Re: [syzbot ci] Re: ext4: deferred iput framework for EA inodes

Hi Yun,

On Tue 07-07-26 19:00:47, Zhou, Yun wrote:
> On 7/3/26 22:02, Jan Kara wrote:
> > 
> > On Thu 02-07-26 15:57:01, syzbot ci wrote:
> > > syzbot ci has tested the following series
> > > 
> > > [v13] ext4: deferred iput framework for EA inodes
> > > https://lore.kernel.org/all/[email protected]
> > > * [PATCH v13 1/4] fs: add iput_if_not_last() helper
> > > * [PATCH v13 2/4] ext4: introduce ext4_put_ea_inode() for safe deferred iput
> > > * [PATCH v13 3/4] ext4: convert all EA inode iput() calls to ext4_put_ea_inode()
> > > * [PATCH v13 4/4] ext4: remove ea_inode_array mechanism in favor of ext4_put_ea_inode()
> > > 
> > > and found the following issue:
> > > WARNING in ext4_journal_check_start
> > 
> > Good catch. We need to flush the pending work with iputs also in
> > ext4_sync_fs() which should also take care of this race with remount
> > read-only (and the same issue with filesystem freezing etc.).
> > 
> 
> Sorry for the late reply—I missed it because I receive too many emails from
> syzbot. I've now reconfigured my filtering rules.
> 
> Regarding the issue reported by syzbot CI, would it be better to add
> flush_delayed_work in __ext4_remount before sb->s_flags |= SB_RDONLY? That
> would make the race window smaller.

Well, if there is remount read-only happening, the operations modifying
filesystem are blocked or made to fail (see mnt_get_write_access() /
sb_prepare_remount_readonly()). So it doesn't really matter where in
ext4_reconfigure() / __ext4_remount() we flush the work - no new items that
are deleting any xattr inodes should be added.

And placing the flushing into ext4_sync_fs() handles multiple similar
problems - sync(2) should make sure all pending deletions are finished,
remount needs to do this (and it calls sync_filesystem() which will end up
calling ext4_sync_fs()), fsfreeze needs to do this (it calls
sync_filesystem() as well).
 
> Also, should I resend the v14 series, or submit a separate patch
> specifically for this fix?

I think you can resend with the fixup. Ted didn't merge the series yet.

								Honza
-- 
Jan Kara <[email protected]>
SUSE Labs, CR