md/raid5: fix interaction of 'replace' and 'recovery'.
authorNeilBrown <neilb@suse.de>
Mon, 22 Jul 2013 02:57:21 +0000 (12:57 +1000)
committerNeilBrown <neilb@suse.de>
Thu, 25 Jul 2013 06:46:57 +0000 (16:46 +1000)
If a device in a RAID4/5/6 is being replaced while another is being
recovered, then the writes to the replacement device currently don't
happen, resulting in corruption when the replacement completes and the
new drive takes over.

This is because the replacement writes are only triggered when
's.replacing' is set and not when the similar 's.sync' is set (which
is the case during resync and recovery - it means all devices need to
be read).

So schedule those writes when s.replacing is set as well.

In this case we cannot use "STRIPE_INSYNC" to record that the
replacement has happened as that is needed for recording that any
parity calculation is complete.  So introduce STRIPE_REPLACED to
record if the replacement has happened.

For safety we should also check that STRIPE_COMPUTE_RUN is not set.
This has a similar effect to the "s.locked == 0" test.  The latter
ensure that now IO has been flagged but not started.  The former
checks if any parity calculation has been flagged by not started.
We must wait for both of these to complete before triggering the
'replace'.

Add a similar test to the subsequent check for "are we finished yet".
This possibly isn't needed (is subsumed in the STRIPE_INSYNC test),
but it makes it more obvious that the REPLACE will happen before we
think we are finished.

Finally if a NeedReplace device is not UPTODATE then that is an
error.  We really must trigger a warning.

This bug was introduced in commit 9a3e1101b827a59ac9036a672f5fa8d5279d0fe2
(md/raid5:  detect and handle replacements during recovery.)
which introduced replacement for raid5.
That was in 3.3-rc3, so any stable kernel since then would benefit
from this fix.

Cc: stable@vger.kernel.org (3.3+)
Reported-by: qindehua <13691222965@163.com>
Tested-by: qindehua <qindehua@163.com>
Signed-off-by: NeilBrown <neilb@suse.de>
drivers/md/raid5.c
drivers/md/raid5.h

index 2bf094a587cb8ff2efb7e1bcabe6f40cd7b52602..78ea44336e75e06d5769b4a6158f2a1e506214e0 100644 (file)
@@ -3462,6 +3462,7 @@ static void handle_stripe(struct stripe_head *sh)
                    test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
                        set_bit(STRIPE_SYNCING, &sh->state);
                        clear_bit(STRIPE_INSYNC, &sh->state);
+                       clear_bit(STRIPE_REPLACED, &sh->state);
                }
                spin_unlock(&sh->stripe_lock);
        }
@@ -3607,19 +3608,23 @@ static void handle_stripe(struct stripe_head *sh)
                        handle_parity_checks5(conf, sh, &s, disks);
        }
 
-       if (s.replacing && s.locked == 0
-           && !test_bit(STRIPE_INSYNC, &sh->state)) {
+       if ((s.replacing || s.syncing) && s.locked == 0
+           && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
+           && !test_bit(STRIPE_REPLACED, &sh->state)) {
                /* Write out to replacement devices where possible */
                for (i = 0; i < conf->raid_disks; i++)
-                       if (test_bit(R5_UPTODATE, &sh->dev[i].flags) &&
-                           test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
+                       if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
+                               WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
                                set_bit(R5_WantReplace, &sh->dev[i].flags);
                                set_bit(R5_LOCKED, &sh->dev[i].flags);
                                s.locked++;
                        }
-               set_bit(STRIPE_INSYNC, &sh->state);
+               if (s.replacing)
+                       set_bit(STRIPE_INSYNC, &sh->state);
+               set_bit(STRIPE_REPLACED, &sh->state);
        }
        if ((s.syncing || s.replacing) && s.locked == 0 &&
+           !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
            test_bit(STRIPE_INSYNC, &sh->state)) {
                md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
                clear_bit(STRIPE_SYNCING, &sh->state);
index b0b663b119a8f324ba7fa3b58953b42c81030fa4..70c49329ca9a23fbbad73061477ac0a0cfe80e6e 100644 (file)
@@ -306,6 +306,7 @@ enum {
        STRIPE_SYNC_REQUESTED,
        STRIPE_SYNCING,
        STRIPE_INSYNC,
+       STRIPE_REPLACED,
        STRIPE_PREREAD_ACTIVE,
        STRIPE_DELAYED,
        STRIPE_DEGRADED,