blk-mq: complete req in softirq context in case of single queue
PATCH v2 1/1 blk-mq: complete req in softirq context in case of single queue
Lot of controllers may have only one irq vector for completing IO request. And usually affinity of the only irq vector is all possible CPUs, however, on most of ARCH, there may be only one specific CPU for handling this interrupt.
So if all IOs are completed in hardirq context, it is inevitable to degrade IO performance because of increased irq latency.
This patch tries to address this issue by allowing to complete request in softirq context, like the legacy IO path.
IOPS is observed as ~13%+ in the following randread test on raid0 over virtio-scsi.
mdadm --create --verbose /dev/md0 --level=0 --chunk=1024 --raid-devices=8 /dev/sdb /dev/sdc /dev/sdd /dev/sde /dev/sdf /dev/sdg /dev/sdh /dev/sdi
fio --time_based --name=benchmark --runtime=30 --filename=/dev/md0 --nrfiles=1 --ioengine=libaio --iodepth=32 --direct=1 --invalidate=1 --verify=0 --verify_fatal=0 --numjobs=32 --rw=randread --blocksize=4k
--- block/blk-mq.c | 14 ++++++++++++++ block/blk-softirq.c | 5 ++--- 2 files changed, 16 insertions(+), 3 deletions(-)
block/blk-mq.c +14 -0
@@ -565,6 +565,20 @@ static void __blk_mq_complete_request(struct request *rq)if (rq->internal_tag !=-1 )blk_mq_sched_completed_request (rq); +/* +* Most of single queue controllers, there is only one irq vector +* for handling IO completion, and the only irq's affinity is set +* as all possible CPUs. On most of ARCHs, this affinity means the +* irq is handled on one specific CPU. +* +* So complete IO reqeust in softirq context in case of single queue +* for not degrading IO performance by irqsoff latency. +*/ +if (rq->q->nr_hw_queues ==1 ) { +__blk_complete_request (rq); +return ; + } +if (!test_bit (QUEUE_FLAG_SAME_COMP , &rq->q->queue_flags)) { rq->q->softirq_done_fn (rq);return ;
block/blk-softirq.c +2 -3
@@ -97,8 +97,8 @@ static int blk_softirq_cpu_dead(unsigned int cpu)void __blk_complete_request (struct request *req) { -int ccpu, cpu;struct request_queue *q = req->q; +int cpu, ccpu = q->mq_ops ? req->mq_ctx->cpu : req->cpu;unsigned long flags;bool shared = false;
@@ -110,8 +110,7 @@ void __blk_complete_request(struct request *req)/* * Select completion CPU */ -if (req->cpu !=-1 ) { - ccpu = req->cpu; +if (test_bit (QUEUE_FLAG_SAME_COMP , &q->queue_flags) && ccpu !=-1 ) {if (!test_bit (QUEUE_FLAG_SAME_FORCE , &q->queue_flags)) shared =cpus_share_cache (cpu, ccpu); }else
Signature · 1 line
2.9.5
Quoted 81 lines · Ming Lei
Lot of controllers may have only one irq vector for completing IO request. And usually affinity of the only irq vector is all possible CPUs, however, on most of ARCH, there may be only one specific CPU for handling this interrupt.
So if all IOs are completed in hardirq context, it is inevitable to degrade IO performance because of increased irq latency.
This patch tries to address this issue by allowing to complete request in softirq context, like the legacy IO path.
IOPS is observed as ~13%+ in the following randread test on raid0 over virtio-scsi.
mdadm --create --verbose
/dev/md0--level=0 --chunk=1024 --raid-devices=8/dev/sdb/dev/sdc/dev/sdd/dev/sde/dev/sdf/dev/sdg/dev/sdh/dev/sdifio --time_based --name=benchmark --runtime=30 --filename=/dev/md0 --nrfiles=1 --ioengine=libaio --iodepth=32 --direct=1 --invalidate=1 --verify=0 --verify_fatal=0 --numjobs=32 --rw=randread --blocksize=4k
Cc Dongli Zhang <[email protected]>Cc Zach Marano <[email protected]>Cc Christoph Hellwig <[email protected]>Cc Bart Van Assche <[email protected]>Cc Jianchao Wang <[email protected]>Signed-off-by Ming Lei <[email protected]>--- block/blk-mq.c | 14 ++++++++++++++ block/blk-softirq.c | 5 ++--- 2 files changed, 16 insertions(+), 3 deletions(-)block/blk-mq.c +14
-0@@ -565,6 +565,20 @@ static void __blk_mq_complete_request(struct request *rq)if (rq->internal_tag !=-1 )blk_mq_sched_completed_request (rq); +/* +* Most of single queue controllers, there is only one irq vector +* for handling IO completion, and the only irq's affinity is set +* as all possible CPUs. On most of ARCHs, this affinity means the +* irq is handled on one specific CPU. +* +* So complete IO reqeust in softirq context in case of single queue +* for not degrading IO performance by irqsoff latency. +*/ +if (rq->q->nr_hw_queues ==1 ) { +__blk_complete_request (rq); +return ; + } +if (!test_bit (QUEUE_FLAG_SAME_COMP , &rq->q->queue_flags)) { rq->q->softirq_done_fn (rq);return ;block/blk-softirq.c +2
-3@@ -97,8 +97,8 @@ static int blk_softirq_cpu_dead(unsigned int cpu)void __blk_complete_request (struct request *req) { -int ccpu, cpu;struct request_queue *q = req->q; +int cpu, ccpu = q->mq_ops ? req->mq_ctx->cpu : req->cpu;unsigned long flags;bool shared = false;@@ -110,8 +110,7 @@ void __blk_complete_request(struct request *req)/* * Select completion CPU */ -if (req->cpu !=-1 ) { - ccpu = req->cpu; +if (test_bit (QUEUE_FLAG_SAME_COMP , &q->queue_flags) && ccpu !=-1 ) {if (!test_bit (QUEUE_FLAG_SAME_FORCE , &q->queue_flags)) shared =cpus_share_cache (cpu, ccpu); }else Signature · 1 line
2.9.5
Ping...
Thanks,
Ming
Jens Axboe
maintainerQuoted 11 lines · Ming Lei
Lot of controllers may have only one irq vector for completing IO request. And usually affinity of the only irq vector is all possible CPUs, however, on most of ARCH, there may be only one specific CPU for handling this interrupt.
So if all IOs are completed in hardirq context, it is inevitable to degrade IO performance because of increased irq latency.
This patch tries to address this issue by allowing to complete request in softirq context, like the legacy IO path.
Quoted 6 lines
IOPS is observed as ~13%+ in the following randread test on raid0 over virtio-scsi.
mdadm --create --verbose
/dev/md0--level=0 --chunk=1024 --raid-devices=8/dev/sdb/dev/sdc/dev/sdd/dev/sde/dev/sdf/dev/sdg/dev/sdh/dev/sdifio --time_based --name=benchmark --runtime=30 --filename=/dev/md0 --nrfiles=1 --ioengine=libaio --iodepth=32 --direct=1 --invalidate=1 --verify=0 --verify_fatal=0 --numjobs=32 --rw=randread --blocksize=4k
Looks fine to me, and imho it makes a lot of sense to apply the same logic for remote completions to single queue as we did for the legacy parts. I have applied it for 4.20.
Signature · 1 line
Jens Axboe