md/raid1: improve handling of read failure during recovery.
authorNeilBrown <neilb@suse.de>
Thu, 28 Jul 2011 01:33:42 +0000 (11:33 +1000)
committerNeilBrown <neilb@suse.de>
Thu, 28 Jul 2011 01:33:42 +0000 (11:33 +1000)
If we cannot read a block from anywhere during recovery, there is
now a better approach than just giving up.
We can record a bad block on each device and keep going - being
careful not to clear the bad block when a write succeeds as it might -
it will be a write of incorrect data.

We have now reached the state where - for raid1 - we only call
md_error if md_set_badblocks has failed.

Signed-off-by: NeilBrown <neilb@suse.de>
Reviewed-by: Namhyung Kim <namhyung@gmail.com>
drivers/md/raid1.c

index e6957151233b90a6a5884cf24272f8f6bf1ecf38..039e3af72929a57360177ed2b4c1cc4c2a8fc46e 100644 (file)
@@ -1392,7 +1392,12 @@ static void end_sync_write(struct bio *bio, int error)
        } else if (is_badblock(conf->mirrors[mirror].rdev,
                               r1_bio->sector,
                               r1_bio->sectors,
-                              &first_bad, &bad_sectors))
+                              &first_bad, &bad_sectors) &&
+                  !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
+                               r1_bio->sector,
+                               r1_bio->sectors,
+                               &first_bad, &bad_sectors)
+               )
                set_bit(R1BIO_MadeGood, &r1_bio->state);
 
        update_head_pos(mirror, r1_bio);
@@ -1473,16 +1478,36 @@ static int fix_sync_read_error(r1bio_t *r1_bio)
 
                if (!success) {
                        char b[BDEVNAME_SIZE];
-                       /* Cannot read from anywhere, array is toast */
-                       md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
+                       int abort = 0;
+                       /* Cannot read from anywhere, this block is lost.
+                        * Record a bad block on each device.  If that doesn't
+                        * work just disable and interrupt the recovery.
+                        * Don't fail devices as that won't really help.
+                        */
                        printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
                               " for block %llu\n",
                               mdname(mddev),
                               bdevname(bio->bi_bdev, b),
                               (unsigned long long)r1_bio->sector);
-                       md_done_sync(mddev, r1_bio->sectors, 0);
-                       put_buf(r1_bio);
-                       return 0;
+                       for (d = 0; d < conf->raid_disks; d++) {
+                               rdev = conf->mirrors[d].rdev;
+                               if (!rdev || test_bit(Faulty, &rdev->flags))
+                                       continue;
+                               if (!rdev_set_badblocks(rdev, sect, s, 0))
+                                       abort = 1;
+                       }
+                       if (abort) {
+                               mddev->recovery_disabled = 1;
+                               set_bit(MD_RECOVERY_INTR, &mddev->recovery);
+                               md_done_sync(mddev, r1_bio->sectors, 0);
+                               put_buf(r1_bio);
+                               return 0;
+                       }
+                       /* Try next page */
+                       sectors -= s;
+                       sect += s;
+                       idx++;
+                       continue;
                }
 
                start = d;
@@ -1879,7 +1904,9 @@ static void raid1d(mddev_t *mddev)
                                        if (bio->bi_end_io == NULL)
                                                continue;
                                        if (test_bit(BIO_UPTODATE,
-                                                    &bio->bi_flags)) {
+                                                    &bio->bi_flags) &&
+                                           test_bit(R1BIO_MadeGood,
+                                                    &r1_bio->state)) {
                                                rdev_clear_badblocks(
                                                        rdev,
                                                        r1_bio->sector,