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.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, gfp_t gfp)
85 if (percpu_ref_tryget_live(&q->mq_usage_counter))
88 if (!(gfp & __GFP_WAIT))
91 ret = swait_event_interruptible(q->mq_freeze_wq,
92 !q->mq_freeze_depth || blk_queue_dying(q));
93 if (blk_queue_dying(q))
100 static void blk_mq_queue_exit(struct request_queue *q)
102 percpu_ref_put(&q->mq_usage_counter);
105 static void blk_mq_usage_counter_release(struct percpu_ref *ref)
107 struct request_queue *q =
108 container_of(ref, struct request_queue, mq_usage_counter);
110 swait_wake_all(&q->mq_freeze_wq);
113 void blk_mq_freeze_queue_start(struct request_queue *q)
117 spin_lock_irq(q->queue_lock);
118 freeze = !q->mq_freeze_depth++;
119 spin_unlock_irq(q->queue_lock);
122 percpu_ref_kill(&q->mq_usage_counter);
123 blk_mq_run_hw_queues(q, false);
126 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
128 static void blk_mq_freeze_queue_wait(struct request_queue *q)
130 swait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter));
134 * Guarantee no request is in use, so we can change any data structure of
135 * the queue afterward.
137 void blk_mq_freeze_queue(struct request_queue *q)
139 blk_mq_freeze_queue_start(q);
140 blk_mq_freeze_queue_wait(q);
142 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
144 void blk_mq_unfreeze_queue(struct request_queue *q)
148 spin_lock_irq(q->queue_lock);
149 wake = !--q->mq_freeze_depth;
150 WARN_ON_ONCE(q->mq_freeze_depth < 0);
151 spin_unlock_irq(q->queue_lock);
153 percpu_ref_reinit(&q->mq_usage_counter);
154 swait_wake_all(&q->mq_freeze_wq);
157 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
159 void blk_mq_wake_waiters(struct request_queue *q)
161 struct blk_mq_hw_ctx *hctx;
164 queue_for_each_hw_ctx(q, hctx, i)
165 if (blk_mq_hw_queue_mapped(hctx))
166 blk_mq_tag_wakeup_all(hctx->tags, true);
169 * If we are called because the queue has now been marked as
170 * dying, we need to ensure that processes currently waiting on
171 * the queue are notified as well.
173 swait_wake_all(&q->mq_freeze_wq);
176 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
178 return blk_mq_has_free_tags(hctx->tags);
180 EXPORT_SYMBOL(blk_mq_can_queue);
182 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
183 struct request *rq, unsigned int rw_flags)
185 if (blk_queue_io_stat(q))
186 rw_flags |= REQ_IO_STAT;
188 INIT_LIST_HEAD(&rq->queuelist);
189 /* csd/requeue_work/fifo_time is initialized before use */
192 rq->cmd_flags |= rw_flags;
193 /* do not touch atomic flags, it needs atomic ops against the timer */
195 INIT_HLIST_NODE(&rq->hash);
196 RB_CLEAR_NODE(&rq->rb_node);
199 rq->start_time = jiffies;
200 #ifdef CONFIG_BLK_CGROUP
202 set_start_time_ns(rq);
203 rq->io_start_time_ns = 0;
205 rq->nr_phys_segments = 0;
206 #if defined(CONFIG_BLK_DEV_INTEGRITY)
207 rq->nr_integrity_segments = 0;
210 /* tag was already set */
220 #ifdef CONFIG_PREEMPT_RT_FULL
221 INIT_WORK(&rq->work, __blk_mq_complete_request_remote_work);
223 INIT_LIST_HEAD(&rq->timeout_list);
227 rq->end_io_data = NULL;
230 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
233 static struct request *
234 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
239 tag = blk_mq_get_tag(data);
240 if (tag != BLK_MQ_TAG_FAIL) {
241 rq = data->hctx->tags->rqs[tag];
243 if (blk_mq_tag_busy(data->hctx)) {
244 rq->cmd_flags = REQ_MQ_INFLIGHT;
245 atomic_inc(&data->hctx->nr_active);
249 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
256 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
259 struct blk_mq_ctx *ctx;
260 struct blk_mq_hw_ctx *hctx;
262 struct blk_mq_alloc_data alloc_data;
265 ret = blk_mq_queue_enter(q, gfp);
269 ctx = blk_mq_get_ctx(q);
270 hctx = q->mq_ops->map_queue(q, ctx->cpu);
271 blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
272 reserved, ctx, hctx);
274 rq = __blk_mq_alloc_request(&alloc_data, rw);
275 if (!rq && (gfp & __GFP_WAIT)) {
276 __blk_mq_run_hw_queue(hctx);
279 ctx = blk_mq_get_ctx(q);
280 hctx = q->mq_ops->map_queue(q, ctx->cpu);
281 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
283 rq = __blk_mq_alloc_request(&alloc_data, rw);
284 ctx = alloc_data.ctx;
288 blk_mq_queue_exit(q);
289 return ERR_PTR(-EWOULDBLOCK);
293 EXPORT_SYMBOL(blk_mq_alloc_request);
295 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
296 struct blk_mq_ctx *ctx, struct request *rq)
298 const int tag = rq->tag;
299 struct request_queue *q = rq->q;
301 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
302 atomic_dec(&hctx->nr_active);
305 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
306 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
307 blk_mq_queue_exit(q);
310 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
312 struct blk_mq_ctx *ctx = rq->mq_ctx;
314 ctx->rq_completed[rq_is_sync(rq)]++;
315 __blk_mq_free_request(hctx, ctx, rq);
318 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
320 void blk_mq_free_request(struct request *rq)
322 struct blk_mq_hw_ctx *hctx;
323 struct request_queue *q = rq->q;
325 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
326 blk_mq_free_hctx_request(hctx, rq);
328 EXPORT_SYMBOL_GPL(blk_mq_free_request);
330 inline void __blk_mq_end_request(struct request *rq, int error)
332 blk_account_io_done(rq);
335 rq->end_io(rq, error);
337 if (unlikely(blk_bidi_rq(rq)))
338 blk_mq_free_request(rq->next_rq);
339 blk_mq_free_request(rq);
342 EXPORT_SYMBOL(__blk_mq_end_request);
344 void blk_mq_end_request(struct request *rq, int error)
346 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
348 __blk_mq_end_request(rq, error);
350 EXPORT_SYMBOL(blk_mq_end_request);
352 #ifdef CONFIG_PREEMPT_RT_FULL
354 void __blk_mq_complete_request_remote_work(struct work_struct *work)
356 struct request *rq = container_of(work, struct request, work);
358 rq->q->softirq_done_fn(rq);
363 static void __blk_mq_complete_request_remote(void *data)
365 struct request *rq = data;
367 rq->q->softirq_done_fn(rq);
372 static void blk_mq_ipi_complete_request(struct request *rq)
374 struct blk_mq_ctx *ctx = rq->mq_ctx;
378 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
379 rq->q->softirq_done_fn(rq);
383 cpu = get_cpu_light();
384 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
385 shared = cpus_share_cache(cpu, ctx->cpu);
387 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
388 #ifdef CONFIG_PREEMPT_RT_FULL
389 schedule_work_on(ctx->cpu, &rq->work);
391 rq->csd.func = __blk_mq_complete_request_remote;
394 smp_call_function_single_async(ctx->cpu, &rq->csd);
397 rq->q->softirq_done_fn(rq);
402 void __blk_mq_complete_request(struct request *rq)
404 struct request_queue *q = rq->q;
406 if (!q->softirq_done_fn)
407 blk_mq_end_request(rq, rq->errors);
409 blk_mq_ipi_complete_request(rq);
413 * blk_mq_complete_request - end I/O on a request
414 * @rq: the request being processed
417 * Ends all I/O on a request. It does not handle partial completions.
418 * The actual completion happens out-of-order, through a IPI handler.
420 void blk_mq_complete_request(struct request *rq)
422 struct request_queue *q = rq->q;
424 if (unlikely(blk_should_fake_timeout(q)))
426 if (!blk_mark_rq_complete(rq))
427 __blk_mq_complete_request(rq);
429 EXPORT_SYMBOL(blk_mq_complete_request);
431 int blk_mq_request_started(struct request *rq)
433 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
435 EXPORT_SYMBOL_GPL(blk_mq_request_started);
437 void blk_mq_start_request(struct request *rq)
439 struct request_queue *q = rq->q;
441 trace_block_rq_issue(q, rq);
443 rq->resid_len = blk_rq_bytes(rq);
444 if (unlikely(blk_bidi_rq(rq)))
445 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
450 * Ensure that ->deadline is visible before set the started
451 * flag and clear the completed flag.
453 smp_mb__before_atomic();
456 * Mark us as started and clear complete. Complete might have been
457 * set if requeue raced with timeout, which then marked it as
458 * complete. So be sure to clear complete again when we start
459 * the request, otherwise we'll ignore the completion event.
461 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
462 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
463 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
464 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
466 if (q->dma_drain_size && blk_rq_bytes(rq)) {
468 * Make sure space for the drain appears. We know we can do
469 * this because max_hw_segments has been adjusted to be one
470 * fewer than the device can handle.
472 rq->nr_phys_segments++;
475 EXPORT_SYMBOL(blk_mq_start_request);
477 static void __blk_mq_requeue_request(struct request *rq)
479 struct request_queue *q = rq->q;
481 trace_block_rq_requeue(q, rq);
483 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
484 if (q->dma_drain_size && blk_rq_bytes(rq))
485 rq->nr_phys_segments--;
489 void blk_mq_requeue_request(struct request *rq)
491 __blk_mq_requeue_request(rq);
493 BUG_ON(blk_queued_rq(rq));
494 blk_mq_add_to_requeue_list(rq, true);
496 EXPORT_SYMBOL(blk_mq_requeue_request);
498 static void blk_mq_requeue_work(struct work_struct *work)
500 struct request_queue *q =
501 container_of(work, struct request_queue, requeue_work);
503 struct request *rq, *next;
506 spin_lock_irqsave(&q->requeue_lock, flags);
507 list_splice_init(&q->requeue_list, &rq_list);
508 spin_unlock_irqrestore(&q->requeue_lock, flags);
510 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
511 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
514 rq->cmd_flags &= ~REQ_SOFTBARRIER;
515 list_del_init(&rq->queuelist);
516 blk_mq_insert_request(rq, true, false, false);
519 while (!list_empty(&rq_list)) {
520 rq = list_entry(rq_list.next, struct request, queuelist);
521 list_del_init(&rq->queuelist);
522 blk_mq_insert_request(rq, false, false, false);
526 * Use the start variant of queue running here, so that running
527 * the requeue work will kick stopped queues.
529 blk_mq_start_hw_queues(q);
532 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
534 struct request_queue *q = rq->q;
538 * We abuse this flag that is otherwise used by the I/O scheduler to
539 * request head insertation from the workqueue.
541 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
543 spin_lock_irqsave(&q->requeue_lock, flags);
545 rq->cmd_flags |= REQ_SOFTBARRIER;
546 list_add(&rq->queuelist, &q->requeue_list);
548 list_add_tail(&rq->queuelist, &q->requeue_list);
550 spin_unlock_irqrestore(&q->requeue_lock, flags);
552 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
554 void blk_mq_cancel_requeue_work(struct request_queue *q)
556 cancel_work_sync(&q->requeue_work);
558 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
560 void blk_mq_kick_requeue_list(struct request_queue *q)
562 kblockd_schedule_work(&q->requeue_work);
564 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
566 void blk_mq_abort_requeue_list(struct request_queue *q)
571 spin_lock_irqsave(&q->requeue_lock, flags);
572 list_splice_init(&q->requeue_list, &rq_list);
573 spin_unlock_irqrestore(&q->requeue_lock, flags);
575 while (!list_empty(&rq_list)) {
578 rq = list_first_entry(&rq_list, struct request, queuelist);
579 list_del_init(&rq->queuelist);
581 blk_mq_end_request(rq, rq->errors);
584 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
586 static inline bool is_flush_request(struct request *rq,
587 struct blk_flush_queue *fq, unsigned int tag)
589 return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
590 fq->flush_rq->tag == tag);
593 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
595 struct request *rq = tags->rqs[tag];
596 /* mq_ctx of flush rq is always cloned from the corresponding req */
597 struct blk_flush_queue *fq = blk_get_flush_queue(rq->q, rq->mq_ctx);
599 if (!is_flush_request(rq, fq, tag))
604 EXPORT_SYMBOL(blk_mq_tag_to_rq);
606 struct blk_mq_timeout_data {
608 unsigned int next_set;
611 void blk_mq_rq_timed_out(struct request *req, bool reserved)
613 struct blk_mq_ops *ops = req->q->mq_ops;
614 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
617 * We know that complete is set at this point. If STARTED isn't set
618 * anymore, then the request isn't active and the "timeout" should
619 * just be ignored. This can happen due to the bitflag ordering.
620 * Timeout first checks if STARTED is set, and if it is, assumes
621 * the request is active. But if we race with completion, then
622 * we both flags will get cleared. So check here again, and ignore
623 * a timeout event with a request that isn't active.
625 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
629 ret = ops->timeout(req, reserved);
633 __blk_mq_complete_request(req);
635 case BLK_EH_RESET_TIMER:
637 blk_clear_rq_complete(req);
639 case BLK_EH_NOT_HANDLED:
642 printk(KERN_ERR "block: bad eh return: %d\n", ret);
647 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
648 struct request *rq, void *priv, bool reserved)
650 struct blk_mq_timeout_data *data = priv;
652 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
654 * If a request wasn't started before the queue was
655 * marked dying, kill it here or it'll go unnoticed.
657 if (unlikely(blk_queue_dying(rq->q))) {
659 blk_mq_complete_request(rq);
663 if (rq->cmd_flags & REQ_NO_TIMEOUT)
666 if (time_after_eq(jiffies, rq->deadline)) {
667 if (!blk_mark_rq_complete(rq))
668 blk_mq_rq_timed_out(rq, reserved);
669 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
670 data->next = rq->deadline;
675 static void blk_mq_rq_timer(unsigned long priv)
677 struct request_queue *q = (struct request_queue *)priv;
678 struct blk_mq_timeout_data data = {
682 struct blk_mq_hw_ctx *hctx;
685 queue_for_each_hw_ctx(q, hctx, i) {
687 * If not software queues are currently mapped to this
688 * hardware queue, there's nothing to check
690 if (!blk_mq_hw_queue_mapped(hctx))
693 blk_mq_tag_busy_iter(hctx, blk_mq_check_expired, &data);
697 data.next = blk_rq_timeout(round_jiffies_up(data.next));
698 mod_timer(&q->timeout, data.next);
700 queue_for_each_hw_ctx(q, hctx, i) {
701 /* the hctx may be unmapped, so check it here */
702 if (blk_mq_hw_queue_mapped(hctx))
703 blk_mq_tag_idle(hctx);
709 * Reverse check our software queue for entries that we could potentially
710 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
711 * too much time checking for merges.
713 static bool blk_mq_attempt_merge(struct request_queue *q,
714 struct blk_mq_ctx *ctx, struct bio *bio)
719 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
725 if (!blk_rq_merge_ok(rq, bio))
728 el_ret = blk_try_merge(rq, bio);
729 if (el_ret == ELEVATOR_BACK_MERGE) {
730 if (bio_attempt_back_merge(q, rq, bio)) {
735 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
736 if (bio_attempt_front_merge(q, rq, bio)) {
748 * Process software queues that have been marked busy, splicing them
749 * to the for-dispatch
751 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
753 struct blk_mq_ctx *ctx;
756 for (i = 0; i < hctx->ctx_map.size; i++) {
757 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
758 unsigned int off, bit;
764 off = i * hctx->ctx_map.bits_per_word;
766 bit = find_next_bit(&bm->word, bm->depth, bit);
767 if (bit >= bm->depth)
770 ctx = hctx->ctxs[bit + off];
771 clear_bit(bit, &bm->word);
772 spin_lock(&ctx->lock);
773 list_splice_tail_init(&ctx->rq_list, list);
774 spin_unlock(&ctx->lock);
782 * Run this hardware queue, pulling any software queues mapped to it in.
783 * Note that this function currently has various problems around ordering
784 * of IO. In particular, we'd like FIFO behaviour on handling existing
785 * items on the hctx->dispatch list. Ignore that for now.
787 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
789 struct request_queue *q = hctx->queue;
792 LIST_HEAD(driver_list);
793 struct list_head *dptr;
796 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
798 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
804 * Touch any software queue that has pending entries.
806 flush_busy_ctxs(hctx, &rq_list);
809 * If we have previous entries on our dispatch list, grab them
810 * and stuff them at the front for more fair dispatch.
812 if (!list_empty_careful(&hctx->dispatch)) {
813 spin_lock(&hctx->lock);
814 if (!list_empty(&hctx->dispatch))
815 list_splice_init(&hctx->dispatch, &rq_list);
816 spin_unlock(&hctx->lock);
820 * Start off with dptr being NULL, so we start the first request
821 * immediately, even if we have more pending.
826 * Now process all the entries, sending them to the driver.
829 while (!list_empty(&rq_list)) {
830 struct blk_mq_queue_data bd;
833 rq = list_first_entry(&rq_list, struct request, queuelist);
834 list_del_init(&rq->queuelist);
838 bd.last = list_empty(&rq_list);
840 ret = q->mq_ops->queue_rq(hctx, &bd);
842 case BLK_MQ_RQ_QUEUE_OK:
845 case BLK_MQ_RQ_QUEUE_BUSY:
846 list_add(&rq->queuelist, &rq_list);
847 __blk_mq_requeue_request(rq);
850 pr_err("blk-mq: bad return on queue: %d\n", ret);
851 case BLK_MQ_RQ_QUEUE_ERROR:
853 blk_mq_end_request(rq, rq->errors);
857 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
861 * We've done the first request. If we have more than 1
862 * left in the list, set dptr to defer issue.
864 if (!dptr && rq_list.next != rq_list.prev)
869 hctx->dispatched[0]++;
870 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
871 hctx->dispatched[ilog2(queued) + 1]++;
874 * Any items that need requeuing? Stuff them into hctx->dispatch,
875 * that is where we will continue on next queue run.
877 if (!list_empty(&rq_list)) {
878 spin_lock(&hctx->lock);
879 list_splice(&rq_list, &hctx->dispatch);
880 spin_unlock(&hctx->lock);
882 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
883 * it's possible the queue is stopped and restarted again
884 * before this. Queue restart will dispatch requests. And since
885 * requests in rq_list aren't added into hctx->dispatch yet,
886 * the requests in rq_list might get lost.
888 * blk_mq_run_hw_queue() already checks the STOPPED bit
890 blk_mq_run_hw_queue(hctx, true);
895 * It'd be great if the workqueue API had a way to pass
896 * in a mask and had some smarts for more clever placement.
897 * For now we just round-robin here, switching for every
898 * BLK_MQ_CPU_WORK_BATCH queued items.
900 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
902 if (hctx->queue->nr_hw_queues == 1)
903 return WORK_CPU_UNBOUND;
905 if (--hctx->next_cpu_batch <= 0) {
906 int cpu = hctx->next_cpu, next_cpu;
908 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
909 if (next_cpu >= nr_cpu_ids)
910 next_cpu = cpumask_first(hctx->cpumask);
912 hctx->next_cpu = next_cpu;
913 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
918 return hctx->next_cpu;
921 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
923 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
924 !blk_mq_hw_queue_mapped(hctx)))
928 int cpu = get_cpu_light();
929 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
930 __blk_mq_run_hw_queue(hctx);
938 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
942 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
944 struct blk_mq_hw_ctx *hctx;
947 queue_for_each_hw_ctx(q, hctx, i) {
948 if ((!blk_mq_hctx_has_pending(hctx) &&
949 list_empty_careful(&hctx->dispatch)) ||
950 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
953 blk_mq_run_hw_queue(hctx, async);
956 EXPORT_SYMBOL(blk_mq_run_hw_queues);
958 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
960 cancel_delayed_work(&hctx->run_work);
961 cancel_delayed_work(&hctx->delay_work);
962 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
964 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
966 void blk_mq_stop_hw_queues(struct request_queue *q)
968 struct blk_mq_hw_ctx *hctx;
971 queue_for_each_hw_ctx(q, hctx, i)
972 blk_mq_stop_hw_queue(hctx);
974 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
976 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
978 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
980 blk_mq_run_hw_queue(hctx, false);
982 EXPORT_SYMBOL(blk_mq_start_hw_queue);
984 void blk_mq_start_hw_queues(struct request_queue *q)
986 struct blk_mq_hw_ctx *hctx;
989 queue_for_each_hw_ctx(q, hctx, i)
990 blk_mq_start_hw_queue(hctx);
992 EXPORT_SYMBOL(blk_mq_start_hw_queues);
994 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
996 struct blk_mq_hw_ctx *hctx;
999 queue_for_each_hw_ctx(q, hctx, i) {
1000 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1003 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1004 blk_mq_run_hw_queue(hctx, async);
1007 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1009 static void blk_mq_run_work_fn(struct work_struct *work)
1011 struct blk_mq_hw_ctx *hctx;
1013 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1015 __blk_mq_run_hw_queue(hctx);
1018 static void blk_mq_delay_work_fn(struct work_struct *work)
1020 struct blk_mq_hw_ctx *hctx;
1022 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
1024 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
1025 __blk_mq_run_hw_queue(hctx);
1028 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1030 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1033 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1034 &hctx->delay_work, msecs_to_jiffies(msecs));
1036 EXPORT_SYMBOL(blk_mq_delay_queue);
1038 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
1039 struct request *rq, bool at_head)
1041 struct blk_mq_ctx *ctx = rq->mq_ctx;
1043 trace_block_rq_insert(hctx->queue, rq);
1046 list_add(&rq->queuelist, &ctx->rq_list);
1048 list_add_tail(&rq->queuelist, &ctx->rq_list);
1050 blk_mq_hctx_mark_pending(hctx, ctx);
1053 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1056 struct request_queue *q = rq->q;
1057 struct blk_mq_hw_ctx *hctx;
1058 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
1060 current_ctx = blk_mq_get_ctx(q);
1061 if (!cpu_online(ctx->cpu))
1062 rq->mq_ctx = ctx = current_ctx;
1064 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1066 spin_lock(&ctx->lock);
1067 __blk_mq_insert_request(hctx, rq, at_head);
1068 spin_unlock(&ctx->lock);
1071 blk_mq_run_hw_queue(hctx, async);
1073 blk_mq_put_ctx(current_ctx);
1076 static void blk_mq_insert_requests(struct request_queue *q,
1077 struct blk_mq_ctx *ctx,
1078 struct list_head *list,
1083 struct blk_mq_hw_ctx *hctx;
1084 struct blk_mq_ctx *current_ctx;
1086 trace_block_unplug(q, depth, !from_schedule);
1088 current_ctx = blk_mq_get_ctx(q);
1090 if (!cpu_online(ctx->cpu))
1092 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1095 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1098 spin_lock(&ctx->lock);
1099 while (!list_empty(list)) {
1102 rq = list_first_entry(list, struct request, queuelist);
1103 list_del_init(&rq->queuelist);
1105 __blk_mq_insert_request(hctx, rq, false);
1107 spin_unlock(&ctx->lock);
1109 blk_mq_run_hw_queue(hctx, from_schedule);
1110 blk_mq_put_ctx(current_ctx);
1113 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1115 struct request *rqa = container_of(a, struct request, queuelist);
1116 struct request *rqb = container_of(b, struct request, queuelist);
1118 return !(rqa->mq_ctx < rqb->mq_ctx ||
1119 (rqa->mq_ctx == rqb->mq_ctx &&
1120 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1123 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1125 struct blk_mq_ctx *this_ctx;
1126 struct request_queue *this_q;
1129 LIST_HEAD(ctx_list);
1132 list_splice_init(&plug->mq_list, &list);
1134 list_sort(NULL, &list, plug_ctx_cmp);
1140 while (!list_empty(&list)) {
1141 rq = list_entry_rq(list.next);
1142 list_del_init(&rq->queuelist);
1144 if (rq->mq_ctx != this_ctx) {
1146 blk_mq_insert_requests(this_q, this_ctx,
1151 this_ctx = rq->mq_ctx;
1157 list_add_tail(&rq->queuelist, &ctx_list);
1161 * If 'this_ctx' is set, we know we have entries to complete
1162 * on 'ctx_list'. Do those.
1165 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1170 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1172 init_request_from_bio(rq, bio);
1174 if (blk_do_io_stat(rq))
1175 blk_account_io_start(rq, 1);
1178 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1180 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1181 !blk_queue_nomerges(hctx->queue);
1184 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1185 struct blk_mq_ctx *ctx,
1186 struct request *rq, struct bio *bio)
1188 if (!hctx_allow_merges(hctx)) {
1189 blk_mq_bio_to_request(rq, bio);
1190 spin_lock(&ctx->lock);
1192 __blk_mq_insert_request(hctx, rq, false);
1193 spin_unlock(&ctx->lock);
1196 struct request_queue *q = hctx->queue;
1198 spin_lock(&ctx->lock);
1199 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1200 blk_mq_bio_to_request(rq, bio);
1204 spin_unlock(&ctx->lock);
1205 __blk_mq_free_request(hctx, ctx, rq);
1210 struct blk_map_ctx {
1211 struct blk_mq_hw_ctx *hctx;
1212 struct blk_mq_ctx *ctx;
1215 static struct request *blk_mq_map_request(struct request_queue *q,
1217 struct blk_map_ctx *data)
1219 struct blk_mq_hw_ctx *hctx;
1220 struct blk_mq_ctx *ctx;
1222 int rw = bio_data_dir(bio);
1223 struct blk_mq_alloc_data alloc_data;
1225 if (unlikely(blk_mq_queue_enter(q, GFP_KERNEL))) {
1226 bio_endio(bio, -EIO);
1230 ctx = blk_mq_get_ctx(q);
1231 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1233 if (rw_is_sync(bio->bi_rw))
1236 trace_block_getrq(q, bio, rw);
1237 blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1239 rq = __blk_mq_alloc_request(&alloc_data, rw);
1240 if (unlikely(!rq)) {
1241 __blk_mq_run_hw_queue(hctx);
1242 blk_mq_put_ctx(ctx);
1243 trace_block_sleeprq(q, bio, rw);
1245 ctx = blk_mq_get_ctx(q);
1246 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1247 blk_mq_set_alloc_data(&alloc_data, q,
1248 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1249 rq = __blk_mq_alloc_request(&alloc_data, rw);
1250 ctx = alloc_data.ctx;
1251 hctx = alloc_data.hctx;
1261 * Multiple hardware queue variant. This will not use per-process plugs,
1262 * but will attempt to bypass the hctx queueing if we can go straight to
1263 * hardware for SYNC IO.
1265 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1267 const int is_sync = rw_is_sync(bio->bi_rw);
1268 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1269 struct blk_map_ctx data;
1272 blk_queue_bounce(q, &bio);
1274 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1275 bio_endio(bio, -EIO);
1279 rq = blk_mq_map_request(q, bio, &data);
1283 if (unlikely(is_flush_fua)) {
1284 blk_mq_bio_to_request(rq, bio);
1285 blk_insert_flush(rq);
1290 * If the driver supports defer issued based on 'last', then
1291 * queue it up like normal since we can potentially save some
1294 if (is_sync && !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1295 struct blk_mq_queue_data bd = {
1302 blk_mq_bio_to_request(rq, bio);
1305 * For OK queue, we are done. For error, kill it. Any other
1306 * error (busy), just add it to our list as we previously
1309 ret = q->mq_ops->queue_rq(data.hctx, &bd);
1310 if (ret == BLK_MQ_RQ_QUEUE_OK)
1313 __blk_mq_requeue_request(rq);
1315 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1317 blk_mq_end_request(rq, rq->errors);
1323 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1325 * For a SYNC request, send it to the hardware immediately. For
1326 * an ASYNC request, just ensure that we run it later on. The
1327 * latter allows for merging opportunities and more efficient
1331 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1334 blk_mq_put_ctx(data.ctx);
1338 * Single hardware queue variant. This will attempt to use any per-process
1339 * plug for merging and IO deferral.
1341 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1343 const int is_sync = rw_is_sync(bio->bi_rw);
1344 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1345 unsigned int use_plug, request_count = 0;
1346 struct blk_map_ctx data;
1350 * If we have multiple hardware queues, just go directly to
1351 * one of those for sync IO.
1353 use_plug = !is_flush_fua && !is_sync;
1355 blk_queue_bounce(q, &bio);
1357 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1358 bio_endio(bio, -EIO);
1362 if (use_plug && !blk_queue_nomerges(q) &&
1363 blk_attempt_plug_merge(q, bio, &request_count))
1366 rq = blk_mq_map_request(q, bio, &data);
1370 if (unlikely(is_flush_fua)) {
1371 blk_mq_bio_to_request(rq, bio);
1372 blk_insert_flush(rq);
1377 * A task plug currently exists. Since this is completely lockless,
1378 * utilize that to temporarily store requests until the task is
1379 * either done or scheduled away.
1382 struct blk_plug *plug = current->plug;
1385 blk_mq_bio_to_request(rq, bio);
1386 if (list_empty(&plug->mq_list))
1387 trace_block_plug(q);
1388 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1389 blk_flush_plug_list(plug, false);
1390 trace_block_plug(q);
1392 list_add_tail(&rq->queuelist, &plug->mq_list);
1393 blk_mq_put_ctx(data.ctx);
1398 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1400 * For a SYNC request, send it to the hardware immediately. For
1401 * an ASYNC request, just ensure that we run it later on. The
1402 * latter allows for merging opportunities and more efficient
1406 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1409 blk_mq_put_ctx(data.ctx);
1413 * Default mapping to a software queue, since we use one per CPU.
1415 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1417 return q->queue_hw_ctx[q->mq_map[cpu]];
1419 EXPORT_SYMBOL(blk_mq_map_queue);
1421 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1422 struct blk_mq_tags *tags, unsigned int hctx_idx)
1426 if (tags->rqs && set->ops->exit_request) {
1429 for (i = 0; i < tags->nr_tags; i++) {
1432 set->ops->exit_request(set->driver_data, tags->rqs[i],
1434 tags->rqs[i] = NULL;
1438 while (!list_empty(&tags->page_list)) {
1439 page = list_first_entry(&tags->page_list, struct page, lru);
1440 list_del_init(&page->lru);
1441 __free_pages(page, page->private);
1446 blk_mq_free_tags(tags);
1449 static size_t order_to_size(unsigned int order)
1451 return (size_t)PAGE_SIZE << order;
1454 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1455 unsigned int hctx_idx)
1457 struct blk_mq_tags *tags;
1458 unsigned int i, j, entries_per_page, max_order = 4;
1459 size_t rq_size, left;
1461 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1463 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1467 INIT_LIST_HEAD(&tags->page_list);
1469 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1470 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1473 blk_mq_free_tags(tags);
1478 * rq_size is the size of the request plus driver payload, rounded
1479 * to the cacheline size
1481 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1483 left = rq_size * set->queue_depth;
1485 for (i = 0; i < set->queue_depth; ) {
1486 int this_order = max_order;
1491 while (left < order_to_size(this_order - 1) && this_order)
1495 page = alloc_pages_node(set->numa_node,
1496 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1502 if (order_to_size(this_order) < rq_size)
1509 page->private = this_order;
1510 list_add_tail(&page->lru, &tags->page_list);
1512 p = page_address(page);
1513 entries_per_page = order_to_size(this_order) / rq_size;
1514 to_do = min(entries_per_page, set->queue_depth - i);
1515 left -= to_do * rq_size;
1516 for (j = 0; j < to_do; j++) {
1518 if (set->ops->init_request) {
1519 if (set->ops->init_request(set->driver_data,
1520 tags->rqs[i], hctx_idx, i,
1522 tags->rqs[i] = NULL;
1535 blk_mq_free_rq_map(set, tags, hctx_idx);
1539 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1544 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1546 unsigned int bpw = 8, total, num_maps, i;
1548 bitmap->bits_per_word = bpw;
1550 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1551 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1557 for (i = 0; i < num_maps; i++) {
1558 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1559 total -= bitmap->map[i].depth;
1565 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1567 struct request_queue *q = hctx->queue;
1568 struct blk_mq_ctx *ctx;
1572 * Move ctx entries to new CPU, if this one is going away.
1574 ctx = __blk_mq_get_ctx(q, cpu);
1576 spin_lock(&ctx->lock);
1577 if (!list_empty(&ctx->rq_list)) {
1578 list_splice_init(&ctx->rq_list, &tmp);
1579 blk_mq_hctx_clear_pending(hctx, ctx);
1581 spin_unlock(&ctx->lock);
1583 if (list_empty(&tmp))
1586 ctx = blk_mq_get_ctx(q);
1587 spin_lock(&ctx->lock);
1589 while (!list_empty(&tmp)) {
1592 rq = list_first_entry(&tmp, struct request, queuelist);
1594 list_move_tail(&rq->queuelist, &ctx->rq_list);
1597 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1598 blk_mq_hctx_mark_pending(hctx, ctx);
1600 spin_unlock(&ctx->lock);
1602 blk_mq_run_hw_queue(hctx, true);
1603 blk_mq_put_ctx(ctx);
1607 static int blk_mq_hctx_notify(void *data, unsigned long action,
1610 struct blk_mq_hw_ctx *hctx = data;
1612 if (action == CPU_POST_DEAD)
1613 return blk_mq_hctx_cpu_offline(hctx, cpu);
1616 * In case of CPU online, tags may be reallocated
1617 * in blk_mq_map_swqueue() after mapping is updated.
1623 /* hctx->ctxs will be freed in queue's release handler */
1624 static void blk_mq_exit_hctx(struct request_queue *q,
1625 struct blk_mq_tag_set *set,
1626 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1628 unsigned flush_start_tag = set->queue_depth;
1630 blk_mq_tag_idle(hctx);
1632 if (set->ops->exit_request)
1633 set->ops->exit_request(set->driver_data,
1634 hctx->fq->flush_rq, hctx_idx,
1635 flush_start_tag + hctx_idx);
1637 if (set->ops->exit_hctx)
1638 set->ops->exit_hctx(hctx, hctx_idx);
1640 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1641 blk_free_flush_queue(hctx->fq);
1642 blk_mq_free_bitmap(&hctx->ctx_map);
1645 static void blk_mq_exit_hw_queues(struct request_queue *q,
1646 struct blk_mq_tag_set *set, int nr_queue)
1648 struct blk_mq_hw_ctx *hctx;
1651 queue_for_each_hw_ctx(q, hctx, i) {
1654 blk_mq_exit_hctx(q, set, hctx, i);
1658 static void blk_mq_free_hw_queues(struct request_queue *q,
1659 struct blk_mq_tag_set *set)
1661 struct blk_mq_hw_ctx *hctx;
1664 queue_for_each_hw_ctx(q, hctx, i)
1665 free_cpumask_var(hctx->cpumask);
1668 static int blk_mq_init_hctx(struct request_queue *q,
1669 struct blk_mq_tag_set *set,
1670 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1673 unsigned flush_start_tag = set->queue_depth;
1675 node = hctx->numa_node;
1676 if (node == NUMA_NO_NODE)
1677 node = hctx->numa_node = set->numa_node;
1679 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1680 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1681 spin_lock_init(&hctx->lock);
1682 INIT_LIST_HEAD(&hctx->dispatch);
1684 hctx->queue_num = hctx_idx;
1685 hctx->flags = set->flags;
1687 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1688 blk_mq_hctx_notify, hctx);
1689 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1691 hctx->tags = set->tags[hctx_idx];
1694 * Allocate space for all possible cpus to avoid allocation at
1697 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1700 goto unregister_cpu_notifier;
1702 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1707 if (set->ops->init_hctx &&
1708 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1711 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1715 if (set->ops->init_request &&
1716 set->ops->init_request(set->driver_data,
1717 hctx->fq->flush_rq, hctx_idx,
1718 flush_start_tag + hctx_idx, node))
1726 if (set->ops->exit_hctx)
1727 set->ops->exit_hctx(hctx, hctx_idx);
1729 blk_mq_free_bitmap(&hctx->ctx_map);
1732 unregister_cpu_notifier:
1733 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1738 static int blk_mq_init_hw_queues(struct request_queue *q,
1739 struct blk_mq_tag_set *set)
1741 struct blk_mq_hw_ctx *hctx;
1745 * Initialize hardware queues
1747 queue_for_each_hw_ctx(q, hctx, i) {
1748 if (blk_mq_init_hctx(q, set, hctx, i))
1752 if (i == q->nr_hw_queues)
1758 blk_mq_exit_hw_queues(q, set, i);
1763 static void blk_mq_init_cpu_queues(struct request_queue *q,
1764 unsigned int nr_hw_queues)
1768 for_each_possible_cpu(i) {
1769 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1770 struct blk_mq_hw_ctx *hctx;
1772 memset(__ctx, 0, sizeof(*__ctx));
1774 spin_lock_init(&__ctx->lock);
1775 INIT_LIST_HEAD(&__ctx->rq_list);
1778 /* If the cpu isn't online, the cpu is mapped to first hctx */
1782 hctx = q->mq_ops->map_queue(q, i);
1785 * Set local node, IFF we have more than one hw queue. If
1786 * not, we remain on the home node of the device
1788 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1789 hctx->numa_node = cpu_to_node(i);
1793 static void blk_mq_map_swqueue(struct request_queue *q)
1796 struct blk_mq_hw_ctx *hctx;
1797 struct blk_mq_ctx *ctx;
1798 struct blk_mq_tag_set *set = q->tag_set;
1800 queue_for_each_hw_ctx(q, hctx, i) {
1801 cpumask_clear(hctx->cpumask);
1806 * Map software to hardware queues
1808 queue_for_each_ctx(q, ctx, i) {
1809 /* If the cpu isn't online, the cpu is mapped to first hctx */
1813 hctx = q->mq_ops->map_queue(q, i);
1814 cpumask_set_cpu(i, hctx->cpumask);
1815 ctx->index_hw = hctx->nr_ctx;
1816 hctx->ctxs[hctx->nr_ctx++] = ctx;
1819 queue_for_each_hw_ctx(q, hctx, i) {
1820 struct blk_mq_ctxmap *map = &hctx->ctx_map;
1823 * If no software queues are mapped to this hardware queue,
1824 * disable it and free the request entries.
1826 if (!hctx->nr_ctx) {
1828 blk_mq_free_rq_map(set, set->tags[i], i);
1829 set->tags[i] = NULL;
1835 /* unmapped hw queue can be remapped after CPU topo changed */
1837 set->tags[i] = blk_mq_init_rq_map(set, i);
1838 hctx->tags = set->tags[i];
1839 WARN_ON(!hctx->tags);
1842 * Set the map size to the number of mapped software queues.
1843 * This is more accurate and more efficient than looping
1844 * over all possibly mapped software queues.
1846 map->size = DIV_ROUND_UP(hctx->nr_ctx, map->bits_per_word);
1849 * Initialize batch roundrobin counts
1851 hctx->next_cpu = cpumask_first(hctx->cpumask);
1852 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1856 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1858 struct blk_mq_hw_ctx *hctx;
1859 struct request_queue *q;
1863 if (set->tag_list.next == set->tag_list.prev)
1868 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1869 blk_mq_freeze_queue(q);
1871 queue_for_each_hw_ctx(q, hctx, i) {
1873 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1875 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1877 blk_mq_unfreeze_queue(q);
1881 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1883 struct blk_mq_tag_set *set = q->tag_set;
1885 mutex_lock(&set->tag_list_lock);
1886 list_del_init(&q->tag_set_list);
1887 blk_mq_update_tag_set_depth(set);
1888 mutex_unlock(&set->tag_list_lock);
1891 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1892 struct request_queue *q)
1896 mutex_lock(&set->tag_list_lock);
1897 list_add_tail(&q->tag_set_list, &set->tag_list);
1898 blk_mq_update_tag_set_depth(set);
1899 mutex_unlock(&set->tag_list_lock);
1903 * It is the actual release handler for mq, but we do it from
1904 * request queue's release handler for avoiding use-after-free
1905 * and headache because q->mq_kobj shouldn't have been introduced,
1906 * but we can't group ctx/kctx kobj without it.
1908 void blk_mq_release(struct request_queue *q)
1910 struct blk_mq_hw_ctx *hctx;
1913 /* hctx kobj stays in hctx */
1914 queue_for_each_hw_ctx(q, hctx, i) {
1921 kfree(q->queue_hw_ctx);
1923 /* ctx kobj stays in queue_ctx */
1924 free_percpu(q->queue_ctx);
1927 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1929 struct request_queue *uninit_q, *q;
1931 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1933 return ERR_PTR(-ENOMEM);
1935 q = blk_mq_init_allocated_queue(set, uninit_q);
1937 blk_cleanup_queue(uninit_q);
1941 EXPORT_SYMBOL(blk_mq_init_queue);
1943 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
1944 struct request_queue *q)
1946 struct blk_mq_hw_ctx **hctxs;
1947 struct blk_mq_ctx __percpu *ctx;
1951 ctx = alloc_percpu(struct blk_mq_ctx);
1953 return ERR_PTR(-ENOMEM);
1955 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1961 map = blk_mq_make_queue_map(set);
1965 for (i = 0; i < set->nr_hw_queues; i++) {
1966 int node = blk_mq_hw_queue_to_node(map, i);
1968 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1973 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1977 atomic_set(&hctxs[i]->nr_active, 0);
1978 hctxs[i]->numa_node = node;
1979 hctxs[i]->queue_num = i;
1983 * Init percpu_ref in atomic mode so that it's faster to shutdown.
1984 * See blk_register_queue() for details.
1986 if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release,
1987 PERCPU_REF_INIT_ATOMIC, GFP_KERNEL))
1990 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1991 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
1993 q->nr_queues = nr_cpu_ids;
1994 q->nr_hw_queues = set->nr_hw_queues;
1998 q->queue_hw_ctx = hctxs;
2000 q->mq_ops = set->ops;
2001 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2003 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2004 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2006 q->sg_reserved_size = INT_MAX;
2008 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
2009 INIT_LIST_HEAD(&q->requeue_list);
2010 spin_lock_init(&q->requeue_lock);
2012 if (q->nr_hw_queues > 1)
2013 blk_queue_make_request(q, blk_mq_make_request);
2015 blk_queue_make_request(q, blk_sq_make_request);
2018 * Do this after blk_queue_make_request() overrides it...
2020 q->nr_requests = set->queue_depth;
2022 if (set->ops->complete)
2023 blk_queue_softirq_done(q, set->ops->complete);
2025 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2027 if (blk_mq_init_hw_queues(q, set))
2030 mutex_lock(&all_q_mutex);
2031 list_add_tail(&q->all_q_node, &all_q_list);
2032 mutex_unlock(&all_q_mutex);
2034 blk_mq_add_queue_tag_set(set, q);
2036 blk_mq_map_swqueue(q);
2042 for (i = 0; i < set->nr_hw_queues; i++) {
2045 free_cpumask_var(hctxs[i]->cpumask);
2052 return ERR_PTR(-ENOMEM);
2054 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2056 void blk_mq_free_queue(struct request_queue *q)
2058 struct blk_mq_tag_set *set = q->tag_set;
2060 blk_mq_del_queue_tag_set(q);
2062 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2063 blk_mq_free_hw_queues(q, set);
2065 percpu_ref_exit(&q->mq_usage_counter);
2071 mutex_lock(&all_q_mutex);
2072 list_del_init(&q->all_q_node);
2073 mutex_unlock(&all_q_mutex);
2076 /* Basically redo blk_mq_init_queue with queue frozen */
2077 static void blk_mq_queue_reinit(struct request_queue *q)
2079 WARN_ON_ONCE(!q->mq_freeze_depth);
2081 blk_mq_sysfs_unregister(q);
2083 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
2086 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2087 * we should change hctx numa_node according to new topology (this
2088 * involves free and re-allocate memory, worthy doing?)
2091 blk_mq_map_swqueue(q);
2093 blk_mq_sysfs_register(q);
2096 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2097 unsigned long action, void *hcpu)
2099 struct request_queue *q;
2102 * Before new mappings are established, hotadded cpu might already
2103 * start handling requests. This doesn't break anything as we map
2104 * offline CPUs to first hardware queue. We will re-init the queue
2105 * below to get optimal settings.
2107 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
2108 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
2111 mutex_lock(&all_q_mutex);
2114 * We need to freeze and reinit all existing queues. Freezing
2115 * involves synchronous wait for an RCU grace period and doing it
2116 * one by one may take a long time. Start freezing all queues in
2117 * one swoop and then wait for the completions so that freezing can
2118 * take place in parallel.
2120 list_for_each_entry(q, &all_q_list, all_q_node)
2121 blk_mq_freeze_queue_start(q);
2122 list_for_each_entry(q, &all_q_list, all_q_node) {
2123 blk_mq_freeze_queue_wait(q);
2126 * timeout handler can't touch hw queue during the
2129 del_timer_sync(&q->timeout);
2132 list_for_each_entry(q, &all_q_list, all_q_node)
2133 blk_mq_queue_reinit(q);
2135 list_for_each_entry(q, &all_q_list, all_q_node)
2136 blk_mq_unfreeze_queue(q);
2138 mutex_unlock(&all_q_mutex);
2142 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2146 for (i = 0; i < set->nr_hw_queues; i++) {
2147 set->tags[i] = blk_mq_init_rq_map(set, i);
2156 blk_mq_free_rq_map(set, set->tags[i], i);
2162 * Allocate the request maps associated with this tag_set. Note that this
2163 * may reduce the depth asked for, if memory is tight. set->queue_depth
2164 * will be updated to reflect the allocated depth.
2166 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2171 depth = set->queue_depth;
2173 err = __blk_mq_alloc_rq_maps(set);
2177 set->queue_depth >>= 1;
2178 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2182 } while (set->queue_depth);
2184 if (!set->queue_depth || err) {
2185 pr_err("blk-mq: failed to allocate request map\n");
2189 if (depth != set->queue_depth)
2190 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2191 depth, set->queue_depth);
2197 * Alloc a tag set to be associated with one or more request queues.
2198 * May fail with EINVAL for various error conditions. May adjust the
2199 * requested depth down, if if it too large. In that case, the set
2200 * value will be stored in set->queue_depth.
2202 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2204 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2206 if (!set->nr_hw_queues)
2208 if (!set->queue_depth)
2210 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2213 if (!set->ops->queue_rq || !set->ops->map_queue)
2216 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2217 pr_info("blk-mq: reduced tag depth to %u\n",
2219 set->queue_depth = BLK_MQ_MAX_DEPTH;
2223 * If a crashdump is active, then we are potentially in a very
2224 * memory constrained environment. Limit us to 1 queue and
2225 * 64 tags to prevent using too much memory.
2227 if (is_kdump_kernel()) {
2228 set->nr_hw_queues = 1;
2229 set->queue_depth = min(64U, set->queue_depth);
2232 set->tags = kmalloc_node(set->nr_hw_queues *
2233 sizeof(struct blk_mq_tags *),
2234 GFP_KERNEL, set->numa_node);
2238 if (blk_mq_alloc_rq_maps(set))
2241 mutex_init(&set->tag_list_lock);
2242 INIT_LIST_HEAD(&set->tag_list);
2250 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2252 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2256 for (i = 0; i < set->nr_hw_queues; i++) {
2258 blk_mq_free_rq_map(set, set->tags[i], i);
2264 EXPORT_SYMBOL(blk_mq_free_tag_set);
2266 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2268 struct blk_mq_tag_set *set = q->tag_set;
2269 struct blk_mq_hw_ctx *hctx;
2272 if (!set || nr > set->queue_depth)
2276 queue_for_each_hw_ctx(q, hctx, i) {
2277 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2283 q->nr_requests = nr;
2288 void blk_mq_disable_hotplug(void)
2290 mutex_lock(&all_q_mutex);
2293 void blk_mq_enable_hotplug(void)
2295 mutex_unlock(&all_q_mutex);
2298 static int __init blk_mq_init(void)
2302 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2306 subsys_initcall(blk_mq_init);