On Sat, 24 Nov 2012 10:18:44 +0100 Torsten Kaiser
[off-list ref] wrote:

After my system got stuck with 3.7.0-rc2 as reported in
http://marc.info/?l=linux-kernel&m=135142236520624 LOCKDEP seem to
blame XFS, because it found 2 possible deadlocks. But after these
locking issues where fixed, my system got stuck again with 3.7.0-rc6
as reported in http://marc.info/?l=linux-kernel&m=135344072325490
Dave Chinner thinks its an issue within md, that it gets stuck and
that will then prevent any further xfs activity, and that I should
report it to the raid mailing list.

The issue seems to be that multiple processes (kswapd0, xfsaild/md4
and flush-9:4) get stuck in md_super_wait() like this:
[<ffffffff816b1224>] schedule+0x24/0x60
[<ffffffff814f9dad>] md_super_wait+0x4d/0x80
[<ffffffff8105ca30>] ? __init_waitqueue_head+0x60/0x60
[<ffffffff81500753>] bitmap_unplug+0x173/0x180
[<ffffffff810b6acf>] ? write_cache_pages+0x12f/0x420
[<ffffffff810b6700>] ? set_page_dirty_lock+0x60/0x60
[<ffffffff814e8eb8>] raid1_unplug+0x98/0x110
[<ffffffff81278a6d>] blk_flush_plug_list+0xad/0x240
[<ffffffff81278c13>] blk_finish_plug+0x13/0x50

The full hung-tasks stack traces and the output from SysRq+W can be
found at http://marc.info/?l=linux-kernel&m=135344072325490 or in the
LKML thread 'Hang in XFS reclaim on 3.7.0-rc3'.

Yes, it does look like an md bug....
Can you test to see if this fixes it?

diff --git a/drivers/md/raid1.c b/drivers/md/raid1.c
index 636bae0..a0f7309 100644
--- a/drivers/md/raid1.c
+++ b/drivers/md/raid1.c

@@ -963,7 +963,7 @@ static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
 	struct r1conf *conf = mddev->private;
 	struct bio *bio;
 
-	if (from_schedule) {
+	if (from_schedule || current->bio_list) {
 		spin_lock_irq(&conf->device_lock);
 		bio_list_merge(&conf->pending_bio_list, &plug->pending);
 		conf->pending_count += plug->pending_cnt;

I tried to understand how this could happen, but I don't see anything
wrong. Only that md_super_wait() looks like an open coded version of
__wait_event() and could be replaced by using it.

yeah.  md_super_wait was much more complex back when we had to support
barrier operations.  When they were removed it was simplified a lot and as
you say it could be simplifier further.  Patches welcome.

http://marc.info/?l=linux-raid&m=135283030027665 looks like the same
issue, but using ext4 instead of xfs.

yes, sure does.

My setup wrt. md is two normal sata disks on a normal ahci controller
(AMD SB850 southbridge).
Both disks are divided into 4 partitions and each one assembled into a
separate raid1.
One (md5) is used for swap, the others hold xfs filesystems for /boot/
(md4), / (md6) and /home/ (md7).

I will try to provide any information you ask, but I can't reproduce
the hang on demand so gathering more information about that state is
not so easy, but I will try.

I'm fairly confident the above patch will fixes it, and in any case it fixes
a real bug.  So if you could just run with it and confirm in a week or so
that the problem hasn't recurred, that might have to do.

Thanks,
NeilBrown