Re: [MD] Crash with 4.12+ kernel and high disk load -- bisected to 4ad23a976413: MD: use per-cpu counter for writes_pending

[Dominik Brodowski <linux@dominikbrodowski.net>]

On Tue, Aug 08, 2017 at 09:36:14AM +0200, Dominik Brodowski wrote:
> On Tue, Aug 08, 2017 at 05:01:28PM +1000, NeilBrown wrote:
> > On Mon, Aug 07 2017, Dominik Brodowski wrote:
> > 
> > > Neil, Shaohua,
> > >
> > > following up on David R's bug message: I have observed something similar
> > > on v4.12.[345] and v4.13-rc4, but not on v4.11. This is a RAID1 (on bare
> > > metal partitions, /dev/sdaX and /dev/sdbY linked together). In case it
> > > matters: Further upwards are cryptsetup, a DM volume group, then logical
> > > volumes, and then filesystems (ext4, but also happened with xfs).
> > >
> > > In a tedious bisect (the bug wasn't as quickly reproducible as I would like,
> > > but happened when I repeatedly created large lvs and filled them with some
> > > content, while compiling kernels in parallel), I was able to track this
> > > down to:
> > >
> > >
> > > commit 4ad23a976413aa57fe5ba7a25953dc35ccca5b71
> > > Author: NeilBrown <neilb@suse.com>
> > > Date:   Wed Mar 15 14:05:14 2017 +1100
> > >
> > >     MD: use per-cpu counter for writes_pending
> > >     
> > >     The 'writes_pending' counter is used to determine when the
> > >     array is stable so that it can be marked in the superblock
> > >     as "Clean".  Consequently it needs to be updated frequently
> > >     but only checked for zero occasionally.  Recent changes to
> > >     raid5 cause the count to be updated even more often - once
> > >     per 4K rather than once per bio.  This provided
> > >     justification for making the updates more efficient.
> > >
> > >     ...
> > 
> > Thanks for the report... and for bisecting and for re-sending...
> > 
> > I believe I have found the problem, and have sent a patch separately.
> > 
> > If mddev->safemode == 1 and mddev->in_sync != 0, md_check_recovery()
> > causes the thread that calls it to spin.
> > Prior to the patch you found, that couldn't happen.  Now it can,
> > so it needs to be handled more carefully.
> > 
> > While I was examining the code, I found another bug - so that is a win!
> > 
> > Thanks,
> > NeilBrown
> 
> 
> Nice catch. Thanks! Will give it (both patches at once) a try on the test
> system immediately.

More than 2 hours of stress-testing shows no issues any more. Very nice.
Thanks!

	Dominik