[dm-devel] dm-integrity: Fix flush with external metadata device

[Mike Snitzer]

On Fri, Jan 08 2021 at 11:12am -0500,
Mikulas Patocka <mpatocka at redhat.com> wrote:

> 
> 
> On Mon, 4 Jan 2021, Mike Snitzer wrote:
> 
> > On Sun, Dec 20 2020 at  8:02am -0500,
> > Lukas Straub <lukasstraub2 at web.de> wrote:
> > 
> > > With an external metadata device, flush requests aren't passed down
> > > to the data device.
> > > 
> > > Fix this by issuing flush in the right places: In integrity_commit
> > > when not in journal mode, in do_journal_write after writing the
> > > contents of the journal to the disk and in dm_integrity_postsuspend.
> > > 
> > > Signed-off-by: Lukas Straub <lukasstraub2 at web.de>
> > > ---
> > >  drivers/md/dm-integrity.c | 8 ++++++++
> > >  1 file changed, 8 insertions(+)
> > > 
> > > diff --git a/drivers/md/dm-integrity.c b/drivers/md/dm-integrity.c
> > > index 5a7a1b90e671..a26ed65869f6 100644
> > > --- a/drivers/md/dm-integrity.c
> > > +++ b/drivers/md/dm-integrity.c
> > > @@ -2196,6 +2196,8 @@ static void integrity_commit(struct work_struct *w)
> > >  	if (unlikely(ic->mode != 'J')) {
> > >  		spin_unlock_irq(&ic->endio_wait.lock);
> > >  		dm_integrity_flush_buffers(ic);
> > > +		if (ic->meta_dev)
> > > +			blkdev_issue_flush(ic->dev->bdev, GFP_NOIO);
> > >  		goto release_flush_bios;
> > >  	}
> > >  
> > > @@ -2410,6 +2412,9 @@ static void do_journal_write(struct dm_integrity_c *ic, unsigned write_start,
> > >  	wait_for_completion_io(&comp.comp);
> > >  
> > >  	dm_integrity_flush_buffers(ic);
> > > +	if (ic->meta_dev)
> > > +		blkdev_issue_flush(ic->dev->bdev, GFP_NOIO);
> > > +
> > >  }
> > >  
> > >  static void integrity_writer(struct work_struct *w)
> > > @@ -2949,6 +2954,9 @@ static void dm_integrity_postsuspend(struct dm_target *ti)
> > >  #endif
> > >  	}
> > >  
> > > +	if (ic->meta_dev)
> > > +		blkdev_issue_flush(ic->dev->bdev, GFP_NOIO);
> > > +
> > >  	BUG_ON(!RB_EMPTY_ROOT(&ic->in_progress));
> > >  
> > >  	ic->journal_uptodate = true;
> > > -- 
> > > 2.20.1
> > 
> > 
> > Seems like a pretty bad oversight... but shouldn't you also make sure to
> > flush the data device _before_ the metadata is flushed?
> > 
> > Mike
> 
> I think, ordering is not a problem.
> 
> A disk may flush its cache spontaneously anytime, so it doesn't matter in 
> which order do we flush them. Similarly a dm-bufio buffer may be flushed 
> anytime - if the machine is running out of memory and a dm-bufio shrinker 
> is called.
> 
> I'll send another patch for this - I've created a patch that flushes the 
> metadata device cache and data device cache in parallel, so that 
> performance degradation is reduced.
> 
> My patch also doesn't use GFP_NOIO allocation - which can in theory 
> deadlock if we are swapping on dm-integrity device.

OK, I see your patch, but my concern about ordering was more to do with
crash consistency.  What if metadata is updated but data isn't?

On the surface, your approach of issuing the flushes in parallel seems
to expose us to inconsistency upon recovery from a crash.
If that isn't the case please explain why not.

Thanks,
Mike