On Fri, Aug 31, 2007 at 04:33:51PM +0200, Peter Zijlstra wrote:
> On Fri, 2007-08-31 at 23:50 +1000, David Chinner wrote:
> > On Fri, Aug 31, 2007 at 08:39:49AM +0200, Peter Zijlstra wrote:
> > > On Thu, 2007-08-30 at 23:43 -0500, Eric Sandeen wrote:
> > > > The xfs filesystem can exceed the current lockdep
> > > > MAX_LOCK_DEPTH, because when deleting an entire cluster of inodes,
> > > > they all get locked in xfs_ifree_cluster(). The normal cluster
> > > > size is 8192 bytes, and with the default (and minimum) inode size
> > > > of 256 bytes, that's up to 32 inodes that get locked. Throw in a
> > > > few other locks along the way, and 40 seems enough to get me through
> > > > all the tests in the xfsqa suite on 4k blocks. (block sizes
> > > > above 8K will still exceed this though, I think)
> > >
> > > As 40 will still not be enough for people with larger block sizes, this
> > > does not seems like a solid solution. Could XFS possibly batch in
> > > smaller (fixed sized) chunks, or does that have significant down sides?
> > The problem is not filesystem block size, it's the xfs inode cluster buffer
> > size / the size of the inodes that determines the lock depth. the common
> > case
> > is 8k/256 = 32 inodes in a buffer, and they all get looked during inode
> > cluster writeback.
> > This inode writeback clustering is one of the reasons XFS doesn't suffer
> > from
> > atime issues as much as other filesystems - it doesn't need to do as much
> > I/O
> > to write back dirty inodes to disk.
> > IOWs, we are not going to make the inode clusters smallers - if anything
> > they
> > are going to get *larger* in future so we do less I/O during inode writeback
> > than we do now.....
> Since they are all trylocks it seems to suggest there is no hard _need_
> to lock a whole inode cluster at once, and could iterate through it with
> less inodes locked.
That's kind of misleading. They are trylocks to prevent deadlocks with other
threads that may be cleaning up a given inode or the inodes may already be
locked for writeback and so locking them a second time would deadlock.
Basically we have to run the lower loops with the inodes locked, and the
trylocks simply avoid the need for us to explicity check if items are locked
to avoid deadlocks.
> Granted I have absolutely no understanding of what I'm talking about :-)
> Trouble is, we'd like to have a sane upper bound on the amount of held
> locks at any one time, obviously this is just wanting, because a lot of
> lock chains also depend on the number of online cpus...
Sure - this is an obvious case where it is valid to take >30 locks at
once in a single thread. In fact, worst case here we are taking twice this
number of locks - we actually take 2 per inode (ilock and flock) so a
full 32 inode cluster free would take >60 locks in the middle of this
function and we should be busting this depth couter limit all the
time. Do semaphores (the flush locks) contribute to the lock depth
SGI Australian Software Group