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>
dbf2576e37