]> pilppa.com Git - linux-2.6-omap-h63xx.git/commitdiff
md: close a livelock window in handle_parity_checks5
authorDan Williams <dan.j.williams@intel.com>
Fri, 11 Apr 2008 04:29:27 +0000 (21:29 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 11 Apr 2008 15:06:44 +0000 (08:06 -0700)
If a failure is detected after a parity check operation has been initiated,
but before it completes handle_parity_checks5 will never quiesce operations on
the stripe.

Explicitly handle this case by "canceling" the parity check, i.e.  clear the
STRIPE_OP_CHECK flags and queue the stripe on the handle list again to refresh
any non-uptodate blocks.

Kernel versions >= 2.6.23 are susceptible.

Cc: <stable@kernel.org>
Cc: NeilBrown <neilb@suse.de>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
drivers/md/raid5.c

index c574cf5efb5c5b4ff54694ad5189411afafb4fab..b162b839a6620921eb258252ee9ddaf1a1d12810 100644 (file)
@@ -2348,25 +2348,15 @@ static void handle_issuing_new_write_requests6(raid5_conf_t *conf,
 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
                                struct stripe_head_state *s, int disks)
 {
+       int canceled_check = 0;
+
        set_bit(STRIPE_HANDLE, &sh->state);
-       /* Take one of the following actions:
-        * 1/ start a check parity operation if (uptodate == disks)
-        * 2/ finish a check parity operation and act on the result
-        * 3/ skip to the writeback section if we previously
-        *    initiated a recovery operation
-        */
-       if (s->failed == 0 &&
-           !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
-               if (!test_and_set_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
-                       BUG_ON(s->uptodate != disks);
-                       clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
-                       sh->ops.count++;
-                       s->uptodate--;
-               } else if (
-                      test_and_clear_bit(STRIPE_OP_CHECK, &sh->ops.complete)) {
-                       clear_bit(STRIPE_OP_CHECK, &sh->ops.ack);
-                       clear_bit(STRIPE_OP_CHECK, &sh->ops.pending);
 
+       /* complete a check operation */
+       if (test_and_clear_bit(STRIPE_OP_CHECK, &sh->ops.complete)) {
+           clear_bit(STRIPE_OP_CHECK, &sh->ops.ack);
+           clear_bit(STRIPE_OP_CHECK, &sh->ops.pending);
+               if (s->failed == 0) {
                        if (sh->ops.zero_sum_result == 0)
                                /* parity is correct (on disc,
                                 * not in buffer any more)
@@ -2391,7 +2381,8 @@ static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
                                        s->uptodate++;
                                }
                        }
-               }
+               } else
+                       canceled_check = 1; /* STRIPE_INSYNC is not set */
        }
 
        /* check if we can clear a parity disk reconstruct */
@@ -2404,12 +2395,28 @@ static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
                clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
        }
 
+       /* start a new check operation if there are no failures, the stripe is
+        * not insync, and a repair is not in flight
+        */
+       if (s->failed == 0 &&
+           !test_bit(STRIPE_INSYNC, &sh->state) &&
+           !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
+               if (!test_and_set_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
+                       BUG_ON(s->uptodate != disks);
+                       clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
+                       sh->ops.count++;
+                       s->uptodate--;
+               }
+       }
+
        /* Wait for check parity and compute block operations to complete
-        * before write-back
+        * before write-back.  If a failure occurred while the check operation
+        * was in flight we need to cycle this stripe through handle_stripe
+        * since the parity block may not be uptodate
         */
-       if (!test_bit(STRIPE_INSYNC, &sh->state) &&
-               !test_bit(STRIPE_OP_CHECK, &sh->ops.pending) &&
-               !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) {
+       if (!canceled_check && !test_bit(STRIPE_INSYNC, &sh->state) &&
+           !test_bit(STRIPE_OP_CHECK, &sh->ops.pending) &&
+           !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) {
                struct r5dev *dev;
                /* either failed parity check, or recovery is happening */
                if (s->failed == 0)