btrfs: scrub: Require mandatory block group RO for dev-replace

[BUG]
For dev-replace test cases with fsstress, like btrfs/06[45] btrfs/071,
looped runs can lead to random failure, where scrub finds csum error.

The possibility is not high, around 1/20 to 1/100, but it's causing data
corruption.

The bug is observable after commit ee61ec895a5b ("btrfs: scrub: Don't
check free space before marking a block group RO")

[CAUSE]
Dev-replace has two source of writes:

- Write duplication
  All writes to source device will also be duplicated to target device.

  Content:	Not yet persisted data/meta

- Scrub copy
  Dev-replace reused scrub code to iterate through existing extents, and
  copy the verified data to target device.

  Content:	Previously persisted data and metadata

The difference in contents makes the following race possible:
	Regular Writer		|	Dev-replace
-----------------------------------------------------------------
  ^                             |
  | Preallocate one data extent |
  | at bytenr X, len 1M		|
  v				|
  ^ Commit transaction		|
  | Now extent [X, X+1M) is in  |
  v commit root			|
 ================== Dev replace starts =========================
  				| ^
				| | Scrub extent [X, X+1M)
				| | Read [X, X+1M)
				| | (The content are mostly garbage
				| |  since it's preallocated)
  ^				| v
  | Write back happens for	|
  | extent [X, X+512K)		|
  | New data writes to both	|
  | source and target dev.	|
  v				|
				| ^
				| | Scrub writes back extent [X, X+1M)
				| | to target device.
				| | This will over write the new data in
				| | [X, X+512K)
				| v

This race can only happen for nocow writes. Thus metadata and data cow
writes are safe, as COW will never overwrite extents of previous
transaction (in commit root).

This behavior can be confirmed by disabling all fallocate related calls
in fsstress (*), then all related tests can pass a 2000 run loop.

*: FSSTRESS_AVOID="-f fallocate=0 -f allocsp=0 -f zero=0 -f insert=0 \
		   -f collapse=0 -f punch=0 -f resvsp=0"
   I didn't expect resvsp ioctl will fallback to fallocate in VFS...

[FIX]
Make dev-replace to require mandatory block group RO, and wait for current
nocow writes before calling scrub_chunk().

This patch will mostly revert commit a69aa3ae3701 ("btrfs: Continue replace
when set_block_ro failed") for dev-replace path.

The side effect is, dev-replace can be more strict on avaialble space, but
definitely worth to avoid data corruption.

Reported-by: Filipe Manana <fdmanana@suse.com>
Fixes: a69aa3ae3701 ("btrfs: Continue replace when set_block_ro failed")
Fixes: ee61ec895a5b ("btrfs: scrub: Don't check free space before marking a block group RO")
Reviewed-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>

diff --git a/fs/btrfs/scrub.c b/fs/btrfs/scrub.c
index 21de630b073024c343d75835c3a61ba0922aac7b..fd266a2d15ec84959189521312827350b6795ffe 100644 (file)
--- a/fs/btrfs/scrub.c
+++ b/fs/btrfs/scrub.c
@@ -3577,17 +3577,27 @@ int scrub_enumerate_chunks(struct scrub_ctx *sctx,
                 * This can easily boost the amount of SYSTEM chunks if cleaner
                 * thread can't be triggered fast enough, and use up all space
                 * of btrfs_super_block::sys_chunk_array
+                *
+                * While for dev replace, we need to try our best to mark block
+                * group RO, to prevent race between:
+                * - Write duplication
+                *   Contains latest data
+                * - Scrub copy
+                *   Contains data from commit tree
+                *
+                * If target block group is not marked RO, nocow writes can
+                * be overwritten by scrub copy, causing data corruption.
+                * So for dev-replace, it's not allowed to continue if a block
+                * group is not RO.
                 */
-               ret = btrfs_inc_block_group_ro(cache, false);
-               scrub_pause_off(fs_info);
-
+               ret = btrfs_inc_block_group_ro(cache, sctx->is_dev_replace);
                if (ret == 0) {
                        ro_set = 1;
-               } else if (ret == -ENOSPC) {
+               } else if (ret == -ENOSPC && !sctx->is_dev_replace) {
                        /*
                         * btrfs_inc_block_group_ro return -ENOSPC when it
                         * failed in creating new chunk for metadata.
-                        * It is not a problem for scrub/replace, because
+                        * It is not a problem for scrub, because
                         * metadata are always cowed, and our scrub paused
                         * commit_transactions.
                         */
@@ -3596,9 +3606,22 @@ int scrub_enumerate_chunks(struct scrub_ctx *sctx,
                        btrfs_warn(fs_info,
                                   "failed setting block group ro: %d", ret);
                        btrfs_put_block_group(cache);
+                       scrub_pause_off(fs_info);
                        break;
                }
 
+               /*
+                * Now the target block is marked RO, wait for nocow writes to
+                * finish before dev-replace.
+                * COW is fine, as COW never overwrites extents in commit tree.
+                */
+               if (sctx->is_dev_replace) {
+                       btrfs_wait_nocow_writers(cache);
+                       btrfs_wait_ordered_roots(fs_info, U64_MAX, cache->start,
+                                       cache->length);
+               }
+
+               scrub_pause_off(fs_info);
                down_write(&dev_replace->rwsem);
                dev_replace->cursor_right = found_key.offset + length;
                dev_replace->cursor_left = found_key.offset;