Re: Deadlocks due to per-process plugging

[Mike Galbraith]

On Mon, 2012-07-16 at 11:59 +0200, Thomas Gleixner wrote: 
> On Mon, 16 Jul 2012, Mike Galbraith wrote:
> > On Mon, 2012-07-16 at 10:59 +0200, Thomas Gleixner wrote: 
> > > On Mon, 16 Jul 2012, Mike Galbraith wrote:
> > > > On Sun, 2012-07-15 at 11:14 +0200, Mike Galbraith wrote: 
> > > > > On Sun, 2012-07-15 at 10:59 +0200, Thomas Gleixner wrote: 
> > > > 
> > > > > > Can you figure out on which lock the stuck thread which did not unplug
> > > > > > due to tsk_is_pi_blocked was blocked?
> > > > > 
> > > > > I'll take a peek.
> > > > 
> > > > Sorry for late reply, took a half day away from box.  Jan had already
> > > > done the full ext3 IO deadlock analysis:
> > > > 
> > > > Again kjournald is waiting for buffer IO on block 4367635 (sector
> > > > 78364838) to finish.  Now it is dbench thread 0xffff88026f330e70 which
> > > > has submitted this buffer for IO and is still holding this buffer behind
> > > > its plug (request for sector 78364822..78364846). The dbench thread is
> > > > waiting on j_checkpoint mutex (apparently it has successfully got the
> > > > mutex in the past, checkpointed some buffers, released the mutex and
> > > > hung when trying to acquire it again in the next loop of
> > > > __log_wait_for_space()). 
> > > 
> > > And what's holding j_checkpoint mutex and not making progress?
> > 
> > Waiting for wakeup from kjournald.
> 
> So kicking the io in __log_wait_for_space()
> 
>                 spin_unlock(&journal->j_state_lock);
> 
> 		--> HERE
> 
>                 mutex_lock(&journal->j_checkpoint_mutex);
> 
> should fix the issue, right ?

Should for this scenario.  Jan said he could think of ways to likely
make the same happen in xfs.

-Mike

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/