Re: raid5: R5_WriteError is not cleared after rdev is marked as Faulty

On Thu, 2 Jan 2014 17:58:33 +0200 Alexander Lyakas [off-list ref]
wrote:

quoted

Hi Neil,
I see the following issue: if more than one stripe-head experiences an
error in raid5_end_write_request, then when first stripe-head is
handled, it eventually fails the disk, but when the second stripe-head
is handled, disk is already marked as Faulty, so R5_WriteError is not
cleared from this second stripe-head, but it still remains in the
stripe cache.

Below are more details on the flow. I added some prints to display the
flags values etc.

# raid6 is rebuilding a disk

# disk fails, as a result, raid5_end_write_request is called on two
stripe-heads:
[23369]md48066*[raid5_end_write_request:2002] sh[4560].flags=0x101A
dev[0].flags=0x20003 rdev[dm-6].flags=0x40
bio=ffff880028744fd8[4592(4560):8] WRITE ERROR
[13527.467363] [3842]md48066*[raid5_end_write_request:2002]
sh[4568].flags=0x101B dev[0].flags=0x20003 rdev[dm-6].flags=0x240
bio=ffff880028742bc0[4600(4568):8] WRITE ERROR

As a result, appropriate r5dev's are marked with R5_WriteError on
these two stripe-heads.

# raid5 calls analyse_stripe on behalf of stripe-head 4560 ((through
handle_active_stripes->handle_stripe->analyse_stripe)

# analyse_stripe performs this code:
        if (rdev && test_bit(R5_WriteError, &dev->flags)) {
            /* This flag does not apply to '.replacement'
             * only to .rdev, so make sure to check that*/
            struct md_rdev *rdev2 = rcu_dereference(
                conf->disks[i].rdev);
            if (rdev2 == rdev)
                clear_bit(R5_Insync, &dev->flags);
            if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
                s->handle_bad_blocks = 1;
it clears the R5_WriteError and sets handle_bad_blocks on the stripe-head

# handle_stripe performs:
    if (s.handle_bad_blocks)
        for (i = disks; i--; ) {
            struct md_rdev *rdev;
            struct r5dev *dev = &sh->dev[i];
            if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
                /* We own a safe reference to the rdev */
                rdev = conf->disks[i].rdev;
                if (!rdev_set_badblocks(rdev, sh->sector,
                            STRIPE_SECTORS, 0))
                    md_error(conf->mddev, rdev);
this marks the rdev as Faulty and fails it (bad blocks are disabled)

# raid5d calls analyse_stripe on stripe-head 4568. However, rdev is
already marked as Faulty, so it does:
        if (rdev && test_bit(Faulty, &rdev->flags))
            rdev = NULL;
and later this condition does not hold:
if (rdev && test_bit(R5_WriteError, &dev->flags)) {
so R5_WriteError is not cleared, but this stripe-head is still valid
in the stripe cache.

Dumping the state of this stripe-head in the stripe-cache (through
custom sysfs entry):
sh[4568].state=0x1010 disks=5 pdidx=2 qdidx=3
  dev[0].flags=0x20001 rdev[???].flags=0x0 (sector=13688)
  dev[1].flags=0x11 rdev[dm-7].flags=0x2 (sector=13720)
  dev[2].flags=0x11 rdev[dm-8].flags=0x2 (sector=0) *P*
  dev[3].flags=0x11 rdev[dm-9].flags=0x2 (sector=0) *Q*
  dev[4].flags=0x11 rdev[dm-10].flags=0x2 (sector=13656)

# Now user recovers the disk and re-adds the disk back into the array.
Rebuild starts, and eventually md_do_sync calls sync_request on this
stripe-head, which calls handle_stripe->analyse_stripe. Now
analyse_stripe sees the stale R5_WriteError flag and fails the disk,
which aborts the rebuild. Here is the stack trace, which detects the
stale R5_WriteError and fails the disk:
[<ffffffffa02d4543>] error+0x153/0x1d0 [raid456]
[<ffffffffa0445449>] md_error.part.39+0x19/0xa0 [md_mod]
[<ffffffffa0445503>] md_error+0x33/0x60 [md_mod]
[<ffffffffa02da07c>] handle_stripe+0x57c/0x2250 [raid456]
[<ffffffffa02dcacb>] sync_request+0x18b/0x390 [raid456]
[<ffffffffa044320c>] md_do_sync+0x76c/0xda0 [md_mod]
[<ffffffffa044084d>] md_thread+0x10d/0x140 [md_mod]

Does this analysis make sense? I don't know how to fix issue though. I
am looking at kernel (3.8.13 - as usual), but looking at your
for-linus branch, it seems like the same would happen.

Looks like:

commit 5d8c71f9e5fbdd95650be00294d238e27a363b5c
Author: Adam Kwolek [off-list ref]
Date:   Fri Dec 9 14:26:11 2011 +1100

    md: raid5 crash during degradation


introduced this problem (in linux 3.3) with an over-simple fix for a problem.

I think that we can not just reverted that patch as

commit 14a75d3e07c784c004b4b44b34af996b8e4ac453
Author: NeilBrown [off-list ref]
Date:   Fri Dec 23 10:17:52 2011 +1100

    md/raid5: preferentially read from replacement device if possible.


has effected a better fix.
So the patch below should work.  Does that make sense to you?

Thanks,
NeilBrown

diff --git a/drivers/md/raid5.c b/drivers/md/raid5.c
index cc055da02e2a..9168173deaf3 100644
--- a/drivers/md/raid5.c
+++ b/drivers/md/raid5.c

@@ -3608,7 +3608,7 @@ static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
                         */
                        set_bit(R5_Insync, &dev->flags);

-               if (rdev && test_bit(R5_WriteError, &dev->flags)) {
+               if (test_bit(R5_WriteError, &dev->flags)) {
                        /* This flag does not apply to '.replacement'
                         * only to .rdev, so make sure to check that*/
                        struct md_rdev *rdev2 = rcu_dereference(

@@ -3621,7 +3621,7 @@ static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
                        } else
                                clear_bit(R5_WriteError, &dev->flags);
                }
-               if (rdev && test_bit(R5_MadeGood, &dev->flags)) {
+               if (test_bit(R5_MadeGood, &dev->flags)) {
                        /* This flag does not apply to '.replacement'
                         * only to .rdev, so make sure to check that*/
                        struct md_rdev *rdev2 = rcu_dereference(