On 10/16/19 8:49 AM, Jan Kara wrote: > On Wed 16-10-19 08:23:51, Eric Sandeen wrote: >> On 10/16/19 4:42 AM, Jan Kara wrote: >>> On Tue 15-10-19 21:36:08, Eric Sandeen wrote: >>>> On 10/15/19 2:37 AM, Jan Kara wrote: >>>>> On Mon 14-10-19 16:30:24, Eric Sandeen wrote: >>>>>> Anything that walks all inodes on sb->s_inodes list without rescheduling >>>>>> risks softlockups. >>>>>> >>>>>> Previous efforts were made in 2 functions, see: >>>>>> >>>>>> c27d82f fs/drop_caches.c: avoid softlockups in drop_pagecache_sb() >>>>>> ac05fbb inode: don't softlockup when evicting inodes >>>>>> >>>>>> but there hasn't been an audit of all walkers, so do that now. This >>>>>> also consistently moves the cond_resched() calls to the bottom of each >>>>>> loop in cases where it already exists. >>>>>> >>>>>> One loop remains: remove_dquot_ref(), because I'm not quite sure how >>>>>> to deal with that one w/o taking the i_lock. >>>>>> >>>>>> Signed-off-by: Eric Sandeen <sandeen@xxxxxxxxxx> >>>>> >>>>> Thanks Eric. The patch looks good to me. You can add: >>>>> >>>>> Reviewed-by: Jan Kara <jack@xxxxxxx> >>>> >>>> thanks >>>> >>>>> BTW, I suppose you need to add Al to pickup the patch? >>>> >>>> Yeah (cc'd now) >>>> >>>> But it was just pointed out to me that if/when the majority of inodes >>>> at umount time have i_count == 0, we'll never hit the resched in >>>> fsnotify_unmount_inodes() and may still have an issue ... >>> >>> Yeah, that's a good point. So that loop will need some further tweaking >>> (like doing iget-iput dance in need_resched() case like in some other >>> places). >> >> Well, it's already got an iget/iput for anything with i_count > 0. But >> as the comment says (and I think it's right...) doing an iget/iput >> on i_count == 0 inodes at this point would be without SB_ACTIVE and the final >> iput here would actually start evicting inodes in /this/ loop, right? > > Yes, it would but since this is just before calling evict_inodes(), I have > currently hard time remembering why evicting inodes like that would be an > issue. Probably just weird to effectively evict all inodes prior to evict_inodes() ;) >> I think we could (ab)use the lru list to construct a "dispose" list for >> fsnotify processing as was done in evict_inodes... [narrator: Eric's idea here is dumb and it won't work] >> or maybe the two should be merged, and fsnotify watches could be handled >> directly in evict_inodes. But that doesn't feel quite right. > > Merging the two would be possible (and faster!) as well but I agree it > feels a bit dirty :) It's starting to look like maybe the only option... I'll see if Al is willing to merge this patch as is for the simple "schedule the big loops" and see about a 2nd patch on top to do more surgery for this case. Thanks, -Eric > Honza >
