[PATCH] xfs: shutdown xfs_sync_worker before the log

[Christoph Hellwig]

On Wed, Jun 06, 2012 at 02:26:47PM +1000, Dave Chinner wrote:
> I think starting by renaming the xfs-syncd workqueue to the
> xfs_mount_wq because there's nothing "sync" related about it's
> functionality any more.

Is there any good reason to keep queueing different work items on the
same queue now that workqueues are incredibly cheap? 

> I'll then kill xfs_syncd_init/stop functions and open code the
> intialisation of the work structures and start them in the
> appropriate places for their functionality. e.g. reclaim work is
> demand started and stops when there's nothing more to do or at
> unmount, the flush work is demand started and we need to complete
> them all at unmount, and the xfssync work is really now "periodic
> log work" so should be started once we complete log recovery
> successfullly and stopped before we tear down the log....

Sounds good.  You probably also want to kill off xfs_sync_fsdata as
it doesn't make any sense with our current sync / ail pushing code
before that.

> Then I can move the xfs_sync_worker to xfs_log.c and rename it.

Before that you probably want to kill the xfs_ail_push_all in it in
favour or xfsaild waking up periodically by itself if there is anything
in the AIL.

> If I then convert xfs_flush_worker to use the generic writeback code
> (writeback_inodes_sb_if_idle) the xfs_sync_data() can go away.

Good plan.  It'll still need the trylock changes for it that don't
really seem to make forward progress on fsdevel.

> That
> means the only user of xfs_inode_ag_iterator is the quotaoff code
> (xfs_qm_dqrele_all_inodes), so it could be moved elsewhere (like
> xfs_inode.c).

fair enough.

> Then xfs_quiesce_data() can be moved to xfs-super.c where it can sit
> alongside the two functions that call it, and the same can be done
> for xfs_quiesce_attr().

xfs_fs_remount shouldn't really need to call it, as do_remount_sb
calls it just before entering ->remount_fs.  Although looking at it
closter do_remount_sb probably needs to move the sync_filesystem
call until after the check for r/o files and preventing new writes.

Independent of that I think xfs_quiesce_data should be merged into
xfs_fs_sync_fs - until do_remount_sb is fixed xfs_fs_remount should
simply call xfs_fs_sync_fs.

This also is a good opportunity to redo the maze of comments describing
the freeze process, which is rather outdated and a bit confusing now.

> That will leave only inode cache reclaim functions in xfs_sync.c.
> These are closely aligned to the inode allocation, freeing and cache
> lookup functions in xfs_iget.c, so I'm thinking of merging the two
> into a single file named xfs_inode_cache.c so both xfs_sync.c and
> xfs_iget.c go away.

Sounds good, although I'd call the file xfs_icache.c - that seems to
fit better with the general theme of naming schemes in XFS.