2 * Block multiqueue core code
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/backing-dev.h>
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
13 #include <linux/init.h>
14 #include <linux/slab.h>
15 #include <linux/workqueue.h>
16 #include <linux/smp.h>
17 #include <linux/llist.h>
18 #include <linux/list_sort.h>
19 #include <linux/cpu.h>
20 #include <linux/cache.h>
21 #include <linux/sched/sysctl.h>
22 #include <linux/delay.h>
23 #include <linux/crash_dump.h>
25 #include <trace/events/block.h>
27 #include <linux/blk-mq.h>
30 #include "blk-mq-tag.h"
32 static DEFINE_MUTEX(all_q_mutex);
33 static LIST_HEAD(all_q_list);
35 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
38 * Check if any of the ctx's have pending work in this hardware queue
40 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
44 for (i = 0; i < hctx->ctx_map.map_size; i++)
45 if (hctx->ctx_map.map[i].word)
51 static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx,
52 struct blk_mq_ctx *ctx)
54 return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word];
57 #define CTX_TO_BIT(hctx, ctx) \
58 ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
61 * Mark this ctx as having pending work in this hardware queue
63 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
64 struct blk_mq_ctx *ctx)
66 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
68 if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word))
69 set_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
72 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
73 struct blk_mq_ctx *ctx)
75 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
77 clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
80 static int blk_mq_queue_enter(struct request_queue *q)
85 if (percpu_ref_tryget_live(&q->mq_usage_counter))
88 ret = wait_event_interruptible(q->mq_freeze_wq,
89 !q->mq_freeze_depth || blk_queue_dying(q));
90 if (blk_queue_dying(q))
97 static void blk_mq_queue_exit(struct request_queue *q)
99 percpu_ref_put(&q->mq_usage_counter);
102 static void blk_mq_usage_counter_release(struct percpu_ref *ref)
104 struct request_queue *q =
105 container_of(ref, struct request_queue, mq_usage_counter);
107 wake_up_all(&q->mq_freeze_wq);
111 * Guarantee no request is in use, so we can change any data structure of
112 * the queue afterward.
114 void blk_mq_freeze_queue(struct request_queue *q)
118 spin_lock_irq(q->queue_lock);
119 freeze = !q->mq_freeze_depth++;
120 spin_unlock_irq(q->queue_lock);
123 percpu_ref_kill(&q->mq_usage_counter);
124 blk_mq_run_queues(q, false);
126 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter));
129 static void blk_mq_unfreeze_queue(struct request_queue *q)
133 spin_lock_irq(q->queue_lock);
134 wake = !--q->mq_freeze_depth;
135 WARN_ON_ONCE(q->mq_freeze_depth < 0);
136 spin_unlock_irq(q->queue_lock);
138 percpu_ref_reinit(&q->mq_usage_counter);
139 wake_up_all(&q->mq_freeze_wq);
143 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
145 return blk_mq_has_free_tags(hctx->tags);
147 EXPORT_SYMBOL(blk_mq_can_queue);
149 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
150 struct request *rq, unsigned int rw_flags)
152 if (blk_queue_io_stat(q))
153 rw_flags |= REQ_IO_STAT;
155 INIT_LIST_HEAD(&rq->queuelist);
156 /* csd/requeue_work/fifo_time is initialized before use */
159 rq->cmd_flags |= rw_flags;
160 /* do not touch atomic flags, it needs atomic ops against the timer */
162 INIT_HLIST_NODE(&rq->hash);
163 RB_CLEAR_NODE(&rq->rb_node);
166 rq->start_time = jiffies;
167 #ifdef CONFIG_BLK_CGROUP
169 set_start_time_ns(rq);
170 rq->io_start_time_ns = 0;
172 rq->nr_phys_segments = 0;
173 #if defined(CONFIG_BLK_DEV_INTEGRITY)
174 rq->nr_integrity_segments = 0;
177 /* tag was already set */
187 INIT_LIST_HEAD(&rq->timeout_list);
191 rq->end_io_data = NULL;
194 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
197 static struct request *
198 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
203 tag = blk_mq_get_tag(data);
204 if (tag != BLK_MQ_TAG_FAIL) {
205 rq = data->hctx->tags->rqs[tag];
207 if (blk_mq_tag_busy(data->hctx)) {
208 rq->cmd_flags = REQ_MQ_INFLIGHT;
209 atomic_inc(&data->hctx->nr_active);
213 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
220 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
223 struct blk_mq_ctx *ctx;
224 struct blk_mq_hw_ctx *hctx;
226 struct blk_mq_alloc_data alloc_data;
229 ret = blk_mq_queue_enter(q);
233 ctx = blk_mq_get_ctx(q);
234 hctx = q->mq_ops->map_queue(q, ctx->cpu);
235 blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
236 reserved, ctx, hctx);
238 rq = __blk_mq_alloc_request(&alloc_data, rw);
239 if (!rq && (gfp & __GFP_WAIT)) {
240 __blk_mq_run_hw_queue(hctx);
243 ctx = blk_mq_get_ctx(q);
244 hctx = q->mq_ops->map_queue(q, ctx->cpu);
245 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
247 rq = __blk_mq_alloc_request(&alloc_data, rw);
248 ctx = alloc_data.ctx;
252 return ERR_PTR(-EWOULDBLOCK);
255 EXPORT_SYMBOL(blk_mq_alloc_request);
257 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
258 struct blk_mq_ctx *ctx, struct request *rq)
260 const int tag = rq->tag;
261 struct request_queue *q = rq->q;
263 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
264 atomic_dec(&hctx->nr_active);
267 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
268 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
269 blk_mq_queue_exit(q);
272 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
274 struct blk_mq_ctx *ctx = rq->mq_ctx;
276 ctx->rq_completed[rq_is_sync(rq)]++;
277 __blk_mq_free_request(hctx, ctx, rq);
280 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
282 void blk_mq_free_request(struct request *rq)
284 struct blk_mq_hw_ctx *hctx;
285 struct request_queue *q = rq->q;
287 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
288 blk_mq_free_hctx_request(hctx, rq);
290 EXPORT_SYMBOL_GPL(blk_mq_free_request);
292 inline void __blk_mq_end_request(struct request *rq, int error)
294 blk_account_io_done(rq);
297 rq->end_io(rq, error);
299 if (unlikely(blk_bidi_rq(rq)))
300 blk_mq_free_request(rq->next_rq);
301 blk_mq_free_request(rq);
304 EXPORT_SYMBOL(__blk_mq_end_request);
306 void blk_mq_end_request(struct request *rq, int error)
308 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
310 __blk_mq_end_request(rq, error);
312 EXPORT_SYMBOL(blk_mq_end_request);
314 static void __blk_mq_complete_request_remote(void *data)
316 struct request *rq = data;
318 rq->q->softirq_done_fn(rq);
321 static void blk_mq_ipi_complete_request(struct request *rq)
323 struct blk_mq_ctx *ctx = rq->mq_ctx;
327 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
328 rq->q->softirq_done_fn(rq);
333 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
334 shared = cpus_share_cache(cpu, ctx->cpu);
336 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
337 rq->csd.func = __blk_mq_complete_request_remote;
340 smp_call_function_single_async(ctx->cpu, &rq->csd);
342 rq->q->softirq_done_fn(rq);
347 void __blk_mq_complete_request(struct request *rq)
349 struct request_queue *q = rq->q;
351 if (!q->softirq_done_fn)
352 blk_mq_end_request(rq, rq->errors);
354 blk_mq_ipi_complete_request(rq);
358 * blk_mq_complete_request - end I/O on a request
359 * @rq: the request being processed
362 * Ends all I/O on a request. It does not handle partial completions.
363 * The actual completion happens out-of-order, through a IPI handler.
365 void blk_mq_complete_request(struct request *rq)
367 struct request_queue *q = rq->q;
369 if (unlikely(blk_should_fake_timeout(q)))
371 if (!blk_mark_rq_complete(rq))
372 __blk_mq_complete_request(rq);
374 EXPORT_SYMBOL(blk_mq_complete_request);
376 void blk_mq_start_request(struct request *rq)
378 struct request_queue *q = rq->q;
380 trace_block_rq_issue(q, rq);
382 rq->resid_len = blk_rq_bytes(rq);
383 if (unlikely(blk_bidi_rq(rq)))
384 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
389 * Ensure that ->deadline is visible before set the started
390 * flag and clear the completed flag.
392 smp_mb__before_atomic();
395 * Mark us as started and clear complete. Complete might have been
396 * set if requeue raced with timeout, which then marked it as
397 * complete. So be sure to clear complete again when we start
398 * the request, otherwise we'll ignore the completion event.
400 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
401 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
402 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
403 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
405 if (q->dma_drain_size && blk_rq_bytes(rq)) {
407 * Make sure space for the drain appears. We know we can do
408 * this because max_hw_segments has been adjusted to be one
409 * fewer than the device can handle.
411 rq->nr_phys_segments++;
414 EXPORT_SYMBOL(blk_mq_start_request);
416 static void __blk_mq_requeue_request(struct request *rq)
418 struct request_queue *q = rq->q;
420 trace_block_rq_requeue(q, rq);
422 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
423 if (q->dma_drain_size && blk_rq_bytes(rq))
424 rq->nr_phys_segments--;
428 void blk_mq_requeue_request(struct request *rq)
430 __blk_mq_requeue_request(rq);
432 BUG_ON(blk_queued_rq(rq));
433 blk_mq_add_to_requeue_list(rq, true);
435 EXPORT_SYMBOL(blk_mq_requeue_request);
437 static void blk_mq_requeue_work(struct work_struct *work)
439 struct request_queue *q =
440 container_of(work, struct request_queue, requeue_work);
442 struct request *rq, *next;
445 spin_lock_irqsave(&q->requeue_lock, flags);
446 list_splice_init(&q->requeue_list, &rq_list);
447 spin_unlock_irqrestore(&q->requeue_lock, flags);
449 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
450 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
453 rq->cmd_flags &= ~REQ_SOFTBARRIER;
454 list_del_init(&rq->queuelist);
455 blk_mq_insert_request(rq, true, false, false);
458 while (!list_empty(&rq_list)) {
459 rq = list_entry(rq_list.next, struct request, queuelist);
460 list_del_init(&rq->queuelist);
461 blk_mq_insert_request(rq, false, false, false);
465 * Use the start variant of queue running here, so that running
466 * the requeue work will kick stopped queues.
468 blk_mq_start_hw_queues(q);
471 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
473 struct request_queue *q = rq->q;
477 * We abuse this flag that is otherwise used by the I/O scheduler to
478 * request head insertation from the workqueue.
480 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
482 spin_lock_irqsave(&q->requeue_lock, flags);
484 rq->cmd_flags |= REQ_SOFTBARRIER;
485 list_add(&rq->queuelist, &q->requeue_list);
487 list_add_tail(&rq->queuelist, &q->requeue_list);
489 spin_unlock_irqrestore(&q->requeue_lock, flags);
491 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
493 void blk_mq_kick_requeue_list(struct request_queue *q)
495 kblockd_schedule_work(&q->requeue_work);
497 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
499 static inline bool is_flush_request(struct request *rq,
500 struct blk_flush_queue *fq, unsigned int tag)
502 return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
503 fq->flush_rq->tag == tag);
506 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
508 struct request *rq = tags->rqs[tag];
509 /* mq_ctx of flush rq is always cloned from the corresponding req */
510 struct blk_flush_queue *fq = blk_get_flush_queue(rq->q, rq->mq_ctx);
512 if (!is_flush_request(rq, fq, tag))
517 EXPORT_SYMBOL(blk_mq_tag_to_rq);
519 struct blk_mq_timeout_data {
521 unsigned int next_set;
524 void blk_mq_rq_timed_out(struct request *req, bool reserved)
526 struct blk_mq_ops *ops = req->q->mq_ops;
527 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
530 * We know that complete is set at this point. If STARTED isn't set
531 * anymore, then the request isn't active and the "timeout" should
532 * just be ignored. This can happen due to the bitflag ordering.
533 * Timeout first checks if STARTED is set, and if it is, assumes
534 * the request is active. But if we race with completion, then
535 * we both flags will get cleared. So check here again, and ignore
536 * a timeout event with a request that isn't active.
538 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
542 ret = ops->timeout(req, reserved);
546 __blk_mq_complete_request(req);
548 case BLK_EH_RESET_TIMER:
550 blk_clear_rq_complete(req);
552 case BLK_EH_NOT_HANDLED:
555 printk(KERN_ERR "block: bad eh return: %d\n", ret);
560 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
561 struct request *rq, void *priv, bool reserved)
563 struct blk_mq_timeout_data *data = priv;
565 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
568 if (time_after_eq(jiffies, rq->deadline)) {
569 if (!blk_mark_rq_complete(rq))
570 blk_mq_rq_timed_out(rq, reserved);
571 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
572 data->next = rq->deadline;
577 static void blk_mq_rq_timer(unsigned long priv)
579 struct request_queue *q = (struct request_queue *)priv;
580 struct blk_mq_timeout_data data = {
584 struct blk_mq_hw_ctx *hctx;
587 queue_for_each_hw_ctx(q, hctx, i) {
589 * If not software queues are currently mapped to this
590 * hardware queue, there's nothing to check
592 if (!hctx->nr_ctx || !hctx->tags)
595 blk_mq_tag_busy_iter(hctx, blk_mq_check_expired, &data);
599 data.next = blk_rq_timeout(round_jiffies_up(data.next));
600 mod_timer(&q->timeout, data.next);
602 queue_for_each_hw_ctx(q, hctx, i)
603 blk_mq_tag_idle(hctx);
608 * Reverse check our software queue for entries that we could potentially
609 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
610 * too much time checking for merges.
612 static bool blk_mq_attempt_merge(struct request_queue *q,
613 struct blk_mq_ctx *ctx, struct bio *bio)
618 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
624 if (!blk_rq_merge_ok(rq, bio))
627 el_ret = blk_try_merge(rq, bio);
628 if (el_ret == ELEVATOR_BACK_MERGE) {
629 if (bio_attempt_back_merge(q, rq, bio)) {
634 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
635 if (bio_attempt_front_merge(q, rq, bio)) {
647 * Process software queues that have been marked busy, splicing them
648 * to the for-dispatch
650 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
652 struct blk_mq_ctx *ctx;
655 for (i = 0; i < hctx->ctx_map.map_size; i++) {
656 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
657 unsigned int off, bit;
663 off = i * hctx->ctx_map.bits_per_word;
665 bit = find_next_bit(&bm->word, bm->depth, bit);
666 if (bit >= bm->depth)
669 ctx = hctx->ctxs[bit + off];
670 clear_bit(bit, &bm->word);
671 spin_lock(&ctx->lock);
672 list_splice_tail_init(&ctx->rq_list, list);
673 spin_unlock(&ctx->lock);
681 * Run this hardware queue, pulling any software queues mapped to it in.
682 * Note that this function currently has various problems around ordering
683 * of IO. In particular, we'd like FIFO behaviour on handling existing
684 * items on the hctx->dispatch list. Ignore that for now.
686 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
688 struct request_queue *q = hctx->queue;
691 LIST_HEAD(driver_list);
692 struct list_head *dptr;
695 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
697 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
703 * Touch any software queue that has pending entries.
705 flush_busy_ctxs(hctx, &rq_list);
708 * If we have previous entries on our dispatch list, grab them
709 * and stuff them at the front for more fair dispatch.
711 if (!list_empty_careful(&hctx->dispatch)) {
712 spin_lock(&hctx->lock);
713 if (!list_empty(&hctx->dispatch))
714 list_splice_init(&hctx->dispatch, &rq_list);
715 spin_unlock(&hctx->lock);
719 * Start off with dptr being NULL, so we start the first request
720 * immediately, even if we have more pending.
725 * Now process all the entries, sending them to the driver.
728 while (!list_empty(&rq_list)) {
729 struct blk_mq_queue_data bd;
732 rq = list_first_entry(&rq_list, struct request, queuelist);
733 list_del_init(&rq->queuelist);
737 bd.last = list_empty(&rq_list);
739 ret = q->mq_ops->queue_rq(hctx, &bd);
741 case BLK_MQ_RQ_QUEUE_OK:
744 case BLK_MQ_RQ_QUEUE_BUSY:
745 list_add(&rq->queuelist, &rq_list);
746 __blk_mq_requeue_request(rq);
749 pr_err("blk-mq: bad return on queue: %d\n", ret);
750 case BLK_MQ_RQ_QUEUE_ERROR:
752 blk_mq_end_request(rq, rq->errors);
756 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
760 * We've done the first request. If we have more than 1
761 * left in the list, set dptr to defer issue.
763 if (!dptr && rq_list.next != rq_list.prev)
768 hctx->dispatched[0]++;
769 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
770 hctx->dispatched[ilog2(queued) + 1]++;
773 * Any items that need requeuing? Stuff them into hctx->dispatch,
774 * that is where we will continue on next queue run.
776 if (!list_empty(&rq_list)) {
777 spin_lock(&hctx->lock);
778 list_splice(&rq_list, &hctx->dispatch);
779 spin_unlock(&hctx->lock);
784 * It'd be great if the workqueue API had a way to pass
785 * in a mask and had some smarts for more clever placement.
786 * For now we just round-robin here, switching for every
787 * BLK_MQ_CPU_WORK_BATCH queued items.
789 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
791 int cpu = hctx->next_cpu;
793 if (--hctx->next_cpu_batch <= 0) {
796 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
797 if (next_cpu >= nr_cpu_ids)
798 next_cpu = cpumask_first(hctx->cpumask);
800 hctx->next_cpu = next_cpu;
801 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
807 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
809 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
814 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
815 __blk_mq_run_hw_queue(hctx);
823 if (hctx->queue->nr_hw_queues == 1)
824 kblockd_schedule_delayed_work(&hctx->run_work, 0);
828 cpu = blk_mq_hctx_next_cpu(hctx);
829 kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0);
833 void blk_mq_run_queues(struct request_queue *q, bool async)
835 struct blk_mq_hw_ctx *hctx;
838 queue_for_each_hw_ctx(q, hctx, i) {
839 if ((!blk_mq_hctx_has_pending(hctx) &&
840 list_empty_careful(&hctx->dispatch)) ||
841 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
844 blk_mq_run_hw_queue(hctx, async);
847 EXPORT_SYMBOL(blk_mq_run_queues);
849 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
851 cancel_delayed_work(&hctx->run_work);
852 cancel_delayed_work(&hctx->delay_work);
853 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
855 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
857 void blk_mq_stop_hw_queues(struct request_queue *q)
859 struct blk_mq_hw_ctx *hctx;
862 queue_for_each_hw_ctx(q, hctx, i)
863 blk_mq_stop_hw_queue(hctx);
865 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
867 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
869 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
871 blk_mq_run_hw_queue(hctx, false);
873 EXPORT_SYMBOL(blk_mq_start_hw_queue);
875 void blk_mq_start_hw_queues(struct request_queue *q)
877 struct blk_mq_hw_ctx *hctx;
880 queue_for_each_hw_ctx(q, hctx, i)
881 blk_mq_start_hw_queue(hctx);
883 EXPORT_SYMBOL(blk_mq_start_hw_queues);
886 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
888 struct blk_mq_hw_ctx *hctx;
891 queue_for_each_hw_ctx(q, hctx, i) {
892 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
895 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
896 blk_mq_run_hw_queue(hctx, async);
899 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
901 static void blk_mq_run_work_fn(struct work_struct *work)
903 struct blk_mq_hw_ctx *hctx;
905 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
907 __blk_mq_run_hw_queue(hctx);
910 static void blk_mq_delay_work_fn(struct work_struct *work)
912 struct blk_mq_hw_ctx *hctx;
914 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
916 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
917 __blk_mq_run_hw_queue(hctx);
920 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
922 unsigned long tmo = msecs_to_jiffies(msecs);
924 if (hctx->queue->nr_hw_queues == 1)
925 kblockd_schedule_delayed_work(&hctx->delay_work, tmo);
929 cpu = blk_mq_hctx_next_cpu(hctx);
930 kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo);
933 EXPORT_SYMBOL(blk_mq_delay_queue);
935 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
936 struct request *rq, bool at_head)
938 struct blk_mq_ctx *ctx = rq->mq_ctx;
940 trace_block_rq_insert(hctx->queue, rq);
943 list_add(&rq->queuelist, &ctx->rq_list);
945 list_add_tail(&rq->queuelist, &ctx->rq_list);
947 blk_mq_hctx_mark_pending(hctx, ctx);
950 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
953 struct request_queue *q = rq->q;
954 struct blk_mq_hw_ctx *hctx;
955 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
957 current_ctx = blk_mq_get_ctx(q);
958 if (!cpu_online(ctx->cpu))
959 rq->mq_ctx = ctx = current_ctx;
961 hctx = q->mq_ops->map_queue(q, ctx->cpu);
963 spin_lock(&ctx->lock);
964 __blk_mq_insert_request(hctx, rq, at_head);
965 spin_unlock(&ctx->lock);
968 blk_mq_run_hw_queue(hctx, async);
970 blk_mq_put_ctx(current_ctx);
973 static void blk_mq_insert_requests(struct request_queue *q,
974 struct blk_mq_ctx *ctx,
975 struct list_head *list,
980 struct blk_mq_hw_ctx *hctx;
981 struct blk_mq_ctx *current_ctx;
983 trace_block_unplug(q, depth, !from_schedule);
985 current_ctx = blk_mq_get_ctx(q);
987 if (!cpu_online(ctx->cpu))
989 hctx = q->mq_ops->map_queue(q, ctx->cpu);
992 * preemption doesn't flush plug list, so it's possible ctx->cpu is
995 spin_lock(&ctx->lock);
996 while (!list_empty(list)) {
999 rq = list_first_entry(list, struct request, queuelist);
1000 list_del_init(&rq->queuelist);
1002 __blk_mq_insert_request(hctx, rq, false);
1004 spin_unlock(&ctx->lock);
1006 blk_mq_run_hw_queue(hctx, from_schedule);
1007 blk_mq_put_ctx(current_ctx);
1010 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1012 struct request *rqa = container_of(a, struct request, queuelist);
1013 struct request *rqb = container_of(b, struct request, queuelist);
1015 return !(rqa->mq_ctx < rqb->mq_ctx ||
1016 (rqa->mq_ctx == rqb->mq_ctx &&
1017 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1020 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1022 struct blk_mq_ctx *this_ctx;
1023 struct request_queue *this_q;
1026 LIST_HEAD(ctx_list);
1029 list_splice_init(&plug->mq_list, &list);
1031 list_sort(NULL, &list, plug_ctx_cmp);
1037 while (!list_empty(&list)) {
1038 rq = list_entry_rq(list.next);
1039 list_del_init(&rq->queuelist);
1041 if (rq->mq_ctx != this_ctx) {
1043 blk_mq_insert_requests(this_q, this_ctx,
1048 this_ctx = rq->mq_ctx;
1054 list_add_tail(&rq->queuelist, &ctx_list);
1058 * If 'this_ctx' is set, we know we have entries to complete
1059 * on 'ctx_list'. Do those.
1062 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1067 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1069 init_request_from_bio(rq, bio);
1071 if (blk_do_io_stat(rq))
1072 blk_account_io_start(rq, 1);
1075 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1077 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1078 !blk_queue_nomerges(hctx->queue);
1081 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1082 struct blk_mq_ctx *ctx,
1083 struct request *rq, struct bio *bio)
1085 if (!hctx_allow_merges(hctx)) {
1086 blk_mq_bio_to_request(rq, bio);
1087 spin_lock(&ctx->lock);
1089 __blk_mq_insert_request(hctx, rq, false);
1090 spin_unlock(&ctx->lock);
1093 struct request_queue *q = hctx->queue;
1095 spin_lock(&ctx->lock);
1096 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1097 blk_mq_bio_to_request(rq, bio);
1101 spin_unlock(&ctx->lock);
1102 __blk_mq_free_request(hctx, ctx, rq);
1107 struct blk_map_ctx {
1108 struct blk_mq_hw_ctx *hctx;
1109 struct blk_mq_ctx *ctx;
1112 static struct request *blk_mq_map_request(struct request_queue *q,
1114 struct blk_map_ctx *data)
1116 struct blk_mq_hw_ctx *hctx;
1117 struct blk_mq_ctx *ctx;
1119 int rw = bio_data_dir(bio);
1120 struct blk_mq_alloc_data alloc_data;
1122 if (unlikely(blk_mq_queue_enter(q))) {
1123 bio_endio(bio, -EIO);
1127 ctx = blk_mq_get_ctx(q);
1128 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1130 if (rw_is_sync(bio->bi_rw))
1133 trace_block_getrq(q, bio, rw);
1134 blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1136 rq = __blk_mq_alloc_request(&alloc_data, rw);
1137 if (unlikely(!rq)) {
1138 __blk_mq_run_hw_queue(hctx);
1139 blk_mq_put_ctx(ctx);
1140 trace_block_sleeprq(q, bio, rw);
1142 ctx = blk_mq_get_ctx(q);
1143 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1144 blk_mq_set_alloc_data(&alloc_data, q,
1145 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1146 rq = __blk_mq_alloc_request(&alloc_data, rw);
1147 ctx = alloc_data.ctx;
1148 hctx = alloc_data.hctx;
1158 * Multiple hardware queue variant. This will not use per-process plugs,
1159 * but will attempt to bypass the hctx queueing if we can go straight to
1160 * hardware for SYNC IO.
1162 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1164 const int is_sync = rw_is_sync(bio->bi_rw);
1165 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1166 struct blk_map_ctx data;
1169 blk_queue_bounce(q, &bio);
1171 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1172 bio_endio(bio, -EIO);
1176 rq = blk_mq_map_request(q, bio, &data);
1180 if (unlikely(is_flush_fua)) {
1181 blk_mq_bio_to_request(rq, bio);
1182 blk_insert_flush(rq);
1187 * If the driver supports defer issued based on 'last', then
1188 * queue it up like normal since we can potentially save some
1191 if (is_sync && !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1192 struct blk_mq_queue_data bd = {
1199 blk_mq_bio_to_request(rq, bio);
1202 * For OK queue, we are done. For error, kill it. Any other
1203 * error (busy), just add it to our list as we previously
1206 ret = q->mq_ops->queue_rq(data.hctx, &bd);
1207 if (ret == BLK_MQ_RQ_QUEUE_OK)
1210 __blk_mq_requeue_request(rq);
1212 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1214 blk_mq_end_request(rq, rq->errors);
1220 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1222 * For a SYNC request, send it to the hardware immediately. For
1223 * an ASYNC request, just ensure that we run it later on. The
1224 * latter allows for merging opportunities and more efficient
1228 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1231 blk_mq_put_ctx(data.ctx);
1235 * Single hardware queue variant. This will attempt to use any per-process
1236 * plug for merging and IO deferral.
1238 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1240 const int is_sync = rw_is_sync(bio->bi_rw);
1241 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1242 unsigned int use_plug, request_count = 0;
1243 struct blk_map_ctx data;
1247 * If we have multiple hardware queues, just go directly to
1248 * one of those for sync IO.
1250 use_plug = !is_flush_fua && !is_sync;
1252 blk_queue_bounce(q, &bio);
1254 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1255 bio_endio(bio, -EIO);
1259 if (use_plug && !blk_queue_nomerges(q) &&
1260 blk_attempt_plug_merge(q, bio, &request_count))
1263 rq = blk_mq_map_request(q, bio, &data);
1267 if (unlikely(is_flush_fua)) {
1268 blk_mq_bio_to_request(rq, bio);
1269 blk_insert_flush(rq);
1274 * A task plug currently exists. Since this is completely lockless,
1275 * utilize that to temporarily store requests until the task is
1276 * either done or scheduled away.
1279 struct blk_plug *plug = current->plug;
1282 blk_mq_bio_to_request(rq, bio);
1283 if (list_empty(&plug->mq_list))
1284 trace_block_plug(q);
1285 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1286 blk_flush_plug_list(plug, false);
1287 trace_block_plug(q);
1289 list_add_tail(&rq->queuelist, &plug->mq_list);
1290 blk_mq_put_ctx(data.ctx);
1295 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1297 * For a SYNC request, send it to the hardware immediately. For
1298 * an ASYNC request, just ensure that we run it later on. The
1299 * latter allows for merging opportunities and more efficient
1303 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1306 blk_mq_put_ctx(data.ctx);
1310 * Default mapping to a software queue, since we use one per CPU.
1312 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1314 return q->queue_hw_ctx[q->mq_map[cpu]];
1316 EXPORT_SYMBOL(blk_mq_map_queue);
1318 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1319 struct blk_mq_tags *tags, unsigned int hctx_idx)
1323 if (tags->rqs && set->ops->exit_request) {
1326 for (i = 0; i < tags->nr_tags; i++) {
1329 set->ops->exit_request(set->driver_data, tags->rqs[i],
1331 tags->rqs[i] = NULL;
1335 while (!list_empty(&tags->page_list)) {
1336 page = list_first_entry(&tags->page_list, struct page, lru);
1337 list_del_init(&page->lru);
1338 __free_pages(page, page->private);
1343 blk_mq_free_tags(tags);
1346 static size_t order_to_size(unsigned int order)
1348 return (size_t)PAGE_SIZE << order;
1351 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1352 unsigned int hctx_idx)
1354 struct blk_mq_tags *tags;
1355 unsigned int i, j, entries_per_page, max_order = 4;
1356 size_t rq_size, left;
1358 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1363 INIT_LIST_HEAD(&tags->page_list);
1365 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1366 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1369 blk_mq_free_tags(tags);
1374 * rq_size is the size of the request plus driver payload, rounded
1375 * to the cacheline size
1377 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1379 left = rq_size * set->queue_depth;
1381 for (i = 0; i < set->queue_depth; ) {
1382 int this_order = max_order;
1387 while (left < order_to_size(this_order - 1) && this_order)
1391 page = alloc_pages_node(set->numa_node,
1392 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1398 if (order_to_size(this_order) < rq_size)
1405 page->private = this_order;
1406 list_add_tail(&page->lru, &tags->page_list);
1408 p = page_address(page);
1409 entries_per_page = order_to_size(this_order) / rq_size;
1410 to_do = min(entries_per_page, set->queue_depth - i);
1411 left -= to_do * rq_size;
1412 for (j = 0; j < to_do; j++) {
1414 tags->rqs[i]->atomic_flags = 0;
1415 tags->rqs[i]->cmd_flags = 0;
1416 if (set->ops->init_request) {
1417 if (set->ops->init_request(set->driver_data,
1418 tags->rqs[i], hctx_idx, i,
1420 tags->rqs[i] = NULL;
1433 blk_mq_free_rq_map(set, tags, hctx_idx);
1437 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1442 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1444 unsigned int bpw = 8, total, num_maps, i;
1446 bitmap->bits_per_word = bpw;
1448 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1449 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1454 bitmap->map_size = num_maps;
1457 for (i = 0; i < num_maps; i++) {
1458 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1459 total -= bitmap->map[i].depth;
1465 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1467 struct request_queue *q = hctx->queue;
1468 struct blk_mq_ctx *ctx;
1472 * Move ctx entries to new CPU, if this one is going away.
1474 ctx = __blk_mq_get_ctx(q, cpu);
1476 spin_lock(&ctx->lock);
1477 if (!list_empty(&ctx->rq_list)) {
1478 list_splice_init(&ctx->rq_list, &tmp);
1479 blk_mq_hctx_clear_pending(hctx, ctx);
1481 spin_unlock(&ctx->lock);
1483 if (list_empty(&tmp))
1486 ctx = blk_mq_get_ctx(q);
1487 spin_lock(&ctx->lock);
1489 while (!list_empty(&tmp)) {
1492 rq = list_first_entry(&tmp, struct request, queuelist);
1494 list_move_tail(&rq->queuelist, &ctx->rq_list);
1497 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1498 blk_mq_hctx_mark_pending(hctx, ctx);
1500 spin_unlock(&ctx->lock);
1502 blk_mq_run_hw_queue(hctx, true);
1503 blk_mq_put_ctx(ctx);
1507 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1509 struct request_queue *q = hctx->queue;
1510 struct blk_mq_tag_set *set = q->tag_set;
1512 if (set->tags[hctx->queue_num])
1515 set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1516 if (!set->tags[hctx->queue_num])
1519 hctx->tags = set->tags[hctx->queue_num];
1523 static int blk_mq_hctx_notify(void *data, unsigned long action,
1526 struct blk_mq_hw_ctx *hctx = data;
1528 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1529 return blk_mq_hctx_cpu_offline(hctx, cpu);
1530 else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1531 return blk_mq_hctx_cpu_online(hctx, cpu);
1536 static void blk_mq_exit_hctx(struct request_queue *q,
1537 struct blk_mq_tag_set *set,
1538 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1540 unsigned flush_start_tag = set->queue_depth;
1542 blk_mq_tag_idle(hctx);
1544 if (set->ops->exit_request)
1545 set->ops->exit_request(set->driver_data,
1546 hctx->fq->flush_rq, hctx_idx,
1547 flush_start_tag + hctx_idx);
1549 if (set->ops->exit_hctx)
1550 set->ops->exit_hctx(hctx, hctx_idx);
1552 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1553 blk_free_flush_queue(hctx->fq);
1555 blk_mq_free_bitmap(&hctx->ctx_map);
1558 static void blk_mq_exit_hw_queues(struct request_queue *q,
1559 struct blk_mq_tag_set *set, int nr_queue)
1561 struct blk_mq_hw_ctx *hctx;
1564 queue_for_each_hw_ctx(q, hctx, i) {
1567 blk_mq_exit_hctx(q, set, hctx, i);
1571 static void blk_mq_free_hw_queues(struct request_queue *q,
1572 struct blk_mq_tag_set *set)
1574 struct blk_mq_hw_ctx *hctx;
1577 queue_for_each_hw_ctx(q, hctx, i) {
1578 free_cpumask_var(hctx->cpumask);
1583 static int blk_mq_init_hctx(struct request_queue *q,
1584 struct blk_mq_tag_set *set,
1585 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1588 unsigned flush_start_tag = set->queue_depth;
1590 node = hctx->numa_node;
1591 if (node == NUMA_NO_NODE)
1592 node = hctx->numa_node = set->numa_node;
1594 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1595 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1596 spin_lock_init(&hctx->lock);
1597 INIT_LIST_HEAD(&hctx->dispatch);
1599 hctx->queue_num = hctx_idx;
1600 hctx->flags = set->flags;
1601 hctx->cmd_size = set->cmd_size;
1603 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1604 blk_mq_hctx_notify, hctx);
1605 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1607 hctx->tags = set->tags[hctx_idx];
1610 * Allocate space for all possible cpus to avoid allocation at
1613 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1616 goto unregister_cpu_notifier;
1618 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1623 if (set->ops->init_hctx &&
1624 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1627 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1631 if (set->ops->init_request &&
1632 set->ops->init_request(set->driver_data,
1633 hctx->fq->flush_rq, hctx_idx,
1634 flush_start_tag + hctx_idx, node))
1642 if (set->ops->exit_hctx)
1643 set->ops->exit_hctx(hctx, hctx_idx);
1645 blk_mq_free_bitmap(&hctx->ctx_map);
1648 unregister_cpu_notifier:
1649 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1654 static int blk_mq_init_hw_queues(struct request_queue *q,
1655 struct blk_mq_tag_set *set)
1657 struct blk_mq_hw_ctx *hctx;
1661 * Initialize hardware queues
1663 queue_for_each_hw_ctx(q, hctx, i) {
1664 if (blk_mq_init_hctx(q, set, hctx, i))
1668 if (i == q->nr_hw_queues)
1674 blk_mq_exit_hw_queues(q, set, i);
1679 static void blk_mq_init_cpu_queues(struct request_queue *q,
1680 unsigned int nr_hw_queues)
1684 for_each_possible_cpu(i) {
1685 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1686 struct blk_mq_hw_ctx *hctx;
1688 memset(__ctx, 0, sizeof(*__ctx));
1690 spin_lock_init(&__ctx->lock);
1691 INIT_LIST_HEAD(&__ctx->rq_list);
1694 /* If the cpu isn't online, the cpu is mapped to first hctx */
1698 hctx = q->mq_ops->map_queue(q, i);
1699 cpumask_set_cpu(i, hctx->cpumask);
1703 * Set local node, IFF we have more than one hw queue. If
1704 * not, we remain on the home node of the device
1706 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1707 hctx->numa_node = cpu_to_node(i);
1711 static void blk_mq_map_swqueue(struct request_queue *q)
1714 struct blk_mq_hw_ctx *hctx;
1715 struct blk_mq_ctx *ctx;
1717 queue_for_each_hw_ctx(q, hctx, i) {
1718 cpumask_clear(hctx->cpumask);
1723 * Map software to hardware queues
1725 queue_for_each_ctx(q, ctx, i) {
1726 /* If the cpu isn't online, the cpu is mapped to first hctx */
1730 hctx = q->mq_ops->map_queue(q, i);
1731 cpumask_set_cpu(i, hctx->cpumask);
1732 ctx->index_hw = hctx->nr_ctx;
1733 hctx->ctxs[hctx->nr_ctx++] = ctx;
1736 queue_for_each_hw_ctx(q, hctx, i) {
1738 * If no software queues are mapped to this hardware queue,
1739 * disable it and free the request entries.
1741 if (!hctx->nr_ctx) {
1742 struct blk_mq_tag_set *set = q->tag_set;
1745 blk_mq_free_rq_map(set, set->tags[i], i);
1746 set->tags[i] = NULL;
1753 * Initialize batch roundrobin counts
1755 hctx->next_cpu = cpumask_first(hctx->cpumask);
1756 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1760 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1762 struct blk_mq_hw_ctx *hctx;
1763 struct request_queue *q;
1767 if (set->tag_list.next == set->tag_list.prev)
1772 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1773 blk_mq_freeze_queue(q);
1775 queue_for_each_hw_ctx(q, hctx, i) {
1777 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1779 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1781 blk_mq_unfreeze_queue(q);
1785 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1787 struct blk_mq_tag_set *set = q->tag_set;
1789 mutex_lock(&set->tag_list_lock);
1790 list_del_init(&q->tag_set_list);
1791 blk_mq_update_tag_set_depth(set);
1792 mutex_unlock(&set->tag_list_lock);
1795 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1796 struct request_queue *q)
1800 mutex_lock(&set->tag_list_lock);
1801 list_add_tail(&q->tag_set_list, &set->tag_list);
1802 blk_mq_update_tag_set_depth(set);
1803 mutex_unlock(&set->tag_list_lock);
1806 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1808 struct blk_mq_hw_ctx **hctxs;
1809 struct blk_mq_ctx __percpu *ctx;
1810 struct request_queue *q;
1814 ctx = alloc_percpu(struct blk_mq_ctx);
1816 return ERR_PTR(-ENOMEM);
1819 * If a crashdump is active, then we are potentially in a very
1820 * memory constrained environment. Limit us to 1 queue and
1821 * 64 tags to prevent using too much memory.
1823 if (is_kdump_kernel()) {
1824 set->nr_hw_queues = 1;
1825 set->queue_depth = min(64U, set->queue_depth);
1828 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1834 map = blk_mq_make_queue_map(set);
1838 for (i = 0; i < set->nr_hw_queues; i++) {
1839 int node = blk_mq_hw_queue_to_node(map, i);
1841 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1846 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1850 atomic_set(&hctxs[i]->nr_active, 0);
1851 hctxs[i]->numa_node = node;
1852 hctxs[i]->queue_num = i;
1855 q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1860 * Init percpu_ref in atomic mode so that it's faster to shutdown.
1861 * See blk_register_queue() for details.
1863 if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release,
1864 PERCPU_REF_INIT_ATOMIC, GFP_KERNEL))
1867 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1868 blk_queue_rq_timeout(q, 30000);
1870 q->nr_queues = nr_cpu_ids;
1871 q->nr_hw_queues = set->nr_hw_queues;
1875 q->queue_hw_ctx = hctxs;
1877 q->mq_ops = set->ops;
1878 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1880 if (!(set->flags & BLK_MQ_F_SG_MERGE))
1881 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1883 q->sg_reserved_size = INT_MAX;
1885 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1886 INIT_LIST_HEAD(&q->requeue_list);
1887 spin_lock_init(&q->requeue_lock);
1889 if (q->nr_hw_queues > 1)
1890 blk_queue_make_request(q, blk_mq_make_request);
1892 blk_queue_make_request(q, blk_sq_make_request);
1895 blk_queue_rq_timeout(q, set->timeout);
1898 * Do this after blk_queue_make_request() overrides it...
1900 q->nr_requests = set->queue_depth;
1902 if (set->ops->complete)
1903 blk_queue_softirq_done(q, set->ops->complete);
1905 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1907 if (blk_mq_init_hw_queues(q, set))
1910 mutex_lock(&all_q_mutex);
1911 list_add_tail(&q->all_q_node, &all_q_list);
1912 mutex_unlock(&all_q_mutex);
1914 blk_mq_add_queue_tag_set(set, q);
1916 blk_mq_map_swqueue(q);
1921 blk_cleanup_queue(q);
1924 for (i = 0; i < set->nr_hw_queues; i++) {
1927 free_cpumask_var(hctxs[i]->cpumask);
1934 return ERR_PTR(-ENOMEM);
1936 EXPORT_SYMBOL(blk_mq_init_queue);
1938 void blk_mq_free_queue(struct request_queue *q)
1940 struct blk_mq_tag_set *set = q->tag_set;
1942 blk_mq_del_queue_tag_set(q);
1944 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
1945 blk_mq_free_hw_queues(q, set);
1947 percpu_ref_exit(&q->mq_usage_counter);
1949 free_percpu(q->queue_ctx);
1950 kfree(q->queue_hw_ctx);
1953 q->queue_ctx = NULL;
1954 q->queue_hw_ctx = NULL;
1957 mutex_lock(&all_q_mutex);
1958 list_del_init(&q->all_q_node);
1959 mutex_unlock(&all_q_mutex);
1962 /* Basically redo blk_mq_init_queue with queue frozen */
1963 static void blk_mq_queue_reinit(struct request_queue *q)
1965 blk_mq_freeze_queue(q);
1967 blk_mq_sysfs_unregister(q);
1969 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1972 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1973 * we should change hctx numa_node according to new topology (this
1974 * involves free and re-allocate memory, worthy doing?)
1977 blk_mq_map_swqueue(q);
1979 blk_mq_sysfs_register(q);
1981 blk_mq_unfreeze_queue(q);
1984 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1985 unsigned long action, void *hcpu)
1987 struct request_queue *q;
1990 * Before new mappings are established, hotadded cpu might already
1991 * start handling requests. This doesn't break anything as we map
1992 * offline CPUs to first hardware queue. We will re-init the queue
1993 * below to get optimal settings.
1995 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1996 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1999 mutex_lock(&all_q_mutex);
2000 list_for_each_entry(q, &all_q_list, all_q_node)
2001 blk_mq_queue_reinit(q);
2002 mutex_unlock(&all_q_mutex);
2006 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2010 for (i = 0; i < set->nr_hw_queues; i++) {
2011 set->tags[i] = blk_mq_init_rq_map(set, i);
2020 blk_mq_free_rq_map(set, set->tags[i], i);
2026 * Allocate the request maps associated with this tag_set. Note that this
2027 * may reduce the depth asked for, if memory is tight. set->queue_depth
2028 * will be updated to reflect the allocated depth.
2030 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2035 depth = set->queue_depth;
2037 err = __blk_mq_alloc_rq_maps(set);
2041 set->queue_depth >>= 1;
2042 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2046 } while (set->queue_depth);
2048 if (!set->queue_depth || err) {
2049 pr_err("blk-mq: failed to allocate request map\n");
2053 if (depth != set->queue_depth)
2054 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2055 depth, set->queue_depth);
2061 * Alloc a tag set to be associated with one or more request queues.
2062 * May fail with EINVAL for various error conditions. May adjust the
2063 * requested depth down, if if it too large. In that case, the set
2064 * value will be stored in set->queue_depth.
2066 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2068 if (!set->nr_hw_queues)
2070 if (!set->queue_depth)
2072 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2075 if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue)
2078 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2079 pr_info("blk-mq: reduced tag depth to %u\n",
2081 set->queue_depth = BLK_MQ_MAX_DEPTH;
2084 set->tags = kmalloc_node(set->nr_hw_queues *
2085 sizeof(struct blk_mq_tags *),
2086 GFP_KERNEL, set->numa_node);
2090 if (blk_mq_alloc_rq_maps(set))
2093 mutex_init(&set->tag_list_lock);
2094 INIT_LIST_HEAD(&set->tag_list);
2102 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2104 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2108 for (i = 0; i < set->nr_hw_queues; i++) {
2110 blk_mq_free_rq_map(set, set->tags[i], i);
2116 EXPORT_SYMBOL(blk_mq_free_tag_set);
2118 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2120 struct blk_mq_tag_set *set = q->tag_set;
2121 struct blk_mq_hw_ctx *hctx;
2124 if (!set || nr > set->queue_depth)
2128 queue_for_each_hw_ctx(q, hctx, i) {
2129 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2135 q->nr_requests = nr;
2140 void blk_mq_disable_hotplug(void)
2142 mutex_lock(&all_q_mutex);
2145 void blk_mq_enable_hotplug(void)
2147 mutex_unlock(&all_q_mutex);
2150 static int __init blk_mq_init(void)
2154 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2158 subsys_initcall(blk_mq_init);