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);
110 void blk_mq_freeze_queue_start(struct request_queue *q)
114 spin_lock_irq(q->queue_lock);
115 freeze = !q->mq_freeze_depth++;
116 spin_unlock_irq(q->queue_lock);
119 percpu_ref_kill(&q->mq_usage_counter);
120 blk_mq_run_queues(q, false);
123 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
125 static void blk_mq_freeze_queue_wait(struct request_queue *q)
127 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter));
131 * Guarantee no request is in use, so we can change any data structure of
132 * the queue afterward.
134 void blk_mq_freeze_queue(struct request_queue *q)
136 blk_mq_freeze_queue_start(q);
137 blk_mq_freeze_queue_wait(q);
140 void blk_mq_unfreeze_queue(struct request_queue *q)
144 spin_lock_irq(q->queue_lock);
145 wake = !--q->mq_freeze_depth;
146 WARN_ON_ONCE(q->mq_freeze_depth < 0);
147 spin_unlock_irq(q->queue_lock);
149 percpu_ref_reinit(&q->mq_usage_counter);
150 wake_up_all(&q->mq_freeze_wq);
153 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
155 void blk_mq_wake_waiters(struct request_queue *q)
157 struct blk_mq_hw_ctx *hctx;
160 queue_for_each_hw_ctx(q, hctx, i)
161 if (blk_mq_hw_queue_mapped(hctx))
162 blk_mq_tag_wakeup_all(hctx->tags, true);
165 * If we are called because the queue has now been marked as
166 * dying, we need to ensure that processes currently waiting on
167 * the queue are notified as well.
169 wake_up_all(&q->mq_freeze_wq);
172 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
174 return blk_mq_has_free_tags(hctx->tags);
176 EXPORT_SYMBOL(blk_mq_can_queue);
178 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
179 struct request *rq, unsigned int rw_flags)
181 if (blk_queue_io_stat(q))
182 rw_flags |= REQ_IO_STAT;
184 INIT_LIST_HEAD(&rq->queuelist);
185 /* csd/requeue_work/fifo_time is initialized before use */
188 rq->cmd_flags |= rw_flags;
189 /* do not touch atomic flags, it needs atomic ops against the timer */
191 INIT_HLIST_NODE(&rq->hash);
192 RB_CLEAR_NODE(&rq->rb_node);
195 rq->start_time = jiffies;
196 #ifdef CONFIG_BLK_CGROUP
198 set_start_time_ns(rq);
199 rq->io_start_time_ns = 0;
201 rq->nr_phys_segments = 0;
202 #if defined(CONFIG_BLK_DEV_INTEGRITY)
203 rq->nr_integrity_segments = 0;
206 /* tag was already set */
216 INIT_LIST_HEAD(&rq->timeout_list);
220 rq->end_io_data = NULL;
223 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
226 static struct request *
227 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
232 tag = blk_mq_get_tag(data);
233 if (tag != BLK_MQ_TAG_FAIL) {
234 rq = data->hctx->tags->rqs[tag];
236 if (blk_mq_tag_busy(data->hctx)) {
237 rq->cmd_flags = REQ_MQ_INFLIGHT;
238 atomic_inc(&data->hctx->nr_active);
242 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
249 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
252 struct blk_mq_ctx *ctx;
253 struct blk_mq_hw_ctx *hctx;
255 struct blk_mq_alloc_data alloc_data;
258 ret = blk_mq_queue_enter(q);
262 ctx = blk_mq_get_ctx(q);
263 hctx = q->mq_ops->map_queue(q, ctx->cpu);
264 blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
265 reserved, ctx, hctx);
267 rq = __blk_mq_alloc_request(&alloc_data, rw);
268 if (!rq && (gfp & __GFP_WAIT)) {
269 __blk_mq_run_hw_queue(hctx);
272 ctx = blk_mq_get_ctx(q);
273 hctx = q->mq_ops->map_queue(q, ctx->cpu);
274 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
276 rq = __blk_mq_alloc_request(&alloc_data, rw);
277 ctx = alloc_data.ctx;
281 blk_mq_queue_exit(q);
282 return ERR_PTR(-EWOULDBLOCK);
286 EXPORT_SYMBOL(blk_mq_alloc_request);
288 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
289 struct blk_mq_ctx *ctx, struct request *rq)
291 const int tag = rq->tag;
292 struct request_queue *q = rq->q;
294 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
295 atomic_dec(&hctx->nr_active);
298 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
299 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
300 blk_mq_queue_exit(q);
303 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
305 struct blk_mq_ctx *ctx = rq->mq_ctx;
307 ctx->rq_completed[rq_is_sync(rq)]++;
308 __blk_mq_free_request(hctx, ctx, rq);
311 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
313 void blk_mq_free_request(struct request *rq)
315 struct blk_mq_hw_ctx *hctx;
316 struct request_queue *q = rq->q;
318 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
319 blk_mq_free_hctx_request(hctx, rq);
321 EXPORT_SYMBOL_GPL(blk_mq_free_request);
323 inline void __blk_mq_end_request(struct request *rq, int error)
325 blk_account_io_done(rq);
328 rq->end_io(rq, error);
330 if (unlikely(blk_bidi_rq(rq)))
331 blk_mq_free_request(rq->next_rq);
332 blk_mq_free_request(rq);
335 EXPORT_SYMBOL(__blk_mq_end_request);
337 void blk_mq_end_request(struct request *rq, int error)
339 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
341 __blk_mq_end_request(rq, error);
343 EXPORT_SYMBOL(blk_mq_end_request);
345 static void __blk_mq_complete_request_remote(void *data)
347 struct request *rq = data;
349 rq->q->softirq_done_fn(rq);
352 static void blk_mq_ipi_complete_request(struct request *rq)
354 struct blk_mq_ctx *ctx = rq->mq_ctx;
358 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
359 rq->q->softirq_done_fn(rq);
364 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
365 shared = cpus_share_cache(cpu, ctx->cpu);
367 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
368 rq->csd.func = __blk_mq_complete_request_remote;
371 smp_call_function_single_async(ctx->cpu, &rq->csd);
373 rq->q->softirq_done_fn(rq);
378 void __blk_mq_complete_request(struct request *rq)
380 struct request_queue *q = rq->q;
382 if (!q->softirq_done_fn)
383 blk_mq_end_request(rq, rq->errors);
385 blk_mq_ipi_complete_request(rq);
389 * blk_mq_complete_request - end I/O on a request
390 * @rq: the request being processed
393 * Ends all I/O on a request. It does not handle partial completions.
394 * The actual completion happens out-of-order, through a IPI handler.
396 void blk_mq_complete_request(struct request *rq)
398 struct request_queue *q = rq->q;
400 if (unlikely(blk_should_fake_timeout(q)))
402 if (!blk_mark_rq_complete(rq))
403 __blk_mq_complete_request(rq);
405 EXPORT_SYMBOL(blk_mq_complete_request);
407 int blk_mq_request_started(struct request *rq)
409 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
411 EXPORT_SYMBOL_GPL(blk_mq_request_started);
413 void blk_mq_start_request(struct request *rq)
415 struct request_queue *q = rq->q;
417 trace_block_rq_issue(q, rq);
419 rq->resid_len = blk_rq_bytes(rq);
420 if (unlikely(blk_bidi_rq(rq)))
421 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
426 * Ensure that ->deadline is visible before set the started
427 * flag and clear the completed flag.
429 smp_mb__before_atomic();
432 * Mark us as started and clear complete. Complete might have been
433 * set if requeue raced with timeout, which then marked it as
434 * complete. So be sure to clear complete again when we start
435 * the request, otherwise we'll ignore the completion event.
437 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
438 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
439 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
440 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
442 if (q->dma_drain_size && blk_rq_bytes(rq)) {
444 * Make sure space for the drain appears. We know we can do
445 * this because max_hw_segments has been adjusted to be one
446 * fewer than the device can handle.
448 rq->nr_phys_segments++;
451 EXPORT_SYMBOL(blk_mq_start_request);
453 static void __blk_mq_requeue_request(struct request *rq)
455 struct request_queue *q = rq->q;
457 trace_block_rq_requeue(q, rq);
459 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
460 if (q->dma_drain_size && blk_rq_bytes(rq))
461 rq->nr_phys_segments--;
465 void blk_mq_requeue_request(struct request *rq)
467 __blk_mq_requeue_request(rq);
469 BUG_ON(blk_queued_rq(rq));
470 blk_mq_add_to_requeue_list(rq, true);
472 EXPORT_SYMBOL(blk_mq_requeue_request);
474 static void blk_mq_requeue_work(struct work_struct *work)
476 struct request_queue *q =
477 container_of(work, struct request_queue, requeue_work);
479 struct request *rq, *next;
482 spin_lock_irqsave(&q->requeue_lock, flags);
483 list_splice_init(&q->requeue_list, &rq_list);
484 spin_unlock_irqrestore(&q->requeue_lock, flags);
486 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
487 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
490 rq->cmd_flags &= ~REQ_SOFTBARRIER;
491 list_del_init(&rq->queuelist);
492 blk_mq_insert_request(rq, true, false, false);
495 while (!list_empty(&rq_list)) {
496 rq = list_entry(rq_list.next, struct request, queuelist);
497 list_del_init(&rq->queuelist);
498 blk_mq_insert_request(rq, false, false, false);
502 * Use the start variant of queue running here, so that running
503 * the requeue work will kick stopped queues.
505 blk_mq_start_hw_queues(q);
508 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
510 struct request_queue *q = rq->q;
514 * We abuse this flag that is otherwise used by the I/O scheduler to
515 * request head insertation from the workqueue.
517 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
519 spin_lock_irqsave(&q->requeue_lock, flags);
521 rq->cmd_flags |= REQ_SOFTBARRIER;
522 list_add(&rq->queuelist, &q->requeue_list);
524 list_add_tail(&rq->queuelist, &q->requeue_list);
526 spin_unlock_irqrestore(&q->requeue_lock, flags);
528 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
530 void blk_mq_kick_requeue_list(struct request_queue *q)
532 kblockd_schedule_work(&q->requeue_work);
534 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
536 static inline bool is_flush_request(struct request *rq,
537 struct blk_flush_queue *fq, unsigned int tag)
539 return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
540 fq->flush_rq->tag == tag);
543 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
545 struct request *rq = tags->rqs[tag];
546 /* mq_ctx of flush rq is always cloned from the corresponding req */
547 struct blk_flush_queue *fq = blk_get_flush_queue(rq->q, rq->mq_ctx);
549 if (!is_flush_request(rq, fq, tag))
554 EXPORT_SYMBOL(blk_mq_tag_to_rq);
556 struct blk_mq_timeout_data {
558 unsigned int next_set;
561 void blk_mq_rq_timed_out(struct request *req, bool reserved)
563 struct blk_mq_ops *ops = req->q->mq_ops;
564 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
567 * We know that complete is set at this point. If STARTED isn't set
568 * anymore, then the request isn't active and the "timeout" should
569 * just be ignored. This can happen due to the bitflag ordering.
570 * Timeout first checks if STARTED is set, and if it is, assumes
571 * the request is active. But if we race with completion, then
572 * we both flags will get cleared. So check here again, and ignore
573 * a timeout event with a request that isn't active.
575 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
579 ret = ops->timeout(req, reserved);
583 __blk_mq_complete_request(req);
585 case BLK_EH_RESET_TIMER:
587 blk_clear_rq_complete(req);
589 case BLK_EH_NOT_HANDLED:
592 printk(KERN_ERR "block: bad eh return: %d\n", ret);
597 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
598 struct request *rq, void *priv, bool reserved)
600 struct blk_mq_timeout_data *data = priv;
602 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
605 if (time_after_eq(jiffies, rq->deadline)) {
606 if (!blk_mark_rq_complete(rq))
607 blk_mq_rq_timed_out(rq, reserved);
608 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
609 data->next = rq->deadline;
614 static void blk_mq_rq_timer(unsigned long priv)
616 struct request_queue *q = (struct request_queue *)priv;
617 struct blk_mq_timeout_data data = {
621 struct blk_mq_hw_ctx *hctx;
624 queue_for_each_hw_ctx(q, hctx, i) {
626 * If not software queues are currently mapped to this
627 * hardware queue, there's nothing to check
629 if (!blk_mq_hw_queue_mapped(hctx))
632 blk_mq_tag_busy_iter(hctx, blk_mq_check_expired, &data);
636 data.next = blk_rq_timeout(round_jiffies_up(data.next));
637 mod_timer(&q->timeout, data.next);
639 queue_for_each_hw_ctx(q, hctx, i)
640 blk_mq_tag_idle(hctx);
645 * Reverse check our software queue for entries that we could potentially
646 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
647 * too much time checking for merges.
649 static bool blk_mq_attempt_merge(struct request_queue *q,
650 struct blk_mq_ctx *ctx, struct bio *bio)
655 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
661 if (!blk_rq_merge_ok(rq, bio))
664 el_ret = blk_try_merge(rq, bio);
665 if (el_ret == ELEVATOR_BACK_MERGE) {
666 if (bio_attempt_back_merge(q, rq, bio)) {
671 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
672 if (bio_attempt_front_merge(q, rq, bio)) {
684 * Process software queues that have been marked busy, splicing them
685 * to the for-dispatch
687 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
689 struct blk_mq_ctx *ctx;
692 for (i = 0; i < hctx->ctx_map.map_size; i++) {
693 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
694 unsigned int off, bit;
700 off = i * hctx->ctx_map.bits_per_word;
702 bit = find_next_bit(&bm->word, bm->depth, bit);
703 if (bit >= bm->depth)
706 ctx = hctx->ctxs[bit + off];
707 clear_bit(bit, &bm->word);
708 spin_lock(&ctx->lock);
709 list_splice_tail_init(&ctx->rq_list, list);
710 spin_unlock(&ctx->lock);
718 * Run this hardware queue, pulling any software queues mapped to it in.
719 * Note that this function currently has various problems around ordering
720 * of IO. In particular, we'd like FIFO behaviour on handling existing
721 * items on the hctx->dispatch list. Ignore that for now.
723 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
725 struct request_queue *q = hctx->queue;
728 LIST_HEAD(driver_list);
729 struct list_head *dptr;
732 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
734 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
740 * Touch any software queue that has pending entries.
742 flush_busy_ctxs(hctx, &rq_list);
745 * If we have previous entries on our dispatch list, grab them
746 * and stuff them at the front for more fair dispatch.
748 if (!list_empty_careful(&hctx->dispatch)) {
749 spin_lock(&hctx->lock);
750 if (!list_empty(&hctx->dispatch))
751 list_splice_init(&hctx->dispatch, &rq_list);
752 spin_unlock(&hctx->lock);
756 * Start off with dptr being NULL, so we start the first request
757 * immediately, even if we have more pending.
762 * Now process all the entries, sending them to the driver.
765 while (!list_empty(&rq_list)) {
766 struct blk_mq_queue_data bd;
769 rq = list_first_entry(&rq_list, struct request, queuelist);
770 list_del_init(&rq->queuelist);
774 bd.last = list_empty(&rq_list);
776 ret = q->mq_ops->queue_rq(hctx, &bd);
778 case BLK_MQ_RQ_QUEUE_OK:
781 case BLK_MQ_RQ_QUEUE_BUSY:
782 list_add(&rq->queuelist, &rq_list);
783 __blk_mq_requeue_request(rq);
786 pr_err("blk-mq: bad return on queue: %d\n", ret);
787 case BLK_MQ_RQ_QUEUE_ERROR:
789 blk_mq_end_request(rq, rq->errors);
793 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
797 * We've done the first request. If we have more than 1
798 * left in the list, set dptr to defer issue.
800 if (!dptr && rq_list.next != rq_list.prev)
805 hctx->dispatched[0]++;
806 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
807 hctx->dispatched[ilog2(queued) + 1]++;
810 * Any items that need requeuing? Stuff them into hctx->dispatch,
811 * that is where we will continue on next queue run.
813 if (!list_empty(&rq_list)) {
814 spin_lock(&hctx->lock);
815 list_splice(&rq_list, &hctx->dispatch);
816 spin_unlock(&hctx->lock);
821 * It'd be great if the workqueue API had a way to pass
822 * in a mask and had some smarts for more clever placement.
823 * For now we just round-robin here, switching for every
824 * BLK_MQ_CPU_WORK_BATCH queued items.
826 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
828 if (hctx->queue->nr_hw_queues == 1)
829 return WORK_CPU_UNBOUND;
831 if (--hctx->next_cpu_batch <= 0) {
832 int cpu = hctx->next_cpu, next_cpu;
834 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
835 if (next_cpu >= nr_cpu_ids)
836 next_cpu = cpumask_first(hctx->cpumask);
838 hctx->next_cpu = next_cpu;
839 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
844 return hctx->next_cpu;
847 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
849 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
850 !blk_mq_hw_queue_mapped(hctx)))
855 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
856 __blk_mq_run_hw_queue(hctx);
864 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
868 void blk_mq_run_queues(struct request_queue *q, bool async)
870 struct blk_mq_hw_ctx *hctx;
873 queue_for_each_hw_ctx(q, hctx, i) {
874 if ((!blk_mq_hctx_has_pending(hctx) &&
875 list_empty_careful(&hctx->dispatch)) ||
876 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
879 blk_mq_run_hw_queue(hctx, async);
882 EXPORT_SYMBOL(blk_mq_run_queues);
884 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
886 cancel_delayed_work(&hctx->run_work);
887 cancel_delayed_work(&hctx->delay_work);
888 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
890 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
892 void blk_mq_stop_hw_queues(struct request_queue *q)
894 struct blk_mq_hw_ctx *hctx;
897 queue_for_each_hw_ctx(q, hctx, i)
898 blk_mq_stop_hw_queue(hctx);
900 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
902 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
904 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
906 blk_mq_run_hw_queue(hctx, false);
908 EXPORT_SYMBOL(blk_mq_start_hw_queue);
910 void blk_mq_start_hw_queues(struct request_queue *q)
912 struct blk_mq_hw_ctx *hctx;
915 queue_for_each_hw_ctx(q, hctx, i)
916 blk_mq_start_hw_queue(hctx);
918 EXPORT_SYMBOL(blk_mq_start_hw_queues);
921 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
923 struct blk_mq_hw_ctx *hctx;
926 queue_for_each_hw_ctx(q, hctx, i) {
927 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
930 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
931 blk_mq_run_hw_queue(hctx, async);
934 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
936 static void blk_mq_run_work_fn(struct work_struct *work)
938 struct blk_mq_hw_ctx *hctx;
940 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
942 __blk_mq_run_hw_queue(hctx);
945 static void blk_mq_delay_work_fn(struct work_struct *work)
947 struct blk_mq_hw_ctx *hctx;
949 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
951 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
952 __blk_mq_run_hw_queue(hctx);
955 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
957 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
960 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
961 &hctx->delay_work, msecs_to_jiffies(msecs));
963 EXPORT_SYMBOL(blk_mq_delay_queue);
965 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
966 struct request *rq, bool at_head)
968 struct blk_mq_ctx *ctx = rq->mq_ctx;
970 trace_block_rq_insert(hctx->queue, rq);
973 list_add(&rq->queuelist, &ctx->rq_list);
975 list_add_tail(&rq->queuelist, &ctx->rq_list);
977 blk_mq_hctx_mark_pending(hctx, ctx);
980 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
983 struct request_queue *q = rq->q;
984 struct blk_mq_hw_ctx *hctx;
985 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
987 current_ctx = blk_mq_get_ctx(q);
988 if (!cpu_online(ctx->cpu))
989 rq->mq_ctx = ctx = current_ctx;
991 hctx = q->mq_ops->map_queue(q, ctx->cpu);
993 spin_lock(&ctx->lock);
994 __blk_mq_insert_request(hctx, rq, at_head);
995 spin_unlock(&ctx->lock);
998 blk_mq_run_hw_queue(hctx, async);
1000 blk_mq_put_ctx(current_ctx);
1003 static void blk_mq_insert_requests(struct request_queue *q,
1004 struct blk_mq_ctx *ctx,
1005 struct list_head *list,
1010 struct blk_mq_hw_ctx *hctx;
1011 struct blk_mq_ctx *current_ctx;
1013 trace_block_unplug(q, depth, !from_schedule);
1015 current_ctx = blk_mq_get_ctx(q);
1017 if (!cpu_online(ctx->cpu))
1019 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1022 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1025 spin_lock(&ctx->lock);
1026 while (!list_empty(list)) {
1029 rq = list_first_entry(list, struct request, queuelist);
1030 list_del_init(&rq->queuelist);
1032 __blk_mq_insert_request(hctx, rq, false);
1034 spin_unlock(&ctx->lock);
1036 blk_mq_run_hw_queue(hctx, from_schedule);
1037 blk_mq_put_ctx(current_ctx);
1040 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1042 struct request *rqa = container_of(a, struct request, queuelist);
1043 struct request *rqb = container_of(b, struct request, queuelist);
1045 return !(rqa->mq_ctx < rqb->mq_ctx ||
1046 (rqa->mq_ctx == rqb->mq_ctx &&
1047 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1050 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1052 struct blk_mq_ctx *this_ctx;
1053 struct request_queue *this_q;
1056 LIST_HEAD(ctx_list);
1059 list_splice_init(&plug->mq_list, &list);
1061 list_sort(NULL, &list, plug_ctx_cmp);
1067 while (!list_empty(&list)) {
1068 rq = list_entry_rq(list.next);
1069 list_del_init(&rq->queuelist);
1071 if (rq->mq_ctx != this_ctx) {
1073 blk_mq_insert_requests(this_q, this_ctx,
1078 this_ctx = rq->mq_ctx;
1084 list_add_tail(&rq->queuelist, &ctx_list);
1088 * If 'this_ctx' is set, we know we have entries to complete
1089 * on 'ctx_list'. Do those.
1092 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1097 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1099 init_request_from_bio(rq, bio);
1101 if (blk_do_io_stat(rq))
1102 blk_account_io_start(rq, 1);
1105 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1107 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1108 !blk_queue_nomerges(hctx->queue);
1111 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1112 struct blk_mq_ctx *ctx,
1113 struct request *rq, struct bio *bio)
1115 if (!hctx_allow_merges(hctx)) {
1116 blk_mq_bio_to_request(rq, bio);
1117 spin_lock(&ctx->lock);
1119 __blk_mq_insert_request(hctx, rq, false);
1120 spin_unlock(&ctx->lock);
1123 struct request_queue *q = hctx->queue;
1125 spin_lock(&ctx->lock);
1126 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1127 blk_mq_bio_to_request(rq, bio);
1131 spin_unlock(&ctx->lock);
1132 __blk_mq_free_request(hctx, ctx, rq);
1137 struct blk_map_ctx {
1138 struct blk_mq_hw_ctx *hctx;
1139 struct blk_mq_ctx *ctx;
1142 static struct request *blk_mq_map_request(struct request_queue *q,
1144 struct blk_map_ctx *data)
1146 struct blk_mq_hw_ctx *hctx;
1147 struct blk_mq_ctx *ctx;
1149 int rw = bio_data_dir(bio);
1150 struct blk_mq_alloc_data alloc_data;
1152 if (unlikely(blk_mq_queue_enter(q))) {
1153 bio_endio(bio, -EIO);
1157 ctx = blk_mq_get_ctx(q);
1158 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1160 if (rw_is_sync(bio->bi_rw))
1163 trace_block_getrq(q, bio, rw);
1164 blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1166 rq = __blk_mq_alloc_request(&alloc_data, rw);
1167 if (unlikely(!rq)) {
1168 __blk_mq_run_hw_queue(hctx);
1169 blk_mq_put_ctx(ctx);
1170 trace_block_sleeprq(q, bio, rw);
1172 ctx = blk_mq_get_ctx(q);
1173 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1174 blk_mq_set_alloc_data(&alloc_data, q,
1175 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1176 rq = __blk_mq_alloc_request(&alloc_data, rw);
1177 ctx = alloc_data.ctx;
1178 hctx = alloc_data.hctx;
1188 * Multiple hardware queue variant. This will not use per-process plugs,
1189 * but will attempt to bypass the hctx queueing if we can go straight to
1190 * hardware for SYNC IO.
1192 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1194 const int is_sync = rw_is_sync(bio->bi_rw);
1195 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1196 struct blk_map_ctx data;
1199 blk_queue_bounce(q, &bio);
1201 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1202 bio_endio(bio, -EIO);
1206 rq = blk_mq_map_request(q, bio, &data);
1210 if (unlikely(is_flush_fua)) {
1211 blk_mq_bio_to_request(rq, bio);
1212 blk_insert_flush(rq);
1217 * If the driver supports defer issued based on 'last', then
1218 * queue it up like normal since we can potentially save some
1221 if (is_sync && !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1222 struct blk_mq_queue_data bd = {
1229 blk_mq_bio_to_request(rq, bio);
1232 * For OK queue, we are done. For error, kill it. Any other
1233 * error (busy), just add it to our list as we previously
1236 ret = q->mq_ops->queue_rq(data.hctx, &bd);
1237 if (ret == BLK_MQ_RQ_QUEUE_OK)
1240 __blk_mq_requeue_request(rq);
1242 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1244 blk_mq_end_request(rq, rq->errors);
1250 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1252 * For a SYNC request, send it to the hardware immediately. For
1253 * an ASYNC request, just ensure that we run it later on. The
1254 * latter allows for merging opportunities and more efficient
1258 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1261 blk_mq_put_ctx(data.ctx);
1265 * Single hardware queue variant. This will attempt to use any per-process
1266 * plug for merging and IO deferral.
1268 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1270 const int is_sync = rw_is_sync(bio->bi_rw);
1271 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1272 unsigned int use_plug, request_count = 0;
1273 struct blk_map_ctx data;
1277 * If we have multiple hardware queues, just go directly to
1278 * one of those for sync IO.
1280 use_plug = !is_flush_fua && !is_sync;
1282 blk_queue_bounce(q, &bio);
1284 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1285 bio_endio(bio, -EIO);
1289 if (use_plug && !blk_queue_nomerges(q) &&
1290 blk_attempt_plug_merge(q, bio, &request_count))
1293 rq = blk_mq_map_request(q, bio, &data);
1297 if (unlikely(is_flush_fua)) {
1298 blk_mq_bio_to_request(rq, bio);
1299 blk_insert_flush(rq);
1304 * A task plug currently exists. Since this is completely lockless,
1305 * utilize that to temporarily store requests until the task is
1306 * either done or scheduled away.
1309 struct blk_plug *plug = current->plug;
1312 blk_mq_bio_to_request(rq, bio);
1313 if (list_empty(&plug->mq_list))
1314 trace_block_plug(q);
1315 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1316 blk_flush_plug_list(plug, false);
1317 trace_block_plug(q);
1319 list_add_tail(&rq->queuelist, &plug->mq_list);
1320 blk_mq_put_ctx(data.ctx);
1325 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1327 * For a SYNC request, send it to the hardware immediately. For
1328 * an ASYNC request, just ensure that we run it later on. The
1329 * latter allows for merging opportunities and more efficient
1333 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1336 blk_mq_put_ctx(data.ctx);
1340 * Default mapping to a software queue, since we use one per CPU.
1342 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1344 return q->queue_hw_ctx[q->mq_map[cpu]];
1346 EXPORT_SYMBOL(blk_mq_map_queue);
1348 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1349 struct blk_mq_tags *tags, unsigned int hctx_idx)
1353 if (tags->rqs && set->ops->exit_request) {
1356 for (i = 0; i < tags->nr_tags; i++) {
1359 set->ops->exit_request(set->driver_data, tags->rqs[i],
1361 tags->rqs[i] = NULL;
1365 while (!list_empty(&tags->page_list)) {
1366 page = list_first_entry(&tags->page_list, struct page, lru);
1367 list_del_init(&page->lru);
1368 __free_pages(page, page->private);
1373 blk_mq_free_tags(tags);
1376 static size_t order_to_size(unsigned int order)
1378 return (size_t)PAGE_SIZE << order;
1381 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1382 unsigned int hctx_idx)
1384 struct blk_mq_tags *tags;
1385 unsigned int i, j, entries_per_page, max_order = 4;
1386 size_t rq_size, left;
1388 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1393 INIT_LIST_HEAD(&tags->page_list);
1395 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1396 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1399 blk_mq_free_tags(tags);
1404 * rq_size is the size of the request plus driver payload, rounded
1405 * to the cacheline size
1407 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1409 left = rq_size * set->queue_depth;
1411 for (i = 0; i < set->queue_depth; ) {
1412 int this_order = max_order;
1417 while (left < order_to_size(this_order - 1) && this_order)
1421 page = alloc_pages_node(set->numa_node,
1422 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1428 if (order_to_size(this_order) < rq_size)
1435 page->private = this_order;
1436 list_add_tail(&page->lru, &tags->page_list);
1438 p = page_address(page);
1439 entries_per_page = order_to_size(this_order) / rq_size;
1440 to_do = min(entries_per_page, set->queue_depth - i);
1441 left -= to_do * rq_size;
1442 for (j = 0; j < to_do; j++) {
1444 tags->rqs[i]->atomic_flags = 0;
1445 tags->rqs[i]->cmd_flags = 0;
1446 if (set->ops->init_request) {
1447 if (set->ops->init_request(set->driver_data,
1448 tags->rqs[i], hctx_idx, i,
1450 tags->rqs[i] = NULL;
1463 blk_mq_free_rq_map(set, tags, hctx_idx);
1467 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1472 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1474 unsigned int bpw = 8, total, num_maps, i;
1476 bitmap->bits_per_word = bpw;
1478 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1479 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1484 bitmap->map_size = num_maps;
1487 for (i = 0; i < num_maps; i++) {
1488 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1489 total -= bitmap->map[i].depth;
1495 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1497 struct request_queue *q = hctx->queue;
1498 struct blk_mq_ctx *ctx;
1502 * Move ctx entries to new CPU, if this one is going away.
1504 ctx = __blk_mq_get_ctx(q, cpu);
1506 spin_lock(&ctx->lock);
1507 if (!list_empty(&ctx->rq_list)) {
1508 list_splice_init(&ctx->rq_list, &tmp);
1509 blk_mq_hctx_clear_pending(hctx, ctx);
1511 spin_unlock(&ctx->lock);
1513 if (list_empty(&tmp))
1516 ctx = blk_mq_get_ctx(q);
1517 spin_lock(&ctx->lock);
1519 while (!list_empty(&tmp)) {
1522 rq = list_first_entry(&tmp, struct request, queuelist);
1524 list_move_tail(&rq->queuelist, &ctx->rq_list);
1527 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1528 blk_mq_hctx_mark_pending(hctx, ctx);
1530 spin_unlock(&ctx->lock);
1532 blk_mq_run_hw_queue(hctx, true);
1533 blk_mq_put_ctx(ctx);
1537 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1539 struct request_queue *q = hctx->queue;
1540 struct blk_mq_tag_set *set = q->tag_set;
1542 if (set->tags[hctx->queue_num])
1545 set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1546 if (!set->tags[hctx->queue_num])
1549 hctx->tags = set->tags[hctx->queue_num];
1553 static int blk_mq_hctx_notify(void *data, unsigned long action,
1556 struct blk_mq_hw_ctx *hctx = data;
1558 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1559 return blk_mq_hctx_cpu_offline(hctx, cpu);
1560 else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1561 return blk_mq_hctx_cpu_online(hctx, cpu);
1566 static void blk_mq_exit_hctx(struct request_queue *q,
1567 struct blk_mq_tag_set *set,
1568 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1570 unsigned flush_start_tag = set->queue_depth;
1572 blk_mq_tag_idle(hctx);
1574 if (set->ops->exit_request)
1575 set->ops->exit_request(set->driver_data,
1576 hctx->fq->flush_rq, hctx_idx,
1577 flush_start_tag + hctx_idx);
1579 if (set->ops->exit_hctx)
1580 set->ops->exit_hctx(hctx, hctx_idx);
1582 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1583 blk_free_flush_queue(hctx->fq);
1585 blk_mq_free_bitmap(&hctx->ctx_map);
1588 static void blk_mq_exit_hw_queues(struct request_queue *q,
1589 struct blk_mq_tag_set *set, int nr_queue)
1591 struct blk_mq_hw_ctx *hctx;
1594 queue_for_each_hw_ctx(q, hctx, i) {
1597 blk_mq_exit_hctx(q, set, hctx, i);
1601 static void blk_mq_free_hw_queues(struct request_queue *q,
1602 struct blk_mq_tag_set *set)
1604 struct blk_mq_hw_ctx *hctx;
1607 queue_for_each_hw_ctx(q, hctx, i) {
1608 free_cpumask_var(hctx->cpumask);
1613 static int blk_mq_init_hctx(struct request_queue *q,
1614 struct blk_mq_tag_set *set,
1615 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1618 unsigned flush_start_tag = set->queue_depth;
1620 node = hctx->numa_node;
1621 if (node == NUMA_NO_NODE)
1622 node = hctx->numa_node = set->numa_node;
1624 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1625 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1626 spin_lock_init(&hctx->lock);
1627 INIT_LIST_HEAD(&hctx->dispatch);
1629 hctx->queue_num = hctx_idx;
1630 hctx->flags = set->flags;
1632 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1633 blk_mq_hctx_notify, hctx);
1634 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1636 hctx->tags = set->tags[hctx_idx];
1639 * Allocate space for all possible cpus to avoid allocation at
1642 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1645 goto unregister_cpu_notifier;
1647 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1652 if (set->ops->init_hctx &&
1653 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1656 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1660 if (set->ops->init_request &&
1661 set->ops->init_request(set->driver_data,
1662 hctx->fq->flush_rq, hctx_idx,
1663 flush_start_tag + hctx_idx, node))
1671 if (set->ops->exit_hctx)
1672 set->ops->exit_hctx(hctx, hctx_idx);
1674 blk_mq_free_bitmap(&hctx->ctx_map);
1677 unregister_cpu_notifier:
1678 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1683 static int blk_mq_init_hw_queues(struct request_queue *q,
1684 struct blk_mq_tag_set *set)
1686 struct blk_mq_hw_ctx *hctx;
1690 * Initialize hardware queues
1692 queue_for_each_hw_ctx(q, hctx, i) {
1693 if (blk_mq_init_hctx(q, set, hctx, i))
1697 if (i == q->nr_hw_queues)
1703 blk_mq_exit_hw_queues(q, set, i);
1708 static void blk_mq_init_cpu_queues(struct request_queue *q,
1709 unsigned int nr_hw_queues)
1713 for_each_possible_cpu(i) {
1714 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1715 struct blk_mq_hw_ctx *hctx;
1717 memset(__ctx, 0, sizeof(*__ctx));
1719 spin_lock_init(&__ctx->lock);
1720 INIT_LIST_HEAD(&__ctx->rq_list);
1723 /* If the cpu isn't online, the cpu is mapped to first hctx */
1727 hctx = q->mq_ops->map_queue(q, i);
1728 cpumask_set_cpu(i, hctx->cpumask);
1732 * Set local node, IFF we have more than one hw queue. If
1733 * not, we remain on the home node of the device
1735 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1736 hctx->numa_node = cpu_to_node(i);
1740 static void blk_mq_map_swqueue(struct request_queue *q)
1743 struct blk_mq_hw_ctx *hctx;
1744 struct blk_mq_ctx *ctx;
1746 queue_for_each_hw_ctx(q, hctx, i) {
1747 cpumask_clear(hctx->cpumask);
1752 * Map software to hardware queues
1754 queue_for_each_ctx(q, ctx, i) {
1755 /* If the cpu isn't online, the cpu is mapped to first hctx */
1759 hctx = q->mq_ops->map_queue(q, i);
1760 cpumask_set_cpu(i, hctx->cpumask);
1761 ctx->index_hw = hctx->nr_ctx;
1762 hctx->ctxs[hctx->nr_ctx++] = ctx;
1765 queue_for_each_hw_ctx(q, hctx, i) {
1767 * If no software queues are mapped to this hardware queue,
1768 * disable it and free the request entries.
1770 if (!hctx->nr_ctx) {
1771 struct blk_mq_tag_set *set = q->tag_set;
1774 blk_mq_free_rq_map(set, set->tags[i], i);
1775 set->tags[i] = NULL;
1782 * Initialize batch roundrobin counts
1784 hctx->next_cpu = cpumask_first(hctx->cpumask);
1785 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1789 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1791 struct blk_mq_hw_ctx *hctx;
1792 struct request_queue *q;
1796 if (set->tag_list.next == set->tag_list.prev)
1801 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1802 blk_mq_freeze_queue(q);
1804 queue_for_each_hw_ctx(q, hctx, i) {
1806 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1808 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1810 blk_mq_unfreeze_queue(q);
1814 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1816 struct blk_mq_tag_set *set = q->tag_set;
1818 mutex_lock(&set->tag_list_lock);
1819 list_del_init(&q->tag_set_list);
1820 blk_mq_update_tag_set_depth(set);
1821 mutex_unlock(&set->tag_list_lock);
1824 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1825 struct request_queue *q)
1829 mutex_lock(&set->tag_list_lock);
1830 list_add_tail(&q->tag_set_list, &set->tag_list);
1831 blk_mq_update_tag_set_depth(set);
1832 mutex_unlock(&set->tag_list_lock);
1835 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1837 struct blk_mq_hw_ctx **hctxs;
1838 struct blk_mq_ctx __percpu *ctx;
1839 struct request_queue *q;
1843 ctx = alloc_percpu(struct blk_mq_ctx);
1845 return ERR_PTR(-ENOMEM);
1847 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1853 map = blk_mq_make_queue_map(set);
1857 for (i = 0; i < set->nr_hw_queues; i++) {
1858 int node = blk_mq_hw_queue_to_node(map, i);
1860 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1865 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1869 atomic_set(&hctxs[i]->nr_active, 0);
1870 hctxs[i]->numa_node = node;
1871 hctxs[i]->queue_num = i;
1874 q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1879 * Init percpu_ref in atomic mode so that it's faster to shutdown.
1880 * See blk_register_queue() for details.
1882 if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release,
1883 PERCPU_REF_INIT_ATOMIC, GFP_KERNEL))
1886 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1887 blk_queue_rq_timeout(q, 30000);
1889 q->nr_queues = nr_cpu_ids;
1890 q->nr_hw_queues = set->nr_hw_queues;
1894 q->queue_hw_ctx = hctxs;
1896 q->mq_ops = set->ops;
1897 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1899 if (!(set->flags & BLK_MQ_F_SG_MERGE))
1900 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1902 q->sg_reserved_size = INT_MAX;
1904 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1905 INIT_LIST_HEAD(&q->requeue_list);
1906 spin_lock_init(&q->requeue_lock);
1908 if (q->nr_hw_queues > 1)
1909 blk_queue_make_request(q, blk_mq_make_request);
1911 blk_queue_make_request(q, blk_sq_make_request);
1914 blk_queue_rq_timeout(q, set->timeout);
1917 * Do this after blk_queue_make_request() overrides it...
1919 q->nr_requests = set->queue_depth;
1921 if (set->ops->complete)
1922 blk_queue_softirq_done(q, set->ops->complete);
1924 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1926 if (blk_mq_init_hw_queues(q, set))
1929 mutex_lock(&all_q_mutex);
1930 list_add_tail(&q->all_q_node, &all_q_list);
1931 mutex_unlock(&all_q_mutex);
1933 blk_mq_add_queue_tag_set(set, q);
1935 blk_mq_map_swqueue(q);
1940 blk_cleanup_queue(q);
1943 for (i = 0; i < set->nr_hw_queues; i++) {
1946 free_cpumask_var(hctxs[i]->cpumask);
1953 return ERR_PTR(-ENOMEM);
1955 EXPORT_SYMBOL(blk_mq_init_queue);
1957 void blk_mq_free_queue(struct request_queue *q)
1959 struct blk_mq_tag_set *set = q->tag_set;
1961 blk_mq_del_queue_tag_set(q);
1963 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
1964 blk_mq_free_hw_queues(q, set);
1966 percpu_ref_exit(&q->mq_usage_counter);
1968 free_percpu(q->queue_ctx);
1969 kfree(q->queue_hw_ctx);
1972 q->queue_ctx = NULL;
1973 q->queue_hw_ctx = NULL;
1976 mutex_lock(&all_q_mutex);
1977 list_del_init(&q->all_q_node);
1978 mutex_unlock(&all_q_mutex);
1981 /* Basically redo blk_mq_init_queue with queue frozen */
1982 static void blk_mq_queue_reinit(struct request_queue *q)
1984 WARN_ON_ONCE(!q->mq_freeze_depth);
1986 blk_mq_sysfs_unregister(q);
1988 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1991 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1992 * we should change hctx numa_node according to new topology (this
1993 * involves free and re-allocate memory, worthy doing?)
1996 blk_mq_map_swqueue(q);
1998 blk_mq_sysfs_register(q);
2001 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2002 unsigned long action, void *hcpu)
2004 struct request_queue *q;
2007 * Before new mappings are established, hotadded cpu might already
2008 * start handling requests. This doesn't break anything as we map
2009 * offline CPUs to first hardware queue. We will re-init the queue
2010 * below to get optimal settings.
2012 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
2013 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
2016 mutex_lock(&all_q_mutex);
2019 * We need to freeze and reinit all existing queues. Freezing
2020 * involves synchronous wait for an RCU grace period and doing it
2021 * one by one may take a long time. Start freezing all queues in
2022 * one swoop and then wait for the completions so that freezing can
2023 * take place in parallel.
2025 list_for_each_entry(q, &all_q_list, all_q_node)
2026 blk_mq_freeze_queue_start(q);
2027 list_for_each_entry(q, &all_q_list, all_q_node)
2028 blk_mq_freeze_queue_wait(q);
2030 list_for_each_entry(q, &all_q_list, all_q_node)
2031 blk_mq_queue_reinit(q);
2033 list_for_each_entry(q, &all_q_list, all_q_node)
2034 blk_mq_unfreeze_queue(q);
2036 mutex_unlock(&all_q_mutex);
2040 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2044 for (i = 0; i < set->nr_hw_queues; i++) {
2045 set->tags[i] = blk_mq_init_rq_map(set, i);
2054 blk_mq_free_rq_map(set, set->tags[i], i);
2060 * Allocate the request maps associated with this tag_set. Note that this
2061 * may reduce the depth asked for, if memory is tight. set->queue_depth
2062 * will be updated to reflect the allocated depth.
2064 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2069 depth = set->queue_depth;
2071 err = __blk_mq_alloc_rq_maps(set);
2075 set->queue_depth >>= 1;
2076 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2080 } while (set->queue_depth);
2082 if (!set->queue_depth || err) {
2083 pr_err("blk-mq: failed to allocate request map\n");
2087 if (depth != set->queue_depth)
2088 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2089 depth, set->queue_depth);
2095 * Alloc a tag set to be associated with one or more request queues.
2096 * May fail with EINVAL for various error conditions. May adjust the
2097 * requested depth down, if if it too large. In that case, the set
2098 * value will be stored in set->queue_depth.
2100 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2102 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2104 if (!set->nr_hw_queues)
2106 if (!set->queue_depth)
2108 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2111 if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue)
2114 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2115 pr_info("blk-mq: reduced tag depth to %u\n",
2117 set->queue_depth = BLK_MQ_MAX_DEPTH;
2121 * If a crashdump is active, then we are potentially in a very
2122 * memory constrained environment. Limit us to 1 queue and
2123 * 64 tags to prevent using too much memory.
2125 if (is_kdump_kernel()) {
2126 set->nr_hw_queues = 1;
2127 set->queue_depth = min(64U, set->queue_depth);
2130 set->tags = kmalloc_node(set->nr_hw_queues *
2131 sizeof(struct blk_mq_tags *),
2132 GFP_KERNEL, set->numa_node);
2136 if (blk_mq_alloc_rq_maps(set))
2139 mutex_init(&set->tag_list_lock);
2140 INIT_LIST_HEAD(&set->tag_list);
2148 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2150 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2154 for (i = 0; i < set->nr_hw_queues; i++) {
2156 blk_mq_free_rq_map(set, set->tags[i], i);
2162 EXPORT_SYMBOL(blk_mq_free_tag_set);
2164 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2166 struct blk_mq_tag_set *set = q->tag_set;
2167 struct blk_mq_hw_ctx *hctx;
2170 if (!set || nr > set->queue_depth)
2174 queue_for_each_hw_ctx(q, hctx, i) {
2175 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2181 q->nr_requests = nr;
2186 void blk_mq_disable_hotplug(void)
2188 mutex_lock(&all_q_mutex);
2191 void blk_mq_enable_hotplug(void)
2193 mutex_unlock(&all_q_mutex);
2196 static int __init blk_mq_init(void)
2200 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2204 subsys_initcall(blk_mq_init);