md/raid10: avoid writing to known bad blocks on known bad drives.
authorNeilBrown <neilb@suse.de>
Thu, 28 Jul 2011 01:39:24 +0000 (11:39 +1000)
committerNeilBrown <neilb@suse.de>
Thu, 28 Jul 2011 01:39:24 +0000 (11:39 +1000)
Writing to known bad blocks on drives that have seen a write error
is asking for trouble.  So try to avoid these blocks.

Signed-off-by: NeilBrown <neilb@suse.de>
drivers/md/raid10.c

index de60899..13077a3 100644 (file)
@@ -807,6 +807,8 @@ static int make_request(mddev_t *mddev, struct bio * bio)
        unsigned long flags;
        mdk_rdev_t *blocked_rdev;
        int plugged;
+       int sectors_handled;
+       int max_sectors;
 
        if (unlikely(bio->bi_rw & REQ_FLUSH)) {
                md_flush_request(mddev, bio);
@@ -895,7 +897,6 @@ static int make_request(mddev_t *mddev, struct bio * bio)
                /*
                 * read balancing logic:
                 */
-               int max_sectors;
                int disk;
                int slot;
 
@@ -925,8 +926,6 @@ read_again:
                        /* Could not read all from this device, so we will
                         * need another r10_bio.
                         */
-                       int sectors_handled;
-
                        sectors_handled = (r10_bio->sectors + max_sectors
                                           - bio->bi_sector);
                        r10_bio->sectors = max_sectors;
@@ -963,13 +962,22 @@ read_again:
        /* first select target devices under rcu_lock and
         * inc refcount on their rdev.  Record them by setting
         * bios[x] to bio
+        * If there are known/acknowledged bad blocks on any device
+        * on which we have seen a write error, we want to avoid
+        * writing to those blocks.  This potentially requires several
+        * writes to write around the bad blocks.  Each set of writes
+        * gets its own r10_bio with a set of bios attached.  The number
+        * of r10_bios is recored in bio->bi_phys_segments just as with
+        * the read case.
         */
        plugged = mddev_check_plugged(mddev);
 
        raid10_find_phys(conf, r10_bio);
- retry_write:
+retry_write:
        blocked_rdev = NULL;
        rcu_read_lock();
+       max_sectors = r10_bio->sectors;
+
        for (i = 0;  i < conf->copies; i++) {
                int d = r10_bio->devs[i].devnum;
                mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
@@ -978,13 +986,55 @@ read_again:
                        blocked_rdev = rdev;
                        break;
                }
-               if (rdev && !test_bit(Faulty, &rdev->flags)) {
-                       atomic_inc(&rdev->nr_pending);
-                       r10_bio->devs[i].bio = bio;
-               } else {
-                       r10_bio->devs[i].bio = NULL;
+               r10_bio->devs[i].bio = NULL;
+               if (!rdev || test_bit(Faulty, &rdev->flags)) {
                        set_bit(R10BIO_Degraded, &r10_bio->state);
+                       continue;
+               }
+               if (test_bit(WriteErrorSeen, &rdev->flags)) {
+                       sector_t first_bad;
+                       sector_t dev_sector = r10_bio->devs[i].addr;
+                       int bad_sectors;
+                       int is_bad;
+
+                       is_bad = is_badblock(rdev, dev_sector,
+                                            max_sectors,
+                                            &first_bad, &bad_sectors);
+                       if (is_bad < 0) {
+                               /* Mustn't write here until the bad block
+                                * is acknowledged
+                                */
+                               atomic_inc(&rdev->nr_pending);
+                               set_bit(BlockedBadBlocks, &rdev->flags);
+                               blocked_rdev = rdev;
+                               break;
+                       }
+                       if (is_bad && first_bad <= dev_sector) {
+                               /* Cannot write here at all */
+                               bad_sectors -= (dev_sector - first_bad);
+                               if (bad_sectors < max_sectors)
+                                       /* Mustn't write more than bad_sectors
+                                        * to other devices yet
+                                        */
+                                       max_sectors = bad_sectors;
+                               /* We don't set R10BIO_Degraded as that
+                                * only applies if the disk is missing,
+                                * so it might be re-added, and we want to
+                                * know to recover this chunk.
+                                * In this case the device is here, and the
+                                * fact that this chunk is not in-sync is
+                                * recorded in the bad block log.
+                                */
+                               continue;
+                       }
+                       if (is_bad) {
+                               int good_sectors = first_bad - dev_sector;
+                               if (good_sectors < max_sectors)
+                                       max_sectors = good_sectors;
+                       }
                }
+               r10_bio->devs[i].bio = bio;
+               atomic_inc(&rdev->nr_pending);
        }
        rcu_read_unlock();
 
@@ -1004,8 +1054,22 @@ read_again:
                goto retry_write;
        }
 
+       if (max_sectors < r10_bio->sectors) {
+               /* We are splitting this into multiple parts, so
+                * we need to prepare for allocating another r10_bio.
+                */
+               r10_bio->sectors = max_sectors;
+               spin_lock_irq(&conf->device_lock);
+               if (bio->bi_phys_segments == 0)
+                       bio->bi_phys_segments = 2;
+               else
+                       bio->bi_phys_segments++;
+               spin_unlock_irq(&conf->device_lock);
+       }
+       sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector;
+
        atomic_set(&r10_bio->remaining, 1);
-       bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
+       bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
 
        for (i = 0; i < conf->copies; i++) {
                struct bio *mbio;
@@ -1014,10 +1078,12 @@ read_again:
                        continue;
 
                mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
+               md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
+                           max_sectors);
                r10_bio->devs[i].bio = mbio;
 
-               mbio->bi_sector = r10_bio->devs[i].addr+
-                       conf->mirrors[d].rdev->data_offset;
+               mbio->bi_sector = (r10_bio->devs[i].addr+
+                                  conf->mirrors[d].rdev->data_offset);
                mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
                mbio->bi_end_io = raid10_end_write_request;
                mbio->bi_rw = WRITE | do_sync | do_fua;
@@ -1042,6 +1108,21 @@ read_again:
        /* In case raid10d snuck in to freeze_array */
        wake_up(&conf->wait_barrier);
 
+       if (sectors_handled < (bio->bi_size >> 9)) {
+               /* We need another r1_bio.  It has already been counted
+                * in bio->bi_phys_segments.
+                */
+               r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
+
+               r10_bio->master_bio = bio;
+               r10_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
+
+               r10_bio->mddev = mddev;
+               r10_bio->sector = bio->bi_sector + sectors_handled;
+               r10_bio->state = 0;
+               goto retry_write;
+       }
+
        if (do_sync || !mddev->bitmap || !plugged)
                md_wakeup_thread(mddev->thread);
        return 0;