->quota_sync

->quota_sync

[Christoph Hellwig]

Hi Jan,

is there any reason the ->quota_sync operations (which always ends up in
dquot_quota_sync, or gfs2_quota_sync) is called before writing back the
inodes?  Given that writeback can perform allocations in filesystem not
using ->page_mkwrite or at least cause delalloc conversions that seems
like the wrong place to me.  Even more so fixing the placement means
we could just call dquot_quota_sync from ->sync_fs, similar how XFS
does it's quota writeout, and thus avoiding the duplicate call to
->sync_fs from inside dquot_quota_sync, as well as getting rid of the
abuse of the quotactl ops from VFS code.

Re: ->quota_sync

[Jan Kara]

  Hi,

On Tue 19-07-11 12:14:06, Christoph Hellwig wrote:
> is there any reason the ->quota_sync operations (which always ends up in
> dquot_quota_sync, or gfs2_quota_sync) is called before writing back the
> inodes?
  Hmm, historically, there was a reason because writing of quota entries
might have needed to allocate new blocks to quota file.  But these days I
don't see any reason since we allocate necessary blocks when quota entry is
first looked up and directly modify block device pages when it is modified.

> Given that writeback can perform allocations in filesystem not
> using ->page_mkwrite or at least cause delalloc conversions that seems
> like the wrong place to me.
  Yes, doing quota writeout after inode writeout would be a more logical
place.

> Even more so fixing the placement means
> we could just call dquot_quota_sync from ->sync_fs, similar how XFS
> does it's quota writeout, and thus avoiding the duplicate call to
> ->sync_fs from inside dquot_quota_sync, as well as getting rid of the
> abuse of the quotactl ops from VFS code.
  True. I'll add these changes to my patches cleaning up sync code.

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