Skip to contentSkip to conversation

block: Do not pull requests from the scheduler when we cannot dispatch them

PATCH Applied Jan Kara 3 participants 7 messages Jun 3, 2021 – Jun 16, 2021 Block layer

lore
Share & tools
lore

Jan Kara

author
#0 lore raw

PATCH v2 1/1 block: Do not pull requests from the scheduler when we cannot dispatch them

Provided the device driver does not implement dispatch budget accounting (which only SCSI does) the loop in __blk_mq_do_dispatch_sched() pulls requests from the IO scheduler as long as it is willing to give out any. That defeats scheduling heuristics inside the scheduler by creating false impression that the device can take more IO when it in fact cannot.

For example with BFQ IO scheduler on top of virtio-blk device setting blkio cgroup weight has barely any impact on observed throughput of async IO because __blk_mq_do_dispatch_sched() always sucks out all the IO queued in BFQ. BFQ first submits IO from higher weight cgroups but when that is all dispatched, it will give out IO of lower weight cgroups as well. And then we have to wait for all this IO to be dispatched to the disk (which means lot of it actually has to complete) before the IO scheduler is queried again for dispatching more requests. This completely destroys any service differentiation.

So grab request tag for a request pulled out of the IO scheduler already in __blk_mq_do_dispatch_sched() and do not pull any more requests if we cannot get it because we are unlikely to be able to dispatch it. That way only single request is going to wait in the dispatch list for some tag to free.

---
 block/blk-mq-sched.c | 12 +++++++++++-
 block/blk-mq.c       |  2 +-
 block/blk-mq.h       |  2 ++
 3 files changed, 14 insertions(+), 2 deletions(-)

Jens, can you please merge the patch? Thanks!

block/blk-mq-sched.c +11 -1
index 996a4b2f73aa..714e678f516a 100644
@@ -168,9 +168,19 @@ static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
 		 * in blk_mq_dispatch_rq_list().
 		 */
 		list_add_tail(&rq->queuelist, &rq_list);
+		count++;
 		if (rq->mq_hctx != hctx)
 			multi_hctxs = true;
-	} while (++count < max_dispatch);
+
+		/*
+		 * If we cannot get tag for the request, stop dequeueing
+		 * requests from the IO scheduler. We are unlikely to be able
+		 * to submit them anyway and it creates false impression for
+		 * scheduling heuristics that the device can take more IO.
+		 */
+		if (!blk_mq_get_driver_tag(rq))
+			break;
+	} while (count < max_dispatch);
 
 	if (!count) {
 		if (run_queue)
block/blk-mq.c +1 -1
index c86c01bfecdb..bc2cf80d2c3b 100644
@@ -1100,7 +1100,7 @@ static bool __blk_mq_get_driver_tag(struct request *rq)
 	return true;
 }
 
-static bool blk_mq_get_driver_tag(struct request *rq)
+bool blk_mq_get_driver_tag(struct request *rq)
 {
 	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
 
block/blk-mq.h +2 -0
index 9ce64bc4a6c8..81a775171be7 100644
@@ -259,6 +259,8 @@ static inline void blk_mq_put_driver_tag(struct request *rq)
 	__blk_mq_put_driver_tag(rq->mq_hctx, rq);
 }
 
+bool blk_mq_get_driver_tag(struct request *rq);
+
 static inline void blk_mq_clear_mq_map(struct blk_mq_queue_map *qmap)
 {
 	int cpu;
Signature · 1 line
2.26.2
maintainer Applied
#1 lore raw
Quoted 17 lines · Jan Kara

Provided the device driver does not implement dispatch budget accounting (which only SCSI does) the loop in __blk_mq_do_dispatch_sched() pulls requests from the IO scheduler as long as it is willing to give out any. That defeats scheduling heuristics inside the scheduler by creating false impression that the device can take more IO when it in fact cannot.

For example with BFQ IO scheduler on top of virtio-blk device setting blkio cgroup weight has barely any impact on observed throughput of async IO because __blk_mq_do_dispatch_sched() always sucks out all the IO queued in BFQ. BFQ first submits IO from higher weight cgroups but when that is all dispatched, it will give out IO of lower weight cgroups as well. And then we have to wait for all this IO to be dispatched to the disk (which means lot of it actually has to complete) before the IO scheduler is queried again for dispatching more requests. This completely destroys any service differentiation.

Quoted 5 lines

So grab request tag for a request pulled out of the IO scheduler already in __blk_mq_do_dispatch_sched() and do not pull any more requests if we cannot get it because we are unlikely to be able to dispatch it. That way only single request is going to wait in the dispatch list for some tag to free.

Applied to for-5.14/block, thanks.

Signature · 1 line
Jens Axboe
#2 lore raw
In reply to Jan Kara
Quoted 48 lines · Jan Kara

Provided the device driver does not implement dispatch budget accounting (which only SCSI does) the loop in __blk_mq_do_dispatch_sched() pulls requests from the IO scheduler as long as it is willing to give out any. That defeats scheduling heuristics inside the scheduler by creating false impression that the device can take more IO when it in fact cannot.

For example with BFQ IO scheduler on top of virtio-blk device setting blkio cgroup weight has barely any impact on observed throughput of async IO because __blk_mq_do_dispatch_sched() always sucks out all the IO queued in BFQ. BFQ first submits IO from higher weight cgroups but when that is all dispatched, it will give out IO of lower weight cgroups as well. And then we have to wait for all this IO to be dispatched to the disk (which means lot of it actually has to complete) before the IO scheduler is queried again for dispatching more requests. This completely destroys any service differentiation.

So grab request tag for a request pulled out of the IO scheduler already in __blk_mq_do_dispatch_sched() and do not pull any more requests if we cannot get it because we are unlikely to be able to dispatch it. That way only single request is going to wait in the dispatch list for some tag to free.

---
 block/blk-mq-sched.c | 12 +++++++++++-
 block/blk-mq.c       |  2 +-
 block/blk-mq.h       |  2 ++
 3 files changed, 14 insertions(+), 2 deletions(-)

Jens, can you please merge the patch? Thanks!

block/blk-mq-sched.c +4 -1
index 996a4b2f73aa..714e678f516a 100644
@@ -168,9 +168,19 @@ static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
 		 * in blk_mq_dispatch_rq_list().
 		 */
 		list_add_tail(&rq->queuelist, &rq_list);
+		count++;
 		if (rq->mq_hctx != hctx)
 			multi_hctxs = true;
-	} while (++count < max_dispatch);
+
+		/*
+		 * If we cannot get tag for the request, stop dequeueing

block/blk-mq-sched.c static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx) · line 186

Quoted 10 lines
block/blk-mq-sched.c +7 -0
+		 * requests from the IO scheduler. We are unlikely to be able
+		 * to submit them anyway and it creates false impression for
+		 * scheduling heuristics that the device can take more IO.
+		 */
+		if (!blk_mq_get_driver_tag(rq))
+			break;
+	} while (count < max_dispatch);
 
 	if (!count) {
 		if (run_queue)

Doesn't this lead to a double accounting of the allocated tags? From what I can see we don't really check if the tag is already allocated in blk_mq_get_driver_tag() ...

Hmm?

Cheers,

Hannes

Signature · 4 lines
Dr. Hannes Reinecke		        Kernel Storage Architect
[email protected]			               +49 911 74053 688
SUSE Software Solutions Germany GmbH, 90409 Nürnberg
GF: F. Imendörffer, HRB 36809 (AG Nürnberg)
author
#3 lore raw
Quoted 57 lines · Hannes Reinecke
Quoted 56 lines · Jan Kara

Provided the device driver does not implement dispatch budget accounting (which only SCSI does) the loop in __blk_mq_do_dispatch_sched() pulls requests from the IO scheduler as long as it is willing to give out any. That defeats scheduling heuristics inside the scheduler by creating false impression that the device can take more IO when it in fact cannot.

For example with BFQ IO scheduler on top of virtio-blk device setting blkio cgroup weight has barely any impact on observed throughput of async IO because __blk_mq_do_dispatch_sched() always sucks out all the IO queued in BFQ. BFQ first submits IO from higher weight cgroups but when that is all dispatched, it will give out IO of lower weight cgroups as well. And then we have to wait for all this IO to be dispatched to the disk (which means lot of it actually has to complete) before the IO scheduler is queried again for dispatching more requests. This completely destroys any service differentiation.

So grab request tag for a request pulled out of the IO scheduler already in __blk_mq_do_dispatch_sched() and do not pull any more requests if we cannot get it because we are unlikely to be able to dispatch it. That way only single request is going to wait in the dispatch list for some tag to free.

---
 block/blk-mq-sched.c | 12 +++++++++++-
 block/blk-mq.c       |  2 +-
 block/blk-mq.h       |  2 ++
 3 files changed, 14 insertions(+), 2 deletions(-)

Jens, can you please merge the patch? Thanks!

block/blk-mq-sched.c +11 -1
index 996a4b2f73aa..714e678f516a 100644
@@ -168,9 +168,19 @@ static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
 		 * in blk_mq_dispatch_rq_list().
 		 */
 		list_add_tail(&rq->queuelist, &rq_list);
+		count++;
 		if (rq->mq_hctx != hctx)
 			multi_hctxs = true;
-	} while (++count < max_dispatch);
+
+		/*
+		 * If we cannot get tag for the request, stop dequeueing
+		 * requests from the IO scheduler. We are unlikely to be able
+		 * to submit them anyway and it creates false impression for
+		 * scheduling heuristics that the device can take more IO.
+		 */
+		if (!blk_mq_get_driver_tag(rq))
+			break;
+	} while (count < max_dispatch);
 
Quoted 6 lines

block/blk-mq-sched.c static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx) · line 186

Quoted 2 lines
block/blk-mq-sched.c +0 -0
 	if (!count) {
 		if (run_queue)

Doesn't this lead to a double accounting of the allocated tags? From what I can see we don't really check if the tag is already allocated in blk_mq_get_driver_tag() ...

I think we do check. blk_mq_get_driver_tag() has:

        if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_get_driver_tag(rq))
                return false;

        if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
                        !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
                rq->rq_flags |= RQF_MQ_INFLIGHT;
                __blk_mq_inc_active_requests(hctx);
        }
        hctx->tags->rqs[rq->tag] = rq;

So once we call it, rq->tag will be != BLK_MQ_NO_TAG and RQF_MQ_INFLIGHT will be set. So neither __blk_mq_get_driver_tag() nor __blk_mq_inc_active_requests() will get repeated.

Honza

Signature · 2 lines
Jan Kara <[email protected]>
SUSE Labs, CR
author
#4 lore raw
In reply to Jan Kara
Quoted 25 lines · Jan Kara

Provided the device driver does not implement dispatch budget accounting (which only SCSI does) the loop in __blk_mq_do_dispatch_sched() pulls requests from the IO scheduler as long as it is willing to give out any. That defeats scheduling heuristics inside the scheduler by creating false impression that the device can take more IO when it in fact cannot.

For example with BFQ IO scheduler on top of virtio-blk device setting blkio cgroup weight has barely any impact on observed throughput of async IO because __blk_mq_do_dispatch_sched() always sucks out all the IO queued in BFQ. BFQ first submits IO from higher weight cgroups but when that is all dispatched, it will give out IO of lower weight cgroups as well. And then we have to wait for all this IO to be dispatched to the disk (which means lot of it actually has to complete) before the IO scheduler is queried again for dispatching more requests. This completely destroys any service differentiation.

So grab request tag for a request pulled out of the IO scheduler already in __blk_mq_do_dispatch_sched() and do not pull any more requests if we cannot get it because we are unlikely to be able to dispatch it. That way only single request is going to wait in the dispatch list for some tag to free.

Quoted 7 lines
---
 block/blk-mq-sched.c | 12 +++++++++++-
 block/blk-mq.c       |  2 +-
 block/blk-mq.h       |  2 ++
 3 files changed, 14 insertions(+), 2 deletions(-)

Jens, can you please merge the patch? Thanks!

Ping Jens?

Honza

Quoted 55 lines
block/blk-mq-sched.c +11 -1
index 996a4b2f73aa..714e678f516a 100644
@@ -168,9 +168,19 @@ static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
 		 * in blk_mq_dispatch_rq_list().
 		 */
 		list_add_tail(&rq->queuelist, &rq_list);
+		count++;
 		if (rq->mq_hctx != hctx)
 			multi_hctxs = true;
-	} while (++count < max_dispatch);
+
+		/*
+		 * If we cannot get tag for the request, stop dequeueing
+		 * requests from the IO scheduler. We are unlikely to be able
+		 * to submit them anyway and it creates false impression for
+		 * scheduling heuristics that the device can take more IO.
+		 */
+		if (!blk_mq_get_driver_tag(rq))
+			break;
+	} while (count < max_dispatch);
 
 	if (!count) {
 		if (run_queue)
block/blk-mq.c +1 -1
index c86c01bfecdb..bc2cf80d2c3b 100644
@@ -1100,7 +1100,7 @@ static bool __blk_mq_get_driver_tag(struct request *rq)
 	return true;
 }
 
-static bool blk_mq_get_driver_tag(struct request *rq)
+bool blk_mq_get_driver_tag(struct request *rq)
 {
 	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
 
block/blk-mq.h +2 -0
index 9ce64bc4a6c8..81a775171be7 100644
@@ -259,6 +259,8 @@ static inline void blk_mq_put_driver_tag(struct request *rq)
 	__blk_mq_put_driver_tag(rq->mq_hctx, rq);
 }
 
+bool blk_mq_get_driver_tag(struct request *rq);
+
 static inline void blk_mq_clear_mq_map(struct blk_mq_queue_map *qmap)
 {
 	int cpu;
Signature · 1 line
2.26.2
Signature · 2 lines
Jan Kara <[email protected]>
SUSE Labs, CR
maintainer
#5 lore raw
Quoted 32 lines · Jan Kara
Quoted 31 lines · Jan Kara

Provided the device driver does not implement dispatch budget accounting (which only SCSI does) the loop in __blk_mq_do_dispatch_sched() pulls requests from the IO scheduler as long as it is willing to give out any. That defeats scheduling heuristics inside the scheduler by creating false impression that the device can take more IO when it in fact cannot.

For example with BFQ IO scheduler on top of virtio-blk device setting blkio cgroup weight has barely any impact on observed throughput of async IO because __blk_mq_do_dispatch_sched() always sucks out all the IO queued in BFQ. BFQ first submits IO from higher weight cgroups but when that is all dispatched, it will give out IO of lower weight cgroups as well. And then we have to wait for all this IO to be dispatched to the disk (which means lot of it actually has to complete) before the IO scheduler is queried again for dispatching more requests. This completely destroys any service differentiation.

So grab request tag for a request pulled out of the IO scheduler already in __blk_mq_do_dispatch_sched() and do not pull any more requests if we cannot get it because we are unlikely to be able to dispatch it. That way only single request is going to wait in the dispatch list for some tag to free.

---
 block/blk-mq-sched.c | 12 +++++++++++-
 block/blk-mq.c       |  2 +-
 block/blk-mq.h       |  2 ++
 3 files changed, 14 insertions(+), 2 deletions(-)
Quoted 3 lines
Quoted 1 line

Jens, can you please merge the patch? Thanks!

Ping Jens?

Didn't I reply? It's merged for a while:

https://git.kernel.dk/cgit/linux-block/commit/?h=for-5.14/block&id=613471549f366cdf4170b81ce0f99f3867ec4d16

Signature · 1 line
Jens Axboe
author
#6 lore raw
Quoted 33 lines · Jens Axboe
Quoted 32 lines · Jan Kara
Quoted 31 lines · Jan Kara

Provided the device driver does not implement dispatch budget accounting (which only SCSI does) the loop in __blk_mq_do_dispatch_sched() pulls requests from the IO scheduler as long as it is willing to give out any. That defeats scheduling heuristics inside the scheduler by creating false impression that the device can take more IO when it in fact cannot.

For example with BFQ IO scheduler on top of virtio-blk device setting blkio cgroup weight has barely any impact on observed throughput of async IO because __blk_mq_do_dispatch_sched() always sucks out all the IO queued in BFQ. BFQ first submits IO from higher weight cgroups but when that is all dispatched, it will give out IO of lower weight cgroups as well. And then we have to wait for all this IO to be dispatched to the disk (which means lot of it actually has to complete) before the IO scheduler is queried again for dispatching more requests. This completely destroys any service differentiation.

So grab request tag for a request pulled out of the IO scheduler already in __blk_mq_do_dispatch_sched() and do not pull any more requests if we cannot get it because we are unlikely to be able to dispatch it. That way only single request is going to wait in the dispatch list for some tag to free.

---
 block/blk-mq-sched.c | 12 +++++++++++-
 block/blk-mq.c       |  2 +-
 block/blk-mq.h       |  2 ++
 3 files changed, 14 insertions(+), 2 deletions(-)
Quoted 5 lines
Quoted 3 lines
Quoted 1 line

Jens, can you please merge the patch? Thanks!

Ping Jens?

Didn't I reply? It's merged for a while:

No, I didn't get email back and didn't notice the patch is already in your tree. Sorry for the noise.

Honza

Quoted 6 lines

https://git.kernel.dk/cgit/linux-block/commit/?h=for-5.14/block&id=613471549f366cdf4170b81ce0f99f3867ec4d16

Signature · 1 line
Jens Axboe
Signature · 2 lines
Jan Kara <[email protected]>
SUSE Labs, CR