md: restart recovery cleanly after device failure.
authorNeilBrown <neilb@suse.de>
Fri, 23 May 2008 20:04:39 +0000 (13:04 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Sat, 24 May 2008 16:56:10 +0000 (09:56 -0700)
When we get any IO error during a recovery (rebuilding a spare), we abort
the recovery and restart it.

For RAID6 (and multi-drive RAID1) it may not be best to restart at the
beginning: when multiple failures can be tolerated, the recovery may be
able to continue and re-doing all that has already been done doesn't make
sense.

We already have the infrastructure to record where a recovery is up to
and restart from there, but it is not being used properly.
This is because:
  - We sometimes abort with MD_RECOVERY_ERR rather than just MD_RECOVERY_INTR,
    which causes the recovery not be be checkpointed.
  - We remove spares and then re-added them which loses important state
    information.

The distinction between MD_RECOVERY_ERR and MD_RECOVERY_INTR really isn't
needed.  If there is an error, the relevant drive will be marked as
Faulty, and that is enough to ensure correct handling of the error.  So we
first remove MD_RECOVERY_ERR, changing some of the uses of it to
MD_RECOVERY_INTR.

Then we cause the attempt to remove a non-faulty device from an array to
fail (unless recovery is impossible as the array is too degraded).  Then
when remove_and_add_spares attempts to remove the devices on which
recovery can continue, it will fail, they will remain in place, and
recovery will continue on them as desired.

Issue:  If we are halfway through rebuilding a spare and another drive
fails, and a new spare is immediately available,  do we want to:
 1/ complete the current rebuild, then go back and rebuild the new spare or
 2/ restart the rebuild from the start and rebuild both devices in
    parallel.

Both options can be argued for.  The code currently takes option 2 as
  a/ this requires least code change
  b/ this results in a minimally-degraded array in minimal time.

Cc: "Eivind Sarto" <ivan@kasenna.com>
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
drivers/md/md.c
drivers/md/multipath.c
drivers/md/raid1.c
drivers/md/raid10.c
drivers/md/raid5.c
include/linux/raid/md_k.h

index 295be1a688065b06b1367ab947f6793c9fce1ecd..51c19f86ff99c3785b5d27838e755d04dabe5fe3 100644 (file)
@@ -5434,7 +5434,7 @@ void md_done_sync(mddev_t *mddev, int blocks, int ok)
        atomic_sub(blocks, &mddev->recovery_active);
        wake_up(&mddev->recovery_wait);
        if (!ok) {
-               set_bit(MD_RECOVERY_ERR, &mddev->recovery);
+               set_bit(MD_RECOVERY_INTR, &mddev->recovery);
                md_wakeup_thread(mddev->thread);
                // stop recovery, signal do_sync ....
        }
@@ -5690,7 +5690,7 @@ void md_do_sync(mddev_t *mddev)
                sectors = mddev->pers->sync_request(mddev, j, &skipped,
                                                  currspeed < speed_min(mddev));
                if (sectors == 0) {
-                       set_bit(MD_RECOVERY_ERR, &mddev->recovery);
+                       set_bit(MD_RECOVERY_INTR, &mddev->recovery);
                        goto out;
                }
 
@@ -5713,8 +5713,7 @@ void md_do_sync(mddev_t *mddev)
 
                last_check = io_sectors;
 
-               if (test_bit(MD_RECOVERY_INTR, &mddev->recovery) ||
-                   test_bit(MD_RECOVERY_ERR, &mddev->recovery))
+               if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
                        break;
 
        repeat:
@@ -5768,8 +5767,7 @@ void md_do_sync(mddev_t *mddev)
        /* tell personality that we are finished */
        mddev->pers->sync_request(mddev, max_sectors, &skipped, 1);
 
-       if (!test_bit(MD_RECOVERY_ERR, &mddev->recovery) &&
-           !test_bit(MD_RECOVERY_CHECK, &mddev->recovery) &&
+       if (!test_bit(MD_RECOVERY_CHECK, &mddev->recovery) &&
            mddev->curr_resync > 2) {
                if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
                        if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
@@ -5838,7 +5836,10 @@ static int remove_and_add_spares(mddev_t *mddev)
                }
 
        if (mddev->degraded) {
-               rdev_for_each(rdev, rtmp, mddev)
+               rdev_for_each(rdev, rtmp, mddev) {
+                       if (rdev->raid_disk >= 0 &&
+                           !test_bit(In_sync, &rdev->flags))
+                               spares++;
                        if (rdev->raid_disk < 0
                            && !test_bit(Faulty, &rdev->flags)) {
                                rdev->recovery_offset = 0;
@@ -5856,6 +5857,7 @@ static int remove_and_add_spares(mddev_t *mddev)
                                } else
                                        break;
                        }
+               }
        }
        return spares;
 }
@@ -5869,7 +5871,7 @@ static int remove_and_add_spares(mddev_t *mddev)
  * to do that as needed.
  * When it is determined that resync is needed, we set MD_RECOVERY_RUNNING in
  * "->recovery" and create a thread at ->sync_thread.
- * When the thread finishes it sets MD_RECOVERY_DONE (and might set MD_RECOVERY_ERR)
+ * When the thread finishes it sets MD_RECOVERY_DONE
  * and wakeups up this thread which will reap the thread and finish up.
  * This thread also removes any faulty devices (with nr_pending == 0).
  *
@@ -5944,8 +5946,7 @@ void md_check_recovery(mddev_t *mddev)
                        /* resync has finished, collect result */
                        md_unregister_thread(mddev->sync_thread);
                        mddev->sync_thread = NULL;
-                       if (!test_bit(MD_RECOVERY_ERR, &mddev->recovery) &&
-                           !test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
+                       if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
                                /* success...*/
                                /* activate any spares */
                                mddev->pers->spare_active(mddev);
@@ -5969,7 +5970,6 @@ void md_check_recovery(mddev_t *mddev)
                 * might be left set
                 */
                clear_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
-               clear_bit(MD_RECOVERY_ERR, &mddev->recovery);
                clear_bit(MD_RECOVERY_INTR, &mddev->recovery);
                clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
 
index 4f4d1f383842c9ca41cb3e89e914cdace48e60c1..e968116e0de9699d2bef0f114ae79f638506bf1c 100644 (file)
@@ -327,7 +327,8 @@ static int multipath_remove_disk(mddev_t *mddev, int number)
        if (rdev) {
                if (test_bit(In_sync, &rdev->flags) ||
                    atomic_read(&rdev->nr_pending)) {
-                       printk(KERN_ERR "hot-remove-disk, slot %d is identified"                                " but is still operational!\n", number);
+                       printk(KERN_ERR "hot-remove-disk, slot %d is identified"
+                              " but is still operational!\n", number);
                        err = -EBUSY;
                        goto abort;
                }
index d0f4021bbc2ecb483c7b30676478c5403f84f018..c610b947218afb73f49982d59dd178a7fdbd0959 100644 (file)
@@ -1027,7 +1027,7 @@ static void error(mddev_t *mddev, mdk_rdev_t *rdev)
                /*
                 * if recovery is running, make sure it aborts.
                 */
-               set_bit(MD_RECOVERY_ERR, &mddev->recovery);
+               set_bit(MD_RECOVERY_INTR, &mddev->recovery);
        } else
                set_bit(Faulty, &rdev->flags);
        set_bit(MD_CHANGE_DEVS, &mddev->flags);
@@ -1148,6 +1148,14 @@ static int raid1_remove_disk(mddev_t *mddev, int number)
                        err = -EBUSY;
                        goto abort;
                }
+               /* Only remove non-faulty devices is recovery
+                * is not possible.
+                */
+               if (!test_bit(Faulty, &rdev->flags) &&
+                   mddev->degraded < conf->raid_disks) {
+                       err = -EBUSY;
+                       goto abort;
+               }
                p->rdev = NULL;
                synchronize_rcu();
                if (atomic_read(&rdev->nr_pending)) {
index 8536ede1e7129c37fc30303e15577c383e68569c..1de17da34a956e8383efb1b42474c3c4d3d15dbe 100644 (file)
@@ -1020,7 +1020,7 @@ static void error(mddev_t *mddev, mdk_rdev_t *rdev)
                /*
                 * if recovery is running, make sure it aborts.
                 */
-               set_bit(MD_RECOVERY_ERR, &mddev->recovery);
+               set_bit(MD_RECOVERY_INTR, &mddev->recovery);
        }
        set_bit(Faulty, &rdev->flags);
        set_bit(MD_CHANGE_DEVS, &mddev->flags);
@@ -1171,6 +1171,14 @@ static int raid10_remove_disk(mddev_t *mddev, int number)
                        err = -EBUSY;
                        goto abort;
                }
+               /* Only remove faulty devices in recovery
+                * is not possible.
+                */
+               if (!test_bit(Faulty, &rdev->flags) &&
+                   enough(conf)) {
+                       err = -EBUSY;
+                       goto abort;
+               }
                p->rdev = NULL;
                synchronize_rcu();
                if (atomic_read(&rdev->nr_pending)) {
@@ -1237,6 +1245,7 @@ static void end_sync_write(struct bio *bio, int error)
 
        if (!uptodate)
                md_error(mddev, conf->mirrors[d].rdev);
+
        update_head_pos(i, r10_bio);
 
        while (atomic_dec_and_test(&r10_bio->remaining)) {
@@ -1844,7 +1853,8 @@ static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, i
                                        if (rb2)
                                                atomic_dec(&rb2->remaining);
                                        r10_bio = rb2;
-                                       if (!test_and_set_bit(MD_RECOVERY_ERR, &mddev->recovery))
+                                       if (!test_and_set_bit(MD_RECOVERY_INTR,
+                                                             &mddev->recovery))
                                                printk(KERN_INFO "raid10: %s: insufficient working devices for recovery.\n",
                                                       mdname(mddev));
                                        break;
index 2f28745dacf974b509ae019231973f4de95c8227..425958a76b84afc9d7608509f6c06954db342745 100644 (file)
@@ -1268,7 +1268,7 @@ static void error(mddev_t *mddev, mdk_rdev_t *rdev)
                        /*
                         * if recovery was running, make sure it aborts.
                         */
-                       set_bit(MD_RECOVERY_ERR, &mddev->recovery);
+                       set_bit(MD_RECOVERY_INTR, &mddev->recovery);
                }
                set_bit(Faulty, &rdev->flags);
                printk (KERN_ALERT
@@ -4574,6 +4574,14 @@ static int raid5_remove_disk(mddev_t *mddev, int number)
                        err = -EBUSY;
                        goto abort;
                }
+               /* Only remove non-faulty devices if recovery
+                * isn't possible.
+                */
+               if (!test_bit(Faulty, &rdev->flags) &&
+                   mddev->degraded <= conf->max_degraded) {
+                       err = -EBUSY;
+                       goto abort;
+               }
                p->rdev = NULL;
                synchronize_rcu();
                if (atomic_read(&rdev->nr_pending)) {
index a6d7ab688edec5699887975d691576b369a97af9..3dea9f545c8f337c516b1713a39e06f59ec767d3 100644 (file)
@@ -188,8 +188,7 @@ struct mddev_s
         * NEEDED:   we might need to start a resync/recover
         * RUNNING:  a thread is running, or about to be started
         * SYNC:     actually doing a resync, not a recovery
-        * ERR:      and IO error was detected - abort the resync/recovery
-        * INTR:     someone requested a (clean) early abort.
+        * INTR:     resync needs to be aborted for some reason
         * DONE:     thread is done and is waiting to be reaped
         * REQUEST:  user-space has requested a sync (used with SYNC)
         * CHECK:    user-space request for for check-only, no repair
@@ -199,7 +198,6 @@ struct mddev_s
         */
 #define        MD_RECOVERY_RUNNING     0
 #define        MD_RECOVERY_SYNC        1
-#define        MD_RECOVERY_ERR         2
 #define        MD_RECOVERY_INTR        3
 #define        MD_RECOVERY_DONE        4
 #define        MD_RECOVERY_NEEDED      5