
parameter, which defaults to 10. This switch happens in two phases,
with the first phase causing future callbacks to be enqueued on CPU 0's
queue, but with all queues continuing to be checked for grace periods
and callback invocation. The second phase checks to see if an RCU grace
period has elapsed and if all remaining RCU-Tasks callbacks are queued
on CPU 0. If so, only CPU 0 is checked for future grace periods and
callback operation.
Of course, the return of contention anywhere during this process will
result in returning to per-CPU callback queueing.
Reported-by: Martin Lau <kafai@fb.com>
Cc: Neeraj Upadhyay <neeraj.iitr10@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Decreasing the number of callback queues is a bit tricky because it
is necessary to handle callbacks that were queued before the number of
queues decreased, but which were not ready to invoke until afterwards.
This commit takes a first step in this direction by maintaining a separate
->percpu_dequeue_lim to control callback dequeueing, in addition to the
existing ->percpu_enqueue_lim which now controls only enqueueing.
Reported-by: Martin Lau <kafai@fb.com>
Cc: Neeraj Upadhyay <neeraj.iitr10@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
The rcupdate.rcu_task_enqueue_lim module parameter allows system
administrators to tune the number of callback queues used by the RCU
Tasks flavors. However if callback storms are infrequent, it would
be better to operate with a single queue on a given system unless and
until that system actually needed more queues. Systems not needing
more queues can then avoid the overhead of checking the extra queues
and especially avoid the overhead of fanning workqueue handlers out to
all CPUs to invoke callbacks.
This commit therefore switches to using all the CPUs' callback queues if
call_rcu_tasks_generic() encounters too much lock contention. The amount
of lock contention to tolerate defaults to 100 contended lock acquisitions
per jiffy, and can be adjusted using the new rcupdate.rcu_task_contend_lim
module parameter.
Such switching is undertaken only if the rcupdate.rcu_task_enqueue_lim
module parameter is negative, which is its default value (-1).
This allows savvy systems administrators to set the number of queues
to some known good value and to not have to worry about the kernel doing
any second guessing.
[ paulmck: Apply feedback from Guillaume Tucker and kernelci. ]
Reported-by: Martin Lau <kafai@fb.com>
Cc: Neeraj Upadhyay <neeraj.iitr10@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
If the caller of of call_rcu_tasks(), call_rcu_tasks_rude(),
or call_rcu_tasks_trace() holds a raw spinlock, and then if
call_rcu_tasks_generic() determines that the grace-period kthread must
be awakened, then the wakeup might acquire a normal spinlock while a
raw spinlock is held. This results in lockdep splats when the
kernel is built with CONFIG_PROVE_RAW_LOCK_NESTING=y.
This commit therefore defers the wakeup using irq_work_queue().
It would be nice to directly invoke wakeup when a raw spinlock is not
held, but there is currently no way to check for this in all kernels.
Reported-by: Martin Lau <kafai@fb.com>
Cc: Neeraj Upadhyay <neeraj.iitr10@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
This commit converts the unconditional raw_spin_lock_rcu_node() lock
acquisition in call_rcu_tasks_generic() to a trylock followed by an
unconditional acquisition if the trylock fails. If the trylock fails,
the failure is counted, but the count is reset to zero on each new jiffy.
This statistic will be used to determine when to move from a single
callback queue to per-CPU callback queues.
Reported-by: Martin Lau <kafai@fb.com>
Cc: Neeraj Upadhyay <neeraj.iitr10@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
This commit adds a rcupdate.rcu_task_enqueue_lim module parameter that
sets the initial number of callback queues to use for the RCU Tasks
family of RCU implementations. This parameter allows testing of various
fanout values.
Reported-by: Martin Lau <kafai@fb.com>
Cc: Neeraj Upadhyay <neeraj.iitr10@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Currently, rcu_barrier_tasks(), rcu_barrier_tasks_rude(),
and rcu_barrier_tasks_trace() simply invoke the corresponding
synchronize_rcu_tasks*() function. This works because there is only
one callback queue.
However, there will soon be multiple callback queues. This commit
therefore scans the queues currently in use, entraining a callback on
each non-empty queue. Sequence numbers and reference counts are used
to synchronize this process in a manner similar to the approach taken
by rcu_barrier().
Reported-by: Martin Lau <kafai@fb.com>
Cc: Neeraj Upadhyay <neeraj.iitr10@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
If there is a flood of callbacks, it is necessary to put multiple
CPUs to work invoking those callbacks. This commit therefore uses a
workqueue-flooding approach to parallelize RCU Tasks callback execution.
Reported-by: Martin Lau <kafai@fb.com>
Cc: Neeraj Upadhyay <neeraj.iitr10@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
This commit adds a rcu_tasks_invoke_cbs() function that invokes all
ready callbacks on all of the per-CPU lists that are currently in use.
Reported-by: Martin Lau <kafai@fb.com>
Cc: Neeraj Upadhyay <neeraj.iitr10@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
This commit adds a rcu_tasks_need_gpcb() function that returns an
indication of whether another grace period is required, and if no grace
period is required, whether there are callbacks that need to be invoked.
The function scans all per-CPU lists currently in use.
Reported-by: Martin Lau <kafai@fb.com>
Cc: Neeraj Upadhyay <neeraj.iitr10@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
This commit adds a ->percpu_enqueue_lim field to the rcu_tasks structure.
This field contains two to the power of the ->percpu_enqueue_shift
field, easing construction of iterators over the per-CPU queues that
might contain RCU Tasks callbacks. Such iterators will be introduced
in later commits.
Reported-by: Martin Lau <kafai@fb.com>
Cc: Neeraj Upadhyay <neeraj.iitr10@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
This commit renames the rcu_tasks_percpu structure's ->cbs_pcpu_lock
to ->lock and then uses spin_lock_rcu_node() and friends to acquire and
release this lock, preparing for upcoming commits that will spread the
grace-period process across multiple CPUs and kthreads.
[ paulmck: Apply feedback from kernel test robot. ]
Reported-by: Martin Lau <kafai@fb.com>
Cc: Neeraj Upadhyay <neeraj.iitr10@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
With the addition of multiple callback-flood kthreads, the maximum number
of callbacks from any one of those kthreads is reported in the rcutorture
run summary. This commit changes this to report the sum of each kthread's
maximum number of callbacks in a given callback-flooding episode.
Cc: Neeraj Upadhyay <neeraj.iitr10@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
The RCU tasks flavors of RCU now need concurrent callback flooding to
test their ability to switch between single-queue mode and per-CPU queue
mode, but their lack of heavy-duty forward-progress features rules out
the use of rcutorture's current callback-flooding code. This commit
therefore provides the ability to limit the intensity of the callback
floods using a new ->cbflood_max field in the rcu_operations structure.
When this field is zero, there is no limit, otherwise, each callback-flood
kthread allocates at most ->cbflood_max callbacks.
Cc: Neeraj Upadhyay <neeraj.iitr10@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
This commit converts the rcutorture.fwd_progress module parameter from
bool to int, so that it specifies the number of callback-flood kthreads.
Values less than zero specify one kthread per CPU, however, the number of
kthreads executing concurrently is limited to the number of online CPUs.
This commit also reverse the order of the need-resched and callback-flood
operations to cause the callback flooding to happen more nearly at the
same time.
Cc: Neeraj Upadhyay <neeraj.iitr10@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
If we use the module stall_cpu option, we may get a soft lockup warning
in case we also don't pass the stall_cpu_block option.
Introduce the stall_no_softlockup option to avoid a soft lockup on
cpu stall even if we don't use the stall_cpu_block option.
Signed-off-by: Wander Lairson Costa <wander@redhat.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Unconditionally log messages corresponding to errors.
Acked-by: Davidlohr Bueso <dbueso@suse.de>
Signed-off-by: Li Zhijian <zhijianx.li@intel.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Unconditionally log messages corresponding to errors.
Acked-by: Davidlohr Bueso <dbueso@suse.de>
Signed-off-by: Li Zhijian <zhijianx.li@intel.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Add '\n' to macros to flush message for each call.
Acked-by: Davidlohr Bueso <dbueso@suse.de>
Signed-off-by: Li Zhijian <zhijianx.li@intel.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
An OOM is a serious error that should be logged even in non-verbose runs.
This commit therefore adds an unconditional SCALEOUT_ERRSTRING() macro
and uses it instead of VERBOSE_SCALEOUT_ERRSTRING() when reporting an OOM.
[ paulmck: Drop do-while from SCALEOUT_ERRSTRING() due to only single statement. ]
Signed-off-by: Li Zhijian <lizhijian@cn.fujitsu.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
This commit moves from a bespoke head and tail pointer in the
rcu_tasks_percpu structure to an rcu_segcblist structure, thus allowing
associating the grace-period sequence number with groups of callbacks.
This in turn will allow callbacks to be invoked independently on
different CPUs.
Reported-by: Martin Lau <kafai@fb.com>
Cc: Neeraj Upadhyay <neeraj.iitr10@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
This commit moves the rcu_tasks structure's ->n_gps grace-period-counter
field to a ->task_gp_seq grce-period sequence number in order to enable
use of the rcu_segcblist structure for the callback lists. This in turn
permits CPUs to lag behind the RCU Tasks grace-period sequence number
without suffering long-term slowdowns in callback invocation.
Reported-by: Martin Lau <kafai@fb.com>
Cc: Neeraj Upadhyay <neeraj.iitr10@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
This commit introduces a ->percpu_enqueue_shift field to the rcu_tasks
structure, and uses it to shift down the CPU number in order to
select a rcu_tasks_percpu structure. This field is currently set to a
sufficiently large shift count to always select the CPU-0 instance of
the rcu_tasks_percpu structure, and later commits will adjust this.
Reported-by: Martin Lau <kafai@fb.com>
Cc: Neeraj Upadhyay <neeraj.iitr10@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Currently, RCU Tasks Trace (as well as the other two flavors of RCU Tasks)
use a single global callback list. This works well and is simple, but
expected changes in workload will cause this list to become a bottleneck.
This commit therefore creates per-CPU callback lists for the various
flavors of RCU Tasks, but continues queueing on a single list, namely
that of CPU 0. Later commits will dynamically vary the number of lists
in use to accommodate dynamic changes in workload.
Reported-by: Martin Lau <kafai@fb.com>
Cc: Neeraj Upadhyay <neeraj.iitr10@gmail.com>
Tested-by: kernel test robot <beibei.si@intel.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
rcu_core() tries to ensure that its self-invocation in case of callbacks
overload only happen in softirq/rcuc mode. Indeed it doesn't make sense
to trigger local RCU core from nocb_cb kthread since it can execute
on a CPU different from the target rdp. Also in case of overload, the
nocb_cb kthread simply iterates a new loop of callbacks processing.
However the "offloaded" check that aims at preventing misplaced
rcu_core() invocations is wrong. First of all that state is volatile
and second: softirq/rcuc can execute while the target rdp is offloaded.
As a result rcu_core() can be invoked on the wrong CPU while in the
process of (de-)offloading.
Fix that with moving the rcu_core() self-invocation to rcu_core() itself,
irrespective of the rdp offloaded state.
Tested-by: Valentin Schneider <valentin.schneider@arm.com>
Tested-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <valentin.schneider@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Joel Fernandes <joel@joelfernandes.org>
Cc: Boqun Feng <boqun.feng@gmail.com>
Cc: Neeraj Upadhyay <neeraju@codeaurora.org>
Cc: Uladzislau Rezki <urezki@gmail.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Time limit only makes sense when callbacks are serviced in softirq mode
because:
_ In case we need to get back to the scheduler,
cond_resched_tasks_rcu_qs() is called after each callback.
_ In case some other softirq vector needs the CPU, the call to
local_bh_enable() before cond_resched_tasks_rcu_qs() takes care about
them via a call to do_softirq().
Therefore, make sure the time limit only applies to softirq mode.
Reviewed-by: Valentin Schneider <valentin.schneider@arm.com>
Tested-by: Valentin Schneider <valentin.schneider@arm.com>
Tested-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <valentin.schneider@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Joel Fernandes <joel@joelfernandes.org>
Cc: Boqun Feng <boqun.feng@gmail.com>
Cc: Neeraj Upadhyay <neeraju@codeaurora.org>
Cc: Uladzislau Rezki <urezki@gmail.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
The callbacks processing time limit makes sure we are not exceeding a
given amount of time executing the queue.
However its "continue" clause bypasses the cond_resched() call on
rcuc and NOCB kthreads, delaying it until we reach the limit, which can
be very long...
Make sure the scheduler has a higher priority than the time limit.
Reviewed-by: Valentin Schneider <valentin.schneider@arm.com>
Tested-by: Valentin Schneider <valentin.schneider@arm.com>
Tested-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <valentin.schneider@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Joel Fernandes <joel@joelfernandes.org>
Cc: Boqun Feng <boqun.feng@gmail.com>
Cc: Neeraj Upadhyay <neeraju@codeaurora.org>
Cc: Uladzislau Rezki <urezki@gmail.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
The current condition to limit the number of callbacks executed in a
row checks the offloaded state of the rdp. Not only is it volatile
but it is also misleading: the rcu_core() may well be executing
callbacks concurrently with NOCB kthreads, and the offloaded state
would then be verified on both cases. As a result the limit would
spuriously not apply anymore on softirq while in the middle of
(de-)offloading process.
Fix and clarify the condition with those constraints in mind:
_ If callbacks are processed either by rcuc or NOCB kthread, the call
to cond_resched_tasks_rcu_qs() is enough to take care of the overload.
_ If instead callbacks are processed by softirqs:
* If need_resched(), exit the callbacks processing
* Otherwise if CPU is idle we can continue
* Otherwise exit because a softirq shouldn't interrupt a task for too
long nor deprive other pending softirq vectors of the CPU.
Tested-by: Valentin Schneider <valentin.schneider@arm.com>
Tested-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <valentin.schneider@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Joel Fernandes <joel@joelfernandes.org>
Cc: Boqun Feng <boqun.feng@gmail.com>
Cc: Neeraj Upadhyay <neeraju@codeaurora.org>
Cc: Uladzislau Rezki <urezki@gmail.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Instead of hardcoding IRQ save and nocb lock, use the consolidated
API (and fix a comment as per Valentin Schneider's suggestion).
Reviewed-by: Valentin Schneider <valentin.schneider@arm.com>
Tested-by: Valentin Schneider <valentin.schneider@arm.com>
Tested-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <valentin.schneider@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Joel Fernandes <joel@joelfernandes.org>
Cc: Boqun Feng <boqun.feng@gmail.com>
Cc: Neeraj Upadhyay <neeraju@codeaurora.org>
Cc: Uladzislau Rezki <urezki@gmail.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
It's not entirely obvious why rdp->qlen_last_fqs_check is updated before
processing the queue only on offloaded rdp. There can be different
effect to that, either in favour of triggering the force quiescent state
path or not. For example:
1) If the number of callbacks has decreased since the last
rdp->qlen_last_fqs_check update (because we recently called
rcu_do_batch() and we executed below qhimark callbacks) and the number
of processed callbacks on a subsequent do_batch() arranges for
exceeding qhimark on non-offloaded but not on offloaded setup, then we
may spare a later run to the force quiescent state
slow path on __call_rcu_nocb_wake(), as compared to the non-offloaded
counterpart scenario.
Here is such an offloaded scenario instance:
qhimark = 1000
rdp->last_qlen_last_fqs_check = 3000
rcu_segcblist_n_cbs(rdp) = 2000
rcu_do_batch() {
if (offloaded)
rdp->last_qlen_fqs_check = rcu_segcblist_n_cbs(rdp) // 2000
// run 1000 callback
rcu_segcblist_n_cbs(rdp) = 1000
// Not updating rdp->qlen_last_fqs_check
if (count < rdp->qlen_last_fqs_check - qhimark)
rdp->qlen_last_fqs_check = count;
}
call_rcu() * 1001 {
__call_rcu_nocb_wake() {
// not taking the fqs slowpath:
// rcu_segcblist_n_cbs(rdp) == 2001
// rdp->qlen_last_fqs_check == 2000
// qhimark == 1000
if (len > rdp->qlen_last_fqs_check + qhimark)
...
}
In the case of a non-offloaded scenario, rdp->qlen_last_fqs_check
would be 1000 and the fqs slowpath would have executed.
2) If the number of callbacks has increased since the last
rdp->qlen_last_fqs_check update (because we recently queued below
qhimark callbacks) and the number of callbacks executed in rcu_do_batch()
doesn't exceed qhimark for either offloaded or non-offloaded setup,
then it's possible that the offloaded scenario later run the force
quiescent state slow path on __call_rcu_nocb_wake() while the
non-offloaded doesn't.
qhimark = 1000
rdp->last_qlen_last_fqs_check = 3000
rcu_segcblist_n_cbs(rdp) = 2000
rcu_do_batch() {
if (offloaded)
rdp->last_qlen_last_fqs_check = rcu_segcblist_n_cbs(rdp) // 2000
// run 100 callbacks
// concurrent queued 100
rcu_segcblist_n_cbs(rdp) = 2000
// Not updating rdp->qlen_last_fqs_check
if (count < rdp->qlen_last_fqs_check - qhimark)
rdp->qlen_last_fqs_check = count;
}
call_rcu() * 1001 {
__call_rcu_nocb_wake() {
// Taking the fqs slowpath:
// rcu_segcblist_n_cbs(rdp) == 3001
// rdp->qlen_last_fqs_check == 2000
// qhimark == 1000
if (len > rdp->qlen_last_fqs_check + qhimark)
...
}
In the case of a non-offloaded scenario, rdp->qlen_last_fqs_check
would be 3000 and the fqs slowpath would have executed.
The reason for updating rdp->qlen_last_fqs_check when invoking callbacks
for offloaded CPUs is that there is usually no point in waking up either
the rcuog or rcuoc kthreads while in this state. After all, both threads
are prohibited from indefinite sleeps.
The exception is when some huge number of callbacks are enqueued while
rcu_do_batch() is in the midst of invoking, in which case interrupting
the rcuog kthread's timed sleep might get more callbacks set up for the
next grace period.
Reported-and-tested-by: Valentin Schneider <valentin.schneider@arm.com>
Tested-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Original-patch-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <valentin.schneider@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Joel Fernandes <joel@joelfernandes.org>
Cc: Boqun Feng <boqun.feng@gmail.com>
Cc: Neeraj Upadhyay <neeraju@codeaurora.org>
Cc: Uladzislau Rezki <urezki@gmail.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
When callbacks are offloaded, the NOCB kthreads handle the callbacks
progression on behalf of rcu_core().
However during the (de-)offloading process, the kthread may not be
entirely up to the task. As a result some callbacks grace period
sequence number may remain stale for a while because rcu_core() won't
take care of them either.
Fix this with forcing callbacks acceleration from rcu_core() as long
as the offloading process isn't complete.
Reported-and-tested-by: Valentin Schneider <valentin.schneider@arm.com>
Tested-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <valentin.schneider@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Joel Fernandes <joel@joelfernandes.org>
Cc: Boqun Feng <boqun.feng@gmail.com>
Cc: Neeraj Upadhyay <neeraju@codeaurora.org>
Cc: Uladzislau Rezki <urezki@gmail.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
While reporting a quiescent state for a given CPU, rcu_core() takes
advantage of the freshly loaded grace period sequence number and the
locked rnp to accelerate the callbacks whose sequence number have been
assigned a stale value.
This action is only necessary when the rdp isn't offloaded, otherwise
the NOCB kthreads already take care of the callbacks progression.
However the check for the offloaded state is volatile because it is
performed outside the IRQs disabled section. It's possible for the
offloading process to preempt rcu_core() at that point on PREEMPT_RT.
This is dangerous because rcu_core() may end up accelerating callbacks
concurrently with NOCB kthreads without appropriate locking.
Fix this with moving the offloaded check inside the rnp locking section.
Reported-and-tested-by: Valentin Schneider <valentin.schneider@arm.com>
Reviewed-by: Valentin Schneider <valentin.schneider@arm.com>
Tested-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Joel Fernandes <joel@joelfernandes.org>
Cc: Boqun Feng <boqun.feng@gmail.com>
Cc: Neeraj Upadhyay <neeraju@codeaurora.org>
Cc: Uladzislau Rezki <urezki@gmail.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
On PREEMPT_RT, if rcu_core() is preempted by the de-offloading process,
some work, such as callbacks acceleration and invocation, may be left
unattended due to the volatile checks on the offloaded state.
In the worst case this work is postponed until the next rcu_pending()
check that can take a jiffy to reach, which can be a problem in case
of callbacks flooding.
Solve that with invoking rcu_core() early in the de-offloading process.
This way any work dismissed by an ongoing rcu_core() call fooled by
a preempting deoffloading process will be caught up by a nearby future
recall to rcu_core(), this time fully aware of the de-offloading state.
Tested-by: Valentin Schneider <valentin.schneider@arm.com>
Tested-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <valentin.schneider@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Joel Fernandes <joel@joelfernandes.org>
Cc: Boqun Feng <boqun.feng@gmail.com>
Cc: Neeraj Upadhyay <neeraju@codeaurora.org>
Cc: Uladzislau Rezki <urezki@gmail.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Currently SEGCBLIST_SOFTIRQ_ONLY is a bit of an exception among the
segcblist flags because it is an exclusive state that doesn't mix up
with the other flags. Remove it in favour of:
_ A flag specifying that rcu_core() needs to perform callbacks execution
and acceleration
and
_ A flag specifying we want the nocb lock to be held in any needed
circumstances
This clarifies the code and is more flexible: It allows to have a state
where rcu_core() runs with locking while offloading hasn't started yet.
This is a necessary step to prepare for triggering rcu_core() at the
very beginning of the de-offloading process so that rcu_core() won't
dismiss work while being preempted by the de-offloading process, at
least not without a pending subsequent rcu_core() that will quickly
catch up.
Reviewed-by: Valentin Schneider <Valentin.Schneider@arm.com>
Tested-by: Valentin Schneider <valentin.schneider@arm.com>
Tested-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Joel Fernandes <joel@joelfernandes.org>
Cc: Boqun Feng <boqun.feng@gmail.com>
Cc: Neeraj Upadhyay <neeraju@codeaurora.org>
Cc: Uladzislau Rezki <urezki@gmail.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
rcu_nocb_lock_irqsave() can be preempted between the call to
rcu_segcblist_is_offloaded() and the actual locking. This matters now
that rcu_core() is preemptible on PREEMPT_RT and the (de-)offloading
process can interrupt the softirq or the rcuc kthread.
As a result we may locklessly call into code that requires nocb locking.
In practice this is a problem while we accelerate callbacks on rcu_core().
Simply disabling interrupts before (instead of after) checking the NOCB
offload state fixes the issue.
Reported-and-tested-by: Valentin Schneider <valentin.schneider@arm.com>
Tested-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <valentin.schneider@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Joel Fernandes <joel@joelfernandes.org>
Cc: Boqun Feng <boqun.feng@gmail.com>
Cc: Neeraj Upadhyay <neeraju@codeaurora.org>
Cc: Uladzislau Rezki <urezki@gmail.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Currently, rcu_advance_cbs_nowake() checks that a grace period is in
progress, however, that grace period could end just after the check.
This commit rechecks that a grace period is still in progress while
holding the rcu_node structure's lock. The grace period cannot end while
the current CPU's rcu_node structure's ->lock is held, thus avoiding
false positives from the WARN_ON_ONCE().
As Daniel Vacek noted, it is not necessary for the rcu_node structure
to have a CPU that has not yet passed through its quiescent state.
Tested-by: Guillaume Morin <guillaume@morinfr.org>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Expedited RCU grace periods invoke sync_rcu_exp_select_node_cpus(), which
takes two passes over the leaf rcu_node structure's CPUs. The first
pass gathers up the current CPU and CPUs that are in dynticks idle mode.
The workqueue will report a quiescent state on their behalf later.
The second pass sends IPIs to the rest of the CPUs, but excludes the
current CPU, incorrectly assuming it has been included in the first
pass's list of CPUs.
Unfortunately the current CPU may have changed between the first and
second pass, due to the fact that the various rcu_node structures'
->lock fields have been dropped, thus momentarily enabling preemption.
This means that if the second pass's CPU was not on the first pass's
list, it will be ignored completely. There will be no IPI sent to
it, and there will be no reporting of quiescent states on its behalf.
Unfortunately, the expedited grace period will nevertheless be waiting
for that CPU to report a quiescent state, but with that CPU having no
reason to believe that such a report is needed.
The result will be an expedited grace period stall.
Fix this by no longer excluding the current CPU from consideration during
the second pass.
Fixes: b9ad4d6ed1 ("rcu: Avoid self-IPI in sync_rcu_exp_select_node_cpus()")
Reviewed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Cc: Uladzislau Rezki <urezki@gmail.com>
Cc: Neeraj Upadhyay <quic_neeraju@quicinc.com>
Cc: Boqun Feng <boqun.feng@gmail.com>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Joel Fernandes <joel@joelfernandes.org>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
In non-preemptible kernels, an unfortunately timed expedited grace period
can result in the rcu_exp_handler() IPI handler setting the rcu_data
structure's cpu_no_qs.b.exp field just as the target CPU enters idle.
There are situations in which this field will not be checked until after
that CPU exits idle. The resulting grace-period latency does not qualify
as "expedited".
This commit therefore checks this field upon non-preemptible idle entry in
the rcu_preempt_deferred_qs() function. It also qualifies the rcu_core()
preempt_count() check with IS_ENABLED(CONFIG_PREEMPT_COUNT) to prevent
false-positive quiescent states from count-free kernels.
Reported-by: Neeraj Upadhyay <neeraju@codeaurora.org>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
If an RCU expedited grace period starts just when a CPU is in the process
of going offline, so that the outgoing CPU has completed its pass through
stop-machine but has not yet completed its final dive into the idle loop,
RCU will attempt to enable that CPU's scheduling-clock tick via a call
to tick_dep_set_cpu(). For this to happen, that CPU has to have been
online when the expedited grace period completed its CPU-selection phase.
This is pointless: The outgoing CPU has interrupts disabled, so it cannot
take a scheduling-clock tick anyway. In addition, the tick_dep_set_cpu()
function's eventual call to irq_work_queue_on() will splat as follows:
smpboot: CPU 1 is now offline
WARNING: CPU: 6 PID: 124 at kernel/irq_work.c:95
+irq_work_queue_on+0x57/0x60
Modules linked in:
CPU: 6 PID: 124 Comm: kworker/6:2 Not tainted 5.15.0-rc1+ #3
Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS
+rel-1.14.0-0-g155821a-rebuilt.opensuse.org 04/01/2014
Workqueue: rcu_gp wait_rcu_exp_gp
RIP: 0010:irq_work_queue_on+0x57/0x60
Code: 8b 05 1d c7 ea 62 a9 00 00 f0 00 75 21 4c 89 ce 44 89 c7 e8
+9b 37 fa ff ba 01 00 00 00 89 d0 c3 4c 89 cf e8 3b ff ff ff eb ee <0f> 0b eb b7
+0f 0b eb db 90 48 c7 c0 98 2a 02 00 65 48 03 05 91
6f
RSP: 0000:ffffb12cc038fe48 EFLAGS: 00010282
RAX: 0000000000000001 RBX: 0000000000005208 RCX: 0000000000000020
RDX: 0000000000000001 RSI: 0000000000000001 RDI: ffff9ad01f45a680
RBP: 000000000004c990 R08: 0000000000000001 R09: ffff9ad01f45a680
R10: ffffb12cc0317db0 R11: 0000000000000001 R12: 00000000fffecee8
R13: 0000000000000001 R14: 0000000000026980 R15: ffffffff9e53ae00
FS: 0000000000000000(0000) GS:ffff9ad01f580000(0000)
+knlGS:0000000000000000
CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 0000000000000000 CR3: 000000000de0c000 CR4: 00000000000006e0
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
Call Trace:
tick_nohz_dep_set_cpu+0x59/0x70
rcu_exp_wait_wake+0x54e/0x870
? sync_rcu_exp_select_cpus+0x1fc/0x390
process_one_work+0x1ef/0x3c0
? process_one_work+0x3c0/0x3c0
worker_thread+0x28/0x3c0
? process_one_work+0x3c0/0x3c0
kthread+0x115/0x140
? set_kthread_struct+0x40/0x40
ret_from_fork+0x22/0x30
---[ end trace c5bf75eb6aa80bc6 ]---
This commit therefore avoids invoking tick_dep_set_cpu() on offlined
CPUs to limit both futility and false-positive splats.
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
