Btrfs: change scrub to support big blocks
authorStefan Behrens <sbehrens@giantdisaster.de>
Tue, 27 Mar 2012 18:21:27 +0000 (14:21 -0400)
committerChris Mason <chris.mason@oracle.com>
Tue, 27 Mar 2012 18:21:27 +0000 (14:21 -0400)
Scrub used to be coded for nodesize == leafsize == sectorsize == PAGE_SIZE.
This is now changed to support sizes for nodesize and leafsize which are
N * PAGE_SIZE.

Signed-off-by: Stefan Behrens <sbehrens@giantdisaster.de>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
fs/btrfs/scrub.c

index e68bab4..5221e07 100644 (file)
  *  - add a mode to also read unallocated space
  */
 
+struct scrub_block;
 struct scrub_dev;
 
 #define SCRUB_PAGES_PER_BIO    16      /* 64k per bio */
 #define SCRUB_BIOS_PER_DEV     16      /* 1 MB per device in flight */
+#define SCRUB_MAX_PAGES_PER_BLOCK      16      /* 64k per node/leaf/sector */
 
 struct scrub_page {
+       struct scrub_block      *sblock;
+       struct page             *page;
+       struct block_device     *bdev;
        u64                     flags;  /* extent flags */
        u64                     generation;
-       int                     mirror_num;
-       int                     have_csum;
+       u64                     logical;
+       u64                     physical;
+       struct {
+               unsigned int    mirror_num:8;
+               unsigned int    have_csum:1;
+               unsigned int    io_error:1;
+       };
        u8                      csum[BTRFS_CSUM_SIZE];
 };
 
@@ -60,12 +70,25 @@ struct scrub_bio {
        int                     err;
        u64                     logical;
        u64                     physical;
-       struct scrub_page       spag[SCRUB_PAGES_PER_BIO];
-       u64                     count;
+       struct scrub_page       *pagev[SCRUB_PAGES_PER_BIO];
+       int                     page_count;
        int                     next_free;
        struct btrfs_work       work;
 };
 
+struct scrub_block {
+       struct scrub_page       pagev[SCRUB_MAX_PAGES_PER_BLOCK];
+       int                     page_count;
+       atomic_t                outstanding_pages;
+       atomic_t                ref_count; /* free mem on transition to zero */
+       struct scrub_dev        *sdev;
+       struct {
+               unsigned int    header_error:1;
+               unsigned int    checksum_error:1;
+               unsigned int    no_io_error_seen:1;
+       };
+};
+
 struct scrub_dev {
        struct scrub_bio        *bios[SCRUB_BIOS_PER_DEV];
        struct btrfs_device     *dev;
@@ -79,6 +102,10 @@ struct scrub_dev {
        struct list_head        csum_list;
        atomic_t                cancel_req;
        int                     readonly;
+       int                     pages_per_bio; /* <= SCRUB_PAGES_PER_BIO */
+       u32                     sectorsize;
+       u32                     nodesize;
+       u32                     leafsize;
        /*
         * statistics
         */
@@ -107,19 +134,41 @@ struct scrub_warning {
        int                     scratch_bufsize;
 };
 
+
+static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
+static int scrub_setup_recheck_block(struct scrub_dev *sdev,
+                                    struct btrfs_mapping_tree *map_tree,
+                                    u64 length, u64 logical,
+                                    struct scrub_block *sblock);
+static int scrub_recheck_block(struct btrfs_fs_info *fs_info,
+                              struct scrub_block *sblock, int is_metadata,
+                              int have_csum, u8 *csum, u64 generation,
+                              u16 csum_size);
+static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
+                                        struct scrub_block *sblock,
+                                        int is_metadata, int have_csum,
+                                        const u8 *csum, u64 generation,
+                                        u16 csum_size);
+static void scrub_complete_bio_end_io(struct bio *bio, int err);
+static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
+                                            struct scrub_block *sblock_good,
+                                            int force_write);
+static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
+                                           struct scrub_block *sblock_good,
+                                           int page_num, int force_write);
+static int scrub_checksum_data(struct scrub_block *sblock);
+static int scrub_checksum_tree_block(struct scrub_block *sblock);
+static int scrub_checksum_super(struct scrub_block *sblock);
+static void scrub_block_get(struct scrub_block *sblock);
+static void scrub_block_put(struct scrub_block *sblock);
+static int scrub_add_page_to_bio(struct scrub_dev *sdev,
+                                struct scrub_page *spage);
+static int scrub_pages(struct scrub_dev *sdev, u64 logical, u64 len,
+                      u64 physical, u64 flags, u64 gen, int mirror_num,
+                      u8 *csum, int force);
 static void scrub_bio_end_io(struct bio *bio, int err);
-static void scrub_checksum(struct btrfs_work *work);
-static int scrub_checksum_data(struct scrub_dev *sdev,
-                              struct scrub_page *spag, void *buffer);
-static int scrub_checksum_tree_block(struct scrub_dev *sdev,
-                                    struct scrub_page *spag, u64 logical,
-                                    void *buffer);
-static int scrub_checksum_super(struct scrub_bio *sbio, void *buffer);
-static int scrub_fixup_check(struct scrub_bio *sbio, int ix);
-static void scrub_fixup_end_io(struct bio *bio, int err);
-static int scrub_fixup_io(int rw, struct block_device *bdev, sector_t sector,
-                         struct page *page);
-static void scrub_fixup(struct scrub_bio *sbio, int ix);
+static void scrub_bio_end_io_worker(struct btrfs_work *work);
+static void scrub_block_complete(struct scrub_block *sblock);
 
 
 static void scrub_free_csums(struct scrub_dev *sdev)
@@ -133,23 +182,6 @@ static void scrub_free_csums(struct scrub_dev *sdev)
        }
 }
 
-static void scrub_free_bio(struct bio *bio)
-{
-       int i;
-       struct page *last_page = NULL;
-
-       if (!bio)
-               return;
-
-       for (i = 0; i < bio->bi_vcnt; ++i) {
-               if (bio->bi_io_vec[i].bv_page == last_page)
-                       continue;
-               last_page = bio->bi_io_vec[i].bv_page;
-               __free_page(last_page);
-       }
-       bio_put(bio);
-}
-
 static noinline_for_stack void scrub_free_dev(struct scrub_dev *sdev)
 {
        int i;
@@ -157,13 +189,23 @@ static noinline_for_stack void scrub_free_dev(struct scrub_dev *sdev)
        if (!sdev)
                return;
 
+       /* this can happen when scrub is cancelled */
+       if (sdev->curr != -1) {
+               struct scrub_bio *sbio = sdev->bios[sdev->curr];
+
+               for (i = 0; i < sbio->page_count; i++) {
+                       BUG_ON(!sbio->pagev[i]);
+                       BUG_ON(!sbio->pagev[i]->page);
+                       scrub_block_put(sbio->pagev[i]->sblock);
+               }
+               bio_put(sbio->bio);
+       }
+
        for (i = 0; i < SCRUB_BIOS_PER_DEV; ++i) {
                struct scrub_bio *sbio = sdev->bios[i];
 
                if (!sbio)
                        break;
-
-               scrub_free_bio(sbio->bio);
                kfree(sbio);
        }
 
@@ -177,11 +219,16 @@ struct scrub_dev *scrub_setup_dev(struct btrfs_device *dev)
        struct scrub_dev *sdev;
        int             i;
        struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
+       int pages_per_bio;
 
+       pages_per_bio = min_t(int, SCRUB_PAGES_PER_BIO,
+                             bio_get_nr_vecs(dev->bdev));
        sdev = kzalloc(sizeof(*sdev), GFP_NOFS);
        if (!sdev)
                goto nomem;
        sdev->dev = dev;
+       sdev->pages_per_bio = pages_per_bio;
+       sdev->curr = -1;
        for (i = 0; i < SCRUB_BIOS_PER_DEV; ++i) {
                struct scrub_bio *sbio;
 
@@ -192,8 +239,8 @@ struct scrub_dev *scrub_setup_dev(struct btrfs_device *dev)
 
                sbio->index = i;
                sbio->sdev = sdev;
-               sbio->count = 0;
-               sbio->work.func = scrub_checksum;
+               sbio->page_count = 0;
+               sbio->work.func = scrub_bio_end_io_worker;
 
                if (i != SCRUB_BIOS_PER_DEV-1)
                        sdev->bios[i]->next_free = i + 1;
@@ -201,7 +248,9 @@ struct scrub_dev *scrub_setup_dev(struct btrfs_device *dev)
                        sdev->bios[i]->next_free = -1;
        }
        sdev->first_free = 0;
-       sdev->curr = -1;
+       sdev->nodesize = dev->dev_root->nodesize;
+       sdev->leafsize = dev->dev_root->leafsize;
+       sdev->sectorsize = dev->dev_root->sectorsize;
        atomic_set(&sdev->in_flight, 0);
        atomic_set(&sdev->fixup_cnt, 0);
        atomic_set(&sdev->cancel_req, 0);
@@ -292,10 +341,9 @@ err:
        return 0;
 }
 
-static void scrub_print_warning(const char *errstr, struct scrub_bio *sbio,
-                               int ix)
+static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
 {
-       struct btrfs_device *dev = sbio->sdev->dev;
+       struct btrfs_device *dev = sblock->sdev->dev;
        struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
        struct btrfs_path *path;
        struct btrfs_key found_key;
@@ -314,8 +362,9 @@ static void scrub_print_warning(const char *errstr, struct scrub_bio *sbio,
 
        swarn.scratch_buf = kmalloc(bufsize, GFP_NOFS);
        swarn.msg_buf = kmalloc(bufsize, GFP_NOFS);
-       swarn.sector = (sbio->physical + ix * PAGE_SIZE) >> 9;
-       swarn.logical = sbio->logical + ix * PAGE_SIZE;
+       BUG_ON(sblock->page_count < 1);
+       swarn.sector = (sblock->pagev[0].physical) >> 9;
+       swarn.logical = sblock->pagev[0].logical;
        swarn.errstr = errstr;
        swarn.dev = dev;
        swarn.msg_bufsize = bufsize;
@@ -530,9 +579,9 @@ out:
                spin_lock(&sdev->stat_lock);
                ++sdev->stat.uncorrectable_errors;
                spin_unlock(&sdev->stat_lock);
-               printk_ratelimited(KERN_ERR "btrfs: unable to fixup "
-                                       "(nodatasum) error at logical %llu\n",
-                                       fixup->logical);
+               printk_ratelimited(KERN_ERR
+                       "btrfs: unable to fixup (nodatasum) error at logical %llu on dev %s\n",
+                       (unsigned long long)fixup->logical, sdev->dev->name);
        }
 
        btrfs_free_path(path);
@@ -549,91 +598,168 @@ out:
 }
 
 /*
- * scrub_recheck_error gets called when either verification of the page
- * failed or the bio failed to read, e.g. with EIO. In the latter case,
- * recheck_error gets called for every page in the bio, even though only
- * one may be bad
+ * scrub_handle_errored_block gets called when either verification of the
+ * pages failed or the bio failed to read, e.g. with EIO. In the latter
+ * case, this function handles all pages in the bio, even though only one
+ * may be bad.
+ * The goal of this function is to repair the errored block by using the
+ * contents of one of the mirrors.
  */
-static int scrub_recheck_error(struct scrub_bio *sbio, int ix)
+static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
 {
-       struct scrub_dev *sdev = sbio->sdev;
-       u64 sector = (sbio->physical + ix * PAGE_SIZE) >> 9;
+       struct scrub_dev *sdev = sblock_to_check->sdev;
+       struct btrfs_fs_info *fs_info;
+       u64 length;
+       u64 logical;
+       u64 generation;
+       unsigned int failed_mirror_index;
+       unsigned int is_metadata;
+       unsigned int have_csum;
+       u8 *csum;
+       struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
+       struct scrub_block *sblock_bad;
+       int ret;
+       int mirror_index;
+       int page_num;
+       int success;
        static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
-                                       DEFAULT_RATELIMIT_BURST);
+                                     DEFAULT_RATELIMIT_BURST);
+
+       BUG_ON(sblock_to_check->page_count < 1);
+       fs_info = sdev->dev->dev_root->fs_info;
+       length = sblock_to_check->page_count * PAGE_SIZE;
+       logical = sblock_to_check->pagev[0].logical;
+       generation = sblock_to_check->pagev[0].generation;
+       BUG_ON(sblock_to_check->pagev[0].mirror_num < 1);
+       failed_mirror_index = sblock_to_check->pagev[0].mirror_num - 1;
+       is_metadata = !(sblock_to_check->pagev[0].flags &
+                       BTRFS_EXTENT_FLAG_DATA);
+       have_csum = sblock_to_check->pagev[0].have_csum;
+       csum = sblock_to_check->pagev[0].csum;
 
-       if (sbio->err) {
-               if (scrub_fixup_io(READ, sbio->sdev->dev->bdev, sector,
-                                  sbio->bio->bi_io_vec[ix].bv_page) == 0) {
-                       if (scrub_fixup_check(sbio, ix) == 0)
-                               return 0;
-               }
-               if (__ratelimit(&_rs))
-                       scrub_print_warning("i/o error", sbio, ix);
-       } else {
-               if (__ratelimit(&_rs))
-                       scrub_print_warning("checksum error", sbio, ix);
+       /*
+        * read all mirrors one after the other. This includes to
+        * re-read the extent or metadata block that failed (that was
+        * the cause that this fixup code is called) another time,
+        * page by page this time in order to know which pages
+        * caused I/O errors and which ones are good (for all mirrors).
+        * It is the goal to handle the situation when more than one
+        * mirror contains I/O errors, but the errors do not
+        * overlap, i.e. the data can be repaired by selecting the
+        * pages from those mirrors without I/O error on the
+        * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
+        * would be that mirror #1 has an I/O error on the first page,
+        * the second page is good, and mirror #2 has an I/O error on
+        * the second page, but the first page is good.
+        * Then the first page of the first mirror can be repaired by
+        * taking the first page of the second mirror, and the
+        * second page of the second mirror can be repaired by
+        * copying the contents of the 2nd page of the 1st mirror.
+        * One more note: if the pages of one mirror contain I/O
+        * errors, the checksum cannot be verified. In order to get
+        * the best data for repairing, the first attempt is to find
+        * a mirror without I/O errors and with a validated checksum.
+        * Only if this is not possible, the pages are picked from
+        * mirrors with I/O errors without considering the checksum.
+        * If the latter is the case, at the end, the checksum of the
+        * repaired area is verified in order to correctly maintain
+        * the statistics.
+        */
+
+       sblocks_for_recheck = kzalloc(BTRFS_MAX_MIRRORS *
+                                    sizeof(*sblocks_for_recheck),
+                                    GFP_NOFS);
+       if (!sblocks_for_recheck) {
+               spin_lock(&sdev->stat_lock);
+               sdev->stat.malloc_errors++;
+               sdev->stat.read_errors++;
+               sdev->stat.uncorrectable_errors++;
+               spin_unlock(&sdev->stat_lock);
+               goto out;
        }
 
-       spin_lock(&sdev->stat_lock);
-       ++sdev->stat.read_errors;
-       spin_unlock(&sdev->stat_lock);
+       /* setup the context, map the logical blocks and alloc the pages */
+       ret = scrub_setup_recheck_block(sdev, &fs_info->mapping_tree, length,
+                                       logical, sblocks_for_recheck);
+       if (ret) {
+               spin_lock(&sdev->stat_lock);
+               sdev->stat.read_errors++;
+               sdev->stat.uncorrectable_errors++;
+               spin_unlock(&sdev->stat_lock);
+               goto out;
+       }
+       BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
+       sblock_bad = sblocks_for_recheck + failed_mirror_index;
 
-       scrub_fixup(sbio, ix);
-       return 1;
-}
+       /* build and submit the bios for the failed mirror, check checksums */
+       ret = scrub_recheck_block(fs_info, sblock_bad, is_metadata, have_csum,
+                                 csum, generation, sdev->csum_size);
+       if (ret) {
+               spin_lock(&sdev->stat_lock);
+               sdev->stat.read_errors++;
+               sdev->stat.uncorrectable_errors++;
+               spin_unlock(&sdev->stat_lock);
+               goto out;
+       }
 
-static int scrub_fixup_check(struct scrub_bio *sbio, int ix)
-{
-       int ret = 1;
-       struct page *page;
-       void *buffer;
-       u64 flags = sbio->spag[ix].flags;
+       if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
+           sblock_bad->no_io_error_seen) {
+               /*
+                * the error disappeared after reading page by page, or
+                * the area was part of a huge bio and other parts of the
+                * bio caused I/O errors, or the block layer merged several
+                * read requests into one and the error is caused by a
+                * different bio (usually one of the two latter cases is
+                * the cause)
+                */
+               spin_lock(&sdev->stat_lock);
+               sdev->stat.unverified_errors++;
+               spin_unlock(&sdev->stat_lock);
 
-       page = sbio->bio->bi_io_vec[ix].bv_page;
-       buffer = kmap_atomic(page, KM_USER0);
-       if (flags & BTRFS_EXTENT_FLAG_DATA) {
-               ret = scrub_checksum_data(sbio->sdev,
-                                         sbio->spag + ix, buffer);
-       } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
-               ret = scrub_checksum_tree_block(sbio->sdev,
-                                               sbio->spag + ix,
-                                               sbio->logical + ix * PAGE_SIZE,
-                                               buffer);
-       } else {
-               WARN_ON(1);
+               goto out;
        }
-       kunmap_atomic(buffer, KM_USER0);
 
-       return ret;
-}
+       if (!sblock_bad->no_io_error_seen) {
+               spin_lock(&sdev->stat_lock);
+               sdev->stat.read_errors++;
+               spin_unlock(&sdev->stat_lock);
+               if (__ratelimit(&_rs))
+                       scrub_print_warning("i/o error", sblock_to_check);
+       } else if (sblock_bad->checksum_error) {
+               spin_lock(&sdev->stat_lock);
+               sdev->stat.csum_errors++;
+               spin_unlock(&sdev->stat_lock);
+               if (__ratelimit(&_rs))
+                       scrub_print_warning("checksum error", sblock_to_check);
+       } else if (sblock_bad->header_error) {
+               spin_lock(&sdev->stat_lock);
+               sdev->stat.verify_errors++;
+               spin_unlock(&sdev->stat_lock);
+               if (__ratelimit(&_rs))
+                       scrub_print_warning("checksum/header error",
+                                           sblock_to_check);
+       }
 
-static void scrub_fixup_end_io(struct bio *bio, int err)
-{
-       complete((struct completion *)bio->bi_private);
-}
+       if (sdev->readonly)
+               goto did_not_correct_error;
+
+       if (!is_metadata && !have_csum) {
+               struct scrub_fixup_nodatasum *fixup_nodatasum;
 
-static void scrub_fixup(struct scrub_bio *sbio, int ix)
-{
-       struct scrub_dev *sdev = sbio->sdev;
-       struct btrfs_fs_info *fs_info = sdev->dev->dev_root->fs_info;
-       struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
-       struct btrfs_bio *bbio = NULL;
-       struct scrub_fixup_nodatasum *fixup;
-       u64 logical = sbio->logical + ix * PAGE_SIZE;
-       u64 length;
-       int i;
-       int ret;
-       DECLARE_COMPLETION_ONSTACK(complete);
-
-       if ((sbio->spag[ix].flags & BTRFS_EXTENT_FLAG_DATA) &&
-           (sbio->spag[ix].have_csum == 0)) {
-               fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
-               if (!fixup)
-                       goto uncorrectable;
-               fixup->sdev = sdev;
-               fixup->logical = logical;
-               fixup->root = fs_info->extent_root;
-               fixup->mirror_num = sbio->spag[ix].mirror_num;
+               /*
+                * !is_metadata and !have_csum, this means that the data
+                * might not be COW'ed, that it might be modified
+                * concurrently. The general strategy to work on the
+                * commit root does not help in the case when COW is not
+                * used.
+                */
+               fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
+               if (!fixup_nodatasum)
+                       goto did_not_correct_error;
+               fixup_nodatasum->sdev = sdev;
+               fixup_nodatasum->logical = logical;
+               fixup_nodatasum->root = fs_info->extent_root;
+               fixup_nodatasum->mirror_num = failed_mirror_index + 1;
                /*
                 * increment scrubs_running to prevent cancel requests from
                 * completing as long as a fixup worker is running. we must also
@@ -648,235 +774,529 @@ static void scrub_fixup(struct scrub_bio *sbio, int ix)
                atomic_inc(&fs_info->scrubs_paused);
                mutex_unlock(&fs_info->scrub_lock);
                atomic_inc(&sdev->fixup_cnt);
-               fixup->work.func = scrub_fixup_nodatasum;
-               btrfs_queue_worker(&fs_info->scrub_workers, &fixup->work);
-               return;
+               fixup_nodatasum->work.func = scrub_fixup_nodatasum;
+               btrfs_queue_worker(&fs_info->scrub_workers,
+                                  &fixup_nodatasum->work);
+               goto out;
        }
 
-       length = PAGE_SIZE;
-       ret = btrfs_map_block(map_tree, REQ_WRITE, logical, &length,
-                             &bbio, 0);
-       if (ret || !bbio || length < PAGE_SIZE) {
-               printk(KERN_ERR
-                      "scrub_fixup: btrfs_map_block failed us for %llu\n",
-                      (unsigned long long)logical);
-               WARN_ON(1);
-               kfree(bbio);
-               return;
+       /*
+        * now build and submit the bios for the other mirrors, check
+        * checksums
+        */
+       for (mirror_index = 0;
+            mirror_index < BTRFS_MAX_MIRRORS &&
+            sblocks_for_recheck[mirror_index].page_count > 0;
+            mirror_index++) {
+               if (mirror_index == failed_mirror_index)
+                       continue;
+
+               /* build and submit the bios, check checksums */
+               ret = scrub_recheck_block(fs_info,
+                                         sblocks_for_recheck + mirror_index,
+                                         is_metadata, have_csum, csum,
+                                         generation, sdev->csum_size);
+               if (ret)
+                       goto did_not_correct_error;
        }
 
-       if (bbio->num_stripes == 1)
-               /* there aren't any replicas */
-               goto uncorrectable;
+       /*
+        * first try to pick the mirror which is completely without I/O
+        * errors and also does not have a checksum error.
+        * If one is found, and if a checksum is present, the full block
+        * that is known to contain an error is rewritten. Afterwards
+        * the block is known to be corrected.
+        * If a mirror is found which is completely correct, and no
+        * checksum is present, only those pages are rewritten that had
+        * an I/O error in the block to be repaired, since it cannot be
+        * determined, which copy of the other pages is better (and it
+        * could happen otherwise that a correct page would be
+        * overwritten by a bad one).
+        */
+       for (mirror_index = 0;
+            mirror_index < BTRFS_MAX_MIRRORS &&
+            sblocks_for_recheck[mirror_index].page_count > 0;
+            mirror_index++) {
+               struct scrub_block *sblock_other = sblocks_for_recheck +
+                                                  mirror_index;
+
+               if (!sblock_other->header_error &&
+                   !sblock_other->checksum_error &&
+                   sblock_other->no_io_error_seen) {
+                       int force_write = is_metadata || have_csum;
+
+                       ret = scrub_repair_block_from_good_copy(sblock_bad,
+                                                               sblock_other,
+                                                               force_write);
+                       if (0 == ret)
+                               goto corrected_error;
+               }
+       }
 
        /*
-        * first find a good copy
+        * in case of I/O errors in the area that is supposed to be
+        * repaired, continue by picking good copies of those pages.
+        * Select the good pages from mirrors to rewrite bad pages from
+        * the area to fix. Afterwards verify the checksum of the block
+        * that is supposed to be repaired. This verification step is
+        * only done for the purpose of statistic counting and for the
+        * final scrub report, whether errors remain.
+        * A perfect algorithm could make use of the checksum and try
+        * all possible combinations of pages from the different mirrors
+        * until the checksum verification succeeds. For example, when
+        * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
+        * of mirror #2 is readable but the final checksum test fails,
+        * then the 2nd page of mirror #3 could be tried, whether now
+        * the final checksum succeedes. But this would be a rare
+        * exception and is therefore not implemented. At least it is
+        * avoided that the good copy is overwritten.
+        * A more useful improvement would be to pick the sectors
+        * without I/O error based on sector sizes (512 bytes on legacy
+        * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
+        * mirror could be repaired by taking 512 byte of a different
+        * mirror, even if other 512 byte sectors in the same PAGE_SIZE
+        * area are unreadable.
         */
-       for (i = 0; i < bbio->num_stripes; ++i) {
-               if (i + 1 == sbio->spag[ix].mirror_num)
-                       continue;
 
-               if (scrub_fixup_io(READ, bbio->stripes[i].dev->bdev,
-                                  bbio->stripes[i].physical >> 9,
-                                  sbio->bio->bi_io_vec[ix].bv_page)) {
-                       /* I/O-error, this is not a good copy */
+       /* can only fix I/O errors from here on */
+       if (sblock_bad->no_io_error_seen)
+               goto did_not_correct_error;
+
+       success = 1;
+       for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
+               struct scrub_page *page_bad = sblock_bad->pagev + page_num;
+
+               if (!page_bad->io_error)
                        continue;
+
+               for (mirror_index = 0;
+                    mirror_index < BTRFS_MAX_MIRRORS &&
+                    sblocks_for_recheck[mirror_index].page_count > 0;
+                    mirror_index++) {
+                       struct scrub_block *sblock_other = sblocks_for_recheck +
+                                                          mirror_index;
+                       struct scrub_page *page_other = sblock_other->pagev +
+                                                       page_num;
+
+                       if (!page_other->io_error) {
+                               ret = scrub_repair_page_from_good_copy(
+                                       sblock_bad, sblock_other, page_num, 0);
+                               if (0 == ret) {
+                                       page_bad->io_error = 0;
+                                       break; /* succeeded for this page */
+                               }
+                       }
                }
 
-               if (scrub_fixup_check(sbio, ix) == 0)
-                       break;
+               if (page_bad->io_error) {
+                       /* did not find a mirror to copy the page from */
+                       success = 0;
+               }
        }
-       if (i == bbio->num_stripes)
-               goto uncorrectable;
 
-       if (!sdev->readonly) {
-               /*
-                * bi_io_vec[ix].bv_page now contains good data, write it back
-                */
-               if (scrub_fixup_io(WRITE, sdev->dev->bdev,
-                                  (sbio->physical + ix * PAGE_SIZE) >> 9,
-                                  sbio->bio->bi_io_vec[ix].bv_page)) {
-                       /* I/O-error, writeback failed, give up */
-                       goto uncorrectable;
+       if (success) {
+               if (is_metadata || have_csum) {
+                       /*
+                        * need to verify the checksum now that all
+                        * sectors on disk are repaired (the write
+                        * request for data to be repaired is on its way).
+                        * Just be lazy and use scrub_recheck_block()
+                        * which re-reads the data before the checksum
+                        * is verified, but most likely the data comes out
+                        * of the page cache.
+                        */
+                       ret = scrub_recheck_block(fs_info, sblock_bad,
+                                                 is_metadata, have_csum, csum,
+                                                 generation, sdev->csum_size);
+                       if (!ret && !sblock_bad->header_error &&
+                           !sblock_bad->checksum_error &&
+                           sblock_bad->no_io_error_seen)
+                               goto corrected_error;
+                       else
+                               goto did_not_correct_error;
+               } else {
+corrected_error:
+                       spin_lock(&sdev->stat_lock);
+                       sdev->stat.corrected_errors++;
+                       spin_unlock(&sdev->stat_lock);
+                       printk_ratelimited(KERN_ERR
+                               "btrfs: fixed up error at logical %llu on dev %s\n",
+                               (unsigned long long)logical, sdev->dev->name);
                }
+       } else {
+did_not_correct_error:
+               spin_lock(&sdev->stat_lock);
+               sdev->stat.uncorrectable_errors++;
+               spin_unlock(&sdev->stat_lock);
+               printk_ratelimited(KERN_ERR
+                       "btrfs: unable to fixup (regular) error at logical %llu on dev %s\n",
+                       (unsigned long long)logical, sdev->dev->name);
        }
 
-       kfree(bbio);
-       spin_lock(&sdev->stat_lock);
-       ++sdev->stat.corrected_errors;
-       spin_unlock(&sdev->stat_lock);
+out:
+       if (sblocks_for_recheck) {
+               for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
+                    mirror_index++) {
+                       struct scrub_block *sblock = sblocks_for_recheck +
+                                                    mirror_index;
+                       int page_index;
+
+                       for (page_index = 0; page_index < SCRUB_PAGES_PER_BIO;
+                            page_index++)
+                               if (sblock->pagev[page_index].page)
+                                       __free_page(
+                                               sblock->pagev[page_index].page);
+               }
+               kfree(sblocks_for_recheck);
+       }
 
-       printk_ratelimited(KERN_ERR "btrfs: fixed up error at logical %llu\n",
-                              (unsigned long long)logical);
-       return;
+       return 0;
+}
 
-uncorrectable:
-       kfree(bbio);
-       spin_lock(&sdev->stat_lock);
-       ++sdev->stat.uncorrectable_errors;
-       spin_unlock(&sdev->stat_lock);
+static int scrub_setup_recheck_block(struct scrub_dev *sdev,
+                                    struct btrfs_mapping_tree *map_tree,
+                                    u64 length, u64 logical,
+                                    struct scrub_block *sblocks_for_recheck)
+{
+       int page_index;
+       int mirror_index;
+       int ret;
+
+       /*
+        * note: the three members sdev, ref_count and outstanding_pages
+        * are not used (and not set) in the blocks that are used for
+        * the recheck procedure
+        */
+
+       page_index = 0;
+       while (length > 0) {
+               u64 sublen = min_t(u64, length, PAGE_SIZE);
+               u64 mapped_length = sublen;
+               struct btrfs_bio *bbio = NULL;
 
-       printk_ratelimited(KERN_ERR "btrfs: unable to fixup (regular) error at "
-                               "logical %llu\n", (unsigned long long)logical);
+               /*
+                * with a length of PAGE_SIZE, each returned stripe
+                * represents one mirror
+                */
+               ret = btrfs_map_block(map_tree, WRITE, logical, &mapped_length,
+                                     &bbio, 0);
+               if (ret || !bbio || mapped_length < sublen) {
+                       kfree(bbio);
+                       return -EIO;
+               }
+
+               BUG_ON(page_index >= SCRUB_PAGES_PER_BIO);
+               for (mirror_index = 0; mirror_index < (int)bbio->num_stripes;
+                    mirror_index++) {
+                       struct scrub_block *sblock;
+                       struct scrub_page *page;
+
+                       if (mirror_index >= BTRFS_MAX_MIRRORS)
+                               continue;
+
+                       sblock = sblocks_for_recheck + mirror_index;
+                       page = sblock->pagev + page_index;
+                       page->logical = logical;
+                       page->physical = bbio->stripes[mirror_index].physical;
+                       page->bdev = bbio->stripes[mirror_index].dev->bdev;
+                       page->mirror_num = mirror_index + 1;
+                       page->page = alloc_page(GFP_NOFS);
+                       if (!page->page) {
+                               spin_lock(&sdev->stat_lock);
+                               sdev->stat.malloc_errors++;
+                               spin_unlock(&sdev->stat_lock);
+                               return -ENOMEM;
+                       }
+                       sblock->page_count++;
+               }
+               kfree(bbio);
+               length -= sublen;
+               logical += sublen;
+               page_index++;
+       }
+
+       return 0;
 }
 
-static int scrub_fixup_io(int rw, struct block_device *bdev, sector_t sector,
-                        struct page *page)
+/*
+ * this function will check the on disk data for checksum errors, header
+ * errors and read I/O errors. If any I/O errors happen, the exact pages
+ * which are errored are marked as being bad. The goal is to enable scrub
+ * to take those pages that are not errored from all the mirrors so that
+ * the pages that are errored in the just handled mirror can be repaired.
+ */
+static int scrub_recheck_block(struct btrfs_fs_info *fs_info,
+                              struct scrub_block *sblock, int is_metadata,
+                              int have_csum, u8 *csum, u64 generation,
+                              u16 csum_size)
 {
-       struct bio *bio = NULL;
-       int ret;
-       DECLARE_COMPLETION_ONSTACK(complete);
+       int page_num;
 
-       bio = bio_alloc(GFP_NOFS, 1);
-       bio->bi_bdev = bdev;
-       bio->bi_sector = sector;
-       bio_add_page(bio, page, PAGE_SIZE, 0);
-       bio->bi_end_io = scrub_fixup_end_io;
-       bio->bi_private = &complete;
-       btrfsic_submit_bio(rw, bio);
+       sblock->no_io_error_seen = 1;
+       sblock->header_error = 0;
+       sblock->checksum_error = 0;
 
-       /* this will also unplug the queue */
-       wait_for_completion(&complete);
+       for (page_num = 0; page_num < sblock->page_count; page_num++) {
+               struct bio *bio;
+               int ret;
+               struct scrub_page *page = sblock->pagev + page_num;
+               DECLARE_COMPLETION_ONSTACK(complete);
+
+               BUG_ON(!page->page);
+               bio = bio_alloc(GFP_NOFS, 1);
+               bio->bi_bdev = page->bdev;
+               bio->bi_sector = page->physical >> 9;
+               bio->bi_end_io = scrub_complete_bio_end_io;
+               bio->bi_private = &complete;
+
+               ret = bio_add_page(bio, page->page, PAGE_SIZE, 0);
+               if (PAGE_SIZE != ret) {
+                       bio_put(bio);
+                       return -EIO;
+               }
+               btrfsic_submit_bio(READ, bio);
 
-       ret = !test_bit(BIO_UPTODATE, &bio->bi_flags);
-       bio_put(bio);
-       return ret;
+               /* this will also unplug the queue */
+               wait_for_completion(&complete);
+
+               page->io_error = !test_bit(BIO_UPTODATE, &bio->bi_flags);
+               if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
+                       sblock->no_io_error_seen = 0;
+               bio_put(bio);
+       }
+
+       if (sblock->no_io_error_seen)
+               scrub_recheck_block_checksum(fs_info, sblock, is_metadata,
+                                            have_csum, csum, generation,
+                                            csum_size);
+
+       return 0;
 }
 
-static void scrub_bio_end_io(struct bio *bio, int err)
+static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
+                                        struct scrub_block *sblock,
+                                        int is_metadata, int have_csum,
+                                        const u8 *csum, u64 generation,
+                                        u16 csum_size)
 {
-       struct scrub_bio *sbio = bio->bi_private;
-       struct scrub_dev *sdev = sbio->sdev;
-       struct btrfs_fs_info *fs_info = sdev->dev->dev_root->fs_info;
+       int page_num;
+       u8 calculated_csum[BTRFS_CSUM_SIZE];
+       u32 crc = ~(u32)0;
+       struct btrfs_root *root = fs_info->extent_root;
+       void *mapped_buffer;
+
+       BUG_ON(!sblock->pagev[0].page);
+       if (is_metadata) {
+               struct btrfs_header *h;
+
+               mapped_buffer = kmap_atomic(sblock->pagev[0].page, KM_USER0);
+               h = (struct btrfs_header *)mapped_buffer;
+
+               if (sblock->pagev[0].logical != le64_to_cpu(h->bytenr) ||
+                   generation != le64_to_cpu(h->generation) ||
+                   memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE) ||
+                   memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
+                          BTRFS_UUID_SIZE))
+                       sblock->header_error = 1;
+               csum = h->csum;
+       } else {
+               if (!have_csum)
+                       return;
 
-       sbio->err = err;
-       sbio->bio = bio;
+               mapped_buffer = kmap_atomic(sblock->pagev[0].page, KM_USER0);
+       }
 
-       btrfs_queue_worker(&fs_info->scrub_workers, &sbio->work);
+       for (page_num = 0;;) {
+               if (page_num == 0 && is_metadata)
+                       crc = btrfs_csum_data(root,
+                               ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE,
+                               crc, PAGE_SIZE - BTRFS_CSUM_SIZE);
+               else
+                       crc = btrfs_csum_data(root, mapped_buffer, crc,
+                                             PAGE_SIZE);
+
+               kunmap_atomic(mapped_buffer, KM_USER0);
+               page_num++;
+               if (page_num >= sblock->page_count)
+                       break;
+               BUG_ON(!sblock->pagev[page_num].page);
+
+               mapped_buffer = kmap_atomic(sblock->pagev[page_num].page,
+                                           KM_USER0);
+       }
+
+       btrfs_csum_final(crc, calculated_csum);
+       if (memcmp(calculated_csum, csum, csum_size))
+               sblock->checksum_error = 1;
 }
 
-static void scrub_checksum(struct btrfs_work *work)
+static void scrub_complete_bio_end_io(struct bio *bio, int err)
 {
-       struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
-       struct scrub_dev *sdev = sbio->sdev;
-       struct page *page;
-       void *buffer;
-       int i;
-       u64 flags;
-       u64 logical;
-       int ret;
+       complete((struct completion *)bio->bi_private);
+}
 
-       if (sbio->err) {
-               ret = 0;
-               for (i = 0; i < sbio->count; ++i)
-                       ret |= scrub_recheck_error(sbio, i);
-               if (!ret) {
-                       spin_lock(&sdev->stat_lock);
-                       ++sdev->stat.unverified_errors;
-                       spin_unlock(&sdev->stat_lock);
-               }
+static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
+                                            struct scrub_block *sblock_good,
+                                            int force_write)
+{
+       int page_num;
+       int ret = 0;
 
-               sbio->bio->bi_flags &= ~(BIO_POOL_MASK - 1);
-               sbio->bio->bi_flags |= 1 << BIO_UPTODATE;
-               sbio->bio->bi_phys_segments = 0;
-               sbio->bio->bi_idx = 0;
+       for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
+               int ret_sub;
 
-               for (i = 0; i < sbio->count; i++) {
-                       struct bio_vec *bi;
-                       bi = &sbio->bio->bi_io_vec[i];
-                       bi->bv_offset = 0;
-                       bi->bv_len = PAGE_SIZE;
-               }
-               goto out;
+               ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
+                                                          sblock_good,
+                                                          page_num,
+                                                          force_write);
+               if (ret_sub)
+                       ret = ret_sub;
        }
-       for (i = 0; i < sbio->count; ++i) {
-               page = sbio->bio->bi_io_vec[i].bv_page;
-               buffer = kmap_atomic(page, KM_USER0);
-               flags = sbio->spag[i].flags;
-               logical = sbio->logical + i * PAGE_SIZE;
-               ret = 0;
-               if (flags & BTRFS_EXTENT_FLAG_DATA) {
-                       ret = scrub_checksum_data(sdev, sbio->spag + i, buffer);
-               } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
-                       ret = scrub_checksum_tree_block(sdev, sbio->spag + i,
-                                                       logical, buffer);
-               } else if (flags & BTRFS_EXTENT_FLAG_SUPER) {
-                       BUG_ON(i);
-                       (void)scrub_checksum_super(sbio, buffer);
-               } else {
-                       WARN_ON(1);
-               }
-               kunmap_atomic(buffer, KM_USER0);
-               if (ret) {
-                       ret = scrub_recheck_error(sbio, i);
-                       if (!ret) {
-                               spin_lock(&sdev->stat_lock);
-                               ++sdev->stat.unverified_errors;
-                               spin_unlock(&sdev->stat_lock);
-                       }
+
+       return ret;
+}
+
+static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
+                                           struct scrub_block *sblock_good,
+                                           int page_num, int force_write)
+{
+       struct scrub_page *page_bad = sblock_bad->pagev + page_num;
+       struct scrub_page *page_good = sblock_good->pagev + page_num;
+
+       BUG_ON(sblock_bad->pagev[page_num].page == NULL);
+       BUG_ON(sblock_good->pagev[page_num].page == NULL);
+       if (force_write || sblock_bad->header_error ||
+           sblock_bad->checksum_error || page_bad->io_error) {
+               struct bio *bio;
+               int ret;
+               DECLARE_COMPLETION_ONSTACK(complete);
+
+               bio = bio_alloc(GFP_NOFS, 1);
+               bio->bi_bdev = page_bad->bdev;
+               bio->bi_sector = page_bad->physical >> 9;
+               bio->bi_end_io = scrub_complete_bio_end_io;
+               bio->bi_private = &complete;
+
+               ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
+               if (PAGE_SIZE != ret) {
+                       bio_put(bio);
+                       return -EIO;
                }
+               btrfsic_submit_bio(WRITE, bio);
+
+               /* this will also unplug the queue */
+               wait_for_completion(&complete);
+               bio_put(bio);
        }
 
-out:
-       scrub_free_bio(sbio->bio);
-       sbio->bio = NULL;
-       spin_lock(&sdev->list_lock);
-       sbio->next_free = sdev->first_free;
-       sdev->first_free = sbio->index;
-       spin_unlock(&sdev->list_lock);
-       atomic_dec(&sdev->in_flight);
-       wake_up(&sdev->list_wait);
+       return 0;
+}
+
+static void scrub_checksum(struct scrub_block *sblock)
+{
+       u64 flags;
+       int ret;
+
+       BUG_ON(sblock->page_count < 1);
+       flags = sblock->pagev[0].flags;
+       ret = 0;
+       if (flags & BTRFS_EXTENT_FLAG_DATA)
+               ret = scrub_checksum_data(sblock);
+       else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
+               ret = scrub_checksum_tree_block(sblock);
+       else if (flags & BTRFS_EXTENT_FLAG_SUPER)
+               (void)scrub_checksum_super(sblock);
+       else
+               WARN_ON(1);
+       if (ret)
+               scrub_handle_errored_block(sblock);
 }
 
-static int scrub_checksum_data(struct scrub_dev *sdev,
-                              struct scrub_page *spag, void *buffer)
+static int scrub_checksum_data(struct scrub_block *sblock)
 {
+       struct scrub_dev *sdev = sblock->sdev;
        u8 csum[BTRFS_CSUM_SIZE];
+       u8 *on_disk_csum;
+       struct page *page;
+       void *buffer;
        u32 crc = ~(u32)0;
        int fail = 0;
        struct btrfs_root *root = sdev->dev->dev_root;
+       u64 len;
+       int index;
 
-       if (!spag->have_csum)
+       BUG_ON(sblock->page_count < 1);
+       if (!sblock->pagev[0].have_csum)
                return 0;
 
-       crc = btrfs_csum_data(root, buffer, crc, PAGE_SIZE);
+       on_disk_csum = sblock->pagev[0].csum;
+       page = sblock->pagev[0].page;
+       buffer = kmap_atomic(page, KM_USER0);
+
+       len = sdev->sectorsize;
+       index = 0;
+       for (;;) {
+               u64 l = min_t(u64, len, PAGE_SIZE);
+
+               crc = btrfs_csum_data(root, buffer, crc, l);
+               kunmap_atomic(buffer, KM_USER0);
+               len -= l;
+               if (len == 0)
+                       break;
+               index++;
+               BUG_ON(index >= sblock->page_count);
+               BUG_ON(!sblock->pagev[index].page);
+               page = sblock->pagev[index].page;
+               buffer = kmap_atomic(page, KM_USER0);
+       }
+
        btrfs_csum_final(crc, csum);
-       if (memcmp(csum, spag->csum, sdev->csum_size))
+       if (memcmp(csum, on_disk_csum, sdev->csum_size))
                fail = 1;
 
-       spin_lock(&sdev->stat_lock);
-       ++sdev->stat.data_extents_scrubbed;
-       sdev->stat.data_bytes_scrubbed += PAGE_SIZE;
-       if (fail)
+       if (fail) {
+               spin_lock(&sdev->stat_lock);
                ++sdev->stat.csum_errors;
-       spin_unlock(&sdev->stat_lock);
+               spin_unlock(&sdev->stat_lock);
+       }
 
        return fail;
 }
 
-static int scrub_checksum_tree_block(struct scrub_dev *sdev,
-                                    struct scrub_page *spag, u64 logical,
-                                    void *buffer)
+static int scrub_checksum_tree_block(struct scrub_block *sblock)
 {
+       struct scrub_dev *sdev = sblock->sdev;
        struct btrfs_header *h;
        struct btrfs_root *root = sdev->dev->dev_root;
        struct btrfs_fs_info *fs_info = root->fs_info;
-       u8 csum[BTRFS_CSUM_SIZE];
+       u8 calculated_csum[BTRFS_CSUM_SIZE];
+       u8 on_disk_csum[BTRFS_CSUM_SIZE];
+       struct page *page;
+       void *mapped_buffer;
+       u64 mapped_size;
+       void *p;
        u32 crc = ~(u32)0;
        int fail = 0;
        int crc_fail = 0;
+       u64 len;
+       int index;
+
+       BUG_ON(sblock->page_count < 1);
+       page = sblock->pagev[0].page;
+       mapped_buffer = kmap_atomic(page, KM_USER0);
+       h = (struct btrfs_header *)mapped_buffer;
+       memcpy(on_disk_csum, h->csum, sdev->csum_size);
 
        /*
         * we don't use the getter functions here, as we
         * a) don't have an extent buffer and
         * b) the page is already kmapped
         */
-       h = (struct btrfs_header *)buffer;
 
-       if (logical != le64_to_cpu(h->bytenr))
+       if (sblock->pagev[0].logical != le64_to_cpu(h->bytenr))
                ++fail;
 
-       if (spag->generation != le64_to_cpu(h->generation))
+       if (sblock->pagev[0].generation != le64_to_cpu(h->generation))
                ++fail;
 
        if (memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
@@ -886,51 +1306,99 @@ static int scrub_checksum_tree_block(struct scrub_dev *sdev,
                   BTRFS_UUID_SIZE))
                ++fail;
 
-       crc = btrfs_csum_data(root, buffer + BTRFS_CSUM_SIZE, crc,
-                             PAGE_SIZE - BTRFS_CSUM_SIZE);
-       btrfs_csum_final(crc, csum);
-       if (memcmp(csum, h->csum, sdev->csum_size))
+       BUG_ON(sdev->nodesize != sdev->leafsize);
+       len = sdev->nodesize - BTRFS_CSUM_SIZE;
+       mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
+       p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
+       index = 0;
+       for (;;) {
+               u64 l = min_t(u64, len, mapped_size);
+
+               crc = btrfs_csum_data(root, p, crc, l);
+               kunmap_atomic(mapped_buffer, KM_USER0);
+               len -= l;
+               if (len == 0)
+                       break;
+               index++;
+               BUG_ON(index >= sblock->page_count);
+               BUG_ON(!sblock->pagev[index].page);
+               page = sblock->pagev[index].page;
+               mapped_buffer = kmap_atomic(page, KM_USER0);
+               mapped_size = PAGE_SIZE;
+               p = mapped_buffer;
+       }
+
+       btrfs_csum_final(crc, calculated_csum);
+       if (memcmp(calculated_csum, on_disk_csum, sdev->csum_size))
                ++crc_fail;
 
-       spin_lock(&sdev->stat_lock);
-       ++sdev->stat.tree_extents_scrubbed;
-       sdev->stat.tree_bytes_scrubbed += PAGE_SIZE;
-       if (crc_fail)
-               ++sdev->stat.csum_errors;
-       if (fail)
-               ++sdev->stat.verify_errors;
-       spin_unlock(&sdev->stat_lock);
+       if (crc_fail || fail) {
+               spin_lock(&sdev->stat_lock);
+               if (crc_fail)
+                       ++sdev->stat.csum_errors;
+               if (fail)
+                       ++sdev->stat.verify_errors;
+               spin_unlock(&sdev->stat_lock);
+       }
 
        return fail || crc_fail;
 }
 
-static int scrub_checksum_super(struct scrub_bio *sbio, void *buffer)
+static int scrub_checksum_super(struct scrub_block *sblock)
 {
        struct btrfs_super_block *s;
-       u64 logical;
-       struct scrub_dev *sdev = sbio->sdev;
+       struct scrub_dev *sdev = sblock->sdev;
        struct btrfs_root *root = sdev->dev->dev_root;
        struct btrfs_fs_info *fs_info = root->fs_info;
-       u8 csum[BTRFS_CSUM_SIZE];
+       u8 calculated_csum[BTRFS_CSUM_SIZE];
+       u8 on_disk_csum[BTRFS_CSUM_SIZE];
+       struct page *page;
+       void *mapped_buffer;
+       u64 mapped_size;
+       void *p;
        u32 crc = ~(u32)0;
        int fail = 0;
+       u64 len;
+       int index;
 
-       s = (struct btrfs_super_block *)buffer;
-       logical = sbio->logical;
+       BUG_ON(sblock->page_count < 1);
+       page = sblock->pagev[0].page;
+       mapped_buffer = kmap_atomic(page, KM_USER0);
+       s = (struct btrfs_super_block *)mapped_buffer;
+       memcpy(on_disk_csum, s->csum, sdev->csum_size);
 
-       if (logical != le64_to_cpu(s->bytenr))
+       if (sblock->pagev[0].logical != le64_to_cpu(s->bytenr))
                ++fail;
 
-       if (sbio->spag[0].generation != le64_to_cpu(s->generation))
+       if (sblock->pagev[0].generation != le64_to_cpu(s->generation))
                ++fail;
 
        if (memcmp(s->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
                ++fail;
 
-       crc = btrfs_csum_data(root, buffer + BTRFS_CSUM_SIZE, crc,
-                             PAGE_SIZE - BTRFS_CSUM_SIZE);
-       btrfs_csum_final(crc, csum);
-       if (memcmp(csum, s->csum, sbio->sdev->csum_size))
+       len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
+       mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
+       p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
+       index = 0;
+       for (;;) {
+               u64 l = min_t(u64, len, mapped_size);
+
+               crc = btrfs_csum_data(root, p, crc, l);
+               kunmap_atomic(mapped_buffer, KM_USER0);
+               len -= l;
+               if (len == 0)
+                       break;
+               index++;
+               BUG_ON(index >= sblock->page_count);
+               BUG_ON(!sblock->pagev[index].page);
+               page = sblock->pagev[index].page;
+               mapped_buffer = kmap_atomic(page, KM_USER0);
+               mapped_size = PAGE_SIZE;
+               p = mapped_buffer;
+       }
+
+       btrfs_csum_final(crc, calculated_csum);
+       if (memcmp(calculated_csum, on_disk_csum, sdev->csum_size))
                ++fail;
 
        if (fail) {
@@ -947,6 +1415,23 @@ static int scrub_checksum_super(struct scrub_bio *sbio, void *buffer)
        return fail;
 }
 
+static void scrub_block_get(struct scrub_block *sblock)
+{
+       atomic_inc(&sblock->ref_count);
+}
+
+static void scrub_block_put(struct scrub_block *sblock)
+{
+       if (atomic_dec_and_test(&sblock->ref_count)) {
+               int i;
+
+               for (i = 0; i < sblock->page_count; i++)
+                       if (sblock->pagev[i].page)
+                               __free_page(sblock->pagev[i].page);
+               kfree(sblock);
+       }
+}
+
 static void scrub_submit(struct scrub_dev *sdev)
 {
        struct scrub_bio *sbio;
@@ -955,19 +1440,17 @@ static void scrub_submit(struct scrub_dev *sdev)
                return;
 
        sbio = sdev->bios[sdev->curr];
-       sbio->err = 0;
        sdev->curr = -1;
        atomic_inc(&sdev->in_flight);
 
        btrfsic_submit_bio(READ, sbio->bio);
 }
 
-static int scrub_page(struct scrub_dev *sdev, u64 logical, u64 len,
-                     u64 physical, u64 flags, u64 gen, int mirror_num,
-                     u8 *csum, int force)
+static int scrub_add_page_to_bio(struct scrub_dev *sdev,
+                                struct scrub_page *spage)
 {
+       struct scrub_block *sblock = spage->sblock;
        struct scrub_bio *sbio;
-       struct page *page;
        int ret;
 
 again:
@@ -980,7 +1463,7 @@ again:
                if (sdev->curr != -1) {
                        sdev->first_free = sdev->bios[sdev->curr]->next_free;
                        sdev->bios[sdev->curr]->next_free = -1;
-                       sdev->bios[sdev->curr]->count = 0;
+                       sdev->bios[sdev->curr]->page_count = 0;
                        spin_unlock(&sdev->list_lock);
                } else {
                        spin_unlock(&sdev->list_lock);
@@ -988,53 +1471,200 @@ again:
                }
        }
        sbio = sdev->bios[sdev->curr];
-       if (sbio->count == 0) {
+       if (sbio->page_count == 0) {
                struct bio *bio;
 
-               sbio->physical = physical;
-               sbio->logical = logical;
-               bio = bio_alloc(GFP_NOFS, SCRUB_PAGES_PER_BIO);
-               if (!bio)
-                       return -ENOMEM;
+               sbio->physical = spage->physical;
+               sbio->logical = spage->logical;
+               bio = sbio->bio;
+               if (!bio) {
+                       bio = bio_alloc(GFP_NOFS, sdev->pages_per_bio);
+                       if (!bio)
+                               return -ENOMEM;
+                       sbio->bio = bio;
+               }
 
                bio->bi_private = sbio;
                bio->bi_end_io = scrub_bio_end_io;
                bio->bi_bdev = sdev->dev->bdev;
-               bio->bi_sector = sbio->physical >> 9;
+               bio->bi_sector = spage->physical >> 9;
                sbio->err = 0;
-               sbio->bio = bio;
-       } else if (sbio->physical + sbio->count * PAGE_SIZE != physical ||
-                  sbio->logical + sbio->count * PAGE_SIZE != logical) {
+       } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
+                  spage->physical ||
+                  sbio->logical + sbio->page_count * PAGE_SIZE !=
+                  spage->logical) {
                scrub_submit(sdev);
                goto again;
        }
-       sbio->spag[sbio->count].flags = flags;
-       sbio->spag[sbio->count].generation = gen;
-       sbio->spag[sbio->count].have_csum = 0;
-       sbio->spag[sbio->count].mirror_num = mirror_num;
-
-       page = alloc_page(GFP_NOFS);
-       if (!page)
-               return -ENOMEM;
 
-       ret = bio_add_page(sbio->bio, page, PAGE_SIZE, 0);
-       if (!ret) {
-               __free_page(page);
+       sbio->pagev[sbio->page_count] = spage;
+       ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
+       if (ret != PAGE_SIZE) {
+               if (sbio->page_count < 1) {
+                       bio_put(sbio->bio);
+                       sbio->bio = NULL;
+                       return -EIO;
+               }
                scrub_submit(sdev);
                goto again;
        }
 
-       if (csum) {
-               sbio->spag[sbio->count].have_csum = 1;
-               memcpy(sbio->spag[sbio->count].csum, csum, sdev->csum_size);
+       scrub_block_get(sblock); /* one for the added page */
+       atomic_inc(&sblock->outstanding_pages);
+       sbio->page_count++;
+       if (sbio->page_count == sdev->pages_per_bio)
+               scrub_submit(sdev);
+
+       return 0;
+}
+
+static int scrub_pages(struct scrub_dev *sdev, u64 logical, u64 len,
+                      u64 physical, u64 flags, u64 gen, int mirror_num,
+                      u8 *csum, int force)
+{
+       struct scrub_block *sblock;
+       int index;
+
+       sblock = kzalloc(sizeof(*sblock), GFP_NOFS);
+       if (!sblock) {
+               spin_lock(&sdev->stat_lock);
+               sdev->stat.malloc_errors++;
+               spin_unlock(&sdev->stat_lock);
+               return -ENOMEM;
+       }
+
+       /* one ref inside this function, plus one for each page later on */
+       atomic_set(&sblock->ref_count, 1);
+       sblock->sdev = sdev;
+       sblock->no_io_error_seen = 1;
+
+       for (index = 0; len > 0; index++) {
+               struct scrub_page *spage = sblock->pagev + index;
+               u64 l = min_t(u64, len, PAGE_SIZE);
+
+               BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
+               spage->page = alloc_page(GFP_NOFS);
+               if (!spage->page) {
+                       spin_lock(&sdev->stat_lock);
+                       sdev->stat.malloc_errors++;
+                       spin_unlock(&sdev->stat_lock);
+                       while (index > 0) {
+                               index--;
+                               __free_page(sblock->pagev[index].page);
+                       }
+                       kfree(sblock);
+                       return -ENOMEM;
+               }
+               spage->sblock = sblock;
+               spage->bdev = sdev->dev->bdev;
+               spage->flags = flags;
+               spage->generation = gen;
+               spage->logical = logical;
+               spage->physical = physical;
+               spage->mirror_num = mirror_num;
+               if (csum) {
+                       spage->have_csum = 1;
+                       memcpy(spage->csum, csum, sdev->csum_size);
+               } else {
+                       spage->have_csum = 0;
+               }
+               sblock->page_count++;
+               len -= l;
+               logical += l;
+               physical += l;
+       }
+
+       BUG_ON(sblock->page_count == 0);
+       for (index = 0; index < sblock->page_count; index++) {
+               struct scrub_page *spage = sblock->pagev + index;
+               int ret;
+
+               ret = scrub_add_page_to_bio(sdev, spage);
+               if (ret) {
+                       scrub_block_put(sblock);
+                       return ret;
+               }
        }
-       ++sbio->count;
-       if (sbio->count == SCRUB_PAGES_PER_BIO || force)
+
+       if (force)
                scrub_submit(sdev);
 
+       /* last one frees, either here or in bio completion for last page */
+       scrub_block_put(sblock);
        return 0;
 }
 
+static void scrub_bio_end_io(struct bio *bio, int err)
+{
+       struct scrub_bio *sbio = bio->bi_private;
+       struct scrub_dev *sdev = sbio->sdev;
+       struct btrfs_fs_info *fs_info = sdev->dev->dev_root->fs_info;
+
+       sbio->err = err;
+       sbio->bio = bio;
+
+       btrfs_queue_worker(&fs_info->scrub_workers, &sbio->work);
+}
+
+static void scrub_bio_end_io_worker(struct btrfs_work *work)
+{
+       struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
+       struct scrub_dev *sdev = sbio->sdev;
+       int i;
+
+       BUG_ON(sbio->page_count > SCRUB_PAGES_PER_BIO);
+       if (sbio->err) {
+               for (i = 0; i < sbio->page_count; i++) {
+                       struct scrub_page *spage = sbio->pagev[i];
+
+                       spage->io_error = 1;
+                       spage->sblock->no_io_error_seen = 0;
+               }
+       }
+
+       /* now complete the scrub_block items that have all pages completed */
+       for (i = 0; i < sbio->page_count; i++) {
+               struct scrub_page *spage = sbio->pagev[i];
+               struct scrub_block *sblock = spage->sblock;
+
+               if (atomic_dec_and_test(&sblock->outstanding_pages))
+                       scrub_block_complete(sblock);
+               scrub_block_put(sblock);
+       }
+
+       if (sbio->err) {
+               /* what is this good for??? */
+               sbio->bio->bi_flags &= ~(BIO_POOL_MASK - 1);
+               sbio->bio->bi_flags |= 1 << BIO_UPTODATE;
+               sbio->bio->bi_phys_segments = 0;
+               sbio->bio->bi_idx = 0;
+
+               for (i = 0; i < sbio->page_count; i++) {
+                       struct bio_vec *bi;
+                       bi = &sbio->bio->bi_io_vec[i];
+                       bi->bv_offset = 0;
+                       bi->bv_len = PAGE_SIZE;
+               }
+       }
+
+       bio_put(sbio->bio);
+       sbio->bio = NULL;
+       spin_lock(&sdev->list_lock);
+       sbio->next_free = sdev->first_free;
+       sdev->first_free = sbio->index;
+       spin_unlock(&sdev->list_lock);
+       atomic_dec(&sdev->in_flight);
+       wake_up(&sdev->list_wait);
+}
+
+static void scrub_block_complete(struct scrub_block *sblock)
+{
+       if (!sblock->no_io_error_seen)
+               scrub_handle_errored_block(sblock);
+       else
+               scrub_checksum(sblock);
+}
+
 static int scrub_find_csum(struct scrub_dev *sdev, u64 logical, u64 len,
                           u8 *csum)
 {
@@ -1042,7 +1672,6 @@ static int scrub_find_csum(struct scrub_dev *sdev, u64 logical, u64 len,
        int ret = 0;
        unsigned long i;
        unsigned long num_sectors;
-       u32 sectorsize = sdev->dev->dev_root->sectorsize;
 
        while (!list_empty(&sdev->csum_list)) {
                sum = list_first_entry(&sdev->csum_list,
@@ -1060,7 +1689,7 @@ static int scrub_find_csum(struct scrub_dev *sdev, u64 logical, u64 len,
        if (!sum)
                return 0;
 
-       num_sectors = sum->len / sectorsize;
+       num_sectors = sum->len / sdev->sectorsize;
        for (i = 0; i < num_sectors; ++i) {
                if (sum->sums[i].bytenr == logical) {
                        memcpy(csum, &sum->sums[i].sum, sdev->csum_size);
@@ -1081,9 +1710,28 @@ static int scrub_extent(struct scrub_dev *sdev, u64 logical, u64 len,
 {
        int ret;
        u8 csum[BTRFS_CSUM_SIZE];
+       u32 blocksize;
+
+       if (flags & BTRFS_EXTENT_FLAG_DATA) {
+               blocksize = sdev->sectorsize;
+               spin_lock(&sdev->stat_lock);
+               sdev->stat.data_extents_scrubbed++;
+               sdev->stat.data_bytes_scrubbed += len;
+               spin_unlock(&sdev->stat_lock);
+       } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
+               BUG_ON(sdev->nodesize != sdev->leafsize);
+               blocksize = sdev->nodesize;
+               spin_lock(&sdev->stat_lock);
+               sdev->stat.tree_extents_scrubbed++;
+               sdev->stat.tree_bytes_scrubbed += len;
+               spin_unlock(&sdev->stat_lock);
+       } else {
+               blocksize = sdev->sectorsize;
+               BUG_ON(1);
+       }
 
        while (len) {
-               u64 l = min_t(u64, len, PAGE_SIZE);
+               u64 l = min_t(u64, len, blocksize);
                int have_csum = 0;
 
                if (flags & BTRFS_EXTENT_FLAG_DATA) {
@@ -1092,8 +1740,8 @@ static int scrub_extent(struct scrub_dev *sdev, u64 logical, u64 len,
                        if (have_csum == 0)
                                ++sdev->stat.no_csum;
                }
-               ret = scrub_page(sdev, logical, l, physical, flags, gen,
-                                mirror_num, have_csum ? csum : NULL, 0);
+               ret = scrub_pages(sdev, logical, l, physical, flags, gen,
+                                 mirror_num, have_csum ? csum : NULL, 0);
                if (ret)
                        return ret;
                len -= l;
@@ -1158,6 +1806,11 @@ static noinline_for_stack int scrub_stripe(struct scrub_dev *sdev,
        if (!path)
                return -ENOMEM;
 
+       /*
+        * work on commit root. The related disk blocks are static as
+        * long as COW is applied. This means, it is save to rewrite
+        * them to repair disk errors without any race conditions
+        */
        path->search_commit_root = 1;
        path->skip_locking = 1;
 
@@ -1511,8 +2164,8 @@ static noinline_for_stack int scrub_supers(struct scrub_dev *sdev)
                if (bytenr + BTRFS_SUPER_INFO_SIZE > device->total_bytes)
                        break;
 
-               ret = scrub_page(sdev, bytenr, PAGE_SIZE, bytenr,
-                                BTRFS_EXTENT_FLAG_SUPER, gen, i, NULL, 1);
+               ret = scrub_pages(sdev, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
+                                    BTRFS_EXTENT_FLAG_SUPER, gen, i, NULL, 1);
                if (ret)
                        return ret;
        }
@@ -1571,10 +2224,30 @@ int btrfs_scrub_dev(struct btrfs_root *root, u64 devid, u64 start, u64 end,
        /*
         * check some assumptions
         */
-       if (root->sectorsize != PAGE_SIZE ||
-           root->sectorsize != root->leafsize ||
-           root->sectorsize != root->nodesize) {
-               printk(KERN_ERR "btrfs_scrub: size assumptions fail\n");
+       if (root->nodesize != root->leafsize) {
+               printk(KERN_ERR
+                      "btrfs_scrub: size assumption nodesize == leafsize (%d == %d) fails\n",
+                      root->nodesize, root->leafsize);
+               return -EINVAL;
+       }
+
+       if (root->nodesize > BTRFS_STRIPE_LEN) {
+               /*
+                * in this case scrub is unable to calculate the checksum
+                * the way scrub is implemented. Do not handle this
+                * situation at all because it won't ever happen.
+                */
+               printk(KERN_ERR
+                      "btrfs_scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails\n",
+                      root->nodesize, BTRFS_STRIPE_LEN);
+               return -EINVAL;
+       }
+
+       if (root->sectorsize != PAGE_SIZE) {
+               /* not supported for data w/o checksums */
+               printk(KERN_ERR
+                      "btrfs_scrub: size assumption sectorsize != PAGE_SIZE (%d != %lld) fails\n",
+                      root->sectorsize, (unsigned long long)PAGE_SIZE);
                return -EINVAL;
        }