> > > OK, I spotted the problem. If we fallback to the on-stack allocation in
> > > bdi_writeback_all(), then we do the wait for the work completion with
> > > the bdi_lock mutex held. This can deadlock with bdi_forker_task(), so if
> > > we require that to be invoked to make progress (happens if a thread
> > > needs to be restarted), then we have a deadlock on that mutex.

> > > I'll cook up a fix for this, but probably not before the morning.

> > Untested fix. I think it should work, but I haven't run it here yet.

> Thanks for your feedback and explanations. Will not be able to test the
> patch before the afternoon. Will then give feedback as soon as I can. Of
> course, I can also test a new fix or series if you send one in the
> morning, after your own testing.

Just ran a quick test with your fix applied and ran into the same
problem: manual sync stays in D state. I could capture SysRq-T, I am
attaching part of the output (which was very big and continuously
displaying).


Hope this helps,

-- 
Damien Wyart