Btrfs: fix device replace of a missing RAID 5/6 device
authorOmar Sandoval <osandov@fb.com>
Fri, 19 Jun 2015 18:52:51 +0000 (11:52 -0700)
committerChris Mason <clm@fb.com>
Sun, 9 Aug 2015 14:34:26 +0000 (07:34 -0700)
The original implementation of device replace on RAID 5/6 seems to have
missed support for replacing a missing device. When this is attempted,
we end up calling bio_add_page() on a bio with a NULL ->bi_bdev, which
crashes when we try to dereference it. This happens because
btrfs_map_block() has no choice but to return us the missing device
because RAID 5/6 don't have any alternate mirrors to read from, and a
missing device has a NULL bdev.

The idea implemented here is to handle the missing device case
separately, which better only happen when we're replacing a missing RAID
5/6 device. We use the new BTRFS_RBIO_REBUILD_MISSING operation to
reconstruct the data from parity, check it with
scrub_recheck_block_checksum(), and write it out with
scrub_write_block_to_dev_replace().

Reported-by: Philip <bugzilla@philip-seeger.de>
Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=96141
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
fs/btrfs/scrub.c

index 038162456cfad0dac07f5e1b56f80c63f1517e84..6bce7f2ff80520d5c6a380553574b360887ae93d 100644 (file)
@@ -125,6 +125,7 @@ struct scrub_block {
                /* It is for the data with checksum */
                unsigned int    data_corrected:1;
        };
+       struct btrfs_work       work;
 };
 
 /* Used for the chunks with parity stripe such RAID5/6 */
@@ -2173,6 +2174,134 @@ again:
        return 0;
 }
 
+static void scrub_missing_raid56_end_io(struct bio *bio, int error)
+{
+       struct scrub_block *sblock = bio->bi_private;
+       struct btrfs_fs_info *fs_info = sblock->sctx->dev_root->fs_info;
+
+       if (error)
+               sblock->no_io_error_seen = 0;
+
+       btrfs_queue_work(fs_info->scrub_workers, &sblock->work);
+}
+
+static void scrub_missing_raid56_worker(struct btrfs_work *work)
+{
+       struct scrub_block *sblock = container_of(work, struct scrub_block, work);
+       struct scrub_ctx *sctx = sblock->sctx;
+       struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
+       unsigned int is_metadata;
+       unsigned int have_csum;
+       u8 *csum;
+       u64 generation;
+       u64 logical;
+       struct btrfs_device *dev;
+
+       is_metadata = !(sblock->pagev[0]->flags & BTRFS_EXTENT_FLAG_DATA);
+       have_csum = sblock->pagev[0]->have_csum;
+       csum = sblock->pagev[0]->csum;
+       generation = sblock->pagev[0]->generation;
+       logical = sblock->pagev[0]->logical;
+       dev = sblock->pagev[0]->dev;
+
+       if (sblock->no_io_error_seen) {
+               scrub_recheck_block_checksum(fs_info, sblock, is_metadata,
+                                            have_csum, csum, generation,
+                                            sctx->csum_size);
+       }
+
+       if (!sblock->no_io_error_seen) {
+               spin_lock(&sctx->stat_lock);
+               sctx->stat.read_errors++;
+               spin_unlock(&sctx->stat_lock);
+               printk_ratelimited_in_rcu(KERN_ERR
+                       "BTRFS: I/O error rebulding logical %llu for dev %s\n",
+                       logical, rcu_str_deref(dev->name));
+       } else if (sblock->header_error || sblock->checksum_error) {
+               spin_lock(&sctx->stat_lock);
+               sctx->stat.uncorrectable_errors++;
+               spin_unlock(&sctx->stat_lock);
+               printk_ratelimited_in_rcu(KERN_ERR
+                       "BTRFS: failed to rebuild valid logical %llu for dev %s\n",
+                       logical, rcu_str_deref(dev->name));
+       } else {
+               scrub_write_block_to_dev_replace(sblock);
+       }
+
+       scrub_block_put(sblock);
+
+       if (sctx->is_dev_replace &&
+           atomic_read(&sctx->wr_ctx.flush_all_writes)) {
+               mutex_lock(&sctx->wr_ctx.wr_lock);
+               scrub_wr_submit(sctx);
+               mutex_unlock(&sctx->wr_ctx.wr_lock);
+       }
+
+       scrub_pending_bio_dec(sctx);
+}
+
+static void scrub_missing_raid56_pages(struct scrub_block *sblock)
+{
+       struct scrub_ctx *sctx = sblock->sctx;
+       struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
+       u64 length = sblock->page_count * PAGE_SIZE;
+       u64 logical = sblock->pagev[0]->logical;
+       struct btrfs_bio *bbio;
+       struct bio *bio;
+       struct btrfs_raid_bio *rbio;
+       int ret;
+       int i;
+
+       ret = btrfs_map_sblock(fs_info, REQ_GET_READ_MIRRORS, logical, &length,
+                              &bbio, 0, 1);
+       if (ret || !bbio || !bbio->raid_map)
+               goto bbio_out;
+
+       if (WARN_ON(!sctx->is_dev_replace ||
+                   !(bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
+               /*
+                * We shouldn't be scrubbing a missing device. Even for dev
+                * replace, we should only get here for RAID 5/6. We either
+                * managed to mount something with no mirrors remaining or
+                * there's a bug in scrub_remap_extent()/btrfs_map_block().
+                */
+               goto bbio_out;
+       }
+
+       bio = btrfs_io_bio_alloc(GFP_NOFS, 0);
+       if (!bio)
+               goto bbio_out;
+
+       bio->bi_iter.bi_sector = logical >> 9;
+       bio->bi_private = sblock;
+       bio->bi_end_io = scrub_missing_raid56_end_io;
+
+       rbio = raid56_alloc_missing_rbio(sctx->dev_root, bio, bbio, length);
+       if (!rbio)
+               goto rbio_out;
+
+       for (i = 0; i < sblock->page_count; i++) {
+               struct scrub_page *spage = sblock->pagev[i];
+
+               raid56_add_scrub_pages(rbio, spage->page, spage->logical);
+       }
+
+       btrfs_init_work(&sblock->work, btrfs_scrub_helper,
+                       scrub_missing_raid56_worker, NULL, NULL);
+       scrub_block_get(sblock);
+       scrub_pending_bio_inc(sctx);
+       raid56_submit_missing_rbio(rbio);
+       return;
+
+rbio_out:
+       bio_put(bio);
+bbio_out:
+       btrfs_put_bbio(bbio);
+       spin_lock(&sctx->stat_lock);
+       sctx->stat.malloc_errors++;
+       spin_unlock(&sctx->stat_lock);
+}
+
 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
                       u64 physical, struct btrfs_device *dev, u64 flags,
                       u64 gen, int mirror_num, u8 *csum, int force,
@@ -2236,19 +2365,27 @@ leave_nomem:
        }
 
        WARN_ON(sblock->page_count == 0);
-       for (index = 0; index < sblock->page_count; index++) {
-               struct scrub_page *spage = sblock->pagev[index];
-               int ret;
+       if (dev->missing) {
+               /*
+                * This case should only be hit for RAID 5/6 device replace. See
+                * the comment in scrub_missing_raid56_pages() for details.
+                */
+               scrub_missing_raid56_pages(sblock);
+       } else {
+               for (index = 0; index < sblock->page_count; index++) {
+                       struct scrub_page *spage = sblock->pagev[index];
+                       int ret;
 
-               ret = scrub_add_page_to_rd_bio(sctx, spage);
-               if (ret) {
-                       scrub_block_put(sblock);
-                       return ret;
+                       ret = scrub_add_page_to_rd_bio(sctx, spage);
+                       if (ret) {
+                               scrub_block_put(sblock);
+                               return ret;
+                       }
                }
-       }
 
-       if (force)
-               scrub_submit(sctx);
+               if (force)
+                       scrub_submit(sctx);
+       }
 
        /* last one frees, either here or in bio completion for last page */
        scrub_block_put(sblock);