On 1/4/22 7:04 AM, Song Liu wrote:

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 | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/drivers/md/raid1.c b/drivers/md/raid1.c index 7dc8026cf6ee..85505424f7a4 100644 --- a/drivers/md/raid1.c +++ b/drivers/md/raid1.c @@ -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)) {
  

Indeed, I missed that md_bitmap_startwrite should be always called for the first clone.

Thanks, Guoqing