[PATCH] md/raid1: fix missing bitmap update w/o WriteMostly devices

[PATCH] md/raid1: fix missing bitmap update w/o WriteMostly devices

[Song Liu]

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)) {
-- 
2.30.2

Re: [PATCH] md/raid1: fix missing bitmap update w/o WriteMostly devices

[Guoqing Jiang]

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