md/raid5: use bad-block log to improve handling of uncorrectable read errors.
authorNeilBrown <neilb@suse.de>
Thu, 28 Jul 2011 01:39:22 +0000 (11:39 +1000)
committerNeilBrown <neilb@suse.de>
Thu, 28 Jul 2011 01:39:22 +0000 (11:39 +1000)
If we get an uncorrectable read error - record a bad block rather than
failing the device.
And if these errors (which may be due to known bad blocks) cause
recovery to be impossible, record a bad block on the recovering
devices, or abort the recovery.

As we might abort a recovery without failing a device we need to teach
RAID5 about recovery_disabled handling.

Signed-off-by: NeilBrown <neilb@suse.de>
drivers/md/raid5.c
drivers/md/raid5.h

index a2d68389ee75d5e1704e9ffa04beb8f73cff068d..5fc621673e6ceaaaa339e98dd680215e3919ed63 100644 (file)
@@ -2232,9 +2232,18 @@ handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
                        rcu_read_lock();
                        rdev = rcu_dereference(conf->disks[i].rdev);
                        if (rdev && test_bit(In_sync, &rdev->flags))
-                               /* multiple read failures in one stripe */
-                               md_error(conf->mddev, rdev);
+                               atomic_inc(&rdev->nr_pending);
+                       else
+                               rdev = NULL;
                        rcu_read_unlock();
+                       if (rdev) {
+                               if (!rdev_set_badblocks(
+                                           rdev,
+                                           sh->sector,
+                                           STRIPE_SECTORS, 0))
+                                       md_error(conf->mddev, rdev);
+                               rdev_dec_pending(rdev, conf->mddev);
+                       }
                }
                spin_lock_irq(&conf->device_lock);
                /* fail all writes first */
@@ -2313,6 +2322,41 @@ handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
                        md_wakeup_thread(conf->mddev->thread);
 }
 
+static void
+handle_failed_sync(raid5_conf_t *conf, struct stripe_head *sh,
+                  struct stripe_head_state *s)
+{
+       int abort = 0;
+       int i;
+
+       md_done_sync(conf->mddev, STRIPE_SECTORS, 0);
+       clear_bit(STRIPE_SYNCING, &sh->state);
+       s->syncing = 0;
+       /* There is nothing more to do for sync/check/repair.
+        * For recover we need to record a bad block on all
+        * non-sync devices, or abort the recovery
+        */
+       if (!test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery))
+               return;
+       /* During recovery devices cannot be removed, so locking and
+        * refcounting of rdevs is not needed
+        */
+       for (i = 0; i < conf->raid_disks; i++) {
+               mdk_rdev_t *rdev = conf->disks[i].rdev;
+               if (!rdev
+                   || test_bit(Faulty, &rdev->flags)
+                   || test_bit(In_sync, &rdev->flags))
+                       continue;
+               if (!rdev_set_badblocks(rdev, sh->sector,
+                                       STRIPE_SECTORS, 0))
+                       abort = 1;
+       }
+       if (abort) {
+               conf->recovery_disabled = conf->mddev->recovery_disabled;
+               set_bit(MD_RECOVERY_INTR, &conf->mddev->recovery);
+       }
+}
+
 /* fetch_block - checks the given member device to see if its data needs
  * to be read or computed to satisfy a request.
  *
@@ -3067,11 +3111,8 @@ static void handle_stripe(struct stripe_head *sh)
         */
        if (s.failed > conf->max_degraded && s.to_read+s.to_write+s.written)
                handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
-       if (s.failed > conf->max_degraded && s.syncing) {
-               md_done_sync(conf->mddev, STRIPE_SECTORS, 0);
-               clear_bit(STRIPE_SYNCING, &sh->state);
-               s.syncing = 0;
-       }
+       if (s.failed > conf->max_degraded && s.syncing)
+               handle_failed_sync(conf, sh, &s);
 
        /*
         * might be able to return some write requests if the parity blocks
@@ -4976,6 +5017,7 @@ static int raid5_remove_disk(mddev_t *mddev, int number)
                 * isn't possible.
                 */
                if (!test_bit(Faulty, &rdev->flags) &&
+                   mddev->recovery_disabled != conf->recovery_disabled &&
                    !has_failed(conf) &&
                    number < conf->raid_disks) {
                        err = -EBUSY;
@@ -5004,6 +5046,9 @@ static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
        int first = 0;
        int last = conf->raid_disks - 1;
 
+       if (mddev->recovery_disabled == conf->recovery_disabled)
+               return -EBUSY;
+
        if (has_failed(conf))
                /* no point adding a device */
                return -EINVAL;
index 68c500af1108c4ae620fe292e3245f212d987b45..c5429d12363654075b752b0ca689c672b9361f02 100644 (file)
@@ -399,7 +399,7 @@ struct raid5_private_data {
                                            * (fresh device added).
                                            * Cleared when a sync completes.
                                            */
-
+       int                     recovery_disabled;
        /* per cpu variables */
        struct raid5_percpu {
                struct page     *spare_page; /* Used when checking P/Q in raid6 */