Re: [Qemu-devel] [Qemu-block] [PATCH v2 2/4] block: Pause block jobs in bdrv_drain_all

[Stefan Hajnoczi <stefanha@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 2285 bytes --]

On Wed, Apr 08, 2015 at 04:56:14PM +0200, Alberto Garcia wrote:
> On Wed, Apr 08, 2015 at 11:37:52AM +0100, Stefan Hajnoczi wrote:
> 
> > > +    QTAILQ_FOREACH(bs, &bdrv_states, device_list) {
> > > +        AioContext *aio_context = bdrv_get_aio_context(bs);
> > > +
> > > +        aio_context_acquire(aio_context);
> > > +        if (bs->job) {
> > > +            block_job_pause(bs->job);
> > > +        }
> > > +        aio_context_release(aio_context);
> > > +    }
> > > +
> > >      while (busy) {
> > >          busy = false;
> > >  
> > > @@ -2044,6 +2054,16 @@ void bdrv_drain_all(void)
> > >              aio_context_release(aio_context);
> > >          }
> > >      }
> > > +
> > > +    QTAILQ_FOREACH(bs, &bdrv_states, device_list) {
> > > +        AioContext *aio_context = bdrv_get_aio_context(bs);
> > > +
> > > +        aio_context_acquire(aio_context);
> > > +        if (bs->job) {
> > > +            block_job_resume(bs->job);
> > > +        }
> > > +        aio_context_release(aio_context);
> > > +    }
> > >  }
> > 
> > There is a tiny chance that we pause a job (which actually just sets
> > job->paused = true but there's no guarantee the job's coroutine
> > reacts to this) right before it terminates.  Then aio_poll() enters
> > the coroutine one last time and the job terminates.
> > 
> > We then reach the resume portion of bdrv_drain_all() and the job no
> > longer exists.  Hopefully nothing started a new job in the meantime.
> > bs->job should either be the original block job or NULL.
> > 
> > This code seems under current assumptions, but I just wanted to
> > raise these issues in case someone sees problems that I've missed.
> 
> Is it possible that a new job is started in the meantime? If that's
> the case this will hit the assertion in block_job_resume().

That is currently not possible since the QEMU monitor does not run while
we're waiting in aio_poll().

Therefore no block job monitor commands could spawn a new job.

If code is added that spawns a job based on an AioContext timer or due
to some other event, then this assumption no longer holds and there is a
problem because block_job_resume() is called on a job that never paused.

But for now there is no problem.

Stefan

[-- Attachment #2: Type: application/pgp-signature, Size: 473 bytes --]