On Mon, 4 Jan 2021, Mike Snitzer wrote:
> On Sun, Dec 20 2020 at 8:02am -0500,
> Lukas Straub <lukasstraub2@xxxxxx> 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@xxxxxx>
> > ---
> > 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.
Mikulas
--
dm-devel mailing list
dm-devel@xxxxxxxxxx
https://www.redhat.com/mailman/listinfo/dm-devel
