Re: [PATCH -next 0/7] fs/buffer: split pagecache lookups into atomic or blocking

[Jan Kara <jack@suse.cz>]

On Wed 16-04-25 12:27:57, Luis Chamberlain wrote:
> On Tue, Apr 15, 2025 at 04:16:28PM -0700, Davidlohr Bueso wrote:
> > Hello,
> > 
> > This is a respin of the series[0] to address the sleep in atomic scenarios for
> > noref migration with large folios, introduced in:
> > 
> >       3c20917120ce61 ("block/bdev: enable large folio support for large logical block sizes")
> > 
> > The main difference is that it removes the first patch and moves the fix (reducing
> > the i_private_lock critical region in the migration path) to the final patch, which
> > also introduces the new BH_Migrate flag. It also simplifies the locking scheme in
> > patch 1 to avoid folio trylocking in the atomic lookup cases. So essentially blocking
> > users will take the folio lock and hence wait for migration, and otherwise nonblocking
> > callers will bail the lookup if a noref migration is on-going. Blocking callers
> > will also benefit from potential performance gains by reducing contention on the
> > spinlock for bdev mappings.
> > 
> > It is noteworthy that this series is probably too big for Linus' tree, so there are
> > two options:
> > 
> >  1. Revert 3c20917120ce61, add this series + 3c20917120ce61 for next. Or,
> 
> Reverting due to a fix series is odd, I'd advocate this series as a set
> of fixes to Linus' tree because clearly folio migration was not complete
> for buffer_migrate_folio_norefs() and this is part of the loose bits to help
> it for large folios. This issue was just hard to reproduce. The enabler
> of large folios on the block device cache is actually commit
> 47dd67532303 ("block/bdev: lift block size restrictions to 64k") which
> goes later after 3c20917120ce61.

I fully agree reverting anything upstream when there's fix series available
is just pointless.
 
> Jan Kara, since you've already added your Reviewed-by for all patches
> do you have any preference how this trickles to Linus?

I think pushing it normally through VFS tree is fine.

> >  2. Cherry pick patch 7 as a fix for Linus' tree, and leave the rest for next.
> >     But that could break lookup callers that have been deemed unfit to bail.
> > 
> > Patch 1: carves a path for callers that can block to take the folio lock.
> > Patch 2: adds sleeping flavors to pagecache lookups, no users.
> > Patches 3-6: converts to the new call, where possible.
> > Patch 7: does the actual sleep in atomic fix.
> > 
> > Thanks!
> 
> kdevops has tested this patch series and compared it to the baseline [0]
> and has found no regressions on ext4.
> 
> Tested-by: kdevops@lists.linux.dev

Cool, thanks for testing.

								Honza
-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR