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, gfp_t gfp)
85 if (percpu_ref_tryget_live(&q->mq_usage_counter))
88 if (!(gfp & __GFP_WAIT))
91 ret = wait_event_interruptible(q->mq_freeze_wq,
92 !q->mq_freeze_depth || blk_queue_dying(q));
93 if (blk_queue_dying(q))
100 static void blk_mq_queue_exit(struct request_queue *q)
102 percpu_ref_put(&q->mq_usage_counter);
105 static void blk_mq_usage_counter_release(struct percpu_ref *ref)
107 struct request_queue *q =
108 container_of(ref, struct request_queue, mq_usage_counter);
110 wake_up_all(&q->mq_freeze_wq);
113 void blk_mq_freeze_queue_start(struct request_queue *q)
117 spin_lock_irq(q->queue_lock);
118 freeze = !q->mq_freeze_depth++;
119 spin_unlock_irq(q->queue_lock);
122 percpu_ref_kill(&q->mq_usage_counter);
123 blk_mq_run_hw_queues(q, false);
126 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
128 static void blk_mq_freeze_queue_wait(struct request_queue *q)
130 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter));
134 * Guarantee no request is in use, so we can change any data structure of
135 * the queue afterward.
137 void blk_mq_freeze_queue(struct request_queue *q)
139 blk_mq_freeze_queue_start(q);
140 blk_mq_freeze_queue_wait(q);
142 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
144 void blk_mq_unfreeze_queue(struct request_queue *q)
148 spin_lock_irq(q->queue_lock);
149 wake = !--q->mq_freeze_depth;
150 WARN_ON_ONCE(q->mq_freeze_depth < 0);
151 spin_unlock_irq(q->queue_lock);
153 percpu_ref_reinit(&q->mq_usage_counter);
154 wake_up_all(&q->mq_freeze_wq);
157 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
159 void blk_mq_wake_waiters(struct request_queue *q)
161 struct blk_mq_hw_ctx *hctx;
164 queue_for_each_hw_ctx(q, hctx, i)
165 if (blk_mq_hw_queue_mapped(hctx))
166 blk_mq_tag_wakeup_all(hctx->tags, true);
169 * If we are called because the queue has now been marked as
170 * dying, we need to ensure that processes currently waiting on
171 * the queue are notified as well.
173 wake_up_all(&q->mq_freeze_wq);
176 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
178 return blk_mq_has_free_tags(hctx->tags);
180 EXPORT_SYMBOL(blk_mq_can_queue);
182 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
183 struct request *rq, unsigned int rw_flags)
185 if (blk_queue_io_stat(q))
186 rw_flags |= REQ_IO_STAT;
188 INIT_LIST_HEAD(&rq->queuelist);
189 /* csd/requeue_work/fifo_time is initialized before use */
192 rq->cmd_flags |= rw_flags;
193 /* do not touch atomic flags, it needs atomic ops against the timer */
195 INIT_HLIST_NODE(&rq->hash);
196 RB_CLEAR_NODE(&rq->rb_node);
199 rq->start_time = jiffies;
200 #ifdef CONFIG_BLK_CGROUP
202 set_start_time_ns(rq);
203 rq->io_start_time_ns = 0;
205 rq->nr_phys_segments = 0;
206 #if defined(CONFIG_BLK_DEV_INTEGRITY)
207 rq->nr_integrity_segments = 0;
210 /* tag was already set */
220 INIT_LIST_HEAD(&rq->timeout_list);
224 rq->end_io_data = NULL;
227 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
230 static struct request *
231 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
236 tag = blk_mq_get_tag(data);
237 if (tag != BLK_MQ_TAG_FAIL) {
238 rq = data->hctx->tags->rqs[tag];
240 if (blk_mq_tag_busy(data->hctx)) {
241 rq->cmd_flags = REQ_MQ_INFLIGHT;
242 atomic_inc(&data->hctx->nr_active);
246 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
253 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
256 struct blk_mq_ctx *ctx;
257 struct blk_mq_hw_ctx *hctx;
259 struct blk_mq_alloc_data alloc_data;
262 ret = blk_mq_queue_enter(q, gfp);
266 ctx = blk_mq_get_ctx(q);
267 hctx = q->mq_ops->map_queue(q, ctx->cpu);
268 blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
269 reserved, ctx, hctx);
271 rq = __blk_mq_alloc_request(&alloc_data, rw);
272 if (!rq && (gfp & __GFP_WAIT)) {
273 __blk_mq_run_hw_queue(hctx);
276 ctx = blk_mq_get_ctx(q);
277 hctx = q->mq_ops->map_queue(q, ctx->cpu);
278 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
280 rq = __blk_mq_alloc_request(&alloc_data, rw);
281 ctx = alloc_data.ctx;
285 blk_mq_queue_exit(q);
286 return ERR_PTR(-EWOULDBLOCK);
290 EXPORT_SYMBOL(blk_mq_alloc_request);
292 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
293 struct blk_mq_ctx *ctx, struct request *rq)
295 const int tag = rq->tag;
296 struct request_queue *q = rq->q;
298 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
299 atomic_dec(&hctx->nr_active);
302 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
303 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
304 blk_mq_queue_exit(q);
307 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
309 struct blk_mq_ctx *ctx = rq->mq_ctx;
311 ctx->rq_completed[rq_is_sync(rq)]++;
312 __blk_mq_free_request(hctx, ctx, rq);
315 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
317 void blk_mq_free_request(struct request *rq)
319 struct blk_mq_hw_ctx *hctx;
320 struct request_queue *q = rq->q;
322 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
323 blk_mq_free_hctx_request(hctx, rq);
325 EXPORT_SYMBOL_GPL(blk_mq_free_request);
327 inline void __blk_mq_end_request(struct request *rq, int error)
329 blk_account_io_done(rq);
332 rq->end_io(rq, error);
334 if (unlikely(blk_bidi_rq(rq)))
335 blk_mq_free_request(rq->next_rq);
336 blk_mq_free_request(rq);
339 EXPORT_SYMBOL(__blk_mq_end_request);
341 void blk_mq_end_request(struct request *rq, int error)
343 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
345 __blk_mq_end_request(rq, error);
347 EXPORT_SYMBOL(blk_mq_end_request);
349 static void __blk_mq_complete_request_remote(void *data)
351 struct request *rq = data;
353 rq->q->softirq_done_fn(rq);
356 static void blk_mq_ipi_complete_request(struct request *rq)
358 struct blk_mq_ctx *ctx = rq->mq_ctx;
362 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
363 rq->q->softirq_done_fn(rq);
368 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
369 shared = cpus_share_cache(cpu, ctx->cpu);
371 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
372 rq->csd.func = __blk_mq_complete_request_remote;
375 smp_call_function_single_async(ctx->cpu, &rq->csd);
377 rq->q->softirq_done_fn(rq);
382 void __blk_mq_complete_request(struct request *rq)
384 struct request_queue *q = rq->q;
386 if (!q->softirq_done_fn)
387 blk_mq_end_request(rq, rq->errors);
389 blk_mq_ipi_complete_request(rq);
393 * blk_mq_complete_request - end I/O on a request
394 * @rq: the request being processed
397 * Ends all I/O on a request. It does not handle partial completions.
398 * The actual completion happens out-of-order, through a IPI handler.
400 void blk_mq_complete_request(struct request *rq)
402 struct request_queue *q = rq->q;
404 if (unlikely(blk_should_fake_timeout(q)))
406 if (!blk_mark_rq_complete(rq))
407 __blk_mq_complete_request(rq);
409 EXPORT_SYMBOL(blk_mq_complete_request);
411 int blk_mq_request_started(struct request *rq)
413 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
415 EXPORT_SYMBOL_GPL(blk_mq_request_started);
417 void blk_mq_start_request(struct request *rq)
419 struct request_queue *q = rq->q;
421 trace_block_rq_issue(q, rq);
423 rq->resid_len = blk_rq_bytes(rq);
424 if (unlikely(blk_bidi_rq(rq)))
425 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
430 * Ensure that ->deadline is visible before set the started
431 * flag and clear the completed flag.
433 smp_mb__before_atomic();
436 * Mark us as started and clear complete. Complete might have been
437 * set if requeue raced with timeout, which then marked it as
438 * complete. So be sure to clear complete again when we start
439 * the request, otherwise we'll ignore the completion event.
441 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
442 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
443 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
444 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
446 if (q->dma_drain_size && blk_rq_bytes(rq)) {
448 * Make sure space for the drain appears. We know we can do
449 * this because max_hw_segments has been adjusted to be one
450 * fewer than the device can handle.
452 rq->nr_phys_segments++;
455 EXPORT_SYMBOL(blk_mq_start_request);
457 static void __blk_mq_requeue_request(struct request *rq)
459 struct request_queue *q = rq->q;
461 trace_block_rq_requeue(q, rq);
463 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
464 if (q->dma_drain_size && blk_rq_bytes(rq))
465 rq->nr_phys_segments--;
469 void blk_mq_requeue_request(struct request *rq)
471 __blk_mq_requeue_request(rq);
473 BUG_ON(blk_queued_rq(rq));
474 blk_mq_add_to_requeue_list(rq, true);
476 EXPORT_SYMBOL(blk_mq_requeue_request);
478 static void blk_mq_requeue_work(struct work_struct *work)
480 struct request_queue *q =
481 container_of(work, struct request_queue, requeue_work);
483 struct request *rq, *next;
486 spin_lock_irqsave(&q->requeue_lock, flags);
487 list_splice_init(&q->requeue_list, &rq_list);
488 spin_unlock_irqrestore(&q->requeue_lock, flags);
490 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
491 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
494 rq->cmd_flags &= ~REQ_SOFTBARRIER;
495 list_del_init(&rq->queuelist);
496 blk_mq_insert_request(rq, true, false, false);
499 while (!list_empty(&rq_list)) {
500 rq = list_entry(rq_list.next, struct request, queuelist);
501 list_del_init(&rq->queuelist);
502 blk_mq_insert_request(rq, false, false, false);
506 * Use the start variant of queue running here, so that running
507 * the requeue work will kick stopped queues.
509 blk_mq_start_hw_queues(q);
512 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
514 struct request_queue *q = rq->q;
518 * We abuse this flag that is otherwise used by the I/O scheduler to
519 * request head insertation from the workqueue.
521 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
523 spin_lock_irqsave(&q->requeue_lock, flags);
525 rq->cmd_flags |= REQ_SOFTBARRIER;
526 list_add(&rq->queuelist, &q->requeue_list);
528 list_add_tail(&rq->queuelist, &q->requeue_list);
530 spin_unlock_irqrestore(&q->requeue_lock, flags);
532 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
534 void blk_mq_cancel_requeue_work(struct request_queue *q)
536 cancel_work_sync(&q->requeue_work);
538 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
540 void blk_mq_kick_requeue_list(struct request_queue *q)
542 kblockd_schedule_work(&q->requeue_work);
544 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
546 void blk_mq_abort_requeue_list(struct request_queue *q)
551 spin_lock_irqsave(&q->requeue_lock, flags);
552 list_splice_init(&q->requeue_list, &rq_list);
553 spin_unlock_irqrestore(&q->requeue_lock, flags);
555 while (!list_empty(&rq_list)) {
558 rq = list_first_entry(&rq_list, struct request, queuelist);
559 list_del_init(&rq->queuelist);
561 blk_mq_end_request(rq, rq->errors);
564 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
566 static inline bool is_flush_request(struct request *rq,
567 struct blk_flush_queue *fq, unsigned int tag)
569 return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
570 fq->flush_rq->tag == tag);
573 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
575 struct request *rq = tags->rqs[tag];
576 /* mq_ctx of flush rq is always cloned from the corresponding req */
577 struct blk_flush_queue *fq = blk_get_flush_queue(rq->q, rq->mq_ctx);
579 if (!is_flush_request(rq, fq, tag))
584 EXPORT_SYMBOL(blk_mq_tag_to_rq);
586 struct blk_mq_timeout_data {
588 unsigned int next_set;
591 void blk_mq_rq_timed_out(struct request *req, bool reserved)
593 struct blk_mq_ops *ops = req->q->mq_ops;
594 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
597 * We know that complete is set at this point. If STARTED isn't set
598 * anymore, then the request isn't active and the "timeout" should
599 * just be ignored. This can happen due to the bitflag ordering.
600 * Timeout first checks if STARTED is set, and if it is, assumes
601 * the request is active. But if we race with completion, then
602 * we both flags will get cleared. So check here again, and ignore
603 * a timeout event with a request that isn't active.
605 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
609 ret = ops->timeout(req, reserved);
613 __blk_mq_complete_request(req);
615 case BLK_EH_RESET_TIMER:
617 blk_clear_rq_complete(req);
619 case BLK_EH_NOT_HANDLED:
622 printk(KERN_ERR "block: bad eh return: %d\n", ret);
627 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
628 struct request *rq, void *priv, bool reserved)
630 struct blk_mq_timeout_data *data = priv;
632 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
634 * If a request wasn't started before the queue was
635 * marked dying, kill it here or it'll go unnoticed.
637 if (unlikely(blk_queue_dying(rq->q))) {
639 blk_mq_complete_request(rq);
643 if (rq->cmd_flags & REQ_NO_TIMEOUT)
646 if (time_after_eq(jiffies, rq->deadline)) {
647 if (!blk_mark_rq_complete(rq))
648 blk_mq_rq_timed_out(rq, reserved);
649 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
650 data->next = rq->deadline;
655 static void blk_mq_rq_timer(unsigned long priv)
657 struct request_queue *q = (struct request_queue *)priv;
658 struct blk_mq_timeout_data data = {
662 struct blk_mq_hw_ctx *hctx;
665 queue_for_each_hw_ctx(q, hctx, i) {
667 * If not software queues are currently mapped to this
668 * hardware queue, there's nothing to check
670 if (!blk_mq_hw_queue_mapped(hctx))
673 blk_mq_tag_busy_iter(hctx, blk_mq_check_expired, &data);
677 data.next = blk_rq_timeout(round_jiffies_up(data.next));
678 mod_timer(&q->timeout, data.next);
680 queue_for_each_hw_ctx(q, hctx, i)
681 blk_mq_tag_idle(hctx);
686 * Reverse check our software queue for entries that we could potentially
687 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
688 * too much time checking for merges.
690 static bool blk_mq_attempt_merge(struct request_queue *q,
691 struct blk_mq_ctx *ctx, struct bio *bio)
696 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
702 if (!blk_rq_merge_ok(rq, bio))
705 el_ret = blk_try_merge(rq, bio);
706 if (el_ret == ELEVATOR_BACK_MERGE) {
707 if (bio_attempt_back_merge(q, rq, bio)) {
712 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
713 if (bio_attempt_front_merge(q, rq, bio)) {
725 * Process software queues that have been marked busy, splicing them
726 * to the for-dispatch
728 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
730 struct blk_mq_ctx *ctx;
733 for (i = 0; i < hctx->ctx_map.map_size; i++) {
734 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
735 unsigned int off, bit;
741 off = i * hctx->ctx_map.bits_per_word;
743 bit = find_next_bit(&bm->word, bm->depth, bit);
744 if (bit >= bm->depth)
747 ctx = hctx->ctxs[bit + off];
748 clear_bit(bit, &bm->word);
749 spin_lock(&ctx->lock);
750 list_splice_tail_init(&ctx->rq_list, list);
751 spin_unlock(&ctx->lock);
759 * Run this hardware queue, pulling any software queues mapped to it in.
760 * Note that this function currently has various problems around ordering
761 * of IO. In particular, we'd like FIFO behaviour on handling existing
762 * items on the hctx->dispatch list. Ignore that for now.
764 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
766 struct request_queue *q = hctx->queue;
769 LIST_HEAD(driver_list);
770 struct list_head *dptr;
773 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
775 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
781 * Touch any software queue that has pending entries.
783 flush_busy_ctxs(hctx, &rq_list);
786 * If we have previous entries on our dispatch list, grab them
787 * and stuff them at the front for more fair dispatch.
789 if (!list_empty_careful(&hctx->dispatch)) {
790 spin_lock(&hctx->lock);
791 if (!list_empty(&hctx->dispatch))
792 list_splice_init(&hctx->dispatch, &rq_list);
793 spin_unlock(&hctx->lock);
797 * Start off with dptr being NULL, so we start the first request
798 * immediately, even if we have more pending.
803 * Now process all the entries, sending them to the driver.
806 while (!list_empty(&rq_list)) {
807 struct blk_mq_queue_data bd;
810 rq = list_first_entry(&rq_list, struct request, queuelist);
811 list_del_init(&rq->queuelist);
815 bd.last = list_empty(&rq_list);
817 ret = q->mq_ops->queue_rq(hctx, &bd);
819 case BLK_MQ_RQ_QUEUE_OK:
822 case BLK_MQ_RQ_QUEUE_BUSY:
823 list_add(&rq->queuelist, &rq_list);
824 __blk_mq_requeue_request(rq);
827 pr_err("blk-mq: bad return on queue: %d\n", ret);
828 case BLK_MQ_RQ_QUEUE_ERROR:
830 blk_mq_end_request(rq, rq->errors);
834 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
838 * We've done the first request. If we have more than 1
839 * left in the list, set dptr to defer issue.
841 if (!dptr && rq_list.next != rq_list.prev)
846 hctx->dispatched[0]++;
847 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
848 hctx->dispatched[ilog2(queued) + 1]++;
851 * Any items that need requeuing? Stuff them into hctx->dispatch,
852 * that is where we will continue on next queue run.
854 if (!list_empty(&rq_list)) {
855 spin_lock(&hctx->lock);
856 list_splice(&rq_list, &hctx->dispatch);
857 spin_unlock(&hctx->lock);
862 * It'd be great if the workqueue API had a way to pass
863 * in a mask and had some smarts for more clever placement.
864 * For now we just round-robin here, switching for every
865 * BLK_MQ_CPU_WORK_BATCH queued items.
867 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
869 if (hctx->queue->nr_hw_queues == 1)
870 return WORK_CPU_UNBOUND;
872 if (--hctx->next_cpu_batch <= 0) {
873 int cpu = hctx->next_cpu, next_cpu;
875 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
876 if (next_cpu >= nr_cpu_ids)
877 next_cpu = cpumask_first(hctx->cpumask);
879 hctx->next_cpu = next_cpu;
880 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
885 return hctx->next_cpu;
888 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
890 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
891 !blk_mq_hw_queue_mapped(hctx)))
896 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
897 __blk_mq_run_hw_queue(hctx);
905 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
909 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
911 struct blk_mq_hw_ctx *hctx;
914 queue_for_each_hw_ctx(q, hctx, i) {
915 if ((!blk_mq_hctx_has_pending(hctx) &&
916 list_empty_careful(&hctx->dispatch)) ||
917 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
920 blk_mq_run_hw_queue(hctx, async);
923 EXPORT_SYMBOL(blk_mq_run_hw_queues);
925 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
927 cancel_delayed_work(&hctx->run_work);
928 cancel_delayed_work(&hctx->delay_work);
929 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
931 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
933 void blk_mq_stop_hw_queues(struct request_queue *q)
935 struct blk_mq_hw_ctx *hctx;
938 queue_for_each_hw_ctx(q, hctx, i)
939 blk_mq_stop_hw_queue(hctx);
941 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
943 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
945 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
947 blk_mq_run_hw_queue(hctx, false);
949 EXPORT_SYMBOL(blk_mq_start_hw_queue);
951 void blk_mq_start_hw_queues(struct request_queue *q)
953 struct blk_mq_hw_ctx *hctx;
956 queue_for_each_hw_ctx(q, hctx, i)
957 blk_mq_start_hw_queue(hctx);
959 EXPORT_SYMBOL(blk_mq_start_hw_queues);
961 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
963 struct blk_mq_hw_ctx *hctx;
966 queue_for_each_hw_ctx(q, hctx, i) {
967 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
970 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
971 blk_mq_run_hw_queue(hctx, async);
974 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
976 static void blk_mq_run_work_fn(struct work_struct *work)
978 struct blk_mq_hw_ctx *hctx;
980 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
982 __blk_mq_run_hw_queue(hctx);
985 static void blk_mq_delay_work_fn(struct work_struct *work)
987 struct blk_mq_hw_ctx *hctx;
989 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
991 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
992 __blk_mq_run_hw_queue(hctx);
995 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
997 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1000 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1001 &hctx->delay_work, msecs_to_jiffies(msecs));
1003 EXPORT_SYMBOL(blk_mq_delay_queue);
1005 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
1006 struct request *rq, bool at_head)
1008 struct blk_mq_ctx *ctx = rq->mq_ctx;
1010 trace_block_rq_insert(hctx->queue, rq);
1013 list_add(&rq->queuelist, &ctx->rq_list);
1015 list_add_tail(&rq->queuelist, &ctx->rq_list);
1017 blk_mq_hctx_mark_pending(hctx, ctx);
1020 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1023 struct request_queue *q = rq->q;
1024 struct blk_mq_hw_ctx *hctx;
1025 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
1027 current_ctx = blk_mq_get_ctx(q);
1028 if (!cpu_online(ctx->cpu))
1029 rq->mq_ctx = ctx = current_ctx;
1031 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1033 spin_lock(&ctx->lock);
1034 __blk_mq_insert_request(hctx, rq, at_head);
1035 spin_unlock(&ctx->lock);
1038 blk_mq_run_hw_queue(hctx, async);
1040 blk_mq_put_ctx(current_ctx);
1043 static void blk_mq_insert_requests(struct request_queue *q,
1044 struct blk_mq_ctx *ctx,
1045 struct list_head *list,
1050 struct blk_mq_hw_ctx *hctx;
1051 struct blk_mq_ctx *current_ctx;
1053 trace_block_unplug(q, depth, !from_schedule);
1055 current_ctx = blk_mq_get_ctx(q);
1057 if (!cpu_online(ctx->cpu))
1059 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1062 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1065 spin_lock(&ctx->lock);
1066 while (!list_empty(list)) {
1069 rq = list_first_entry(list, struct request, queuelist);
1070 list_del_init(&rq->queuelist);
1072 __blk_mq_insert_request(hctx, rq, false);
1074 spin_unlock(&ctx->lock);
1076 blk_mq_run_hw_queue(hctx, from_schedule);
1077 blk_mq_put_ctx(current_ctx);
1080 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1082 struct request *rqa = container_of(a, struct request, queuelist);
1083 struct request *rqb = container_of(b, struct request, queuelist);
1085 return !(rqa->mq_ctx < rqb->mq_ctx ||
1086 (rqa->mq_ctx == rqb->mq_ctx &&
1087 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1090 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1092 struct blk_mq_ctx *this_ctx;
1093 struct request_queue *this_q;
1096 LIST_HEAD(ctx_list);
1099 list_splice_init(&plug->mq_list, &list);
1101 list_sort(NULL, &list, plug_ctx_cmp);
1107 while (!list_empty(&list)) {
1108 rq = list_entry_rq(list.next);
1109 list_del_init(&rq->queuelist);
1111 if (rq->mq_ctx != this_ctx) {
1113 blk_mq_insert_requests(this_q, this_ctx,
1118 this_ctx = rq->mq_ctx;
1124 list_add_tail(&rq->queuelist, &ctx_list);
1128 * If 'this_ctx' is set, we know we have entries to complete
1129 * on 'ctx_list'. Do those.
1132 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1137 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1139 init_request_from_bio(rq, bio);
1141 if (blk_do_io_stat(rq))
1142 blk_account_io_start(rq, 1);
1145 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1147 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1148 !blk_queue_nomerges(hctx->queue);
1151 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1152 struct blk_mq_ctx *ctx,
1153 struct request *rq, struct bio *bio)
1155 if (!hctx_allow_merges(hctx)) {
1156 blk_mq_bio_to_request(rq, bio);
1157 spin_lock(&ctx->lock);
1159 __blk_mq_insert_request(hctx, rq, false);
1160 spin_unlock(&ctx->lock);
1163 struct request_queue *q = hctx->queue;
1165 spin_lock(&ctx->lock);
1166 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1167 blk_mq_bio_to_request(rq, bio);
1171 spin_unlock(&ctx->lock);
1172 __blk_mq_free_request(hctx, ctx, rq);
1177 struct blk_map_ctx {
1178 struct blk_mq_hw_ctx *hctx;
1179 struct blk_mq_ctx *ctx;
1182 static struct request *blk_mq_map_request(struct request_queue *q,
1184 struct blk_map_ctx *data)
1186 struct blk_mq_hw_ctx *hctx;
1187 struct blk_mq_ctx *ctx;
1189 int rw = bio_data_dir(bio);
1190 struct blk_mq_alloc_data alloc_data;
1192 if (unlikely(blk_mq_queue_enter(q, GFP_KERNEL))) {
1193 bio_endio(bio, -EIO);
1197 ctx = blk_mq_get_ctx(q);
1198 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1200 if (rw_is_sync(bio->bi_rw))
1203 trace_block_getrq(q, bio, rw);
1204 blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1206 rq = __blk_mq_alloc_request(&alloc_data, rw);
1207 if (unlikely(!rq)) {
1208 __blk_mq_run_hw_queue(hctx);
1209 blk_mq_put_ctx(ctx);
1210 trace_block_sleeprq(q, bio, rw);
1212 ctx = blk_mq_get_ctx(q);
1213 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1214 blk_mq_set_alloc_data(&alloc_data, q,
1215 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1216 rq = __blk_mq_alloc_request(&alloc_data, rw);
1217 ctx = alloc_data.ctx;
1218 hctx = alloc_data.hctx;
1228 * Multiple hardware queue variant. This will not use per-process plugs,
1229 * but will attempt to bypass the hctx queueing if we can go straight to
1230 * hardware for SYNC IO.
1232 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1234 const int is_sync = rw_is_sync(bio->bi_rw);
1235 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1236 struct blk_map_ctx data;
1239 blk_queue_bounce(q, &bio);
1241 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1242 bio_endio(bio, -EIO);
1246 rq = blk_mq_map_request(q, bio, &data);
1250 if (unlikely(is_flush_fua)) {
1251 blk_mq_bio_to_request(rq, bio);
1252 blk_insert_flush(rq);
1257 * If the driver supports defer issued based on 'last', then
1258 * queue it up like normal since we can potentially save some
1261 if (is_sync && !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1262 struct blk_mq_queue_data bd = {
1269 blk_mq_bio_to_request(rq, bio);
1272 * For OK queue, we are done. For error, kill it. Any other
1273 * error (busy), just add it to our list as we previously
1276 ret = q->mq_ops->queue_rq(data.hctx, &bd);
1277 if (ret == BLK_MQ_RQ_QUEUE_OK)
1280 __blk_mq_requeue_request(rq);
1282 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1284 blk_mq_end_request(rq, rq->errors);
1290 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1292 * For a SYNC request, send it to the hardware immediately. For
1293 * an ASYNC request, just ensure that we run it later on. The
1294 * latter allows for merging opportunities and more efficient
1298 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1301 blk_mq_put_ctx(data.ctx);
1305 * Single hardware queue variant. This will attempt to use any per-process
1306 * plug for merging and IO deferral.
1308 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1310 const int is_sync = rw_is_sync(bio->bi_rw);
1311 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1312 unsigned int use_plug, request_count = 0;
1313 struct blk_map_ctx data;
1317 * If we have multiple hardware queues, just go directly to
1318 * one of those for sync IO.
1320 use_plug = !is_flush_fua && !is_sync;
1322 blk_queue_bounce(q, &bio);
1324 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1325 bio_endio(bio, -EIO);
1329 if (use_plug && !blk_queue_nomerges(q) &&
1330 blk_attempt_plug_merge(q, bio, &request_count))
1333 rq = blk_mq_map_request(q, bio, &data);
1337 if (unlikely(is_flush_fua)) {
1338 blk_mq_bio_to_request(rq, bio);
1339 blk_insert_flush(rq);
1344 * A task plug currently exists. Since this is completely lockless,
1345 * utilize that to temporarily store requests until the task is
1346 * either done or scheduled away.
1349 struct blk_plug *plug = current->plug;
1352 blk_mq_bio_to_request(rq, bio);
1353 if (list_empty(&plug->mq_list))
1354 trace_block_plug(q);
1355 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1356 blk_flush_plug_list(plug, false);
1357 trace_block_plug(q);
1359 list_add_tail(&rq->queuelist, &plug->mq_list);
1360 blk_mq_put_ctx(data.ctx);
1365 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1367 * For a SYNC request, send it to the hardware immediately. For
1368 * an ASYNC request, just ensure that we run it later on. The
1369 * latter allows for merging opportunities and more efficient
1373 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1376 blk_mq_put_ctx(data.ctx);
1380 * Default mapping to a software queue, since we use one per CPU.
1382 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1384 return q->queue_hw_ctx[q->mq_map[cpu]];
1386 EXPORT_SYMBOL(blk_mq_map_queue);
1388 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1389 struct blk_mq_tags *tags, unsigned int hctx_idx)
1393 if (tags->rqs && set->ops->exit_request) {
1396 for (i = 0; i < tags->nr_tags; i++) {
1399 set->ops->exit_request(set->driver_data, tags->rqs[i],
1401 tags->rqs[i] = NULL;
1405 while (!list_empty(&tags->page_list)) {
1406 page = list_first_entry(&tags->page_list, struct page, lru);
1407 list_del_init(&page->lru);
1408 __free_pages(page, page->private);
1413 blk_mq_free_tags(tags);
1416 static size_t order_to_size(unsigned int order)
1418 return (size_t)PAGE_SIZE << order;
1421 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1422 unsigned int hctx_idx)
1424 struct blk_mq_tags *tags;
1425 unsigned int i, j, entries_per_page, max_order = 4;
1426 size_t rq_size, left;
1428 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1430 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1434 INIT_LIST_HEAD(&tags->page_list);
1436 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1437 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1440 blk_mq_free_tags(tags);
1445 * rq_size is the size of the request plus driver payload, rounded
1446 * to the cacheline size
1448 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1450 left = rq_size * set->queue_depth;
1452 for (i = 0; i < set->queue_depth; ) {
1453 int this_order = max_order;
1458 while (left < order_to_size(this_order - 1) && this_order)
1462 page = alloc_pages_node(set->numa_node,
1463 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1469 if (order_to_size(this_order) < rq_size)
1476 page->private = this_order;
1477 list_add_tail(&page->lru, &tags->page_list);
1479 p = page_address(page);
1480 entries_per_page = order_to_size(this_order) / rq_size;
1481 to_do = min(entries_per_page, set->queue_depth - i);
1482 left -= to_do * rq_size;
1483 for (j = 0; j < to_do; j++) {
1485 tags->rqs[i]->atomic_flags = 0;
1486 tags->rqs[i]->cmd_flags = 0;
1487 if (set->ops->init_request) {
1488 if (set->ops->init_request(set->driver_data,
1489 tags->rqs[i], hctx_idx, i,
1491 tags->rqs[i] = NULL;
1504 blk_mq_free_rq_map(set, tags, hctx_idx);
1508 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1513 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1515 unsigned int bpw = 8, total, num_maps, i;
1517 bitmap->bits_per_word = bpw;
1519 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1520 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1526 for (i = 0; i < num_maps; i++) {
1527 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1528 total -= bitmap->map[i].depth;
1534 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1536 struct request_queue *q = hctx->queue;
1537 struct blk_mq_ctx *ctx;
1541 * Move ctx entries to new CPU, if this one is going away.
1543 ctx = __blk_mq_get_ctx(q, cpu);
1545 spin_lock(&ctx->lock);
1546 if (!list_empty(&ctx->rq_list)) {
1547 list_splice_init(&ctx->rq_list, &tmp);
1548 blk_mq_hctx_clear_pending(hctx, ctx);
1550 spin_unlock(&ctx->lock);
1552 if (list_empty(&tmp))
1555 ctx = blk_mq_get_ctx(q);
1556 spin_lock(&ctx->lock);
1558 while (!list_empty(&tmp)) {
1561 rq = list_first_entry(&tmp, struct request, queuelist);
1563 list_move_tail(&rq->queuelist, &ctx->rq_list);
1566 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1567 blk_mq_hctx_mark_pending(hctx, ctx);
1569 spin_unlock(&ctx->lock);
1571 blk_mq_run_hw_queue(hctx, true);
1572 blk_mq_put_ctx(ctx);
1576 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1578 struct request_queue *q = hctx->queue;
1579 struct blk_mq_tag_set *set = q->tag_set;
1581 if (set->tags[hctx->queue_num])
1584 set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1585 if (!set->tags[hctx->queue_num])
1588 hctx->tags = set->tags[hctx->queue_num];
1592 static int blk_mq_hctx_notify(void *data, unsigned long action,
1595 struct blk_mq_hw_ctx *hctx = data;
1597 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1598 return blk_mq_hctx_cpu_offline(hctx, cpu);
1599 else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1600 return blk_mq_hctx_cpu_online(hctx, cpu);
1605 static void blk_mq_exit_hctx(struct request_queue *q,
1606 struct blk_mq_tag_set *set,
1607 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1609 unsigned flush_start_tag = set->queue_depth;
1611 blk_mq_tag_idle(hctx);
1613 if (set->ops->exit_request)
1614 set->ops->exit_request(set->driver_data,
1615 hctx->fq->flush_rq, hctx_idx,
1616 flush_start_tag + hctx_idx);
1618 if (set->ops->exit_hctx)
1619 set->ops->exit_hctx(hctx, hctx_idx);
1621 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1622 blk_free_flush_queue(hctx->fq);
1624 blk_mq_free_bitmap(&hctx->ctx_map);
1627 static void blk_mq_exit_hw_queues(struct request_queue *q,
1628 struct blk_mq_tag_set *set, int nr_queue)
1630 struct blk_mq_hw_ctx *hctx;
1633 queue_for_each_hw_ctx(q, hctx, i) {
1636 blk_mq_exit_hctx(q, set, hctx, i);
1640 static void blk_mq_free_hw_queues(struct request_queue *q,
1641 struct blk_mq_tag_set *set)
1643 struct blk_mq_hw_ctx *hctx;
1646 queue_for_each_hw_ctx(q, hctx, i)
1647 free_cpumask_var(hctx->cpumask);
1650 static int blk_mq_init_hctx(struct request_queue *q,
1651 struct blk_mq_tag_set *set,
1652 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1655 unsigned flush_start_tag = set->queue_depth;
1657 node = hctx->numa_node;
1658 if (node == NUMA_NO_NODE)
1659 node = hctx->numa_node = set->numa_node;
1661 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1662 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1663 spin_lock_init(&hctx->lock);
1664 INIT_LIST_HEAD(&hctx->dispatch);
1666 hctx->queue_num = hctx_idx;
1667 hctx->flags = set->flags;
1669 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1670 blk_mq_hctx_notify, hctx);
1671 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1673 hctx->tags = set->tags[hctx_idx];
1676 * Allocate space for all possible cpus to avoid allocation at
1679 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1682 goto unregister_cpu_notifier;
1684 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1689 if (set->ops->init_hctx &&
1690 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1693 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1697 if (set->ops->init_request &&
1698 set->ops->init_request(set->driver_data,
1699 hctx->fq->flush_rq, hctx_idx,
1700 flush_start_tag + hctx_idx, node))
1708 if (set->ops->exit_hctx)
1709 set->ops->exit_hctx(hctx, hctx_idx);
1711 blk_mq_free_bitmap(&hctx->ctx_map);
1714 unregister_cpu_notifier:
1715 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1720 static int blk_mq_init_hw_queues(struct request_queue *q,
1721 struct blk_mq_tag_set *set)
1723 struct blk_mq_hw_ctx *hctx;
1727 * Initialize hardware queues
1729 queue_for_each_hw_ctx(q, hctx, i) {
1730 if (blk_mq_init_hctx(q, set, hctx, i))
1734 if (i == q->nr_hw_queues)
1740 blk_mq_exit_hw_queues(q, set, i);
1745 static void blk_mq_init_cpu_queues(struct request_queue *q,
1746 unsigned int nr_hw_queues)
1750 for_each_possible_cpu(i) {
1751 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1752 struct blk_mq_hw_ctx *hctx;
1754 memset(__ctx, 0, sizeof(*__ctx));
1756 spin_lock_init(&__ctx->lock);
1757 INIT_LIST_HEAD(&__ctx->rq_list);
1760 /* If the cpu isn't online, the cpu is mapped to first hctx */
1764 hctx = q->mq_ops->map_queue(q, i);
1767 * Set local node, IFF we have more than one hw queue. If
1768 * not, we remain on the home node of the device
1770 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1771 hctx->numa_node = cpu_to_node(i);
1775 static void blk_mq_map_swqueue(struct request_queue *q)
1778 struct blk_mq_hw_ctx *hctx;
1779 struct blk_mq_ctx *ctx;
1781 queue_for_each_hw_ctx(q, hctx, i) {
1782 cpumask_clear(hctx->cpumask);
1787 * Map software to hardware queues
1789 queue_for_each_ctx(q, ctx, i) {
1790 /* If the cpu isn't online, the cpu is mapped to first hctx */
1794 hctx = q->mq_ops->map_queue(q, i);
1795 cpumask_set_cpu(i, hctx->cpumask);
1796 ctx->index_hw = hctx->nr_ctx;
1797 hctx->ctxs[hctx->nr_ctx++] = ctx;
1800 queue_for_each_hw_ctx(q, hctx, i) {
1801 struct blk_mq_ctxmap *map = &hctx->ctx_map;
1804 * If no software queues are mapped to this hardware queue,
1805 * disable it and free the request entries.
1807 if (!hctx->nr_ctx) {
1808 struct blk_mq_tag_set *set = q->tag_set;
1811 blk_mq_free_rq_map(set, set->tags[i], i);
1812 set->tags[i] = NULL;
1819 * Set the map size to the number of mapped software queues.
1820 * This is more accurate and more efficient than looping
1821 * over all possibly mapped software queues.
1823 map->map_size = hctx->nr_ctx / map->bits_per_word;
1826 * Initialize batch roundrobin counts
1828 hctx->next_cpu = cpumask_first(hctx->cpumask);
1829 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1833 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1835 struct blk_mq_hw_ctx *hctx;
1836 struct request_queue *q;
1840 if (set->tag_list.next == set->tag_list.prev)
1845 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1846 blk_mq_freeze_queue(q);
1848 queue_for_each_hw_ctx(q, hctx, i) {
1850 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1852 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1854 blk_mq_unfreeze_queue(q);
1858 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1860 struct blk_mq_tag_set *set = q->tag_set;
1862 mutex_lock(&set->tag_list_lock);
1863 list_del_init(&q->tag_set_list);
1864 blk_mq_update_tag_set_depth(set);
1865 mutex_unlock(&set->tag_list_lock);
1868 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1869 struct request_queue *q)
1873 mutex_lock(&set->tag_list_lock);
1874 list_add_tail(&q->tag_set_list, &set->tag_list);
1875 blk_mq_update_tag_set_depth(set);
1876 mutex_unlock(&set->tag_list_lock);
1880 * It is the actual release handler for mq, but we do it from
1881 * request queue's release handler for avoiding use-after-free
1882 * and headache because q->mq_kobj shouldn't have been introduced,
1883 * but we can't group ctx/kctx kobj without it.
1885 void blk_mq_release(struct request_queue *q)
1887 struct blk_mq_hw_ctx *hctx;
1890 /* hctx kobj stays in hctx */
1891 queue_for_each_hw_ctx(q, hctx, i)
1894 kfree(q->queue_hw_ctx);
1896 /* ctx kobj stays in queue_ctx */
1897 free_percpu(q->queue_ctx);
1900 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1902 struct request_queue *uninit_q, *q;
1904 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1906 return ERR_PTR(-ENOMEM);
1908 q = blk_mq_init_allocated_queue(set, uninit_q);
1910 blk_cleanup_queue(uninit_q);
1914 EXPORT_SYMBOL(blk_mq_init_queue);
1916 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
1917 struct request_queue *q)
1919 struct blk_mq_hw_ctx **hctxs;
1920 struct blk_mq_ctx __percpu *ctx;
1924 ctx = alloc_percpu(struct blk_mq_ctx);
1926 return ERR_PTR(-ENOMEM);
1928 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1934 map = blk_mq_make_queue_map(set);
1938 for (i = 0; i < set->nr_hw_queues; i++) {
1939 int node = blk_mq_hw_queue_to_node(map, i);
1941 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1946 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1950 atomic_set(&hctxs[i]->nr_active, 0);
1951 hctxs[i]->numa_node = node;
1952 hctxs[i]->queue_num = i;
1956 * Init percpu_ref in atomic mode so that it's faster to shutdown.
1957 * See blk_register_queue() for details.
1959 if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release,
1960 PERCPU_REF_INIT_ATOMIC, GFP_KERNEL))
1963 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1964 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30000);
1966 q->nr_queues = nr_cpu_ids;
1967 q->nr_hw_queues = set->nr_hw_queues;
1971 q->queue_hw_ctx = hctxs;
1973 q->mq_ops = set->ops;
1974 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1976 if (!(set->flags & BLK_MQ_F_SG_MERGE))
1977 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1979 q->sg_reserved_size = INT_MAX;
1981 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1982 INIT_LIST_HEAD(&q->requeue_list);
1983 spin_lock_init(&q->requeue_lock);
1985 if (q->nr_hw_queues > 1)
1986 blk_queue_make_request(q, blk_mq_make_request);
1988 blk_queue_make_request(q, blk_sq_make_request);
1991 * Do this after blk_queue_make_request() overrides it...
1993 q->nr_requests = set->queue_depth;
1995 if (set->ops->complete)
1996 blk_queue_softirq_done(q, set->ops->complete);
1998 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2000 if (blk_mq_init_hw_queues(q, set))
2003 mutex_lock(&all_q_mutex);
2004 list_add_tail(&q->all_q_node, &all_q_list);
2005 mutex_unlock(&all_q_mutex);
2007 blk_mq_add_queue_tag_set(set, q);
2009 blk_mq_map_swqueue(q);
2015 for (i = 0; i < set->nr_hw_queues; i++) {
2018 free_cpumask_var(hctxs[i]->cpumask);
2025 return ERR_PTR(-ENOMEM);
2027 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2029 void blk_mq_free_queue(struct request_queue *q)
2031 struct blk_mq_tag_set *set = q->tag_set;
2033 blk_mq_del_queue_tag_set(q);
2035 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2036 blk_mq_free_hw_queues(q, set);
2038 percpu_ref_exit(&q->mq_usage_counter);
2044 mutex_lock(&all_q_mutex);
2045 list_del_init(&q->all_q_node);
2046 mutex_unlock(&all_q_mutex);
2049 /* Basically redo blk_mq_init_queue with queue frozen */
2050 static void blk_mq_queue_reinit(struct request_queue *q)
2052 WARN_ON_ONCE(!q->mq_freeze_depth);
2054 blk_mq_sysfs_unregister(q);
2056 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
2059 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2060 * we should change hctx numa_node according to new topology (this
2061 * involves free and re-allocate memory, worthy doing?)
2064 blk_mq_map_swqueue(q);
2066 blk_mq_sysfs_register(q);
2069 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2070 unsigned long action, void *hcpu)
2072 struct request_queue *q;
2075 * Before new mappings are established, hotadded cpu might already
2076 * start handling requests. This doesn't break anything as we map
2077 * offline CPUs to first hardware queue. We will re-init the queue
2078 * below to get optimal settings.
2080 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
2081 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
2084 mutex_lock(&all_q_mutex);
2087 * We need to freeze and reinit all existing queues. Freezing
2088 * involves synchronous wait for an RCU grace period and doing it
2089 * one by one may take a long time. Start freezing all queues in
2090 * one swoop and then wait for the completions so that freezing can
2091 * take place in parallel.
2093 list_for_each_entry(q, &all_q_list, all_q_node)
2094 blk_mq_freeze_queue_start(q);
2095 list_for_each_entry(q, &all_q_list, all_q_node)
2096 blk_mq_freeze_queue_wait(q);
2098 list_for_each_entry(q, &all_q_list, all_q_node)
2099 blk_mq_queue_reinit(q);
2101 list_for_each_entry(q, &all_q_list, all_q_node)
2102 blk_mq_unfreeze_queue(q);
2104 mutex_unlock(&all_q_mutex);
2108 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2112 for (i = 0; i < set->nr_hw_queues; i++) {
2113 set->tags[i] = blk_mq_init_rq_map(set, i);
2122 blk_mq_free_rq_map(set, set->tags[i], i);
2128 * Allocate the request maps associated with this tag_set. Note that this
2129 * may reduce the depth asked for, if memory is tight. set->queue_depth
2130 * will be updated to reflect the allocated depth.
2132 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2137 depth = set->queue_depth;
2139 err = __blk_mq_alloc_rq_maps(set);
2143 set->queue_depth >>= 1;
2144 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2148 } while (set->queue_depth);
2150 if (!set->queue_depth || err) {
2151 pr_err("blk-mq: failed to allocate request map\n");
2155 if (depth != set->queue_depth)
2156 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2157 depth, set->queue_depth);
2163 * Alloc a tag set to be associated with one or more request queues.
2164 * May fail with EINVAL for various error conditions. May adjust the
2165 * requested depth down, if if it too large. In that case, the set
2166 * value will be stored in set->queue_depth.
2168 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2170 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2172 if (!set->nr_hw_queues)
2174 if (!set->queue_depth)
2176 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2179 if (!set->ops->queue_rq || !set->ops->map_queue)
2182 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2183 pr_info("blk-mq: reduced tag depth to %u\n",
2185 set->queue_depth = BLK_MQ_MAX_DEPTH;
2189 * If a crashdump is active, then we are potentially in a very
2190 * memory constrained environment. Limit us to 1 queue and
2191 * 64 tags to prevent using too much memory.
2193 if (is_kdump_kernel()) {
2194 set->nr_hw_queues = 1;
2195 set->queue_depth = min(64U, set->queue_depth);
2198 set->tags = kmalloc_node(set->nr_hw_queues *
2199 sizeof(struct blk_mq_tags *),
2200 GFP_KERNEL, set->numa_node);
2204 if (blk_mq_alloc_rq_maps(set))
2207 mutex_init(&set->tag_list_lock);
2208 INIT_LIST_HEAD(&set->tag_list);
2216 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2218 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2222 for (i = 0; i < set->nr_hw_queues; i++) {
2224 blk_mq_free_rq_map(set, set->tags[i], i);
2230 EXPORT_SYMBOL(blk_mq_free_tag_set);
2232 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2234 struct blk_mq_tag_set *set = q->tag_set;
2235 struct blk_mq_hw_ctx *hctx;
2238 if (!set || nr > set->queue_depth)
2242 queue_for_each_hw_ctx(q, hctx, i) {
2243 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2249 q->nr_requests = nr;
2254 void blk_mq_disable_hotplug(void)
2256 mutex_lock(&all_q_mutex);
2259 void blk_mq_enable_hotplug(void)
2261 mutex_unlock(&all_q_mutex);
2264 static int __init blk_mq_init(void)
2268 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2272 subsys_initcall(blk_mq_init);