Skip to contentSkip to conversation

block: Do not merge if merging is disabled

PATCH In review Bart Van Assche 3 participants 6 messages Jul 6, 2023 – Jul 7, 2023 Block layer

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 block: Do not merge if merging is disabled

While testing the performance impact of zoned write pipelining, I noticed that merging happens even if merging has been disabled via sysfs. Fix this.

Cc Christoph Hellwig <[email protected]>
Cc Ming Lei <[email protected]>
Cc Damien Le Moal <[email protected]>
---
 block/blk-mq-sched.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
block/blk-mq-sched.c +2 -1
index 67c95f31b15b..8883721f419a 100644
@@ -375,7 +375,8 @@ bool blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio,
 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq,
 				   struct list_head *free)
 {
-	return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq, free);
+	return !blk_queue_nomerges(q) && rq_mergeable(rq) &&
+		elv_attempt_insert_merge(q, rq, free);
 }
 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
 
Reviewed-by on v1 1/1
#1 lore raw
Quoted 8 lines · Bart Van Assche

While testing the performance impact of zoned write pipelining, I noticed that merging happens even if merging has been disabled via sysfs. Fix this.

Cc Christoph Hellwig <[email protected]>
Cc Ming Lei <[email protected]>
Cc Damien Le Moal <[email protected]>

Looks good.

Quoted 18 lines
---
 block/blk-mq-sched.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
block/blk-mq-sched.c +2 -1
index 67c95f31b15b..8883721f419a 100644
@@ -375,7 +375,8 @@ bool blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio,
 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq,
 				   struct list_head *free)
 {
-	return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq, free);
+	return !blk_queue_nomerges(q) && rq_mergeable(rq) &&
+		elv_attempt_insert_merge(q, rq, free);
 }
 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
 
Signature · 2 lines
Damien Le Moal
Western Digital Research
#2 lore raw
In reply to Bart Van Assche
Quoted 17 lines · Bart Van Assche

While testing the performance impact of zoned write pipelining, I noticed that merging happens even if merging has been disabled via sysfs. Fix this.

Cc Christoph Hellwig <[email protected]>
Cc Ming Lei <[email protected]>
Cc Damien Le Moal <[email protected]>
---
 block/blk-mq-sched.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
block/blk-mq-sched.c +0 -0
index 67c95f31b15b..8883721f419a 100644

block/blk-mq-sched.c bool blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio, · line 381

Quoted 8 lines
block/blk-mq-sched.c +2 -1
@@ -375,7 +375,8 @@ bool blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio,
 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq,
 				   struct list_head *free)
 {
-	return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq, free);
+	return !blk_queue_nomerges(q) && rq_mergeable(rq) &&
+		elv_attempt_insert_merge(q, rq, free);
 }
 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);

elv_attempt_insert_merge() does check blk_queue_nomerges() at its entry, so this patch fix nothing.

Given blk_mq_sched_try_insert_merge is only called from bfq and deadline, it may not matter to apply this optimization.

Thanks,
Ming

#3 lore raw
Quoted 2 lines · Ming Lei

Given blk_mq_sched_try_insert_merge is only called from bfq and deadline, it may not matter to apply this optimization.

Without this patch, the documentation of the "nomerges" sysfs attribute is incorrect. I need this patch because I want the ability to disable merging even if an I/O scheduler has been selected. As mentioned in the patch description, I discovered this while I was writing a shell script that submits various I/O workloads to a block device.

Bart.

#4 lore raw
Quoted 10 lines · Bart Van Assche
Quoted 2 lines · Ming Lei

Given blk_mq_sched_try_insert_merge is only called from bfq and deadline, it may not matter to apply this optimization.

Without this patch, the documentation of the "nomerges" sysfs attribute is incorrect. I need this patch because I want the ability to disable merging even if an I/O scheduler has been selected. As mentioned in the patch description, I discovered this while I was writing a shell script that submits various I/O workloads to a block device.

Ming's point still stands I think: blk_queue_nomerges(q) is the first thing checked in elv_attempt_insert_merge(). So your patch should be a no-op and disabling merging through sysfs should still be effective. Why is your patch changing anything ?

Moving blk_mq_sched_try_insert_merge() call to rq_mergeable(rq) inside elv_attempt_insert_merge() would also make a lot of sense I think. With that, blk_mq_sched_try_insert_merge() would be reduced to calling only elv_attempt_insert_merge(), which means that elv_attempt_insert_merge() could go away.

Signature · 2 lines
Damien Le Moal
Western Digital Research
#5 lore raw
Quoted 10 lines · Damien Le Moal

Ming's point still stands I think: blk_queue_nomerges(q) is the first thing checked in elv_attempt_insert_merge(). So your patch should be a no-op and disabling merging through sysfs should still be effective. Why is your patch changing anything ?

Moving blk_mq_sched_try_insert_merge() call to rq_mergeable(rq) inside elv_attempt_insert_merge() would also make a lot of sense I think. With that, blk_mq_sched_try_insert_merge() would be reduced to calling only elv_attempt_insert_merge(), which means that elv_attempt_insert_merge() could go away.

Let's drop this patch. Since this patch was developed against an older kernel version, let me check whether this patch is perhaps only needed for the kernel version it was developed against.

Thanks,

Bart.