Re: [PATCH] task_work: Consume only item at a time while invoking the callbacks.

[Frederic Weisbecker <frederic@kernel.org>]

Le Wed, Feb 26, 2025 at 01:50:48PM +0100, Oleg Nesterov a écrit :
> On 02/21, Sebastian Andrzej Siewior wrote:
> >
> > Yi and syzbot managed to hang the task within task_run().
> >
> > The problem is
> >  task_work_run() -> __fput() -> perf_release() ->
> >  perf_event_release_kernel() -> _free_event() ->
> >  perf_pending_task_sync() -> task_work_cancel() failed ->
> >  rcuwait_wait_event().
> >
> > Once task_work_run() is running, the list of callbacks removed from the
> > task_struct and from this point on task_work_cancel() can't remove any
> > pending and not yet started work items.
> 
> But can this patch really solve the problem?
> 
> Suppose we have two tasks, T1 and T2.
> 
> T1 does fd = perf_event_open(pid => T2->pid);
> T2 does fd = perf_event_open(pid => T1->pid);
> 
> Now, both T1 and T2 do close(fd), call task_work_run(), dequeue the
> ____fput work, and finally call __fput(). Now suppose that both perf
> events fire before T1/T2 call perf_event_release_kernel/_free_event.
> 
> Now, T1 and T2 will hang forever in perf_pending_task_sync() waiting
> for each other. task_work_cancel(current) can't succeed with or without
> this patch.
> 
> No?

Duh!

So indeed, the wait/wake based solution is too fragile. Are we back to the
old serialized workqueue days flavour of deadlocks with task work?

Anyway the perf_pending_task()'s put_event() based solution thing should fix
that scenario too.

Thanks.