]> err.no Git - linux-2.6/commitdiff
[PATCH] md: allow checkpoint of recovery with version-1 superblock
authorNeilBrown <neilb@suse.de>
Mon, 26 Jun 2006 07:27:40 +0000 (00:27 -0700)
committerLinus Torvalds <torvalds@g5.osdl.org>
Mon, 26 Jun 2006 16:58:37 +0000 (09:58 -0700)
For a while we have had checkpointing of resync.  The version-1 superblock
allows recovery to be checkpointed as well, and this patch implements that.

Due to early carelessness we need to add a feature flag to signal that the
recovery_offset field is in use, otherwise older kernels would assume that a
partially recovered array is in fact fully recovered.

Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
drivers/md/md.c
drivers/md/raid1.c
drivers/md/raid10.c
drivers/md/raid5.c
include/linux/raid/md_k.h
include/linux/raid/md_p.h

index 71d46eb2c43870b94b1080872028608f311e7624..a296edd7e1c34059d734752613fa05ba7aa6780f 100644 (file)
@@ -1175,7 +1175,11 @@ static int super_1_validate(mddev_t *mddev, mdk_rdev_t *rdev)
                        set_bit(Faulty, &rdev->flags);
                        break;
                default:
-                       set_bit(In_sync, &rdev->flags);
+                       if ((le32_to_cpu(sb->feature_map) &
+                            MD_FEATURE_RECOVERY_OFFSET))
+                               rdev->recovery_offset = le64_to_cpu(sb->recovery_offset);
+                       else
+                               set_bit(In_sync, &rdev->flags);
                        rdev->raid_disk = role;
                        break;
                }
@@ -1199,6 +1203,7 @@ static void super_1_sync(mddev_t *mddev, mdk_rdev_t *rdev)
 
        sb->feature_map = 0;
        sb->pad0 = 0;
+       sb->recovery_offset = cpu_to_le64(0);
        memset(sb->pad1, 0, sizeof(sb->pad1));
        memset(sb->pad2, 0, sizeof(sb->pad2));
        memset(sb->pad3, 0, sizeof(sb->pad3));
@@ -1219,6 +1224,14 @@ static void super_1_sync(mddev_t *mddev, mdk_rdev_t *rdev)
                sb->bitmap_offset = cpu_to_le32((__u32)mddev->bitmap_offset);
                sb->feature_map = cpu_to_le32(MD_FEATURE_BITMAP_OFFSET);
        }
+
+       if (rdev->raid_disk >= 0 &&
+           !test_bit(In_sync, &rdev->flags) &&
+           rdev->recovery_offset > 0) {
+               sb->feature_map |= cpu_to_le32(MD_FEATURE_RECOVERY_OFFSET);
+               sb->recovery_offset = cpu_to_le64(rdev->recovery_offset);
+       }
+
        if (mddev->reshape_position != MaxSector) {
                sb->feature_map |= cpu_to_le32(MD_FEATURE_RESHAPE_ACTIVE);
                sb->reshape_position = cpu_to_le64(mddev->reshape_position);
@@ -1243,11 +1256,12 @@ static void super_1_sync(mddev_t *mddev, mdk_rdev_t *rdev)
                        sb->dev_roles[i] = cpu_to_le16(0xfffe);
                else if (test_bit(In_sync, &rdev2->flags))
                        sb->dev_roles[i] = cpu_to_le16(rdev2->raid_disk);
+               else if (rdev2->raid_disk >= 0 && rdev2->recovery_offset > 0)
+                       sb->dev_roles[i] = cpu_to_le16(rdev2->raid_disk);
                else
                        sb->dev_roles[i] = cpu_to_le16(0xffff);
        }
 
-       sb->recovery_offset = cpu_to_le64(0); /* not supported yet */
        sb->sb_csum = calc_sb_1_csum(sb);
 }
 
@@ -2603,8 +2617,6 @@ static struct kobject *md_probe(dev_t dev, int *part, void *data)
        return NULL;
 }
 
-void md_wakeup_thread(mdk_thread_t *thread);
-
 static void md_safemode_timeout(unsigned long data)
 {
        mddev_t *mddev = (mddev_t *) data;
@@ -2786,6 +2798,36 @@ static int do_md_run(mddev_t * mddev)
        mddev->queue->queuedata = mddev;
        mddev->queue->make_request_fn = mddev->pers->make_request;
 
+       /* If there is a partially-recovered drive we need to
+        * start recovery here.  If we leave it to md_check_recovery,
+        * it will remove the drives and not do the right thing
+        */
+       if (mddev->degraded) {
+               struct list_head *rtmp;
+               int spares = 0;
+               ITERATE_RDEV(mddev,rdev,rtmp)
+                       if (rdev->raid_disk >= 0 &&
+                           !test_bit(In_sync, &rdev->flags) &&
+                           !test_bit(Faulty, &rdev->flags))
+                               /* complete an interrupted recovery */
+                               spares++;
+               if (spares && mddev->pers->sync_request) {
+                       mddev->recovery = 0;
+                       set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
+                       mddev->sync_thread = md_register_thread(md_do_sync,
+                                                               mddev,
+                                                               "%s_resync");
+                       if (!mddev->sync_thread) {
+                               printk(KERN_ERR "%s: could not start resync"
+                                      " thread...\n",
+                                      mdname(mddev));
+                               /* leave the spares where they are, it shouldn't hurt */
+                               mddev->recovery = 0;
+                       } else
+                               md_wakeup_thread(mddev->sync_thread);
+               }
+       }
+
        mddev->changed = 1;
        md_new_event(mddev);
        return 0;
@@ -2819,6 +2861,7 @@ static int restart_array(mddev_t *mddev)
                 */
                set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
                md_wakeup_thread(mddev->thread);
+               md_wakeup_thread(mddev->sync_thread);
                err = 0;
        } else {
                printk(KERN_ERR "md: %s has no personality assigned.\n",
@@ -2842,6 +2885,7 @@ static int do_md_stop(mddev_t * mddev, int ro)
                }
 
                if (mddev->sync_thread) {
+                       set_bit(MD_RECOVERY_FROZEN, &mddev->recovery);
                        set_bit(MD_RECOVERY_INTR, &mddev->recovery);
                        md_unregister_thread(mddev->sync_thread);
                        mddev->sync_thread = NULL;
@@ -2871,13 +2915,14 @@ static int do_md_stop(mddev_t * mddev, int ro)
                        if (mddev->ro)
                                mddev->ro = 0;
                }
-               if (!mddev->in_sync) {
+               if (!mddev->in_sync || mddev->sb_dirty) {
                        /* mark array as shutdown cleanly */
                        mddev->in_sync = 1;
                        md_update_sb(mddev);
                }
                if (ro)
                        set_disk_ro(disk, 1);
+               clear_bit(MD_RECOVERY_FROZEN, &mddev->recovery);
        }
 
        /*
@@ -4665,10 +4710,14 @@ void md_do_sync(mddev_t *mddev)
        struct list_head *tmp;
        sector_t last_check;
        int skipped = 0;
+       struct list_head *rtmp;
+       mdk_rdev_t *rdev;
 
        /* just incase thread restarts... */
        if (test_bit(MD_RECOVERY_DONE, &mddev->recovery))
                return;
+       if (mddev->ro) /* never try to sync a read-only array */
+               return;
 
        /* we overload curr_resync somewhat here.
         * 0 == not engaged in resync at all
@@ -4727,17 +4776,30 @@ void md_do_sync(mddev_t *mddev)
                }
        } while (mddev->curr_resync < 2);
 
+       j = 0;
        if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
                /* resync follows the size requested by the personality,
                 * which defaults to physical size, but can be virtual size
                 */
                max_sectors = mddev->resync_max_sectors;
                mddev->resync_mismatches = 0;
+               /* we don't use the checkpoint if there's a bitmap */
+               if (!mddev->bitmap &&
+                   !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
+                       j = mddev->recovery_cp;
        } else if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
                max_sectors = mddev->size << 1;
-       else
+       else {
                /* recovery follows the physical size of devices */
                max_sectors = mddev->size << 1;
+               j = MaxSector;
+               ITERATE_RDEV(mddev,rdev,rtmp)
+                       if (rdev->raid_disk >= 0 &&
+                           !test_bit(Faulty, &rdev->flags) &&
+                           !test_bit(In_sync, &rdev->flags) &&
+                           rdev->recovery_offset < j)
+                               j = rdev->recovery_offset;
+       }
 
        printk(KERN_INFO "md: syncing RAID array %s\n", mdname(mddev));
        printk(KERN_INFO "md: minimum _guaranteed_ reconstruction speed:"
@@ -4747,12 +4809,7 @@ void md_do_sync(mddev_t *mddev)
               speed_max(mddev));
 
        is_mddev_idle(mddev); /* this also initializes IO event counters */
-       /* we don't use the checkpoint if there's a bitmap */
-       if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && !mddev->bitmap
-           && ! test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
-               j = mddev->recovery_cp;
-       else
-               j = 0;
+
        io_sectors = 0;
        for (m = 0; m < SYNC_MARKS; m++) {
                mark[m] = jiffies;
@@ -4873,15 +4930,28 @@ void md_do_sync(mddev_t *mddev)
        if (!test_bit(MD_RECOVERY_ERR, &mddev->recovery) &&
            test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
            !test_bit(MD_RECOVERY_CHECK, &mddev->recovery) &&
-           mddev->curr_resync > 2 &&
-           mddev->curr_resync >= mddev->recovery_cp) {
-               if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
-                       printk(KERN_INFO 
-                               "md: checkpointing recovery of %s.\n",
-                               mdname(mddev));
-                       mddev->recovery_cp = mddev->curr_resync;
-               } else
-                       mddev->recovery_cp = MaxSector;
+           mddev->curr_resync > 2) {
+               if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
+                       if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
+                               if (mddev->curr_resync >= mddev->recovery_cp) {
+                                       printk(KERN_INFO
+                                              "md: checkpointing recovery of %s.\n",
+                                              mdname(mddev));
+                                       mddev->recovery_cp = mddev->curr_resync;
+                               }
+                       } else
+                               mddev->recovery_cp = MaxSector;
+               } else {
+                       if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery))
+                               mddev->curr_resync = MaxSector;
+                       ITERATE_RDEV(mddev,rdev,rtmp)
+                               if (rdev->raid_disk >= 0 &&
+                                   !test_bit(Faulty, &rdev->flags) &&
+                                   !test_bit(In_sync, &rdev->flags) &&
+                                   rdev->recovery_offset < mddev->curr_resync)
+                                       rdev->recovery_offset = mddev->curr_resync;
+                       mddev->sb_dirty = 1;
+               }
        }
 
  skip:
@@ -5002,6 +5072,8 @@ void md_check_recovery(mddev_t *mddev)
                clear_bit(MD_RECOVERY_INTR, &mddev->recovery);
                clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
 
+               if (test_bit(MD_RECOVERY_FROZEN, &mddev->recovery))
+                       goto unlock;
                /* no recovery is running.
                 * remove any failed drives, then
                 * add spares if possible.
@@ -5024,6 +5096,7 @@ void md_check_recovery(mddev_t *mddev)
                        ITERATE_RDEV(mddev,rdev,rtmp)
                                if (rdev->raid_disk < 0
                                    && !test_bit(Faulty, &rdev->flags)) {
+                                       rdev->recovery_offset = 0;
                                        if (mddev->pers->hot_add_disk(mddev,rdev)) {
                                                char nm[20];
                                                sprintf(nm, "rd%d", rdev->raid_disk);
index 4dd53fcb791a502e76595cd11aa95ab0ed351ed2..b3cfae41f7690d0cc8e09ba25896ebffc3b4e5ef 100644 (file)
@@ -1888,7 +1888,8 @@ static int run(mddev_t *mddev)
 
                disk = conf->mirrors + i;
 
-               if (!disk->rdev) {
+               if (!disk->rdev ||
+                   !test_bit(In_sync, &disk->rdev->flags)) {
                        disk->head_position = 0;
                        mddev->degraded++;
                }
index affeaefd4033f0fb0dec900d0c0d36b1f6402de6..2ca18770575fdc6df1327b8b9b2a1abccc960413 100644 (file)
@@ -2015,7 +2015,8 @@ static int run(mddev_t *mddev)
 
                disk = conf->mirrors + i;
 
-               if (!disk->rdev) {
+               if (!disk->rdev ||
+                   !test_bit(In_sync, &rdev->flags)) {
                        disk->head_position = 0;
                        mddev->degraded++;
                }
index 9ba73074df0475e795c71de4d783e9090df4da8b..970eb03ec6c0aa289662be59aa4268b6140c4119 100644 (file)
@@ -3555,6 +3555,7 @@ static int raid5_start_reshape(mddev_t *mddev)
                                set_bit(In_sync, &rdev->flags);
                                conf->working_disks++;
                                added_devices++;
+                               rdev->recovery_offset = 0;
                                sprintf(nm, "rd%d", rdev->raid_disk);
                                sysfs_create_link(&mddev->kobj, &rdev->kobj, nm);
                        } else
index db2ca2d9066e000670a5db52a4071e1578611298..682574f3bd364fee9324b805a12e53fae76467eb 100644 (file)
@@ -88,6 +88,10 @@ struct mdk_rdev_s
                                         * array and could again if we did a partial
                                         * resync from the bitmap
                                         */
+       sector_t        recovery_offset;/* If this device has been partially
+                                        * recovered, this is where we were
+                                        * up to.
+                                        */
 
        atomic_t        nr_pending;     /* number of pending requests.
                                         * only maintained for arrays that
@@ -183,6 +187,8 @@ struct mddev_s
 #define        MD_RECOVERY_REQUESTED   6
 #define        MD_RECOVERY_CHECK       7
 #define MD_RECOVERY_RESHAPE    8
+#define        MD_RECOVERY_FROZEN      9
+
        unsigned long                   recovery;
 
        int                             in_sync;        /* know to not need resync */
index f1fbae7e390ec4f53d67fa54e244153768449061..b6ebc69bae54d7b22402732355279d474701f699 100644 (file)
@@ -265,9 +265,12 @@ struct mdp_superblock_1 {
 
 /* feature_map bits */
 #define MD_FEATURE_BITMAP_OFFSET       1
+#define        MD_FEATURE_RECOVERY_OFFSET      2 /* recovery_offset is present and
+                                          * must be honoured
+                                          */
 #define        MD_FEATURE_RESHAPE_ACTIVE       4
 
-#define        MD_FEATURE_ALL                  5
+#define        MD_FEATURE_ALL                  (1|2|4)
 
 #endif