Re: [PATCH v2 1/8] migrate: fix skipping metadata buffer heads on migration

[Luis Chamberlain <mcgrof@kernel.org>]

On Tue, Apr 15, 2025 at 06:23:54PM +0200, Jan Kara wrote:
> On Tue 15-04-25 08:47:51, Luis Chamberlain wrote:
> > On Tue, Apr 15, 2025 at 11:05:38AM +0200, Christian Brauner wrote:
> > > On Mon, Apr 14, 2025 at 03:19:33PM -0700, Luis Chamberlain wrote:
> > > > On Mon, Apr 14, 2025 at 02:09:46PM -0700, Luis Chamberlain wrote:
> > > > > On Thu, Apr 10, 2025 at 02:05:38PM +0200, Jan Kara wrote:
> > > > > > > @@ -859,12 +862,12 @@ static int __buffer_migrate_folio(struct address_space *mapping,
> > > > > > >  			}
> > > > > > >  			bh = bh->b_this_page;
> > > > > > >  		} while (bh != head);
> > > > > > > +		spin_unlock(&mapping->i_private_lock);
> > > > > > 
> > > > > > No, you've just broken all simple filesystems (like ext2) with this patch.
> > > > > > You can reduce the spinlock critical section only after providing
> > > > > > alternative way to protect them from migration. So this should probably
> > > > > > happen at the end of the series.
> > > > > 
> > > > > So you're OK with this spin lock move with the other series in place?
> > > > > 
> > > > > And so we punt the hard-to-reproduce corruption issue as future work
> > > > > to do? Becuase the other alternative for now is to just disable
> > > > > migration for jbd2:
> > > > > 
> > > > > diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
> > > > > index 1dc09ed5d403..ef1c3ef68877 100644
> > > > > --- a/fs/ext4/inode.c
> > > > > +++ b/fs/ext4/inode.c
> > > > > @@ -3631,7 +3631,6 @@ static const struct address_space_operations ext4_journalled_aops = {
> > > > >  	.bmap			= ext4_bmap,
> > > > >  	.invalidate_folio	= ext4_journalled_invalidate_folio,
> > > > >  	.release_folio		= ext4_release_folio,
> > > > > -	.migrate_folio		= buffer_migrate_folio_norefs,
> > > > >  	.is_partially_uptodate  = block_is_partially_uptodate,
> > > > >  	.error_remove_folio	= generic_error_remove_folio,
> > > > >  	.swap_activate		= ext4_iomap_swap_activate,
> > > > 
> > > > BTW I ask because.. are your expectations that the next v3 series also
> > > > be a target for Linus tree as part of a fix for this spinlock
> > > > replacement?
> > > 
> > > Since this is fixing potential filesystem corruption I will upstream
> > > whatever we need to do to fix this. Ideally we have a minimal fix to
> > > upstream now and a comprehensive fix and cleanup for v6.16.
> > 
> > Despite our efforts we don't yet have an agreement on how to fix the
> > ext4 corruption, becuase Jan noted the buffer_meta() check in this patch
> > is too broad and would affect other filesystems (I have yet to
> > understand how, but will review).
> > 
> > And so while we have agreement we can remove the spin lock to fix the
> > sleeping while atomic incurred by large folios for buffer heads by this
> > patch series, the removal of the spin lock would happen at the end of
> > this series.
> > 
> > And so the ext4 corruption is an existing issue as-is today, its
> > separate from the spin lock removal goal to fix the sleeping while
> > atomic..
> 
> I agree. Ext4 corruption problems are separate from sleeping in atomic
> issues.

Glad that's clear.

> > However this series might be quite big for an rc2 or rc3 fix for that spin
> > lock removal issue. It should bring in substantial performance benefits
> > though, so it might be worthy to consider. We can re-run tests with the
> > adjustment to remove the spin lock until the last patch in this series.
> > 
> > The alternative is to revert the spin lock addition commit for Linus'
> > tree, ie commit ebdf4de5642fb6 ("mm: migrate: fix reference check race
> > between __find_get_block() and migration") and note that it in fact does
> > not fix the ext4 corruption as we've noted, and in fact causes an issue
> > with sleeping while atomic with support for large folios for buffer
> > heads. If we do that then we  punt this series for the next development
> > window, and it would just not have the spin lock removal on the last
> > patch.
> 
> Well, the commit ebdf4de5642fb6 is 6 years old. At that time there were no
> large folios (in fact there were no folios at all ;)) in the page cache and
> it does work quite well (I didn't see a corruption report from real users
> since then).

It is still a work around.

> So I don't like removing that commit because it makes a
> "reproducible with a heavy stress test" problem become a "reproduced by
> real world workloads" problem.

So how about just patch 2 and 8 in this series, with the spin lock
removal happening on the last patch for Linus tree?

  Luis