On Wed, 29 Feb 2012 18:44:13 -0600 Ray Morris [off-list ref] wrote:
I am attempting to debug a hang in raid1 and possibly one raid5.
I have experienced the same problem with many kernel versions
over a couple of years, and with disparate hardware.
My current plan, if noone more experienced suggests I do otherwise, is
to compile a kernel with some printk() in strategic locations and try to
narrow down the problem. I know very little about kernel work, so I am
seeking suggestions from those who know better than I.
In the case logged below, it appears to hang at raise_barrier in md2_resync
at raise_barrier, then further access to the device hangs. I'm just a Perl
programmer who dabbles in C, but my intuition said I check that if perhaps
lower_barrier had been called with conf->barrier already at zero, so that's
the one printk I've added so far. It may take a week or more before it
crashes again, so is there any more debugging I should add before waiting
for it to hang?
Also below is some older logging from similar symptoms with raid5,
hanging at raid5_quiesce. I got rid of the raid5 in hopes of getting
rid of the problem, but if anyone has suggestions on how to further
debug that I maybe be able to add a raid5 array.
The load when I've noticed it is rsync to LVM volumes with snapshots.
After some discussion, lvm-devel suggested linux-raid would be the next
logical step. Tested kernels include 2.6.32-220.4.1.el6.x86_64
2.6.32.26-175.fc12.x86_64, vmlinuz-2.6.32.9-70.fc12.x86_64, and others.
Since I already have updated the kernel several times in the last couple
of years, I figured I'd try some debugging with the current EL 6 kernel.
Anyway, any thoughts on how to debug, where to stick some printk, other
debugging functions?
I might know what is happening.
It is kind-a complicated and involved the magic code in
block/blk-core.c:generic_make_request which turns recursive calls into tail
recursion.
The fs sends a request to dm.
dm split it in 2 for some reason and sends them both to md.
This involves them getting queued in generic_make_request.
The first gets actioned by md/raid1 and converted into a request to the
underlying device (it must be a read request for this to happen - so just one
device). This gets added to the queue and is counted in nr_pending.
At this point sync_request is called by another thread and it tries to
raise_battier. It gets past the first hurdle, increments ->barrier, and
waits for nr_pending to hit zero.
Now the second request from dm to md is passed to raid1.c:make_request where
it tries to wait_barrier. This blocks because ->barrier is up, and we have a
deadlock - the request to the underlying device will not progress until this
md request progresses, and it is stuck.
This patch might fix it. Maybe. If it compiles.
NeilBrown
Index: linux-2.6.32-SLE11-SP1/drivers/md/raid1.c
===================================================================
--- linux-2.6.32-SLE11-SP1.orig/drivers/md/raid1.c 2012-03-01 12:28:05.000000000 +1100
+++ linux-2.6.32-SLE11-SP1/drivers/md/raid1.c 2012-03-01 12:28:22.427992913 +1100
@@ -695,7 +695,11 @@ static void wait_barrier(conf_t *conf)
spin_lock_irq(&conf->resync_lock);
if (conf->barrier) {
conf->nr_waiting++;
- wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
+ wait_event_lock_irq(conf->wait_barrier,
+ !conf->barrier ||
+ (current->bio_tail &&
+ current->bio_list &&
+ conf->nr_pending),
conf->resync_lock,
raid1_unplug(conf->mddev->queue));
conf->nr_waiting--;