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 static 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);
124 static void blk_mq_freeze_queue_wait(struct request_queue *q)
126 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter));
130 * Guarantee no request is in use, so we can change any data structure of
131 * the queue afterward.
133 void blk_mq_freeze_queue(struct request_queue *q)
135 blk_mq_freeze_queue_start(q);
136 blk_mq_freeze_queue_wait(q);
139 static void blk_mq_unfreeze_queue(struct request_queue *q)
143 spin_lock_irq(q->queue_lock);
144 wake = !--q->mq_freeze_depth;
145 WARN_ON_ONCE(q->mq_freeze_depth < 0);
146 spin_unlock_irq(q->queue_lock);
148 percpu_ref_reinit(&q->mq_usage_counter);
149 wake_up_all(&q->mq_freeze_wq);
153 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
155 return blk_mq_has_free_tags(hctx->tags);
157 EXPORT_SYMBOL(blk_mq_can_queue);
159 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
160 struct request *rq, unsigned int rw_flags)
162 if (blk_queue_io_stat(q))
163 rw_flags |= REQ_IO_STAT;
165 INIT_LIST_HEAD(&rq->queuelist);
166 /* csd/requeue_work/fifo_time is initialized before use */
169 rq->cmd_flags |= rw_flags;
170 /* do not touch atomic flags, it needs atomic ops against the timer */
172 INIT_HLIST_NODE(&rq->hash);
173 RB_CLEAR_NODE(&rq->rb_node);
176 rq->start_time = jiffies;
177 #ifdef CONFIG_BLK_CGROUP
179 set_start_time_ns(rq);
180 rq->io_start_time_ns = 0;
182 rq->nr_phys_segments = 0;
183 #if defined(CONFIG_BLK_DEV_INTEGRITY)
184 rq->nr_integrity_segments = 0;
187 /* tag was already set */
197 INIT_LIST_HEAD(&rq->timeout_list);
201 rq->end_io_data = NULL;
204 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
207 static struct request *
208 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
213 tag = blk_mq_get_tag(data);
214 if (tag != BLK_MQ_TAG_FAIL) {
215 rq = data->hctx->tags->rqs[tag];
217 if (blk_mq_tag_busy(data->hctx)) {
218 rq->cmd_flags = REQ_MQ_INFLIGHT;
219 atomic_inc(&data->hctx->nr_active);
223 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
230 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
233 struct blk_mq_ctx *ctx;
234 struct blk_mq_hw_ctx *hctx;
236 struct blk_mq_alloc_data alloc_data;
239 ret = blk_mq_queue_enter(q);
243 ctx = blk_mq_get_ctx(q);
244 hctx = q->mq_ops->map_queue(q, ctx->cpu);
245 blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
246 reserved, ctx, hctx);
248 rq = __blk_mq_alloc_request(&alloc_data, rw);
249 if (!rq && (gfp & __GFP_WAIT)) {
250 __blk_mq_run_hw_queue(hctx);
253 ctx = blk_mq_get_ctx(q);
254 hctx = q->mq_ops->map_queue(q, ctx->cpu);
255 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
257 rq = __blk_mq_alloc_request(&alloc_data, rw);
258 ctx = alloc_data.ctx;
262 return ERR_PTR(-EWOULDBLOCK);
265 EXPORT_SYMBOL(blk_mq_alloc_request);
267 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
268 struct blk_mq_ctx *ctx, struct request *rq)
270 const int tag = rq->tag;
271 struct request_queue *q = rq->q;
273 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
274 atomic_dec(&hctx->nr_active);
277 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
278 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
279 blk_mq_queue_exit(q);
282 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
284 struct blk_mq_ctx *ctx = rq->mq_ctx;
286 ctx->rq_completed[rq_is_sync(rq)]++;
287 __blk_mq_free_request(hctx, ctx, rq);
290 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
292 void blk_mq_free_request(struct request *rq)
294 struct blk_mq_hw_ctx *hctx;
295 struct request_queue *q = rq->q;
297 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
298 blk_mq_free_hctx_request(hctx, rq);
300 EXPORT_SYMBOL_GPL(blk_mq_free_request);
302 inline void __blk_mq_end_request(struct request *rq, int error)
304 blk_account_io_done(rq);
307 rq->end_io(rq, error);
309 if (unlikely(blk_bidi_rq(rq)))
310 blk_mq_free_request(rq->next_rq);
311 blk_mq_free_request(rq);
314 EXPORT_SYMBOL(__blk_mq_end_request);
316 void blk_mq_end_request(struct request *rq, int error)
318 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
320 __blk_mq_end_request(rq, error);
322 EXPORT_SYMBOL(blk_mq_end_request);
324 static void __blk_mq_complete_request_remote(void *data)
326 struct request *rq = data;
328 rq->q->softirq_done_fn(rq);
331 static void blk_mq_ipi_complete_request(struct request *rq)
333 struct blk_mq_ctx *ctx = rq->mq_ctx;
337 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
338 rq->q->softirq_done_fn(rq);
343 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
344 shared = cpus_share_cache(cpu, ctx->cpu);
346 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
347 rq->csd.func = __blk_mq_complete_request_remote;
350 smp_call_function_single_async(ctx->cpu, &rq->csd);
352 rq->q->softirq_done_fn(rq);
357 void __blk_mq_complete_request(struct request *rq)
359 struct request_queue *q = rq->q;
361 if (!q->softirq_done_fn)
362 blk_mq_end_request(rq, rq->errors);
364 blk_mq_ipi_complete_request(rq);
368 * blk_mq_complete_request - end I/O on a request
369 * @rq: the request being processed
372 * Ends all I/O on a request. It does not handle partial completions.
373 * The actual completion happens out-of-order, through a IPI handler.
375 void blk_mq_complete_request(struct request *rq)
377 struct request_queue *q = rq->q;
379 if (unlikely(blk_should_fake_timeout(q)))
381 if (!blk_mark_rq_complete(rq))
382 __blk_mq_complete_request(rq);
384 EXPORT_SYMBOL(blk_mq_complete_request);
386 void blk_mq_start_request(struct request *rq)
388 struct request_queue *q = rq->q;
390 trace_block_rq_issue(q, rq);
392 rq->resid_len = blk_rq_bytes(rq);
393 if (unlikely(blk_bidi_rq(rq)))
394 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
399 * Ensure that ->deadline is visible before set the started
400 * flag and clear the completed flag.
402 smp_mb__before_atomic();
405 * Mark us as started and clear complete. Complete might have been
406 * set if requeue raced with timeout, which then marked it as
407 * complete. So be sure to clear complete again when we start
408 * the request, otherwise we'll ignore the completion event.
410 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
411 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
412 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
413 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
415 if (q->dma_drain_size && blk_rq_bytes(rq)) {
417 * Make sure space for the drain appears. We know we can do
418 * this because max_hw_segments has been adjusted to be one
419 * fewer than the device can handle.
421 rq->nr_phys_segments++;
424 EXPORT_SYMBOL(blk_mq_start_request);
426 static void __blk_mq_requeue_request(struct request *rq)
428 struct request_queue *q = rq->q;
430 trace_block_rq_requeue(q, rq);
432 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
433 if (q->dma_drain_size && blk_rq_bytes(rq))
434 rq->nr_phys_segments--;
438 void blk_mq_requeue_request(struct request *rq)
440 __blk_mq_requeue_request(rq);
442 BUG_ON(blk_queued_rq(rq));
443 blk_mq_add_to_requeue_list(rq, true);
445 EXPORT_SYMBOL(blk_mq_requeue_request);
447 static void blk_mq_requeue_work(struct work_struct *work)
449 struct request_queue *q =
450 container_of(work, struct request_queue, requeue_work);
452 struct request *rq, *next;
455 spin_lock_irqsave(&q->requeue_lock, flags);
456 list_splice_init(&q->requeue_list, &rq_list);
457 spin_unlock_irqrestore(&q->requeue_lock, flags);
459 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
460 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
463 rq->cmd_flags &= ~REQ_SOFTBARRIER;
464 list_del_init(&rq->queuelist);
465 blk_mq_insert_request(rq, true, false, false);
468 while (!list_empty(&rq_list)) {
469 rq = list_entry(rq_list.next, struct request, queuelist);
470 list_del_init(&rq->queuelist);
471 blk_mq_insert_request(rq, false, false, false);
475 * Use the start variant of queue running here, so that running
476 * the requeue work will kick stopped queues.
478 blk_mq_start_hw_queues(q);
481 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
483 struct request_queue *q = rq->q;
487 * We abuse this flag that is otherwise used by the I/O scheduler to
488 * request head insertation from the workqueue.
490 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
492 spin_lock_irqsave(&q->requeue_lock, flags);
494 rq->cmd_flags |= REQ_SOFTBARRIER;
495 list_add(&rq->queuelist, &q->requeue_list);
497 list_add_tail(&rq->queuelist, &q->requeue_list);
499 spin_unlock_irqrestore(&q->requeue_lock, flags);
501 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
503 void blk_mq_kick_requeue_list(struct request_queue *q)
505 kblockd_schedule_work(&q->requeue_work);
507 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
509 static inline bool is_flush_request(struct request *rq,
510 struct blk_flush_queue *fq, unsigned int tag)
512 return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
513 fq->flush_rq->tag == tag);
516 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
518 struct request *rq = tags->rqs[tag];
519 /* mq_ctx of flush rq is always cloned from the corresponding req */
520 struct blk_flush_queue *fq = blk_get_flush_queue(rq->q, rq->mq_ctx);
522 if (!is_flush_request(rq, fq, tag))
527 EXPORT_SYMBOL(blk_mq_tag_to_rq);
529 struct blk_mq_timeout_data {
531 unsigned int next_set;
534 void blk_mq_rq_timed_out(struct request *req, bool reserved)
536 struct blk_mq_ops *ops = req->q->mq_ops;
537 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
540 * We know that complete is set at this point. If STARTED isn't set
541 * anymore, then the request isn't active and the "timeout" should
542 * just be ignored. This can happen due to the bitflag ordering.
543 * Timeout first checks if STARTED is set, and if it is, assumes
544 * the request is active. But if we race with completion, then
545 * we both flags will get cleared. So check here again, and ignore
546 * a timeout event with a request that isn't active.
548 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
552 ret = ops->timeout(req, reserved);
556 __blk_mq_complete_request(req);
558 case BLK_EH_RESET_TIMER:
560 blk_clear_rq_complete(req);
562 case BLK_EH_NOT_HANDLED:
565 printk(KERN_ERR "block: bad eh return: %d\n", ret);
570 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
571 struct request *rq, void *priv, bool reserved)
573 struct blk_mq_timeout_data *data = priv;
575 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
578 if (time_after_eq(jiffies, rq->deadline)) {
579 if (!blk_mark_rq_complete(rq))
580 blk_mq_rq_timed_out(rq, reserved);
581 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
582 data->next = rq->deadline;
587 static void blk_mq_rq_timer(unsigned long priv)
589 struct request_queue *q = (struct request_queue *)priv;
590 struct blk_mq_timeout_data data = {
594 struct blk_mq_hw_ctx *hctx;
597 queue_for_each_hw_ctx(q, hctx, i) {
599 * If not software queues are currently mapped to this
600 * hardware queue, there's nothing to check
602 if (!hctx->nr_ctx || !hctx->tags)
605 blk_mq_tag_busy_iter(hctx, blk_mq_check_expired, &data);
609 data.next = blk_rq_timeout(round_jiffies_up(data.next));
610 mod_timer(&q->timeout, data.next);
612 queue_for_each_hw_ctx(q, hctx, i)
613 blk_mq_tag_idle(hctx);
618 * Reverse check our software queue for entries that we could potentially
619 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
620 * too much time checking for merges.
622 static bool blk_mq_attempt_merge(struct request_queue *q,
623 struct blk_mq_ctx *ctx, struct bio *bio)
628 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
634 if (!blk_rq_merge_ok(rq, bio))
637 el_ret = blk_try_merge(rq, bio);
638 if (el_ret == ELEVATOR_BACK_MERGE) {
639 if (bio_attempt_back_merge(q, rq, bio)) {
644 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
645 if (bio_attempt_front_merge(q, rq, bio)) {
657 * Process software queues that have been marked busy, splicing them
658 * to the for-dispatch
660 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
662 struct blk_mq_ctx *ctx;
665 for (i = 0; i < hctx->ctx_map.map_size; i++) {
666 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
667 unsigned int off, bit;
673 off = i * hctx->ctx_map.bits_per_word;
675 bit = find_next_bit(&bm->word, bm->depth, bit);
676 if (bit >= bm->depth)
679 ctx = hctx->ctxs[bit + off];
680 clear_bit(bit, &bm->word);
681 spin_lock(&ctx->lock);
682 list_splice_tail_init(&ctx->rq_list, list);
683 spin_unlock(&ctx->lock);
691 * Run this hardware queue, pulling any software queues mapped to it in.
692 * Note that this function currently has various problems around ordering
693 * of IO. In particular, we'd like FIFO behaviour on handling existing
694 * items on the hctx->dispatch list. Ignore that for now.
696 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
698 struct request_queue *q = hctx->queue;
701 LIST_HEAD(driver_list);
702 struct list_head *dptr;
705 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
707 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
713 * Touch any software queue that has pending entries.
715 flush_busy_ctxs(hctx, &rq_list);
718 * If we have previous entries on our dispatch list, grab them
719 * and stuff them at the front for more fair dispatch.
721 if (!list_empty_careful(&hctx->dispatch)) {
722 spin_lock(&hctx->lock);
723 if (!list_empty(&hctx->dispatch))
724 list_splice_init(&hctx->dispatch, &rq_list);
725 spin_unlock(&hctx->lock);
729 * Start off with dptr being NULL, so we start the first request
730 * immediately, even if we have more pending.
735 * Now process all the entries, sending them to the driver.
738 while (!list_empty(&rq_list)) {
739 struct blk_mq_queue_data bd;
742 rq = list_first_entry(&rq_list, struct request, queuelist);
743 list_del_init(&rq->queuelist);
747 bd.last = list_empty(&rq_list);
749 ret = q->mq_ops->queue_rq(hctx, &bd);
751 case BLK_MQ_RQ_QUEUE_OK:
754 case BLK_MQ_RQ_QUEUE_BUSY:
755 list_add(&rq->queuelist, &rq_list);
756 __blk_mq_requeue_request(rq);
759 pr_err("blk-mq: bad return on queue: %d\n", ret);
760 case BLK_MQ_RQ_QUEUE_ERROR:
762 blk_mq_end_request(rq, rq->errors);
766 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
770 * We've done the first request. If we have more than 1
771 * left in the list, set dptr to defer issue.
773 if (!dptr && rq_list.next != rq_list.prev)
778 hctx->dispatched[0]++;
779 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
780 hctx->dispatched[ilog2(queued) + 1]++;
783 * Any items that need requeuing? Stuff them into hctx->dispatch,
784 * that is where we will continue on next queue run.
786 if (!list_empty(&rq_list)) {
787 spin_lock(&hctx->lock);
788 list_splice(&rq_list, &hctx->dispatch);
789 spin_unlock(&hctx->lock);
794 * It'd be great if the workqueue API had a way to pass
795 * in a mask and had some smarts for more clever placement.
796 * For now we just round-robin here, switching for every
797 * BLK_MQ_CPU_WORK_BATCH queued items.
799 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
801 if (hctx->queue->nr_hw_queues == 1)
802 return WORK_CPU_UNBOUND;
804 if (--hctx->next_cpu_batch <= 0) {
805 int cpu = hctx->next_cpu, next_cpu;
807 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
808 if (next_cpu >= nr_cpu_ids)
809 next_cpu = cpumask_first(hctx->cpumask);
811 hctx->next_cpu = next_cpu;
812 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
817 return hctx->next_cpu;
820 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
822 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
827 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
828 __blk_mq_run_hw_queue(hctx);
836 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
840 void blk_mq_run_queues(struct request_queue *q, bool async)
842 struct blk_mq_hw_ctx *hctx;
845 queue_for_each_hw_ctx(q, hctx, i) {
846 if ((!blk_mq_hctx_has_pending(hctx) &&
847 list_empty_careful(&hctx->dispatch)) ||
848 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
851 blk_mq_run_hw_queue(hctx, async);
854 EXPORT_SYMBOL(blk_mq_run_queues);
856 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
858 cancel_delayed_work(&hctx->run_work);
859 cancel_delayed_work(&hctx->delay_work);
860 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
862 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
864 void blk_mq_stop_hw_queues(struct request_queue *q)
866 struct blk_mq_hw_ctx *hctx;
869 queue_for_each_hw_ctx(q, hctx, i)
870 blk_mq_stop_hw_queue(hctx);
872 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
874 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
876 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
878 blk_mq_run_hw_queue(hctx, false);
880 EXPORT_SYMBOL(blk_mq_start_hw_queue);
882 void blk_mq_start_hw_queues(struct request_queue *q)
884 struct blk_mq_hw_ctx *hctx;
887 queue_for_each_hw_ctx(q, hctx, i)
888 blk_mq_start_hw_queue(hctx);
890 EXPORT_SYMBOL(blk_mq_start_hw_queues);
893 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
895 struct blk_mq_hw_ctx *hctx;
898 queue_for_each_hw_ctx(q, hctx, i) {
899 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
902 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
903 blk_mq_run_hw_queue(hctx, async);
906 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
908 static void blk_mq_run_work_fn(struct work_struct *work)
910 struct blk_mq_hw_ctx *hctx;
912 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
914 __blk_mq_run_hw_queue(hctx);
917 static void blk_mq_delay_work_fn(struct work_struct *work)
919 struct blk_mq_hw_ctx *hctx;
921 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
923 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
924 __blk_mq_run_hw_queue(hctx);
927 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
929 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
930 &hctx->delay_work, msecs_to_jiffies(msecs));
932 EXPORT_SYMBOL(blk_mq_delay_queue);
934 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
935 struct request *rq, bool at_head)
937 struct blk_mq_ctx *ctx = rq->mq_ctx;
939 trace_block_rq_insert(hctx->queue, rq);
942 list_add(&rq->queuelist, &ctx->rq_list);
944 list_add_tail(&rq->queuelist, &ctx->rq_list);
946 blk_mq_hctx_mark_pending(hctx, ctx);
949 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
952 struct request_queue *q = rq->q;
953 struct blk_mq_hw_ctx *hctx;
954 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
956 current_ctx = blk_mq_get_ctx(q);
957 if (!cpu_online(ctx->cpu))
958 rq->mq_ctx = ctx = current_ctx;
960 hctx = q->mq_ops->map_queue(q, ctx->cpu);
962 spin_lock(&ctx->lock);
963 __blk_mq_insert_request(hctx, rq, at_head);
964 spin_unlock(&ctx->lock);
967 blk_mq_run_hw_queue(hctx, async);
969 blk_mq_put_ctx(current_ctx);
972 static void blk_mq_insert_requests(struct request_queue *q,
973 struct blk_mq_ctx *ctx,
974 struct list_head *list,
979 struct blk_mq_hw_ctx *hctx;
980 struct blk_mq_ctx *current_ctx;
982 trace_block_unplug(q, depth, !from_schedule);
984 current_ctx = blk_mq_get_ctx(q);
986 if (!cpu_online(ctx->cpu))
988 hctx = q->mq_ops->map_queue(q, ctx->cpu);
991 * preemption doesn't flush plug list, so it's possible ctx->cpu is
994 spin_lock(&ctx->lock);
995 while (!list_empty(list)) {
998 rq = list_first_entry(list, struct request, queuelist);
999 list_del_init(&rq->queuelist);
1001 __blk_mq_insert_request(hctx, rq, false);
1003 spin_unlock(&ctx->lock);
1005 blk_mq_run_hw_queue(hctx, from_schedule);
1006 blk_mq_put_ctx(current_ctx);
1009 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1011 struct request *rqa = container_of(a, struct request, queuelist);
1012 struct request *rqb = container_of(b, struct request, queuelist);
1014 return !(rqa->mq_ctx < rqb->mq_ctx ||
1015 (rqa->mq_ctx == rqb->mq_ctx &&
1016 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1019 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1021 struct blk_mq_ctx *this_ctx;
1022 struct request_queue *this_q;
1025 LIST_HEAD(ctx_list);
1028 list_splice_init(&plug->mq_list, &list);
1030 list_sort(NULL, &list, plug_ctx_cmp);
1036 while (!list_empty(&list)) {
1037 rq = list_entry_rq(list.next);
1038 list_del_init(&rq->queuelist);
1040 if (rq->mq_ctx != this_ctx) {
1042 blk_mq_insert_requests(this_q, this_ctx,
1047 this_ctx = rq->mq_ctx;
1053 list_add_tail(&rq->queuelist, &ctx_list);
1057 * If 'this_ctx' is set, we know we have entries to complete
1058 * on 'ctx_list'. Do those.
1061 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1066 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1068 init_request_from_bio(rq, bio);
1070 if (blk_do_io_stat(rq))
1071 blk_account_io_start(rq, 1);
1074 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1076 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1077 !blk_queue_nomerges(hctx->queue);
1080 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1081 struct blk_mq_ctx *ctx,
1082 struct request *rq, struct bio *bio)
1084 if (!hctx_allow_merges(hctx)) {
1085 blk_mq_bio_to_request(rq, bio);
1086 spin_lock(&ctx->lock);
1088 __blk_mq_insert_request(hctx, rq, false);
1089 spin_unlock(&ctx->lock);
1092 struct request_queue *q = hctx->queue;
1094 spin_lock(&ctx->lock);
1095 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1096 blk_mq_bio_to_request(rq, bio);
1100 spin_unlock(&ctx->lock);
1101 __blk_mq_free_request(hctx, ctx, rq);
1106 struct blk_map_ctx {
1107 struct blk_mq_hw_ctx *hctx;
1108 struct blk_mq_ctx *ctx;
1111 static struct request *blk_mq_map_request(struct request_queue *q,
1113 struct blk_map_ctx *data)
1115 struct blk_mq_hw_ctx *hctx;
1116 struct blk_mq_ctx *ctx;
1118 int rw = bio_data_dir(bio);
1119 struct blk_mq_alloc_data alloc_data;
1121 if (unlikely(blk_mq_queue_enter(q))) {
1122 bio_endio(bio, -EIO);
1126 ctx = blk_mq_get_ctx(q);
1127 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1129 if (rw_is_sync(bio->bi_rw))
1132 trace_block_getrq(q, bio, rw);
1133 blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1135 rq = __blk_mq_alloc_request(&alloc_data, rw);
1136 if (unlikely(!rq)) {
1137 __blk_mq_run_hw_queue(hctx);
1138 blk_mq_put_ctx(ctx);
1139 trace_block_sleeprq(q, bio, rw);
1141 ctx = blk_mq_get_ctx(q);
1142 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1143 blk_mq_set_alloc_data(&alloc_data, q,
1144 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1145 rq = __blk_mq_alloc_request(&alloc_data, rw);
1146 ctx = alloc_data.ctx;
1147 hctx = alloc_data.hctx;
1157 * Multiple hardware queue variant. This will not use per-process plugs,
1158 * but will attempt to bypass the hctx queueing if we can go straight to
1159 * hardware for SYNC IO.
1161 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1163 const int is_sync = rw_is_sync(bio->bi_rw);
1164 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1165 struct blk_map_ctx data;
1168 blk_queue_bounce(q, &bio);
1170 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1171 bio_endio(bio, -EIO);
1175 rq = blk_mq_map_request(q, bio, &data);
1179 if (unlikely(is_flush_fua)) {
1180 blk_mq_bio_to_request(rq, bio);
1181 blk_insert_flush(rq);
1186 * If the driver supports defer issued based on 'last', then
1187 * queue it up like normal since we can potentially save some
1190 if (is_sync && !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1191 struct blk_mq_queue_data bd = {
1198 blk_mq_bio_to_request(rq, bio);
1201 * For OK queue, we are done. For error, kill it. Any other
1202 * error (busy), just add it to our list as we previously
1205 ret = q->mq_ops->queue_rq(data.hctx, &bd);
1206 if (ret == BLK_MQ_RQ_QUEUE_OK)
1209 __blk_mq_requeue_request(rq);
1211 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1213 blk_mq_end_request(rq, rq->errors);
1219 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1221 * For a SYNC request, send it to the hardware immediately. For
1222 * an ASYNC request, just ensure that we run it later on. The
1223 * latter allows for merging opportunities and more efficient
1227 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1230 blk_mq_put_ctx(data.ctx);
1234 * Single hardware queue variant. This will attempt to use any per-process
1235 * plug for merging and IO deferral.
1237 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1239 const int is_sync = rw_is_sync(bio->bi_rw);
1240 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1241 unsigned int use_plug, request_count = 0;
1242 struct blk_map_ctx data;
1246 * If we have multiple hardware queues, just go directly to
1247 * one of those for sync IO.
1249 use_plug = !is_flush_fua && !is_sync;
1251 blk_queue_bounce(q, &bio);
1253 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1254 bio_endio(bio, -EIO);
1258 if (use_plug && !blk_queue_nomerges(q) &&
1259 blk_attempt_plug_merge(q, bio, &request_count))
1262 rq = blk_mq_map_request(q, bio, &data);
1266 if (unlikely(is_flush_fua)) {
1267 blk_mq_bio_to_request(rq, bio);
1268 blk_insert_flush(rq);
1273 * A task plug currently exists. Since this is completely lockless,
1274 * utilize that to temporarily store requests until the task is
1275 * either done or scheduled away.
1278 struct blk_plug *plug = current->plug;
1281 blk_mq_bio_to_request(rq, bio);
1282 if (list_empty(&plug->mq_list))
1283 trace_block_plug(q);
1284 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1285 blk_flush_plug_list(plug, false);
1286 trace_block_plug(q);
1288 list_add_tail(&rq->queuelist, &plug->mq_list);
1289 blk_mq_put_ctx(data.ctx);
1294 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1296 * For a SYNC request, send it to the hardware immediately. For
1297 * an ASYNC request, just ensure that we run it later on. The
1298 * latter allows for merging opportunities and more efficient
1302 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1305 blk_mq_put_ctx(data.ctx);
1309 * Default mapping to a software queue, since we use one per CPU.
1311 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1313 return q->queue_hw_ctx[q->mq_map[cpu]];
1315 EXPORT_SYMBOL(blk_mq_map_queue);
1317 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1318 struct blk_mq_tags *tags, unsigned int hctx_idx)
1322 if (tags->rqs && set->ops->exit_request) {
1325 for (i = 0; i < tags->nr_tags; i++) {
1328 set->ops->exit_request(set->driver_data, tags->rqs[i],
1330 tags->rqs[i] = NULL;
1334 while (!list_empty(&tags->page_list)) {
1335 page = list_first_entry(&tags->page_list, struct page, lru);
1336 list_del_init(&page->lru);
1337 __free_pages(page, page->private);
1342 blk_mq_free_tags(tags);
1345 static size_t order_to_size(unsigned int order)
1347 return (size_t)PAGE_SIZE << order;
1350 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1351 unsigned int hctx_idx)
1353 struct blk_mq_tags *tags;
1354 unsigned int i, j, entries_per_page, max_order = 4;
1355 size_t rq_size, left;
1357 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1362 INIT_LIST_HEAD(&tags->page_list);
1364 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1365 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1368 blk_mq_free_tags(tags);
1373 * rq_size is the size of the request plus driver payload, rounded
1374 * to the cacheline size
1376 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1378 left = rq_size * set->queue_depth;
1380 for (i = 0; i < set->queue_depth; ) {
1381 int this_order = max_order;
1386 while (left < order_to_size(this_order - 1) && this_order)
1390 page = alloc_pages_node(set->numa_node,
1391 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1397 if (order_to_size(this_order) < rq_size)
1404 page->private = this_order;
1405 list_add_tail(&page->lru, &tags->page_list);
1407 p = page_address(page);
1408 entries_per_page = order_to_size(this_order) / rq_size;
1409 to_do = min(entries_per_page, set->queue_depth - i);
1410 left -= to_do * rq_size;
1411 for (j = 0; j < to_do; j++) {
1413 tags->rqs[i]->atomic_flags = 0;
1414 tags->rqs[i]->cmd_flags = 0;
1415 if (set->ops->init_request) {
1416 if (set->ops->init_request(set->driver_data,
1417 tags->rqs[i], hctx_idx, i,
1419 tags->rqs[i] = NULL;
1432 blk_mq_free_rq_map(set, tags, hctx_idx);
1436 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1441 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1443 unsigned int bpw = 8, total, num_maps, i;
1445 bitmap->bits_per_word = bpw;
1447 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1448 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1453 bitmap->map_size = num_maps;
1456 for (i = 0; i < num_maps; i++) {
1457 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1458 total -= bitmap->map[i].depth;
1464 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1466 struct request_queue *q = hctx->queue;
1467 struct blk_mq_ctx *ctx;
1471 * Move ctx entries to new CPU, if this one is going away.
1473 ctx = __blk_mq_get_ctx(q, cpu);
1475 spin_lock(&ctx->lock);
1476 if (!list_empty(&ctx->rq_list)) {
1477 list_splice_init(&ctx->rq_list, &tmp);
1478 blk_mq_hctx_clear_pending(hctx, ctx);
1480 spin_unlock(&ctx->lock);
1482 if (list_empty(&tmp))
1485 ctx = blk_mq_get_ctx(q);
1486 spin_lock(&ctx->lock);
1488 while (!list_empty(&tmp)) {
1491 rq = list_first_entry(&tmp, struct request, queuelist);
1493 list_move_tail(&rq->queuelist, &ctx->rq_list);
1496 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1497 blk_mq_hctx_mark_pending(hctx, ctx);
1499 spin_unlock(&ctx->lock);
1501 blk_mq_run_hw_queue(hctx, true);
1502 blk_mq_put_ctx(ctx);
1506 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1508 struct request_queue *q = hctx->queue;
1509 struct blk_mq_tag_set *set = q->tag_set;
1511 if (set->tags[hctx->queue_num])
1514 set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1515 if (!set->tags[hctx->queue_num])
1518 hctx->tags = set->tags[hctx->queue_num];
1522 static int blk_mq_hctx_notify(void *data, unsigned long action,
1525 struct blk_mq_hw_ctx *hctx = data;
1527 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1528 return blk_mq_hctx_cpu_offline(hctx, cpu);
1529 else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1530 return blk_mq_hctx_cpu_online(hctx, cpu);
1535 static void blk_mq_exit_hctx(struct request_queue *q,
1536 struct blk_mq_tag_set *set,
1537 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1539 unsigned flush_start_tag = set->queue_depth;
1541 blk_mq_tag_idle(hctx);
1543 if (set->ops->exit_request)
1544 set->ops->exit_request(set->driver_data,
1545 hctx->fq->flush_rq, hctx_idx,
1546 flush_start_tag + hctx_idx);
1548 if (set->ops->exit_hctx)
1549 set->ops->exit_hctx(hctx, hctx_idx);
1551 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1552 blk_free_flush_queue(hctx->fq);
1554 blk_mq_free_bitmap(&hctx->ctx_map);
1557 static void blk_mq_exit_hw_queues(struct request_queue *q,
1558 struct blk_mq_tag_set *set, int nr_queue)
1560 struct blk_mq_hw_ctx *hctx;
1563 queue_for_each_hw_ctx(q, hctx, i) {
1566 blk_mq_exit_hctx(q, set, hctx, i);
1570 static void blk_mq_free_hw_queues(struct request_queue *q,
1571 struct blk_mq_tag_set *set)
1573 struct blk_mq_hw_ctx *hctx;
1576 queue_for_each_hw_ctx(q, hctx, i) {
1577 free_cpumask_var(hctx->cpumask);
1582 static int blk_mq_init_hctx(struct request_queue *q,
1583 struct blk_mq_tag_set *set,
1584 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1587 unsigned flush_start_tag = set->queue_depth;
1589 node = hctx->numa_node;
1590 if (node == NUMA_NO_NODE)
1591 node = hctx->numa_node = set->numa_node;
1593 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1594 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1595 spin_lock_init(&hctx->lock);
1596 INIT_LIST_HEAD(&hctx->dispatch);
1598 hctx->queue_num = hctx_idx;
1599 hctx->flags = set->flags;
1600 hctx->cmd_size = set->cmd_size;
1602 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1603 blk_mq_hctx_notify, hctx);
1604 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1606 hctx->tags = set->tags[hctx_idx];
1609 * Allocate space for all possible cpus to avoid allocation at
1612 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1615 goto unregister_cpu_notifier;
1617 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1622 if (set->ops->init_hctx &&
1623 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1626 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1630 if (set->ops->init_request &&
1631 set->ops->init_request(set->driver_data,
1632 hctx->fq->flush_rq, hctx_idx,
1633 flush_start_tag + hctx_idx, node))
1641 if (set->ops->exit_hctx)
1642 set->ops->exit_hctx(hctx, hctx_idx);
1644 blk_mq_free_bitmap(&hctx->ctx_map);
1647 unregister_cpu_notifier:
1648 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1653 static int blk_mq_init_hw_queues(struct request_queue *q,
1654 struct blk_mq_tag_set *set)
1656 struct blk_mq_hw_ctx *hctx;
1660 * Initialize hardware queues
1662 queue_for_each_hw_ctx(q, hctx, i) {
1663 if (blk_mq_init_hctx(q, set, hctx, i))
1667 if (i == q->nr_hw_queues)
1673 blk_mq_exit_hw_queues(q, set, i);
1678 static void blk_mq_init_cpu_queues(struct request_queue *q,
1679 unsigned int nr_hw_queues)
1683 for_each_possible_cpu(i) {
1684 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1685 struct blk_mq_hw_ctx *hctx;
1687 memset(__ctx, 0, sizeof(*__ctx));
1689 spin_lock_init(&__ctx->lock);
1690 INIT_LIST_HEAD(&__ctx->rq_list);
1693 /* If the cpu isn't online, the cpu is mapped to first hctx */
1697 hctx = q->mq_ops->map_queue(q, i);
1698 cpumask_set_cpu(i, hctx->cpumask);
1702 * Set local node, IFF we have more than one hw queue. If
1703 * not, we remain on the home node of the device
1705 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1706 hctx->numa_node = cpu_to_node(i);
1710 static void blk_mq_map_swqueue(struct request_queue *q)
1713 struct blk_mq_hw_ctx *hctx;
1714 struct blk_mq_ctx *ctx;
1716 queue_for_each_hw_ctx(q, hctx, i) {
1717 cpumask_clear(hctx->cpumask);
1722 * Map software to hardware queues
1724 queue_for_each_ctx(q, ctx, i) {
1725 /* If the cpu isn't online, the cpu is mapped to first hctx */
1729 hctx = q->mq_ops->map_queue(q, i);
1730 cpumask_set_cpu(i, hctx->cpumask);
1731 ctx->index_hw = hctx->nr_ctx;
1732 hctx->ctxs[hctx->nr_ctx++] = ctx;
1735 queue_for_each_hw_ctx(q, hctx, i) {
1737 * If no software queues are mapped to this hardware queue,
1738 * disable it and free the request entries.
1740 if (!hctx->nr_ctx) {
1741 struct blk_mq_tag_set *set = q->tag_set;
1744 blk_mq_free_rq_map(set, set->tags[i], i);
1745 set->tags[i] = NULL;
1752 * Initialize batch roundrobin counts
1754 hctx->next_cpu = cpumask_first(hctx->cpumask);
1755 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1759 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1761 struct blk_mq_hw_ctx *hctx;
1762 struct request_queue *q;
1766 if (set->tag_list.next == set->tag_list.prev)
1771 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1772 blk_mq_freeze_queue(q);
1774 queue_for_each_hw_ctx(q, hctx, i) {
1776 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1778 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1780 blk_mq_unfreeze_queue(q);
1784 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1786 struct blk_mq_tag_set *set = q->tag_set;
1788 mutex_lock(&set->tag_list_lock);
1789 list_del_init(&q->tag_set_list);
1790 blk_mq_update_tag_set_depth(set);
1791 mutex_unlock(&set->tag_list_lock);
1794 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1795 struct request_queue *q)
1799 mutex_lock(&set->tag_list_lock);
1800 list_add_tail(&q->tag_set_list, &set->tag_list);
1801 blk_mq_update_tag_set_depth(set);
1802 mutex_unlock(&set->tag_list_lock);
1805 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1807 struct blk_mq_hw_ctx **hctxs;
1808 struct blk_mq_ctx __percpu *ctx;
1809 struct request_queue *q;
1813 ctx = alloc_percpu(struct blk_mq_ctx);
1815 return ERR_PTR(-ENOMEM);
1818 * If a crashdump is active, then we are potentially in a very
1819 * memory constrained environment. Limit us to 1 queue and
1820 * 64 tags to prevent using too much memory.
1822 if (is_kdump_kernel()) {
1823 set->nr_hw_queues = 1;
1824 set->queue_depth = min(64U, set->queue_depth);
1827 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1833 map = blk_mq_make_queue_map(set);
1837 for (i = 0; i < set->nr_hw_queues; i++) {
1838 int node = blk_mq_hw_queue_to_node(map, i);
1840 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1845 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1849 atomic_set(&hctxs[i]->nr_active, 0);
1850 hctxs[i]->numa_node = node;
1851 hctxs[i]->queue_num = i;
1854 q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1859 * Init percpu_ref in atomic mode so that it's faster to shutdown.
1860 * See blk_register_queue() for details.
1862 if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release,
1863 PERCPU_REF_INIT_ATOMIC, GFP_KERNEL))
1866 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1867 blk_queue_rq_timeout(q, 30000);
1869 q->nr_queues = nr_cpu_ids;
1870 q->nr_hw_queues = set->nr_hw_queues;
1874 q->queue_hw_ctx = hctxs;
1876 q->mq_ops = set->ops;
1877 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1879 if (!(set->flags & BLK_MQ_F_SG_MERGE))
1880 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1882 q->sg_reserved_size = INT_MAX;
1884 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1885 INIT_LIST_HEAD(&q->requeue_list);
1886 spin_lock_init(&q->requeue_lock);
1888 if (q->nr_hw_queues > 1)
1889 blk_queue_make_request(q, blk_mq_make_request);
1891 blk_queue_make_request(q, blk_sq_make_request);
1894 blk_queue_rq_timeout(q, set->timeout);
1897 * Do this after blk_queue_make_request() overrides it...
1899 q->nr_requests = set->queue_depth;
1901 if (set->ops->complete)
1902 blk_queue_softirq_done(q, set->ops->complete);
1904 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1906 if (blk_mq_init_hw_queues(q, set))
1909 mutex_lock(&all_q_mutex);
1910 list_add_tail(&q->all_q_node, &all_q_list);
1911 mutex_unlock(&all_q_mutex);
1913 blk_mq_add_queue_tag_set(set, q);
1915 blk_mq_map_swqueue(q);
1920 blk_cleanup_queue(q);
1923 for (i = 0; i < set->nr_hw_queues; i++) {
1926 free_cpumask_var(hctxs[i]->cpumask);
1933 return ERR_PTR(-ENOMEM);
1935 EXPORT_SYMBOL(blk_mq_init_queue);
1937 void blk_mq_free_queue(struct request_queue *q)
1939 struct blk_mq_tag_set *set = q->tag_set;
1941 blk_mq_del_queue_tag_set(q);
1943 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
1944 blk_mq_free_hw_queues(q, set);
1946 percpu_ref_exit(&q->mq_usage_counter);
1948 free_percpu(q->queue_ctx);
1949 kfree(q->queue_hw_ctx);
1952 q->queue_ctx = NULL;
1953 q->queue_hw_ctx = NULL;
1956 mutex_lock(&all_q_mutex);
1957 list_del_init(&q->all_q_node);
1958 mutex_unlock(&all_q_mutex);
1961 /* Basically redo blk_mq_init_queue with queue frozen */
1962 static void blk_mq_queue_reinit(struct request_queue *q)
1964 WARN_ON_ONCE(!q->mq_freeze_depth);
1966 blk_mq_sysfs_unregister(q);
1968 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1971 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1972 * we should change hctx numa_node according to new topology (this
1973 * involves free and re-allocate memory, worthy doing?)
1976 blk_mq_map_swqueue(q);
1978 blk_mq_sysfs_register(q);
1981 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1982 unsigned long action, void *hcpu)
1984 struct request_queue *q;
1987 * Before new mappings are established, hotadded cpu might already
1988 * start handling requests. This doesn't break anything as we map
1989 * offline CPUs to first hardware queue. We will re-init the queue
1990 * below to get optimal settings.
1992 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1993 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1996 mutex_lock(&all_q_mutex);
1999 * We need to freeze and reinit all existing queues. Freezing
2000 * involves synchronous wait for an RCU grace period and doing it
2001 * one by one may take a long time. Start freezing all queues in
2002 * one swoop and then wait for the completions so that freezing can
2003 * take place in parallel.
2005 list_for_each_entry(q, &all_q_list, all_q_node)
2006 blk_mq_freeze_queue_start(q);
2007 list_for_each_entry(q, &all_q_list, all_q_node)
2008 blk_mq_freeze_queue_wait(q);
2010 list_for_each_entry(q, &all_q_list, all_q_node)
2011 blk_mq_queue_reinit(q);
2013 list_for_each_entry(q, &all_q_list, all_q_node)
2014 blk_mq_unfreeze_queue(q);
2016 mutex_unlock(&all_q_mutex);
2020 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2024 for (i = 0; i < set->nr_hw_queues; i++) {
2025 set->tags[i] = blk_mq_init_rq_map(set, i);
2034 blk_mq_free_rq_map(set, set->tags[i], i);
2040 * Allocate the request maps associated with this tag_set. Note that this
2041 * may reduce the depth asked for, if memory is tight. set->queue_depth
2042 * will be updated to reflect the allocated depth.
2044 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2049 depth = set->queue_depth;
2051 err = __blk_mq_alloc_rq_maps(set);
2055 set->queue_depth >>= 1;
2056 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2060 } while (set->queue_depth);
2062 if (!set->queue_depth || err) {
2063 pr_err("blk-mq: failed to allocate request map\n");
2067 if (depth != set->queue_depth)
2068 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2069 depth, set->queue_depth);
2075 * Alloc a tag set to be associated with one or more request queues.
2076 * May fail with EINVAL for various error conditions. May adjust the
2077 * requested depth down, if if it too large. In that case, the set
2078 * value will be stored in set->queue_depth.
2080 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2082 if (!set->nr_hw_queues)
2084 if (!set->queue_depth)
2086 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2089 if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue)
2092 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2093 pr_info("blk-mq: reduced tag depth to %u\n",
2095 set->queue_depth = BLK_MQ_MAX_DEPTH;
2098 set->tags = kmalloc_node(set->nr_hw_queues *
2099 sizeof(struct blk_mq_tags *),
2100 GFP_KERNEL, set->numa_node);
2104 if (blk_mq_alloc_rq_maps(set))
2107 mutex_init(&set->tag_list_lock);
2108 INIT_LIST_HEAD(&set->tag_list);
2116 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2118 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2122 for (i = 0; i < set->nr_hw_queues; i++) {
2124 blk_mq_free_rq_map(set, set->tags[i], i);
2130 EXPORT_SYMBOL(blk_mq_free_tag_set);
2132 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2134 struct blk_mq_tag_set *set = q->tag_set;
2135 struct blk_mq_hw_ctx *hctx;
2138 if (!set || nr > set->queue_depth)
2142 queue_for_each_hw_ctx(q, hctx, i) {
2143 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2149 q->nr_requests = nr;
2154 void blk_mq_disable_hotplug(void)
2156 mutex_lock(&all_q_mutex);
2159 void blk_mq_enable_hotplug(void)
2161 mutex_unlock(&all_q_mutex);
2164 static int __init blk_mq_init(void)
2168 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2172 subsys_initcall(blk_mq_init);