On Fri, Oct 21, 2011 at 09:22:40AM -0400, Christoph Hellwig wrote:
> On Thu, Oct 20, 2011 at 03:42:14PM -0700, Simon Kirby wrote:
> > [<ffffffff8126f205>] xfs_reclaim_inode+0x85/0x2b0
> > [<ffffffff8126f5b0>] xfs_reclaim_inodes_ag+0x180/0x2f0
> > [<ffffffff8126f74e>] xfs_reclaim_inodes_nr+0x2e/0x40
> > [<ffffffff8126ccf0>] xfs_fs_free_cached_objects+0x10/0x20
> > [<ffffffff81119a70>] prune_super+0x110/0x1b0
> > [<ffffffff810e4fa5>] shrink_slab+0x1e5/0x2a0
> > [<ffffffff810e5821>] kswapd+0x7c1/0xba0
> > [<ffffffff8107ada6>] kthread+0x96/0xb0
> > [<ffffffff816c0474>] kernel_thread_helper+0x4/0x10
> > [<ffffffffffffffff>] 0xffffffffffffffff
> We're stuck in synchronous inode reclaim.
> > All of the other processes that get stuck have this stack:
> > [<ffffffff81080587>] down+0x47/0x50
> > [<ffffffff8125e816>] xfs_buf_lock+0x66/0xd0
> > [<ffffffff812603ad>] _xfs_buf_find+0x16d/0x270
> > [<ffffffff81260517>] xfs_buf_get+0x67/0x1a0
> > [<ffffffff8126067a>] xfs_buf_read+0x2a/0x120
> > [<ffffffff812b876f>] xfs_trans_read_buf+0x28f/0x3f0
> > [<ffffffff8129e161>] xfs_read_agi+0x71/0x100
> They are waiting for the AGI buffer to become unlocked. The only reason
> it is held locked for longer time is when it is under I/O.
> > By the way, xfs_reclaim_inode+0x85 (133) disassembles as:
> > ...So the next function is wait_for_completion(), which is marked
> > __sched and thus doesn't show up in the trace.
> So we're waiting for the inode to be flushed, aka I/O again.
But I don't seem to see any queued I/O, hmm.
> What is interesting here is that we're always blocking on the AGI
> buffer - which is used during unlinks of inodes, and thus gets hit
> fairly heavily for a workload that does a lot of unlinks.
I don't think we do too many unlinks, but there are quite a bit of
renames over existing files (dovecot-2.0 w/mdbox).
> > When the clog happens, "iostat -x -k 1" shows no reads from the XFS
> > devices, though writes keep happening. "vmstat 1" matches. I tried
> > switching schedulers from CFQ to deadline -- no difference. Queue depth
> > is empty on the devices and nothing is actually clogged up at the device
> > -- it's not actually plugged at the controller or disk. I did a sysreq-w
> > while this was happening. About 10 seconds later, everything unclogs and
> > continues. Sysreq-W output below. I poked around at the various XFS
> > tracepoints in /sys/kernel/debug/tracing/events/xfs, but I'm not sure
> > which tracepoints to use and many of them scroll too fast to see
> > anything. Any suggestions?
> Given that you are doing a lot of unlinks I wonder if it is related
> to the recent ail pushing issues in that area. While your symptoms
> looks completely different we could be blocking on the flush completion
> for an inode that gets stuck in the AIL.
> Can you run with latest 3.0-stable plus the patches at:
> If this doesn't help I'll probably need to come up with some tracing
> patches for you.
It seemes 3.0.7+gregkh's stable-queue queue-3.0 patches seems to be
running fine without blocking at all on this SSD box, so that should
narrow it down significantly.
Hmm, looking at git diff --stat v3.0.7..v3.1-rc10 fs/xfs , maybe not.. :)
Maybe 3.1 fs/xfs would transplant into 3.0 or vice-versa?