On Tue, Oct 13, 2020 at 03:39:03PM +0200, Pavel Reichl wrote: > > > On 10/13/20 1:04 PM, Brian Foster wrote: > > On Mon, Oct 12, 2020 at 10:44:38PM +0200, Pavel Reichl wrote: > >> > >> > >> On 10/12/20 6:04 PM, Brian Foster wrote: > >>> ... > >>>> @@ -2863,8 +2875,20 @@ xfs_btree_split( > >>>> args.done = &done; > >>>> args.kswapd = current_is_kswapd(); > >>>> INIT_WORK_ONSTACK(&args.work, xfs_btree_split_worker); > >>>> + /* > >>>> + * Update lockdep's ownership information to reflect that we > >>>> + * will be transferring the ilock from this thread to the > >>>> + * worker. > >>>> + */ > >>> > >>> Can we update this comment to explain why we need to do this? E.g., I'm > >>> assuming there's a lockdep splat somewhere down in the split worker > >>> without it, but it's not immediately clear where and so it might not be > >>> obvious if we're ever able to remove this. > >> > >> Hi, would something like this work for you? > >> > >> /* > >> + * Update lockdep's ownership information to reflect that we > >> + * will be transferring the ilock from this thread to the > >> + * worker (xfs_btree_split_worker() run via queue_work()). > >> + * If the ownership transfer would not happen lockdep would > >> + * assert in the worker thread because the ilock would be owned > >> + * by the original thread. > >> + */ > >> > > > > That doesn't really answer the question. Do you have a record of the > > lockdep error message that occurs without this state transfer, by > > chance? > > > > Brian > > Hi, please see this mail from Darrick - he hit the issue first - http://mail.spinics.net/lists/linux-xfs/msg38967.html > Ah, I see.. thanks. I was thinking there was some kind of lock imbalance warning or something, but in reality I think something like the following is sufficient: "Update lockdep's ownership information to reflect transfer of the ilock from the current task to the worker. Otherwise assertions that the lock is held (such as when logging the inode) might fail due to incorrect task owner state." Brian > > > >> > > >
