Messages in this thread | | | Date | Sun, 17 Oct 2010 18:05:19 +1100 | From | Nick Piggin <> | Subject | Re: [PATCH 17/19] fs: Reduce inode I_FREEING and factor inode disposal |
| |
On Sun, Oct 17, 2010 at 05:52:45PM +1100, Dave Chinner wrote: > On Sun, Oct 17, 2010 at 04:13:10PM +1100, Nick Piggin wrote: > > On Sun, Oct 17, 2010 at 03:35:14PM +1100, Nick Piggin wrote: > > > On Sun, Oct 17, 2010 at 03:13:13PM +1100, Dave Chinner wrote: > > > > On Sun, Oct 17, 2010 at 01:49:23PM +1100, Nick Piggin wrote: > > > > > On Sat, Oct 16, 2010 at 09:30:47PM -0400, Christoph Hellwig wrote: > > > > > > > * inode->i_lock is *always* the innermost lock. > > > > > > > * > > > > > > > + * inode->i_lock is *always* the innermost lock. > > > > > > > + * > > > > > > > > > > > > No need to repeat, we got it.. > > > > > > > > > > Except that I didn't see where you fixed all the places where it is > > > > > *not* the innermost lock. Like for example places that take dcache_lock > > > > > inside i_lock. > > > > > > > > I can't find any code outside of ceph where the dcache_lock is used > > > > within 200 lines of code of the inode->i_lock. The ceph code is not > > > > nesting them, though. > > > > > > You mustn't have looked very hard? From ceph: > > > > > > spin_unlock(&dcache_lock); > > > spin_unlock(&inode->i_lock); > > > > > > (and yes, acquisition side does go in i_lock->dcache_lock order) > > Sorry, easy to miss with a quick grep when the locks are taken in > different functions.
Easy to see they're nested when they're dropped in adjacent lines. That should give you a clue to go and check their lock order.
> Anyway, this one looks difficult to fix without knowing something > about Ceph and wtf it is doing there. It's one to punt to the > maintainer to solve as it's not critical to this patch set.
I thought the raison detre for your starting to write your own vfs scale branch was because you objected to i_lock not being an "innermost" lock (not that it was before my patch).
So I don't get it. If your patch mandates it to be an innermost lock, then you absolutely do need to fix the filesystems before changing the lock order.
> > A really quick grep reveals cifs is using GlobalSMBSeslock inside i_lock > > too. > > I'm having a grep-fail day. Where is that one?
Uh, inside one of the 6 places that i_lock is taken in cifs. The only non-trivial one, not surprisingly.
| |