The following behavior is inconsistent:
* For request-based dm queues the default value of rq_affinity is 1.
* For bio-based dm queues the default value of rq_affinity is 0.
The default value for request-based dm queues is 1 because of the following
code in blk_mq_init_allocated_queue():
q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
>From <linux/blkdev.h>:
#define QUEUE_FLAG_MQ_DEFAULT ((1UL << QUEUE_FLAG_IO_STAT) | \
(1UL << QUEUE_FLAG_SAME_COMP) | \
(1UL << QUEUE_FLAG_NOWAIT))
The default value of rq_affinity for bio-based dm queues is 0 because the
dm alloc_dev() function does not set any of the QUEUE_FLAG_SAME_* flags. I
think the different default values are the result of an oversight when
blk-mq support was added in the device mapper code. Hence this patch that
changes the default value of rq_affinity from 0 to 1 for bio-based dm
queues.
This patch reduces the boot time from 12.23 to 12.20 seconds on my test
setup, a Pixel 2023 development board. The storage controller on that test
setup supports a single completion interrupt and hence benefits from
redirecting I/O completions to a CPU core that is closer to the submitter.
Cc: Mikulas Patocka <[email protected]>
Cc: Eric Biggers <[email protected]>
Cc: Jaegeuk Kim <[email protected]>
Cc: Daniel Lee <[email protected]>
Cc: [email protected]
Fixes: bfebd1cdb497 ("dm: add full blk-mq support to request-based DM")
Signed-off-by: Bart Van Assche <[email protected]>
---
drivers/md/dm.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/drivers/md/dm.c b/drivers/md/dm.c
index 56aa2a8b9d71..9af216c11cf7 100644
--- a/drivers/md/dm.c
+++ b/drivers/md/dm.c
@@ -2106,6 +2106,7 @@ static struct mapped_device *alloc_dev(int minor)
if (IS_ERR(md->disk))
goto bad;
md->queue = md->disk->queue;
+ blk_queue_flag_set(QUEUE_FLAG_SAME_COMP, md->queue);
init_waitqueue_head(&md->wait);
INIT_WORK(&md->work, dm_wq_work);
