[dm-devel] dm-snap deadlock in pending_complete()

[Mikulas Patocka]

On Wed, 12 Aug 2015, NeilBrown wrote:

> On Tue, 11 Aug 2015 05:14:33 -0400 (EDT) Mikulas Patocka
> <mpatocka at redhat.com> wrote:
> 
> > Hi
> > 
> > On Mon, 10 Aug 2015, NeilBrown wrote:
> > 
> > > 
> > > Hi Mikulas,
> > >  I have a customer hitting the deadlock you described over a year ago
> > >  in:
> > > 
> > > Subject: [dm-devel] [PATCH] block: flush queued bios when the process
> > >          blocks
> > 
> > Ask block layer maintainers to accept that patch.
> 
> Unfortunately I don't really like the patch ... or the bioset rescue
> workqueues that it is based on.   Sorry.
> 
> So I might keep looking for a more localised approach....

The problem here is that other dm targets may deadlock in a similar way 
too - for example, dm-thin could deadlock on pmd->pool_lock.

The cause of the bug is bio queuing on current->bio_list. There is an 
assumption that if a dm target submits a bio to a lower-level target, the 
bio finishes in finite time. Queuing on current->bio_list breaks the 
assumption, bios can be held indefinitelly on current->bio_list.

The patch that flushes current->bio_list is the correct way to fix it - it 
makes sure that a bio can't be held indefinitely.

Another way to fix it would be to abandon current->bio_list --- but then, 
there would be problem with stack overflow on deeply nested targets.

Mikulas