On Tue, 3 Sep 2013 15:02:28 +0800 Shaohua Li <shli@xxxxxxxxxx> wrote: > On Tue, Sep 03, 2013 at 04:08:58PM +1000, NeilBrown wrote: > > On Wed, 28 Aug 2013 14:39:53 +0800 Shaohua Li <shli@xxxxxxxxxx> wrote: > > > > > On Wed, Aug 28, 2013 at 02:32:52PM +1000, NeilBrown wrote: > > > > On Tue, 27 Aug 2013 16:53:30 +0800 Shaohua Li <shli@xxxxxxxxxx> wrote: > > > > > > > > > On Tue, Aug 27, 2013 at 01:17:52PM +1000, NeilBrown wrote: > > > > > > > > > > > > > > > Then get_active_stripe wouldn't need to worry about device_lock at all and > > > > > > would only need to get the hash lock for the particular sector. That should > > > > > > make it a lot simpler. > > > > > > > > > > did you mean get_active_stripe() doesn't need device_lock for any code path? > > > > > How could it be safe? device_lock still protects something like handle_list, > > > > > delayed_list, which release_stripe() will use while a get_active_stripe can run > > > > > concurrently. > > > > > > > > Yes you will still need device_lock to protect list_del_init(&sh->lru), > > > > as well as the hash lock. > > > > Do you need device_lock anywhere else in there? > > > > > > That's what I mean. So I need get both device_lock and hash_lock. To not > > > deadlock, I need release hash_lock and relock device_lock/hash_lock. Since I > > > release lock, I need recheck if I can find the stripe in hash again. So the > > > seqcount locking doesn't simplify things here. I thought the seqlock only fixes > > > one race. Did I miss anything? > > > > Can you order the locks so that you take the hash_lock first, then the > > device_lock? That would be a lot simpler. > > Looks impossible. For example, in handle_active_stripes() we release several > stripes, we can't take hash_lock first. "impossible" just takes a little longer :-) do_release_stripe gets called with only device_lock held. It gets passed an (initially) empty list_head too. If it wants to add the stripe to an inactive list it puts it on the given list_head instead. release_stripe(), after calling do_release_stripe() calls some function to grab the appropriate hash_lock for each stripe in the list_head and add it to that inactive list. release_stripe_list() might collect some stripes from from __release_stripe that need to go on an inactive list. It arranges for them to be put on the right list, with the right lock, next time device_lock is dropped. That might be in handle_active_stripes() activate_bit_delay might similarly collect stripes, which are handled the same way as those collected by release_stripe_list. etc. i.e. the hash_locks protect the various inactive lists. device_lock protects all the others. If we need to add something to an inactive list while holding device_lock we delay until device_lock can be dropped. > > > > I saw your tree only has seqcount_write lock in one place, but there are still > > > other places which changing quiesce, degraded. I thought we still need lock all > > > locks like what I did. > > > > Can you be specific? I thought I had convinced my self that I covered > > everything that was necessary, but I might have missed something. > > For example, raid5_quiesce() will change quiesce which get_active_stripe() will > use. So my point is get_active_stripe() still need get device_lock. Appears you > agree get_active_stripe() need get device_lock. Maybe I confused your > comments. raid5_quiesce might reasonably take all of the hash_locks and then the device_lock - it is expected to be a rare event and can afford to be heavy handed. get_active_stripe() should only take device_lock for list_del_init(&sh->lru). What else have I missed? Thanks, NeilBrown
Attachment:
signature.asc
Description: PGP signature
