From: "yukuai (C)" <yukuai3@huawei.com>
To: <axboe@kernel.dk>, <ming.lei@redhat.com>
Cc: <linux-block@vger.kernel.org>, <linux-kernel@vger.kernel.org>,
<yi.zhang@huawei.com>
Subject: Re: [PATCH] blk-mq: allow hardware queue to get more tag while sharing a tag set
Date: Mon, 12 Jul 2021 11:19:25 +0800 [thread overview]
Message-ID: <16b775f7-b700-68dd-7e36-c28292a7b274@huawei.com> (raw)
In-Reply-To: <20210712031818.31918-1-yukuai3@huawei.com>
On 2021/07/12 11:18, Yu Kuai wrote:
> If there are multiple active queues while sharing a tag set, it's not
> necessary to limit the available tags as same share for each active queue
> if no one ever failed to get driver tag. And fall back to same share if
> someone do failed to get driver tag.
>
> This modification will be beneficial if total queue_depth of disks
> on the same host is less than total tags.
Sorry this should be "more than", I'll correct this in next iteration
if anyone think this patch is acceptable.
Thanks,
Yu Kuai
>
> Signed-off-by: Yu Kuai <yukuai3@huawei.com>
> ---
> block/blk-mq-debugfs.c | 2 ++
> block/blk-mq-tag.c | 43 +++++++++++++++++++++++++++++++++++++++++-
> block/blk-mq-tag.h | 27 ++++++++++++++++++++++++--
> block/blk-mq.c | 13 ++++++++++---
> block/blk-mq.h | 8 ++++++++
> include/linux/blk-mq.h | 4 ++++
> include/linux/blkdev.h | 1 +
> 7 files changed, 92 insertions(+), 6 deletions(-)
>
> diff --git a/block/blk-mq-debugfs.c b/block/blk-mq-debugfs.c
> index 4b66d2776eda..35f1f01d93ae 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 86f87346232a..618624b359d6 100644
> --- a/block/blk-mq-tag.c
> +++ b/block/blk-mq-tag.c
> @@ -40,6 +40,22 @@ 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
> */
> @@ -74,6 +90,24 @@ 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)) {
> + if (!test_and_clear_bit(QUEUE_FLAG_HCTX_WAIT,
> + &q->queue_flags))
> + return;
> + atomic_dec(&set->pending_queues_shared_sbitmap);
> + } else {
> + if (!test_and_clear_bit(BLK_MQ_S_DTAG_WAIT, &hctx->state))
> + return;
> + atomic_dec(&tags->pending_queues);
> + }
> +}
> +
> static int __blk_mq_get_tag(struct blk_mq_alloc_data *data,
> struct sbitmap_queue *bt)
> {
> @@ -112,8 +146,12 @@ 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 {
> @@ -140,6 +178,9 @@ 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();
>
> diff --git a/block/blk-mq-tag.h b/block/blk-mq-tag.h
> index 8ed55af08427..badcf3693749 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;
> @@ -69,8 +74,10 @@ enum {
> BLK_MQ_TAG_MAX = BLK_MQ_NO_TAG - 1,
> };
>
> -extern bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *);
> -extern void __blk_mq_tag_idle(struct blk_mq_hw_ctx *);
> +extern bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx);
> +extern void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx);
> +extern void __blk_mq_dtag_busy(struct blk_mq_hw_ctx *hctx);
> +extern void __blk_mq_dtag_idle(struct blk_mq_hw_ctx *hctx);
>
> static inline bool blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
> {
> @@ -88,6 +95,22 @@ 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 2c4ac51e54eb..1bb52bd71da8 100644
> --- a/block/blk-mq.c
> +++ b/block/blk-mq.c
> @@ -991,8 +991,10 @@ 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);
> @@ -1097,8 +1099,10 @@ static bool __blk_mq_get_driver_tag(struct request *rq)
> }
>
> tag = __sbitmap_queue_get(bt);
> - if (tag == BLK_MQ_NO_TAG)
> + if (tag == BLK_MQ_NO_TAG) {
> + blk_mq_dtag_busy(rq->mq_hctx);
> return false;
> + }
>
> rq->tag = tag + tag_offset;
> return true;
> @@ -2676,8 +2680,10 @@ static void blk_mq_exit_hctx(struct request_queue *q,
> {
> struct request *flush_rq = hctx->fq->flush_rq;
>
> - 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_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
> set->queue_depth, flush_rq);
> @@ -3536,6 +3542,7 @@ int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
>
> if (blk_mq_is_sbitmap_shared(set->flags)) {
> atomic_set(&set->active_queues_shared_sbitmap, 0);
> + atomic_set(&set->pending_queues_shared_sbitmap, 0);
>
> if (blk_mq_init_shared_sbitmap(set)) {
> ret = -ENOMEM;
> diff --git a/block/blk-mq.h b/block/blk-mq.h
> index d08779f77a26..9e646ade81a8 100644
> --- a/block/blk-mq.h
> +++ b/block/blk-mq.h
> @@ -337,10 +337,18 @@ static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
>
> if (!test_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags))
> return true;
> +
> + if (!atomic_read(&set->pending_queues_shared_sbitmap))
> + return true;
> +
> users = atomic_read(&set->active_queues_shared_sbitmap);
> } else {
> if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
> return true;
> +
> + if (!atomic_read(&hctx->tags->pending_queues))
> + return true;
> +
> users = atomic_read(&hctx->tags->active_queues);
> }
>
> diff --git a/include/linux/blk-mq.h b/include/linux/blk-mq.h
> index 1d18447ebebc..3bc0faf0e2cf 100644
> --- a/include/linux/blk-mq.h
> +++ b/include/linux/blk-mq.h
> @@ -256,6 +256,7 @@ 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;
> @@ -415,6 +416,9 @@ 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 = 1,
> +
> BLK_MQ_MAX_DEPTH = 10240,
>
> BLK_MQ_CPU_WORK_BATCH = 8,
> diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
> index 3177181c4326..55e0965c9c3c 100644
> --- a/include/linux/blkdev.h
> +++ b/include/linux/blkdev.h
> @@ -603,6 +603,7 @@ 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) | \
>
next prev parent reply other threads:[~2021-07-12 3:19 UTC|newest]
Thread overview: 13+ messages / expand[flat|nested] mbox.gz Atom feed top
2021-07-12 3:18 [PATCH] blk-mq: allow hardware queue to get more tag while sharing a tag set Yu Kuai
2021-07-12 3:19 ` yukuai (C) [this message]
2021-07-20 12:33 ` yukuai (C)
2021-07-31 7:13 ` yukuai (C)
2021-07-31 17:15 ` Bart Van Assche
2021-08-02 13:34 ` yukuai (C)
2021-08-02 16:17 ` Bart Van Assche
2021-08-03 2:57 ` yukuai (C)
2021-08-03 18:38 ` Bart Van Assche
2021-08-06 1:50 ` yukuai (C)
2021-08-06 2:43 ` Bart Van Assche
2021-08-14 9:43 ` yukuai (C)
2021-08-16 4:16 ` Bart Van Assche
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=16b775f7-b700-68dd-7e36-c28292a7b274@huawei.com \
--to=yukuai3@huawei.com \
--cc=axboe@kernel.dk \
--cc=linux-block@vger.kernel.org \
--cc=linux-kernel@vger.kernel.org \
--cc=ming.lei@redhat.com \
--cc=yi.zhang@huawei.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).