btrfs: always write superblocks synchronously
diff mbox

Message ID 20170502162813.4299-1-dave@stgolabs.net
State New
Headers show

Commit Message

Davidlohr Bueso May 2, 2017, 4:28 p.m. UTC
Commit b685d3d65ac7 "block: treat REQ_FUA and REQ_PREFLUSH as
synchronous" removed REQ_SYNC flag from WRITE_FUA implementation.
Since REQ_FUA and REQ_FLUSH flags are stripped from submitted IO
when the disk doesn't have volatile write cache and thus effectively
make the write async. This was seen to cause performance hits up
to 90% regression in disk IO related benchmarks such as reaim and
dbench[1].

Fix the problem by making sure the first superblock write is also
treated as synchronous since they can block progress of the
journalling (commit, log syncs) machinery and thus the whole filesystem.

[1] https://www.spinics.net/lists/linux-ext4/msg56238.html

Fixes: b685d3d65ac (block: treat REQ_FUA and REQ_PREFLUSH as synchronous)
Cc: stable <stable@vger.kernel.org>
Cc: Jan Kara <jack@suze.cz>
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
---
 fs/btrfs/disk-io.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

Comments

Jan Kara May 3, 2017, 8:36 a.m. UTC | #1
On Tue 02-05-17 09:28:13, Davidlohr Bueso wrote:
> Commit b685d3d65ac7 "block: treat REQ_FUA and REQ_PREFLUSH as
> synchronous" removed REQ_SYNC flag from WRITE_FUA implementation.
> Since REQ_FUA and REQ_FLUSH flags are stripped from submitted IO
> when the disk doesn't have volatile write cache and thus effectively
> make the write async. This was seen to cause performance hits up
> to 90% regression in disk IO related benchmarks such as reaim and
> dbench[1].
> 
> Fix the problem by making sure the first superblock write is also
> treated as synchronous since they can block progress of the
> journalling (commit, log syncs) machinery and thus the whole filesystem.
> 
> [1] https://www.spinics.net/lists/linux-ext4/msg56238.html
> 
> Fixes: b685d3d65ac (block: treat REQ_FUA and REQ_PREFLUSH as synchronous)
> Cc: stable <stable@vger.kernel.org>
> Cc: Jan Kara <jack@suze.cz>
> Signed-off-by: Davidlohr Bueso <dbueso@suse.de>

I wasn't patient enough and already sent the fix as part of my series
fixing other filesystems [1]. It also fixes one more place in btrfs that
needs REQ_SYNC to return to the original behavior.

[1] https://www.spinics.net/lists/linux-btrfs/msg65022.html

								Honza

> ---
>  fs/btrfs/disk-io.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/fs/btrfs/disk-io.c b/fs/btrfs/disk-io.c
> index 061c1d1f774f..51b2fd8ceccb 100644
> --- a/fs/btrfs/disk-io.c
> +++ b/fs/btrfs/disk-io.c
> @@ -3468,7 +3468,7 @@ static int write_dev_supers(struct btrfs_device *device,
>  		 * to go down lazy.
>  		 */
>  		if (i == 0)
> -			ret = btrfsic_submit_bh(REQ_OP_WRITE, REQ_FUA, bh);
> +			ret = btrfsic_submit_bh(REQ_OP_WRITE, REQ_FUA | REQ_SYNC, bh);
>  		else
>  			ret = btrfsic_submit_bh(REQ_OP_WRITE, REQ_SYNC, bh);
>  		if (ret)
> -- 
> 2.12.0
>
Chris Mason May 3, 2017, 2:55 p.m. UTC | #2
On 05/03/2017 04:36 AM, Jan Kara wrote:
> On Tue 02-05-17 09:28:13, Davidlohr Bueso wrote:
>> Commit b685d3d65ac7 "block: treat REQ_FUA and REQ_PREFLUSH as
>> synchronous" removed REQ_SYNC flag from WRITE_FUA implementation.
>> Since REQ_FUA and REQ_FLUSH flags are stripped from submitted IO
>> when the disk doesn't have volatile write cache and thus effectively
>> make the write async. This was seen to cause performance hits up
>> to 90% regression in disk IO related benchmarks such as reaim and
>> dbench[1].
>>
>> Fix the problem by making sure the first superblock write is also
>> treated as synchronous since they can block progress of the
>> journalling (commit, log syncs) machinery and thus the whole filesystem.
>>

>>
>> Fixes: b685d3d65ac (block: treat REQ_FUA and REQ_PREFLUSH as synchronous)
>> Cc: stable <stable@vger.kernel.org>
>> Cc: Jan Kara <jack@suze.cz>
>> Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
>
> I wasn't patient enough and already sent the fix as part of my series
> fixing other filesystems [1]. It also fixes one more place in btrfs that
> needs REQ_SYNC to return to the original behavior.
>


Thanks guys.

-chris

--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Patch
diff mbox

diff --git a/fs/btrfs/disk-io.c b/fs/btrfs/disk-io.c
index 061c1d1f774f..51b2fd8ceccb 100644
--- a/fs/btrfs/disk-io.c
+++ b/fs/btrfs/disk-io.c
@@ -3468,7 +3468,7 @@  static int write_dev_supers(struct btrfs_device *device,
 		 * to go down lazy.
 		 */
 		if (i == 0)
-			ret = btrfsic_submit_bh(REQ_OP_WRITE, REQ_FUA, bh);
+			ret = btrfsic_submit_bh(REQ_OP_WRITE, REQ_FUA | REQ_SYNC, bh);
 		else
 			ret = btrfsic_submit_bh(REQ_OP_WRITE, REQ_SYNC, bh);
 		if (ret)