On Wed, Apr 27, 2011 at 12:46:51AM -0700, Christian Kujau wrote: > On Wed, 27 Apr 2011 at 12:26, Dave Chinner wrote: > > What this shows is that VFS inode cache memory usage increases until > > about the 550 sample mark before the VM starts to reclaim it with > > extreme prejudice. At that point, I'd expect the XFS inode cache to > > then shrink, and it doesn't. I've got no idea why the either the > > Do you remember any XFS changes past 2.6.38 that could be related to > something like this? There's plenty of changes that coul dbe the cause - we've changed the inode reclaim to run in the background out of a workqueue as well as via the shrinker, so it could even be workqueue starvation causing the the problem... hmmmm. Speaking of which - have you changed any of the XFS tunables in /proc/sys/fs/xfs/ on your machine (specifically xfssyncd_centisecs)? > Bisecting is pretty slow on this machine. Could I somehow try to run > 2.6.39-rc4 but w/o the XFS changes merged after 2.6.38? (Does someone know > how to do this via git?) Not easy because there are tree-wide changes that need to be preserved (e.g. block layer plugging changes) while others around it would need to be reverted.... > > Can you check if there are any blocked tasks nearing OOM (i.e. "echo > > w > /proc/sysrq-trigger") so we can see if XFS inode reclaim is > > stuck somewhere? > > Will do, tomorrow. > > Should I open a regression bug, so we don't loose track of this thing? Whatever you want. Cheers, Dave. -- Dave Chinner david@xxxxxxxxxxxxx _______________________________________________ xfs mailing list xfs@xxxxxxxxxxx http://oss.sgi.com/mailman/listinfo/xfs
