Re: [RFC] workqueue: remove manual lockdep uses to detect deadlocks
From: Peter Zijlstra
Date: Tue Aug 29 2017 - 14:57:36 EST
On Sat, Aug 26, 2017 at 12:49:26AM +0900, Byungchul Park wrote:
> > However, how would it distinguish things like flushing another work
>
> I think it must be distinguished with what it actually waits for, e.i.
> completion
> variables instead of work or wq. I will make it next week and let you know.
So no. The existing annotations are strictly better than relying on
cross-release.
As you know the problem with cross-release is that it is timing
dependent. You need to actually observe the problematic sequence before
it can warn, and only the whole instance->class mapping saves us from
actually hitting the deadlock.
Cross-release can result in deadlocks without warnings. If you were to
run:
mutex_lock(A);
mutex_lock(A);
complete(C);
wait_for_completion(C);
You'd deadlock without issue. Only if we observe this:
mutex_lock(A);
wait_for_completion(C);
mutex_lock(A);
complete(C);
Where we acquire A after wait_for_completion() but before complete()
will we observe the deadlock.
The same would be true for using cross-release for workqueues as well,
something like:
W:
mutex_lock(A)
mutex_lock(A)
flush_work(W)
would go unreported whereas the current workqueue annotation will
generate a splat.
This does not mean cross-release isn't worth it, its better than nothing,
but its strictly weaker than traditional annotations.
So where a traditional annotation is possible, we should use them.