Message ID | 20201226102808.2534966-2-yukuai3@huawei.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | fix the performance fluctuation due to shared tagset | expand |
Hi Yu Kuai, On Sat, Dec 26, 2020 at 06:28:06PM +0800, Yu Kuai wrote: > When sharing a tag set, if most disks are issuing small amount of IO, and > only a few is issuing a large amount of IO. Current approach is to limit > the max amount of tags a disk can get equally to the average of total > tags. Thus the few heavy load disk can't get enough tags while many tags > are still free in the tag set. Yeah, current approach just allocates same share for each active queue which is evaluated in each timeout period. That said you are trying to improve the following case: - heavy IO on one or several disks, and the average share for these disks become bottleneck of IO performance - small amount IO on other disks attached to the same host, and all IOs are submitted to disk in <30 second period. Just wondering if you may share the workload you are trying to optimize, or it is just one improvement in theory? And what is the disk(hdd, ssd or nvme) and host? And how many disks in your setting? And how deep the tagset depth is? Thanks, Ming
Hi, On 2020/12/27 19:58, Ming Lei wrote: > Hi Yu Kuai, > > On Sat, Dec 26, 2020 at 06:28:06PM +0800, Yu Kuai wrote: >> When sharing a tag set, if most disks are issuing small amount of IO, and >> only a few is issuing a large amount of IO. Current approach is to limit >> the max amount of tags a disk can get equally to the average of total >> tags. Thus the few heavy load disk can't get enough tags while many tags >> are still free in the tag set. > > Yeah, current approach just allocates same share for each active queue > which is evaluated in each timeout period. > > That said you are trying to improve the following case: > - heavy IO on one or several disks, and the average share for these > disks become bottleneck of IO performance > - small amount IO on other disks attached to the same host, and all IOs are > submitted to disk in <30 second period. > > Just wondering if you may share the workload you are trying to optimize, > or it is just one improvement in theory? And what is the disk(hdd, ssd > or nvme) and host? And how many disks in your setting? And how deep the tagset > depth is? The details of the environment that we found the problem are as follows: total driver tags: 128 number of disks: 13 (network drive, and they form a dm-multipath) default queue_depth: 32 disk performance: when test with 4k randread and single thread, iops is 300. And can up to 4000 with 32 thread. test cmd: fio -ioengine=psync -numjobs=32 ... We found that mpath will issue sg_io periodically(about 15s),which lead to active_queues setting to 13 for about 5s in every 15s. By the way, I'm not sure this is a common scenario, however, sq don't have such problem, Thanks Yu Kuai
On Mon, Dec 28, 2020 at 09:56:15AM +0800, yukuai (C) wrote: > Hi, > > On 2020/12/27 19:58, Ming Lei wrote: > > Hi Yu Kuai, > > > > On Sat, Dec 26, 2020 at 06:28:06PM +0800, Yu Kuai wrote: > > > When sharing a tag set, if most disks are issuing small amount of IO, and > > > only a few is issuing a large amount of IO. Current approach is to limit > > > the max amount of tags a disk can get equally to the average of total > > > tags. Thus the few heavy load disk can't get enough tags while many tags > > > are still free in the tag set. > > > > Yeah, current approach just allocates same share for each active queue > > which is evaluated in each timeout period. > > > > That said you are trying to improve the following case: > > - heavy IO on one or several disks, and the average share for these > > disks become bottleneck of IO performance > > - small amount IO on other disks attached to the same host, and all IOs are > > submitted to disk in <30 second period. > > > > Just wondering if you may share the workload you are trying to optimize, > > or it is just one improvement in theory? And what is the disk(hdd, ssd > > or nvme) and host? And how many disks in your setting? And how deep the tagset > > depth is? > > The details of the environment that we found the problem are as follows: > > total driver tags: 128 Looks the tagset depth is a bit low. > number of disks: 13 (network drive, and they form a dm-multipath) > default queue_depth: 32 Another candidate solution may be to always return true from hctx_may_queue() for this kind of queue because queue_depth has provided fair allocation for each LUN, and looks not necessary to do that again. > disk performance: when test with 4k randread and single thread, iops is > 300. And can up to 4000 with 32 thread. > test cmd: fio -ioengine=psync -numjobs=32 ... > > We found that mpath will issue sg_io periodically(about 15s),which lead > to active_queues setting to 13 for about 5s in every 15s. BTW, I just observe sg_io on rhel8 & rhel7 on mpath over scsi_debug, looks not see any such activity. > > By the way, I'm not sure this is a common scenario, however, sq don't > have such problem, If it is done by mpath at default setting, I think it can be thought as one common case. Thanks, Ming
Hi On 2020/12/28 16:28, Ming Lei wrote: > Another candidate solution may be to always return true from hctx_may_queue() > for this kind of queue because queue_depth has provided fair allocation for > each LUN, and looks not necessary to do that again. If always return true from hctx_may_queue() in this case, for example, we set queue_depth to 128(if can't, the biggger, the better) for all disks, and test with numjobs=64. The result should be one disk with high iops, and the rest very low. So I think it's better to ensure the max tags a disk can get in this case. Thanks! Yu Kuai
On Mon, Dec 28, 2020 at 05:02:50PM +0800, yukuai (C) wrote: > Hi > > On 2020/12/28 16:28, Ming Lei wrote: > > Another candidate solution may be to always return true from hctx_may_queue() > > for this kind of queue because queue_depth has provided fair allocation for > > each LUN, and looks not necessary to do that again. > > If always return true from hctx_may_queue() in this case, for example, > we set queue_depth to 128(if can't, the biggger, the better) for all > disks, and test with numjobs=64. The result should be one disk with high > iops, and the rest very low. So I think it's better to ensure the max > tags a disk can get in this case. Just wondering why you try to set 128 via sysfs for all disks? If you do that, you should know the potential result given the whole tags queue depth is just 128. Thanks, Ming
Hi On 2020/12/29 9:15, Ming Lei wrote: > Just wondering why you try to set 128 via sysfs for all disks? If you do that, > you should know the potential result given the whole tags queue depth is just > 128. It's just a extreme example to show the unexpected result of "always return true from hctx_may_queue()". Thanks, Yu Kuai
diff --git a/block/blk-mq-debugfs.c b/block/blk-mq-debugfs.c index 9336a6f8d6ef..fed193b10406 100644 --- a/block/blk-mq-debugfs.c +++ b/block/blk-mq-debugfs.c @@ -450,6 +450,8 @@ static void blk_mq_debugfs_tags_show(struct seq_file *m, seq_printf(m, "nr_reserved_tags=%u\n", tags->nr_reserved_tags); seq_printf(m, "active_queues=%d\n", atomic_read(&tags->active_queues)); + seq_printf(m, "pending_queues=%d\n", + atomic_read(&tags->pending_queues)); seq_puts(m, "\nbitmap_tags:\n"); sbitmap_queue_show(tags->bitmap_tags, m); diff --git a/block/blk-mq-tag.c b/block/blk-mq-tag.c index 9c92053e704d..6dcd6dd9123a 100644 --- a/block/blk-mq-tag.c +++ b/block/blk-mq-tag.c @@ -39,6 +39,22 @@ bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx) return true; } +void __blk_mq_dtag_busy(struct blk_mq_hw_ctx *hctx) +{ + if (blk_mq_is_sbitmap_shared(hctx->flags)) { + struct request_queue *q = hctx->queue; + struct blk_mq_tag_set *set = q->tag_set; + + if (!test_bit(QUEUE_FLAG_HCTX_WAIT, &q->queue_flags) && + !test_and_set_bit(QUEUE_FLAG_HCTX_WAIT, &q->queue_flags)) + atomic_inc(&set->pending_queues_shared_sbitmap); + } else { + if (!test_bit(BLK_MQ_S_DTAG_WAIT, &hctx->state) && + !test_and_set_bit(BLK_MQ_S_DTAG_WAIT, &hctx->state)) + atomic_inc(&hctx->tags->pending_queues); + } +} + /* * Wakeup all potentially sleeping on tags */ @@ -73,6 +89,19 @@ void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx) blk_mq_tag_wakeup_all(tags, false); } +void __blk_mq_dtag_idle(struct blk_mq_hw_ctx *hctx) +{ + struct blk_mq_tags *tags = hctx->tags; + struct request_queue *q = hctx->queue; + struct blk_mq_tag_set *set = q->tag_set; + + if (blk_mq_is_sbitmap_shared(hctx->flags) && + test_and_clear_bit(QUEUE_FLAG_HCTX_WAIT, &q->queue_flags)) + atomic_dec(&set->pending_queues_shared_sbitmap); + else if (test_and_clear_bit(BLK_MQ_S_DTAG_WAIT, &hctx->state)) + atomic_dec(&tags->pending_queues); +} + static int __blk_mq_get_tag(struct blk_mq_alloc_data *data, struct sbitmap_queue *bt) { @@ -111,8 +140,11 @@ unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data) if (tag != BLK_MQ_NO_TAG) goto found_tag; - if (data->flags & BLK_MQ_REQ_NOWAIT) + if (data->flags & BLK_MQ_REQ_NOWAIT) { + if (!data->q->elevator) + blk_mq_dtag_busy(data->hctx); return BLK_MQ_NO_TAG; + } ws = bt_wait_ptr(bt, data->hctx); do { @@ -139,6 +171,9 @@ unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data) if (tag != BLK_MQ_NO_TAG) break; + if (!data->q->elevator) + blk_mq_dtag_busy(data->hctx); + bt_prev = bt; io_schedule(); @@ -459,6 +494,7 @@ static int blk_mq_init_bitmap_tags(struct blk_mq_tags *tags, tags->bitmap_tags = &tags->__bitmap_tags; tags->breserved_tags = &tags->__breserved_tags; + atomic_set(&tags->pending_queues, 0); return 0; free_bitmap_tags: diff --git a/block/blk-mq-tag.h b/block/blk-mq-tag.h index 7d3e6b333a4a..52d08a92f683 100644 --- a/block/blk-mq-tag.h +++ b/block/blk-mq-tag.h @@ -10,6 +10,11 @@ struct blk_mq_tags { unsigned int nr_reserved_tags; atomic_t active_queues; + /* + * if multiple queues share a tag set, pending_queues record the + * number of queues that can't get driver tag. + */ + atomic_t pending_queues; struct sbitmap_queue *bitmap_tags; struct sbitmap_queue *breserved_tags; @@ -62,6 +67,8 @@ enum { extern bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *); extern void __blk_mq_tag_idle(struct blk_mq_hw_ctx *); +extern void __blk_mq_dtag_busy(struct blk_mq_hw_ctx *); +extern void __blk_mq_dtag_idle(struct blk_mq_hw_ctx *); static inline bool blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx) { @@ -79,6 +86,22 @@ static inline void blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx) __blk_mq_tag_idle(hctx); } +static inline void blk_mq_dtag_busy(struct blk_mq_hw_ctx *hctx) +{ + if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) + return; + + __blk_mq_dtag_busy(hctx); +} + +static inline void blk_mq_dtag_idle(struct blk_mq_hw_ctx *hctx) +{ + if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) + return; + + __blk_mq_dtag_idle(hctx); +} + static inline bool blk_mq_tag_is_reserved(struct blk_mq_tags *tags, unsigned int tag) { diff --git a/block/blk-mq.c b/block/blk-mq.c index f285a9123a8b..2b8fa49bccb4 100644 --- a/block/blk-mq.c +++ b/block/blk-mq.c @@ -1013,8 +1013,10 @@ static void blk_mq_timeout_work(struct work_struct *work) */ queue_for_each_hw_ctx(q, hctx, i) { /* the hctx may be unmapped, so check it here */ - if (blk_mq_hw_queue_mapped(hctx)) + if (blk_mq_hw_queue_mapped(hctx)) { blk_mq_tag_idle(hctx); + blk_mq_dtag_idle(hctx); + } } } blk_queue_exit(q); @@ -2564,8 +2566,10 @@ static void blk_mq_exit_hctx(struct request_queue *q, struct blk_mq_tag_set *set, struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) { - if (blk_mq_hw_queue_mapped(hctx)) + if (blk_mq_hw_queue_mapped(hctx)) { blk_mq_tag_idle(hctx); + blk_mq_dtag_idle(hctx); + } if (set->ops->exit_request) set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); diff --git a/block/blk-mq.h b/block/blk-mq.h index c1458d9502f1..f7212bacfa56 100644 --- a/block/blk-mq.h +++ b/block/blk-mq.h @@ -114,7 +114,7 @@ static inline struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, type = HCTX_TYPE_POLL; else if ((flags & REQ_OP_MASK) == REQ_OP_READ) type = HCTX_TYPE_READ; - + return ctx->hctxs[type]; } @@ -212,8 +212,8 @@ static inline void __blk_mq_dec_active_requests(struct blk_mq_hw_ctx *hctx) { if (blk_mq_is_sbitmap_shared(hctx->flags)) atomic_dec(&hctx->queue->nr_active_requests_shared_sbitmap); - else - atomic_dec(&hctx->nr_active); + else if (!atomic_dec_return(&hctx->nr_active)) + blk_mq_dtag_idle(hctx); } static inline int __blk_mq_active_requests(struct blk_mq_hw_ctx *hctx) @@ -300,6 +300,9 @@ static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx, if (bt->sb.depth == 1) return true; + if (!atomic_read(&hctx->tags->pending_queues)) + return true; + if (blk_mq_is_sbitmap_shared(hctx->flags)) { struct request_queue *q = hctx->queue; struct blk_mq_tag_set *set = q->tag_set; diff --git a/include/linux/blk-mq.h b/include/linux/blk-mq.h index d705b174d346..2a473865ee7f 100644 --- a/include/linux/blk-mq.h +++ b/include/linux/blk-mq.h @@ -260,6 +260,7 @@ struct blk_mq_tag_set { unsigned int flags; void *driver_data; atomic_t active_queues_shared_sbitmap; + atomic_t pending_queues_shared_sbitmap; struct sbitmap_queue __bitmap_tags; struct sbitmap_queue __breserved_tags; @@ -410,6 +411,9 @@ enum { /* hw queue is inactive after all its CPUs become offline */ BLK_MQ_S_INACTIVE = 3, + /* hw queue is waiting for driver tag */ + BLK_MQ_S_DTAG_WAIT = 4, + BLK_MQ_MAX_DEPTH = 10240, BLK_MQ_CPU_WORK_BATCH = 8, diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h index f94ee3089e01..445249e00b9d 100644 --- a/include/linux/blkdev.h +++ b/include/linux/blkdev.h @@ -621,6 +621,7 @@ struct request_queue { #define QUEUE_FLAG_RQ_ALLOC_TIME 27 /* record rq->alloc_time_ns */ #define QUEUE_FLAG_HCTX_ACTIVE 28 /* at least one blk-mq hctx is active */ #define QUEUE_FLAG_NOWAIT 29 /* device supports NOWAIT */ +#define QUEUE_FLAG_HCTX_WAIT 30 /* at least one blk-mq hctx can't get driver tag */ #define QUEUE_FLAG_MQ_DEFAULT ((1 << QUEUE_FLAG_IO_STAT) | \ (1 << QUEUE_FLAG_SAME_COMP) | \