On Thu, Apr 13, 2017 at 07:54:20AM +0800, Fam Zheng wrote:
> On Wed, 04/12 18:22, Jeff Cody wrote:
> > On Wed, Apr 12, 2017 at 05:38:17PM -0400, John Snow wrote:
> > >
> > >
> > > On 04/12/2017 04:46 PM, Jeff Cody wrote:
> > > >
> > > > This occurs on v2.9.0-rc4, but not on v2.8.0.
> > > >
> > > > When running QEMU with an iothread, and then performing a block-mirror,
> > > > if
> > > > we do a system-reset after the BLOCK_JOB_READY event has emitted, qemu
> > > > becomes deadlocked.
> > > >
> > > > The block job is not paused, nor cancelled, so we are stuck in the while
> > > > loop in block_job_detach_aio_context:
> > > >
> > > > static void block_job_detach_aio_context(void *opaque)
> > > > {
> > > > BlockJob *job = opaque;
> > > >
> > > > /* In case the job terminates during aio_poll()... */
> > > > block_job_ref(job);
> > > >
> > > > block_job_pause(job);
> > > >
> > > > while (!job->paused && !job->completed) {
> > > > block_job_drain(job);
> > > > }
> > > >
> > >
> > > Looks like when block_job_drain calls block_job_enter from this context
> > > (the main thread, since we're trying to do a system_reset...), we cannot
> > > enter the coroutine because it's the wrong context, so we schedule an
> > > entry instead with
> > >
> > > aio_co_schedule(ctx, co);
> > >
> > > But that entry never happens, so the job never wakes up and we never
> > > make enough progress in the coroutine to gracefully pause, so we wedge
> > > here.
> > >
> >
> >
> > John Snow and I debugged this some over IRC. Here is a summary:
> >
> > Simply put, with iothreads the aio context is different. When
> > block_job_detach_aio_context() is called from the main thread via the system
> > reset (from main_loop_should_exit()), it calls block_job_drain() in a while
> > loop, with job->busy and job->completed as exit conditions.
> >
> > block_job_drain() attempts to enter the coroutine (thus allowing job->busy
> > or job->completed to change). However, since the aio context is different
> > with iothreads, we schedule the coroutine entry rather than directly
> > entering it.
> >
> > This means the job coroutine is never going to be re-entered, because we are
> > waiting for it to complete in a while loop from the main thread, which is
> > blocking the qemu timers which would run the scheduled coroutine... hence,
> > we become stuck.
>
> John and I confirmed that this can be fixed by this pending patch:
>
> [PATCH for-2.9 4/5] block: Drain BH in bdrv_drained_begin
>
> https://lists.gnu.org/archive/html/qemu-devel/2017-04/msg01018.html
>
> It didn't make it into 2.9-rc4 because of limited time. :(
>
> Looks like there is no -rc5, we'll have to document this as a known issue.
> Users should "block-job-complete/cancel" as soon as possible to avoid such a
> hang.
>
I'd argue for including a fix for 2.9, since this is both a regression, and
a hard lock without possible recovery short of restarting the QEMU process.
-Jeff
