Now that manager_mutex's role has changed from synchronizing manager
role to excluding hotplug against manager, the name is misleading.
As it is protecting the CPU-association of the gcwq now, rename it to
assoc_mutex.
This patch is pure rename and doesn't introduce any functional change.
tj: Updated comments and description.
Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Now both worker destruction and idle rebinding remove the worker from
idle list while it's still idle, so list_empty(&worker->entry) can be
used to test whether either is pending and WORKER_DIE to distinguish
between the two instead making WORKER_REBIND unnecessary.
Use list_empty(&worker->entry) to determine whether destruction or
rebinding is pending. This simplifies worker state transitions.
WORKER_REBIND is not needed anymore. Remove it.
tj: Updated comments and description.
Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Because the old unbind/rebinding implementation wasn't atomic w.r.t.
GCWQ_DISASSOCIATED manipulation which is protected by
global_cwq->lock, we had to use two flags, WORKER_UNBOUND and
WORKER_REBIND, to avoid incorrectly losing all NOT_RUNNING bits with
back-to-back CPU hotplug operations; otherwise, completion of
rebinding while another unbinding is in progress could clear UNBIND
prematurely.
Now that both unbind/rebinding are atomic w.r.t. GCWQ_DISASSOCIATED,
there's no need to use two flags. Just one is enough. Don't use
WORKER_REBIND for busy rebinding.
tj: Updated description.
Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Currently rebind_workers() uses rebinds idle workers synchronously
before proceeding to requesting busy workers to rebind. This is
necessary because all workers on @worker_pool->idle_list must be bound
before concurrency management local wake-ups from the busy workers
take place.
Unfortunately, the synchronous idle rebinding is quite complicated.
This patch reimplements idle rebinding to simplify the code path.
Rather than trying to make all idle workers bound before rebinding
busy workers, we simply remove all to-be-bound idle workers from the
idle list and let them add themselves back after completing rebinding
(successful or not).
As only workers which finished rebinding can on on the idle worker
list, the idle worker list is guaranteed to have only bound workers
unless CPU went down again and local wake-ups are safe.
After the change, @worker_pool->nr_idle may deviate than the actual
number of idle workers on @worker_pool->idle_list. More specifically,
nr_idle may be non-zero while ->idle_list is empty. All users of
->nr_idle and ->idle_list are audited. The only affected one is
too_many_workers() which is updated to check %false if ->idle_list is
empty regardless of ->nr_idle.
After this patch, rebind_workers() no longer performs the nasty
idle-rebind retries which require temporary release of gcwq->lock, and
both unbinding and rebinding are atomic w.r.t. global_cwq->lock.
worker->idle_rebind and global_cwq->rebind_hold are now unnecessary
and removed along with the definition of struct idle_rebind.
Changed from V1:
1) remove unlikely from too_many_workers(), ->idle_list can be empty
anytime, even before this patch, no reason to use unlikely.
2) fix a small rebasing mistake.
(which is from rebasing the orignal fixing patch to for-next)
3) add a lot of comments.
4) clear WORKER_REBIND unconditionaly in idle_worker_rebind()
tj: Updated comments and description.
Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
This merge is necessary as Lai's CPU hotplug restructuring series
depends on the CPU hotplug bug fixes in for-3.6-fixes.
The merge creates one trivial conflict between the following two
commits.
96e65306b8 "workqueue: UNBOUND -> REBIND morphing in rebind_workers() should be atomic"
e2b6a6d570 "workqueue: use system_highpri_wq for highpri workers in rebind_workers()"
Both add local variable definitions to the same block and can be
merged in any order.
Signed-off-by: Tejun Heo <tj@kernel.org>
To simplify both normal and CPU hotplug paths, worker management is
prevented while CPU hoplug is in progress. This is achieved by CPU
hotplug holding the same exclusion mechanism used by workers to ensure
there's only one manager per pool.
If someone else seems to be performing the manager role, workers
proceed to execute work items. CPU hotplug using the same mechanism
can lead to idle worker depletion because all workers could proceed to
execute work items while CPU hotplug is in progress and CPU hotplug
itself wouldn't actually perform the worker management duty - it
doesn't guarantee that there's an idle worker left when it releases
management.
This idle worker depletion, under extreme circumstances, can break
forward-progress guarantee and thus lead to deadlock.
This patch fixes the bug by using separate mechanisms for manager
exclusion among workers and hotplug exclusion. For manager exclusion,
POOL_MANAGING_WORKERS which was restored by the previous patch is
used. pool->manager_mutex is now only used for exclusion between the
elected manager and CPU hotplug. The elected manager won't proceed
without holding pool->manager_mutex.
This ensures that the worker which won the manager position can't skip
managing while CPU hotplug is in progress. It will block on
manager_mutex and perform management after CPU hotplug is complete.
Note that hotplug may happen while waiting for manager_mutex. A
manager isn't either on idle or busy list and thus the hoplug code
can't unbind/rebind it. Make the manager handle its own un/rebinding.
tj: Updated comment and description.
Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
This patch restores POOL_MANAGING_WORKERS which was replaced by
pool->manager_mutex by 6037315269 "workqueue: use mutex for global_cwq
manager exclusion".
There's a subtle idle worker depletion bug across CPU hotplug events
and we need to distinguish an actual manager and CPU hotplug
preventing management. POOL_MANAGING_WORKERS will be used for the
former and manager_mutex the later.
This patch just lays POOL_MANAGING_WORKERS on top of the existing
manager_mutex and doesn't introduce any synchronization changes. The
next patch will update it.
Note that this patch fixes a non-critical anomaly where
too_many_workers() may return %true spuriously while CPU hotplug is in
progress. While the issue could schedule idle timer spuriously, it
didn't trigger any actual misbehavior.
tj: Rewrote patch description.
Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Currently, rebind_workers() and idle_worker_rebind() are two-way
interlocked. rebind_workers() waits for idle workers to finish
rebinding and rebound idle workers wait for rebind_workers() to finish
rebinding busy workers before proceeding.
Unfortunately, this isn't enough. The second wait from idle workers
is implemented as follows.
wait_event(gcwq->rebind_hold, !(worker->flags & WORKER_REBIND));
rebind_workers() clears WORKER_REBIND, wakes up the idle workers and
then returns. If CPU hotplug cycle happens again before one of the
idle workers finishes the above wait_event(), rebind_workers() will
repeat the first part of the handshake - set WORKER_REBIND again and
wait for the idle worker to finish rebinding - and this leads to
deadlock because the idle worker would be waiting for WORKER_REBIND to
clear.
This is fixed by adding another interlocking step at the end -
rebind_workers() now waits for all the idle workers to finish the
above WORKER_REBIND wait before returning. This ensures that all
rebinding steps are complete on all idle workers before the next
hotplug cycle can happen.
This problem was diagnosed by Lai Jiangshan who also posted a patch to
fix the issue, upon which this patch is based.
This is the minimal fix and further patches are scheduled for the next
merge window to simplify the CPU hotplug path.
Signed-off-by: Tejun Heo <tj@kernel.org>
Original-patch-by: Lai Jiangshan <laijs@cn.fujitsu.com>
LKML-Reference: <1346516916-1991-3-git-send-email-laijs@cn.fujitsu.com>
This doesn't make any functional difference and is purely to help the
next patch to be simpler.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Lai Jiangshan <laijs@cn.fujitsu.com>
The compiler may compile the following code into TWO write/modify
instructions.
worker->flags &= ~WORKER_UNBOUND;
worker->flags |= WORKER_REBIND;
so the other CPU may temporarily see worker->flags which doesn't have
either WORKER_UNBOUND or WORKER_REBIND set and perform local wakeup
prematurely.
Fix it by using single explicit assignment via ACCESS_ONCE().
Because idle workers have another WORKER_NOT_RUNNING flag, this bug
doesn't exist for them; however, update it to use the same pattern for
consistency.
tj: Applied the change to idle workers too and updated comments and
patch description a bit.
Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: stable@vger.kernel.org
cancel_delayed_work() can't be called from IRQ handlers due to its use
of del_timer_sync() and can't cancel work items which are already
transferred from timer to worklist.
Also, unlike other flush and cancel functions, a canceled delayed_work
would still point to the last associated cpu_workqueue. If the
workqueue is destroyed afterwards and the work item is re-used on a
different workqueue, the queueing code can oops trying to dereference
already freed cpu_workqueue.
This patch reimplements cancel_delayed_work() using
try_to_grab_pending() and set_work_cpu_and_clear_pending(). This
allows the function to be called from IRQ handlers and makes its
behavior consistent with other flush / cancel functions.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Up to now, for delayed_works, try_to_grab_pending() couldn't be used
from IRQ handlers because IRQs may happen while
delayed_work_timer_fn() is in progress leading to indefinite -EAGAIN.
This patch makes delayed_work use the new TIMER_IRQSAFE flag for
delayed_work->timer. This makes try_to_grab_pending() and thus
mod_delayed_work_on() safe to call from IRQ handlers.
Signed-off-by: Tejun Heo <tj@kernel.org>
Now that all workqueues are non-reentrant, system[_freezable]_wq() are
equivalent to system_nrt[_freezable]_wq(). Replace the latter with
wrappers around system[_freezable]_wq(). The wrapping goes through
inline functions so that __deprecated can be added easily.
Signed-off-by: Tejun Heo <tj@kernel.org>
Now that all workqueues are non-reentrant, flush[_delayed]_work_sync()
are equivalent to flush[_delayed]_work(). Drop the separate
implementation and make them thin wrappers around
flush[_delayed]_work().
* start_flush_work() no longer takes @wait_executing as the only left
user - flush_work() - always sets it to %true.
* __cancel_work_timer() uses flush_work() instead of wait_on_work().
Signed-off-by: Tejun Heo <tj@kernel.org>
By default, each per-cpu part of a bound workqueue operates separately
and a work item may be executing concurrently on different CPUs. The
behavior avoids some cross-cpu traffic but leads to subtle weirdities
and not-so-subtle contortions in the API.
* There's no sane usefulness in allowing a single work item to be
executed concurrently on multiple CPUs. People just get the
behavior unintentionally and get surprised after learning about it.
Most either explicitly synchronize or use non-reentrant/ordered
workqueue but this is error-prone.
* flush_work() can't wait for multiple instances of the same work item
on different CPUs. If a work item is executing on cpu0 and then
queued on cpu1, flush_work() can only wait for the one on cpu1.
Unfortunately, work items can easily cross CPU boundaries
unintentionally when the queueing thread gets migrated. This means
that if multiple queuers compete, flush_work() can't even guarantee
that the instance queued right before it is finished before
returning.
* flush_work_sync() was added to work around some of the deficiencies
of flush_work(). In addition to the usual flushing, it ensures that
all currently executing instances are finished before returning.
This operation is expensive as it has to walk all CPUs and at the
same time fails to address competing queuer case.
Incorrectly using flush_work() when flush_work_sync() is necessary
is an easy error to make and can lead to bugs which are difficult to
reproduce.
* Similar problems exist for flush_delayed_work[_sync]().
Other than the cross-cpu access concern, there's no benefit in
allowing parallel execution and it's plain silly to have this level of
contortion for workqueue which is widely used from core code to
extremely obscure drivers.
This patch makes all workqueues non-reentrant. If a work item is
executing on a different CPU when queueing is requested, it is always
queued to that CPU. This guarantees that any given work item can be
executing on one CPU at maximum and if a work item is queued and
executing, both are on the same CPU.
The only behavior change which may affect workqueue users negatively
is that non-reentrancy overrides the affinity specified by
queue_work_on(). On a reentrant workqueue, the affinity specified by
queue_work_on() is always followed. Now, if the work item is
executing on one of the CPUs, the work item will be queued there
regardless of the requested affinity. I've reviewed all workqueue
users which request explicit affinity, and, fortunately, none seems to
be crazy enough to exploit parallel execution of the same work item.
This adds an additional busy_hash lookup if the work item was
previously queued on a different CPU. This shouldn't be noticeable
under any sane workload. Work item queueing isn't a very
high-frequency operation and they don't jump across CPUs all the time.
In a micro benchmark to exaggerate this difference - measuring the
time it takes for two work items to repeatedly jump between two CPUs a
number (10M) of times with busy_hash table densely populated, the
difference was around 3%.
While the overhead is measureable, it is only visible in pathological
cases and the difference isn't huge. This change brings much needed
sanity to workqueue and makes its behavior consistent with timer. I
think this is the right tradeoff to make.
This enables significant simplification of workqueue API.
Simplification patches will follow.
Signed-off-by: Tejun Heo <tj@kernel.org>
To speed cpu down processing up, use system_highpri_wq.
As scheduling priority of workers on it is higher than system_wq and
it is not contended by other normal works on this cpu, work on it
is processed faster than system_wq.
tj: CPU up/downs care quite a bit about latency these days. This
shouldn't hurt anything and makes sense.
Signed-off-by: Joonsoo Kim <js1304@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
In rebind_workers(), we do inserting a work to rebind to cpu for busy workers.
Currently, in this case, we use only system_wq. This makes a possible
error situation as there is mismatch between cwq->pool and worker->pool.
To prevent this, we should use system_highpri_wq for highpri worker
to match theses. This implements it.
tj: Rephrased comment a bit.
Signed-off-by: Joonsoo Kim <js1304@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Commit 3270476a6c ('workqueue: reimplement
WQ_HIGHPRI using a separate worker_pool') introduce separate worker pool
for HIGHPRI. When we handle busyworkers for gcwq, it can be normal worker
or highpri worker. But, we don't consider this difference in rebind_workers(),
we use just system_wq for highpri worker. It makes mismatch between
cwq->pool and worker->pool.
It doesn't make error in current implementation, but possible in the future.
Now, we introduce system_highpri_wq to use proper cwq for highpri workers
in rebind_workers(). Following patch fix this issue properly.
tj: Even apart from rebinding, having system_highpri_wq generally
makes sense.
Signed-off-by: Joonsoo Kim <js1304@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
We assign cpu id into work struct's data field in __queue_delayed_work_on().
In current implementation, when work is come in first time,
current running cpu id is assigned.
If we do __queue_delayed_work_on() with CPU A on CPU B,
__queue_work() invoked in delayed_work_timer_fn() go into
the following sub-optimal path in case of WQ_NON_REENTRANT.
gcwq = get_gcwq(cpu);
if (wq->flags & WQ_NON_REENTRANT &&
(last_gcwq = get_work_gcwq(work)) && last_gcwq != gcwq) {
Change lcpu to @cpu and rechange lcpu to local cpu if lcpu is WORK_CPU_UNBOUND.
It is sufficient to prevent to go into sub-optimal path.
tj: Slightly rephrased the comment.
Signed-off-by: Joonsoo Kim <js1304@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
When we do tracing workqueue_queue_work(), it records requested cpu.
But, if !(@wq->flag & WQ_UNBOUND) and @cpu is WORK_CPU_UNBOUND,
requested cpu is changed as local cpu.
In case of @wq->flag & WQ_UNBOUND, above change is not occured,
therefore it is reasonable to correct it.
Use temporary local variable for storing requested cpu.
Signed-off-by: Joonsoo Kim <js1304@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Commit 3270476a6c ('workqueue: reimplement
WQ_HIGHPRI using a separate worker_pool') introduce separate worker_pool
for HIGHPRI. Although there is NR_WORKER_POOLS enum value which represent
size of pools, definition of worker_pool in gcwq doesn't use it.
Using it makes code robust and prevent future mistakes.
So change code to use this enum value.
Signed-off-by: Joonsoo Kim <js1304@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Any operation which clears PENDING should be preceded by a wmb to
guarantee that the next PENDING owner sees all the changes made before
PENDING release.
There are only two places where PENDING is cleared -
set_work_cpu_and_clear_pending() and clear_work_data(). The caller of
the former already does smp_wmb() but the latter doesn't have any.
Move the wmb above set_work_cpu_and_clear_pending() into it and add
one to clear_work_data().
There hasn't been any report related to this issue, and, given how
clear_work_data() is used, it is extremely unlikely to have caused any
actual problems on any architecture.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Oleg Nesterov <oleg@redhat.com>
delayed_work encodes the workqueue to use and the last CPU in
delayed_work->work.data while it's on timer. The target CPU is
implicitly recorded as the CPU the timer is queued on and
delayed_work_timer_fn() queues delayed_work->work to the CPU it is
running on.
Unfortunately, this leaves flush_delayed_work[_sync]() no way to find
out which CPU the delayed_work was queued for when they try to
re-queue after killing the timer. Currently, it chooses the local CPU
flush is running on. This can unexpectedly move a delayed_work queued
on a specific CPU to another CPU and lead to subtle errors.
There isn't much point in trying to save several bytes in struct
delayed_work, which is already close to a hundred bytes on 64bit with
all debug options turned off. This patch adds delayed_work->cpu to
remember the CPU it's queued for.
Note that if the timer is migrated during CPU down, the work item
could be queued to the downed global_cwq after this change. As a
detached global_cwq behaves like an unbound one, this doesn't change
much for the delayed_work.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
