md/raid5: Use correct data_offset for all IO.
authorNeilBrown <neilb@suse.de>
Sun, 20 May 2012 23:27:00 +0000 (09:27 +1000)
committerNeilBrown <neilb@suse.de>
Sun, 20 May 2012 23:27:00 +0000 (09:27 +1000)
As there can now be two different data_offsets - an 'old' and
a 'new' - we need to carefully choose between them.

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

index 3705585d756700739866350cdf498138b3902f33..71d1de909ba50666964ebbe2be983a1bdfe125cf 100644 (file)
@@ -488,6 +488,27 @@ get_active_stripe(struct r5conf *conf, sector_t sector,
        return sh;
 }
 
+/* Determine if 'data_offset' or 'new_data_offset' should be used
+ * in this stripe_head.
+ */
+static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
+{
+       sector_t progress = conf->reshape_progress;
+       /* Need a memory barrier to make sure we see the value
+        * of conf->generation, or ->data_offset that was set before
+        * reshape_progress was updated.
+        */
+       smp_rmb();
+       if (progress == MaxSector)
+               return 0;
+       if (sh->generation == conf->generation - 1)
+               return 0;
+       /* We are in a reshape, and this is a new-generation stripe,
+        * so use new_data_offset.
+        */
+       return 1;
+}
+
 static void
 raid5_end_read_request(struct bio *bi, int error);
 static void
@@ -603,7 +624,12 @@ static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
                                __func__, (unsigned long long)sh->sector,
                                bi->bi_rw, i);
                        atomic_inc(&sh->count);
-                       bi->bi_sector = sh->sector + rdev->data_offset;
+                       if (use_new_offset(conf, sh))
+                               bi->bi_sector = (sh->sector
+                                                + rdev->new_data_offset);
+                       else
+                               bi->bi_sector = (sh->sector
+                                                + rdev->data_offset);
                        bi->bi_flags = 1 << BIO_UPTODATE;
                        bi->bi_idx = 0;
                        bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
@@ -627,7 +653,12 @@ static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
                                __func__, (unsigned long long)sh->sector,
                                rbi->bi_rw, i);
                        atomic_inc(&sh->count);
-                       rbi->bi_sector = sh->sector + rrdev->data_offset;
+                       if (use_new_offset(conf, sh))
+                               rbi->bi_sector = (sh->sector
+                                                 + rrdev->new_data_offset);
+                       else
+                               rbi->bi_sector = (sh->sector
+                                                 + rrdev->data_offset);
                        rbi->bi_flags = 1 << BIO_UPTODATE;
                        rbi->bi_idx = 0;
                        rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
@@ -1648,7 +1679,7 @@ static void raid5_end_read_request(struct bio * bi, int error)
        int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
        char b[BDEVNAME_SIZE];
        struct md_rdev *rdev = NULL;
-
+       sector_t s;
 
        for (i=0 ; i<disks; i++)
                if (bi == &sh->dev[i].req)
@@ -1671,6 +1702,10 @@ static void raid5_end_read_request(struct bio * bi, int error)
        if (!rdev)
                rdev = conf->disks[i].rdev;
 
+       if (use_new_offset(conf, sh))
+               s = sh->sector + rdev->new_data_offset;
+       else
+               s = sh->sector + rdev->data_offset;
        if (uptodate) {
                set_bit(R5_UPTODATE, &sh->dev[i].flags);
                if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
@@ -1683,8 +1718,7 @@ static void raid5_end_read_request(struct bio * bi, int error)
                                "md/raid:%s: read error corrected"
                                " (%lu sectors at %llu on %s)\n",
                                mdname(conf->mddev), STRIPE_SECTORS,
-                               (unsigned long long)(sh->sector
-                                                    + rdev->data_offset),
+                               (unsigned long long)s,
                                bdevname(rdev->bdev, b));
                        atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
                        clear_bit(R5_ReadError, &sh->dev[i].flags);
@@ -1704,8 +1738,7 @@ static void raid5_end_read_request(struct bio * bi, int error)
                                "md/raid:%s: read error on replacement device "
                                "(sector %llu on %s).\n",
                                mdname(conf->mddev),
-                               (unsigned long long)(sh->sector
-                                                    + rdev->data_offset),
+                               (unsigned long long)s,
                                bdn);
                else if (conf->mddev->degraded >= conf->max_degraded)
                        printk_ratelimited(
@@ -1713,8 +1746,7 @@ static void raid5_end_read_request(struct bio * bi, int error)
                                "md/raid:%s: read error not correctable "
                                "(sector %llu on %s).\n",
                                mdname(conf->mddev),
-                               (unsigned long long)(sh->sector
-                                                    + rdev->data_offset),
+                               (unsigned long long)s,
                                bdn);
                else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
                        /* Oh, no!!! */
@@ -1723,8 +1755,7 @@ static void raid5_end_read_request(struct bio * bi, int error)
                                "md/raid:%s: read error NOT corrected!! "
                                "(sector %llu on %s).\n",
                                mdname(conf->mddev),
-                               (unsigned long long)(sh->sector
-                                                    + rdev->data_offset),
+                               (unsigned long long)s,
                                bdn);
                else if (atomic_read(&rdev->read_errors)
                         > conf->max_nr_stripes)
@@ -3842,6 +3873,7 @@ static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
                raid_bio->bi_next = (void*)rdev;
                align_bi->bi_bdev =  rdev->bdev;
                align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
+               /* No reshape active, so we can trust rdev->data_offset */
                align_bi->bi_sector += rdev->data_offset;
 
                if (!bio_fits_rdev(align_bi) ||
@@ -5182,9 +5214,12 @@ static int run(struct mddev *mddev)
                blk_queue_io_opt(mddev->queue, chunk_size *
                                 (conf->raid_disks - conf->max_degraded));
 
-               rdev_for_each(rdev, mddev)
+               rdev_for_each(rdev, mddev) {
                        disk_stack_limits(mddev->gendisk, rdev->bdev,
                                          rdev->data_offset << 9);
+                       disk_stack_limits(mddev->gendisk, rdev->bdev,
+                                         rdev->new_data_offset << 9);
+               }
        }
 
        return 0;
@@ -5539,12 +5574,16 @@ static int raid5_start_reshape(struct mddev *mddev)
        conf->chunk_sectors = mddev->new_chunk_sectors;
        conf->prev_algo = conf->algorithm;
        conf->algorithm = mddev->new_layout;
+       conf->generation++;
+       /* Code that selects data_offset needs to see the generation update
+        * if reshape_progress has been set - so a memory barrier needed.
+        */
+       smp_mb();
        if (mddev->reshape_backwards)
                conf->reshape_progress = raid5_size(mddev, 0, 0);
        else
                conf->reshape_progress = 0;
        conf->reshape_safe = conf->reshape_progress;
-       conf->generation++;
        spin_unlock_irq(&conf->device_lock);
 
        /* Add some new drives, as many as will fit.
@@ -5596,6 +5635,9 @@ static int raid5_start_reshape(struct mddev *mddev)
                mddev->recovery = 0;
                spin_lock_irq(&conf->device_lock);
                mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
+               rdev_for_each(rdev, mddev)
+                       rdev->new_data_offset = rdev->data_offset;
+               smp_wmb();
                conf->reshape_progress = MaxSector;
                mddev->reshape_position = MaxSector;
                spin_unlock_irq(&conf->device_lock);
@@ -5614,9 +5656,13 @@ static void end_reshape(struct r5conf *conf)
 {
 
        if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
+               struct md_rdev *rdev;
 
                spin_lock_irq(&conf->device_lock);
                conf->previous_raid_disks = conf->raid_disks;
+               rdev_for_each(rdev, conf->mddev)
+                       rdev->data_offset = rdev->new_data_offset;
+               smp_wmb();
                conf->reshape_progress = MaxSector;
                spin_unlock_irq(&conf->device_lock);
                wake_up(&conf->wait_for_overlap);