| Message ID | 20210520112528.16250-1-jack@suse.cz (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | block: Do not pull requests from the scheduler when we cannot dispatch them | expand |
On Thu, May 20, 2021 at 01:25:28PM +0200, Jan Kara wrote: > 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. So hctx->dispatch_busy isn't set as true in this case? > > 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. > > Signed-off-by: Jan Kara <jack@suse.cz> > --- > block/blk-mq-sched.c | 12 +++++++++++- > block/blk-mq.c | 2 +- > block/blk-mq.h | 2 ++ > 3 files changed, 14 insertions(+), 2 deletions(-) > > diff --git a/block/blk-mq-sched.c b/block/blk-mq-sched.c > index 996a4b2f73aa..714e678f516a 100644 > --- a/block/blk-mq-sched.c > +++ b/block/blk-mq-sched.c > @@ -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; At default BFQ's queue depth is same with virtblk_queue_depth, both are 256, so looks you use non-default setting? Also in case of running out of driver tag, hctx->dispatch_busy should have been set as true for avoiding batching dequeuing, does the following patch make a difference for you? diff --git a/block/blk-mq-sched.c b/block/blk-mq-sched.c index 045b6878b8c5..c2ce3091ad6e 100644 --- a/block/blk-mq-sched.c +++ b/block/blk-mq-sched.c @@ -107,6 +107,13 @@ static bool blk_mq_dispatch_hctx_list(struct list_head *rq_list) #define BLK_MQ_BUDGET_DELAY 3 /* ms units */ +static int blk_mq_sched_max_disaptch(struct blk_mq_hw_ctx *hctx) +{ + if (!hctx->dispatch_busy) + return hctx->queue->nr_requests; + return 1; +} + /* * Only SCSI implements .get_budget and .put_budget, and SCSI restarts * its queue by itself in its completion handler, so we don't need to @@ -121,15 +128,9 @@ static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx) struct elevator_queue *e = q->elevator; bool multi_hctxs = false, run_queue = false; bool dispatched = false, busy = false; - unsigned int max_dispatch; LIST_HEAD(rq_list); int count = 0; - if (hctx->dispatch_busy) - max_dispatch = 1; - else - max_dispatch = hctx->queue->nr_requests; - do { struct request *rq; int budget_token; @@ -170,7 +171,7 @@ static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx) list_add_tail(&rq->queuelist, &rq_list); if (rq->mq_hctx != hctx) multi_hctxs = true; - } while (++count < max_dispatch); + } while (++count < blk_mq_sched_max_disaptch(hctx)); if (!count) { if (run_queue) Thanks, Ming
On Fri 21-05-21 09:29:33, Ming Lei wrote: > On Thu, May 20, 2021 at 01:25:28PM +0200, Jan Kara wrote: > > 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. > > So hctx->dispatch_busy isn't set as true in this case? No. blk_mq_update_dispatch_busy() has: if (hctx->queue->elevator) return; > > 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. > > > > Signed-off-by: Jan Kara <jack@suse.cz> > > --- > > block/blk-mq-sched.c | 12 +++++++++++- > > block/blk-mq.c | 2 +- > > block/blk-mq.h | 2 ++ > > 3 files changed, 14 insertions(+), 2 deletions(-) > > > > diff --git a/block/blk-mq-sched.c b/block/blk-mq-sched.c > > index 996a4b2f73aa..714e678f516a 100644 > > --- a/block/blk-mq-sched.c > > +++ b/block/blk-mq-sched.c > > @@ -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; > > At default BFQ's queue depth is same with virtblk_queue_depth, both are > 256, so looks you use non-default setting? Ah yes, I forgot to mention that. I actually had nr_requests set to 1024 as otherwise all requests get sucked into virtio-blk and no IO scheduling happens either. > Also in case of running out of driver tag, hctx->dispatch_busy should have > been set as true for avoiding batching dequeuing, does the following > patch make a difference for you? I'll try it with modifying blk_mq_update_dispatch_busy() to be updated when elevator is in use... Honza > > > diff --git a/block/blk-mq-sched.c b/block/blk-mq-sched.c > index 045b6878b8c5..c2ce3091ad6e 100644 > --- a/block/blk-mq-sched.c > +++ b/block/blk-mq-sched.c > @@ -107,6 +107,13 @@ static bool blk_mq_dispatch_hctx_list(struct list_head *rq_list) > > #define BLK_MQ_BUDGET_DELAY 3 /* ms units */ > > +static int blk_mq_sched_max_disaptch(struct blk_mq_hw_ctx *hctx) > +{ > + if (!hctx->dispatch_busy) > + return hctx->queue->nr_requests; > + return 1; > +} > + > /* > * Only SCSI implements .get_budget and .put_budget, and SCSI restarts > * its queue by itself in its completion handler, so we don't need to > @@ -121,15 +128,9 @@ static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx) > struct elevator_queue *e = q->elevator; > bool multi_hctxs = false, run_queue = false; > bool dispatched = false, busy = false; > - unsigned int max_dispatch; > LIST_HEAD(rq_list); > int count = 0; > > - if (hctx->dispatch_busy) > - max_dispatch = 1; > - else > - max_dispatch = hctx->queue->nr_requests; > - > do { > struct request *rq; > int budget_token; > @@ -170,7 +171,7 @@ static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx) > list_add_tail(&rq->queuelist, &rq_list); > if (rq->mq_hctx != hctx) > multi_hctxs = true; > - } while (++count < max_dispatch); > + } while (++count < blk_mq_sched_max_disaptch(hctx)); > > if (!count) { > if (run_queue) > > > Thanks, > Ming >
On Fri 21-05-21 13:20:16, Jan Kara wrote: > On Fri 21-05-21 09:29:33, Ming Lei wrote: > > On Thu, May 20, 2021 at 01:25:28PM +0200, Jan Kara wrote: > > > 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. > > > > So hctx->dispatch_busy isn't set as true in this case? > > No. blk_mq_update_dispatch_busy() has: > > if (hctx->queue->elevator) > return; > > > > 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. > > > > > > Signed-off-by: Jan Kara <jack@suse.cz> > > > --- > > > block/blk-mq-sched.c | 12 +++++++++++- > > > block/blk-mq.c | 2 +- > > > block/blk-mq.h | 2 ++ > > > 3 files changed, 14 insertions(+), 2 deletions(-) > > > > > > diff --git a/block/blk-mq-sched.c b/block/blk-mq-sched.c > > > index 996a4b2f73aa..714e678f516a 100644 > > > --- a/block/blk-mq-sched.c > > > +++ b/block/blk-mq-sched.c > > > @@ -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; > > > > At default BFQ's queue depth is same with virtblk_queue_depth, both are > > 256, so looks you use non-default setting? > > Ah yes, I forgot to mention that. I actually had nr_requests set to 1024 as > otherwise all requests get sucked into virtio-blk and no IO scheduling > happens either. > > > Also in case of running out of driver tag, hctx->dispatch_busy should have > > been set as true for avoiding batching dequeuing, does the following > > patch make a difference for you? > > I'll try it with modifying blk_mq_update_dispatch_busy() to be updated when > elevator is in use... OK, tried your patch with the modification and it seems to work OK! Thanks! Honza
On Fri, May 21, 2021 at 01:20:16PM +0200, Jan Kara wrote: > On Fri 21-05-21 09:29:33, Ming Lei wrote: > > On Thu, May 20, 2021 at 01:25:28PM +0200, Jan Kara wrote: > > > 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. > > > > So hctx->dispatch_busy isn't set as true in this case? > > No. blk_mq_update_dispatch_busy() has: > > if (hctx->queue->elevator) > return; ooops, the above check should have been killed in commit 6e6fcbc27e77 ("blk-mq: support batching dispatch in case of io"), :-( Thanks, Ming
On Thu, May 20, 2021 at 01:25:28PM +0200, Jan Kara wrote: > 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. > > Signed-off-by: Jan Kara <jack@suse.cz> > --- > block/blk-mq-sched.c | 12 +++++++++++- > block/blk-mq.c | 2 +- > block/blk-mq.h | 2 ++ > 3 files changed, 14 insertions(+), 2 deletions(-) > > diff --git a/block/blk-mq-sched.c b/block/blk-mq-sched.c > index 996a4b2f73aa..714e678f516a 100644 > --- a/block/blk-mq-sched.c > +++ b/block/blk-mq-sched.c > @@ -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) > diff --git a/block/blk-mq.c b/block/blk-mq.c > index c86c01bfecdb..bc2cf80d2c3b 100644 > --- a/block/blk-mq.c > +++ b/block/blk-mq.c > @@ -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; > > diff --git a/block/blk-mq.h b/block/blk-mq.h > index 9ce64bc4a6c8..81a775171be7 100644 > --- a/block/blk-mq.h > +++ b/block/blk-mq.h > @@ -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; Thinking of further, looks this patch is fine, and it is safe to use driver tag allocation result to decide if more requests need to be dequeued since run queue always be followed when breaking from the loop. Also I can observe that io.bfq.weight is improved on virtio-blk, so Reviewed-by: Ming Lei <ming.lei@redhat.com> Thanks, Ming
On Wed 02-06-21 17:25:52, Ming Lei wrote: > On Thu, May 20, 2021 at 01:25:28PM +0200, Jan Kara wrote: > > 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. > > > > Signed-off-by: Jan Kara <jack@suse.cz> > > --- > > block/blk-mq-sched.c | 12 +++++++++++- > > block/blk-mq.c | 2 +- > > block/blk-mq.h | 2 ++ > > 3 files changed, 14 insertions(+), 2 deletions(-) > > > > diff --git a/block/blk-mq-sched.c b/block/blk-mq-sched.c > > index 996a4b2f73aa..714e678f516a 100644 > > --- a/block/blk-mq-sched.c > > +++ b/block/blk-mq-sched.c > > @@ -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) > > diff --git a/block/blk-mq.c b/block/blk-mq.c > > index c86c01bfecdb..bc2cf80d2c3b 100644 > > --- a/block/blk-mq.c > > +++ b/block/blk-mq.c > > @@ -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; > > > > diff --git a/block/blk-mq.h b/block/blk-mq.h > > index 9ce64bc4a6c8..81a775171be7 100644 > > --- a/block/blk-mq.h > > +++ b/block/blk-mq.h > > @@ -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; > > Thinking of further, looks this patch is fine, and it is safe to use driver tag > allocation result to decide if more requests need to be dequeued since run queue > always be followed when breaking from the loop. Also I can observe that > io.bfq.weight is improved on virtio-blk, so > > Reviewed-by: Ming Lei <ming.lei@redhat.com> OK, thanks for your review! Honza
diff --git a/block/blk-mq-sched.c b/block/blk-mq-sched.c index 996a4b2f73aa..714e678f516a 100644 --- a/block/blk-mq-sched.c +++ b/block/blk-mq-sched.c @@ -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) diff --git a/block/blk-mq.c b/block/blk-mq.c index c86c01bfecdb..bc2cf80d2c3b 100644 --- a/block/blk-mq.c +++ b/block/blk-mq.c @@ -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; diff --git a/block/blk-mq.h b/block/blk-mq.h index 9ce64bc4a6c8..81a775171be7 100644 --- a/block/blk-mq.h +++ b/block/blk-mq.h @@ -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;
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. Signed-off-by: Jan Kara <jack@suse.cz> --- block/blk-mq-sched.c | 12 +++++++++++- block/blk-mq.c | 2 +- block/blk-mq.h | 2 ++ 3 files changed, 14 insertions(+), 2 deletions(-)