]> git.ipfire.org Git - thirdparty/kernel/stable.git/commitdiff
md/raid1: fix missing bitmap update w/o WriteMostly devices
authorSong Liu <song@kernel.org>
Mon, 3 Jan 2022 21:49:36 +0000 (13:49 -0800)
committerSong Liu <song@kernel.org>
Mon, 3 Jan 2022 22:13:48 +0000 (14:13 -0800)
commit [1] causes missing bitmap updates when there isn't any WriteMostly
devices.

Detailed steps to reproduce by Norbert (which somehow didn't make to lore):

   # setup md10 (raid1) with two drives (1 GByte sparse files)
   dd if=/dev/zero of=disk1 bs=1024k seek=1024 count=0
   dd if=/dev/zero of=disk2 bs=1024k seek=1024 count=0

   losetup /dev/loop11 disk1
   losetup /dev/loop12 disk2

   mdadm --create /dev/md10 --level=1 --raid-devices=2 /dev/loop11 /dev/loop12

   # add bitmap (aka write-intent log)
   mdadm /dev/md10 --grow --bitmap=internal

   echo check > /sys/block/md10/md/sync_action

   root:# cat /sys/block/md10/md/mismatch_cnt
   0
   root:#

   # remove member drive disk2 (loop12)
   mdadm /dev/md10 -f loop12 ; mdadm /dev/md10 -r loop12

   # modify degraded md device
   dd if=/dev/urandom of=/dev/md10 bs=512 count=1

   # no blocks recorded as out of sync on the remaining member disk1/loop11
   root:# mdadm -X /dev/loop11 | grep Bitmap
             Bitmap : 16 bits (chunks), 0 dirty (0.0%)
   root:#

   # re-add disk2, nothing synced because of empty bitmap
   mdadm /dev/md10 --re-add /dev/loop12

   # check integrity again
   echo check > /sys/block/md10/md/sync_action

   # disk1 and disk2 are no longer in sync, reads return differend data
   root:# cat /sys/block/md10/md/mismatch_cnt
   128
   root:#

   # clean up
   mdadm -S /dev/md10
   losetup -d /dev/loop11
   losetup -d /dev/loop12
   rm disk1 disk2

Fix this by moving the WriteMostly check to the if condition for
alloc_behind_master_bio().

[1] commit fd3b6975e9c1 ("md/raid1: only allocate write behind bio for WriteMostly device")
Fixes: fd3b6975e9c1 ("md/raid1: only allocate write behind bio for WriteMostly device")
Cc: stable@vger.kernel.org # v5.12+
Cc: Guoqing Jiang <guoqing.jiang@linux.dev>
Cc: Jens Axboe <axboe@kernel.dk>
Reported-by: Norbert Warmuth <nwarmuth@t-online.de>
Suggested-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Song Liu <song@kernel.org>
drivers/md/raid1.c

index 7dc8026cf6ee04077980056357447d2b710ee440..85505424f7a48a543ea81d6fad1be86f3b25ebb0 100644 (file)
@@ -1496,12 +1496,13 @@ static void raid1_write_request(struct mddev *mddev, struct bio *bio,
                if (!r1_bio->bios[i])
                        continue;
 
-               if (first_clone && test_bit(WriteMostly, &rdev->flags)) {
+               if (first_clone) {
                        /* do behind I/O ?
                         * Not if there are too many, or cannot
                         * allocate memory, or a reader on WriteMostly
                         * is waiting for behind writes to flush */
                        if (bitmap &&
+                           test_bit(WriteMostly, &rdev->flags) &&
                            (atomic_read(&bitmap->behind_writes)
                             < mddev->bitmap_info.max_write_behind) &&
                            !waitqueue_active(&bitmap->behind_wait)) {