Re: [PATCH] mm: always respect QUEUE_FLAG_STABLE_WRITES on the block device

[Ilya Dryomov]

On Thu, May 4, 2023 at 4:16 PM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
>
> On Thu, May 04, 2023 at 03:55:15PM +0200, Christoph Hellwig wrote:
> > On Thu, May 04, 2023 at 12:56:24PM +0200, Ilya Dryomov wrote:
> > > Commit 1cb039f3dc16 ("bdi: replace BDI_CAP_STABLE_WRITES with a queue
> > > and a sb flag") introduced a regression for the raw block device use
> > > case.  Capturing QUEUE_FLAG_STABLE_WRITES flag in set_bdev_super() has
> > > the effect of respecting it only when there is a filesystem mounted on
> > > top of the block device.  If a filesystem is not mounted, block devices
> > > that do integrity checking return sporadic checksum errors.
> >
> > With "If a file system is not mounted" you want to say "when accessing
> > a block device directly" here, right?  The two are not exclusive..
> >
> > > Additionally, this commit made the corresponding sysfs knob writeable
> > > for debugging purposes.  However, because QUEUE_FLAG_STABLE_WRITES flag
> > > is captured when the filesystem is mounted and isn't consulted after
> > > that anywhere outside of swap code, changing it doesn't take immediate
> > > effect even though dumping the knob shows the new value.  With no way
> > > to dump SB_I_STABLE_WRITES flag, this is needlessly confusing.
> >
> > But very much intentional.  s_bdev often is not the only device
> > in a file system, and we should never reference if from core
> > helpers.
> >
> > So I think we should go with something like this:
> >
> > diff --git a/mm/page-writeback.c b/mm/page-writeback.c
> > index db794399900734..aa36cc2a4530c1 100644
> > --- a/mm/page-writeback.c
> > +++ b/mm/page-writeback.c
> > @@ -3129,7 +3129,11 @@ EXPORT_SYMBOL_GPL(folio_wait_writeback_killable);
> >   */
> >  void folio_wait_stable(struct folio *folio)
> >  {
> > -     if (folio_inode(folio)->i_sb->s_iflags & SB_I_STABLE_WRITES)
> > +     struct inode *inode = folio_inode(folio);
> > +     struct super_block *sb = inode->i_sb;
> > +
> > +     if ((sb->s_iflags & SB_I_STABLE_WRITES) ||
> > +         (sb_is_blkdev_sb(sb) && bdev_stable_writes(I_BDEV(inode))))
> >               folio_wait_writeback(folio);
> >  }
> >  EXPORT_SYMBOL_GPL(folio_wait_stable);
>
> I hate both of these patches ;-)  What we should do is add
> AS_STABLE_WRITES, have the appropriate places call
> mapping_set_stable_writes() and then folio_wait_stable() becomes
>
>         if (mapping_test_stable_writes(folio->mapping))
>                 folio_wait_writeback(folio);
>
> and we remove all the dereferences (mapping->host->i_sb->s_iflags, plus
> whatever else is going on there)

Hi Matthew,

We would still need something resembling Christoph's suggestion for
5.10 and 5.15 (at least).  Since this fixes a regression, would you
support merging the "ugly" version to facilitate backports or would
you rather see the AS/mapping-based refactor first?

Thanks,

                Ilya