Thanks for the info.
On Mon, Jul 18, 2005 at 03:45:52PM +0100, jim@xxxxxxxxxxxxxxxxxx wrote:
> I got a deadlock on 2.6.10 where (due to some fs corruption somewhere
> -- not the point of this e-mail) xfs_trans_delete_ail called
> xfs_do_force_shutdown holding the AIL_LOCK. Later on,
> xfs_trans_tail_ail was called, which went for AIL_LOCK again...
> The code path in question (though perhaps there are other possible ones)
> looks like:
> xfs_trans_delete_ail holds AIL_LOCK
> -> calls xfs_do_force_shutdown
> -> calls xfs_log_force_umount
> -> calls xlog_state_sync_all
> -> calls xlog_state_release_iclog
> -> calls xlog_assign_tail_lsn
> -> calls xfs_trans_tail_ail
> -> tries to take AIL_LOCK
Yes, we should be dropping the AIL_LOCK before calling xfs_force_shutdown()
(instead of afterwards).
I'll check in the fix shortly.
(I believe others are currently looking into a scenario in which the item to be
deleted is missing from the AIL.)