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>
24 #include <trace/events/block.h>
26 #include <linux/blk-mq.h>
29 #include "blk-mq-tag.h"
31 static DEFINE_MUTEX(all_q_mutex);
32 static LIST_HEAD(all_q_list);
34 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
37 * Check if any of the ctx's have pending work in this hardware queue
39 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
43 for (i = 0; i < hctx->ctx_map.map_size; i++)
44 if (hctx->ctx_map.map[i].word)
50 static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx,
51 struct blk_mq_ctx *ctx)
53 return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word];
56 #define CTX_TO_BIT(hctx, ctx) \
57 ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
60 * Mark this ctx as having pending work in this hardware queue
62 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
63 struct blk_mq_ctx *ctx)
65 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
67 if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word))
68 set_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
71 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
72 struct blk_mq_ctx *ctx)
74 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
76 clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
79 static int blk_mq_queue_enter(struct request_queue *q)
84 if (percpu_ref_tryget_live(&q->mq_usage_counter))
87 ret = wait_event_interruptible(q->mq_freeze_wq,
88 !q->mq_freeze_depth || blk_queue_dying(q));
89 if (blk_queue_dying(q))
96 static void blk_mq_queue_exit(struct request_queue *q)
98 percpu_ref_put(&q->mq_usage_counter);
101 static void blk_mq_usage_counter_release(struct percpu_ref *ref)
103 struct request_queue *q =
104 container_of(ref, struct request_queue, mq_usage_counter);
106 wake_up_all(&q->mq_freeze_wq);
110 * Guarantee no request is in use, so we can change any data structure of
111 * the queue afterward.
113 void blk_mq_freeze_queue(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_queues(q, false);
125 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter));
128 static void blk_mq_unfreeze_queue(struct request_queue *q)
132 spin_lock_irq(q->queue_lock);
133 wake = !--q->mq_freeze_depth;
134 WARN_ON_ONCE(q->mq_freeze_depth < 0);
135 spin_unlock_irq(q->queue_lock);
137 percpu_ref_reinit(&q->mq_usage_counter);
138 wake_up_all(&q->mq_freeze_wq);
142 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
144 return blk_mq_has_free_tags(hctx->tags);
146 EXPORT_SYMBOL(blk_mq_can_queue);
148 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
149 struct request *rq, unsigned int rw_flags)
151 if (blk_queue_io_stat(q))
152 rw_flags |= REQ_IO_STAT;
154 INIT_LIST_HEAD(&rq->queuelist);
155 /* csd/requeue_work/fifo_time is initialized before use */
158 rq->cmd_flags |= rw_flags;
159 /* do not touch atomic flags, it needs atomic ops against the timer */
161 INIT_HLIST_NODE(&rq->hash);
162 RB_CLEAR_NODE(&rq->rb_node);
165 rq->start_time = jiffies;
166 #ifdef CONFIG_BLK_CGROUP
168 set_start_time_ns(rq);
169 rq->io_start_time_ns = 0;
171 rq->nr_phys_segments = 0;
172 #if defined(CONFIG_BLK_DEV_INTEGRITY)
173 rq->nr_integrity_segments = 0;
176 /* tag was already set */
186 INIT_LIST_HEAD(&rq->timeout_list);
190 rq->end_io_data = NULL;
193 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
196 static struct request *
197 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
202 tag = blk_mq_get_tag(data);
203 if (tag != BLK_MQ_TAG_FAIL) {
204 rq = data->hctx->tags->rqs[tag];
206 if (blk_mq_tag_busy(data->hctx)) {
207 rq->cmd_flags = REQ_MQ_INFLIGHT;
208 atomic_inc(&data->hctx->nr_active);
212 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
219 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
222 struct blk_mq_ctx *ctx;
223 struct blk_mq_hw_ctx *hctx;
225 struct blk_mq_alloc_data alloc_data;
227 if (blk_mq_queue_enter(q))
230 ctx = blk_mq_get_ctx(q);
231 hctx = q->mq_ops->map_queue(q, ctx->cpu);
232 blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
233 reserved, ctx, hctx);
235 rq = __blk_mq_alloc_request(&alloc_data, rw);
236 if (!rq && (gfp & __GFP_WAIT)) {
237 __blk_mq_run_hw_queue(hctx);
240 ctx = blk_mq_get_ctx(q);
241 hctx = q->mq_ops->map_queue(q, ctx->cpu);
242 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
244 rq = __blk_mq_alloc_request(&alloc_data, rw);
245 ctx = alloc_data.ctx;
250 EXPORT_SYMBOL(blk_mq_alloc_request);
252 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
253 struct blk_mq_ctx *ctx, struct request *rq)
255 const int tag = rq->tag;
256 struct request_queue *q = rq->q;
258 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
259 atomic_dec(&hctx->nr_active);
262 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
263 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
264 blk_mq_queue_exit(q);
267 void blk_mq_free_request(struct request *rq)
269 struct blk_mq_ctx *ctx = rq->mq_ctx;
270 struct blk_mq_hw_ctx *hctx;
271 struct request_queue *q = rq->q;
273 ctx->rq_completed[rq_is_sync(rq)]++;
275 hctx = q->mq_ops->map_queue(q, ctx->cpu);
276 __blk_mq_free_request(hctx, ctx, rq);
280 * Clone all relevant state from a request that has been put on hold in
281 * the flush state machine into the preallocated flush request that hangs
282 * off the request queue.
284 * For a driver the flush request should be invisible, that's why we are
285 * impersonating the original request here.
287 void blk_mq_clone_flush_request(struct request *flush_rq,
288 struct request *orig_rq)
290 struct blk_mq_hw_ctx *hctx =
291 orig_rq->q->mq_ops->map_queue(orig_rq->q, orig_rq->mq_ctx->cpu);
293 flush_rq->mq_ctx = orig_rq->mq_ctx;
294 flush_rq->tag = orig_rq->tag;
295 memcpy(blk_mq_rq_to_pdu(flush_rq), blk_mq_rq_to_pdu(orig_rq),
299 inline void __blk_mq_end_io(struct request *rq, int error)
301 blk_account_io_done(rq);
304 rq->end_io(rq, error);
306 if (unlikely(blk_bidi_rq(rq)))
307 blk_mq_free_request(rq->next_rq);
308 blk_mq_free_request(rq);
311 EXPORT_SYMBOL(__blk_mq_end_io);
313 void blk_mq_end_io(struct request *rq, int error)
315 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
317 __blk_mq_end_io(rq, error);
319 EXPORT_SYMBOL(blk_mq_end_io);
321 static void __blk_mq_complete_request_remote(void *data)
323 struct request *rq = data;
325 rq->q->softirq_done_fn(rq);
328 static void blk_mq_ipi_complete_request(struct request *rq)
330 struct blk_mq_ctx *ctx = rq->mq_ctx;
334 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
335 rq->q->softirq_done_fn(rq);
340 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
341 shared = cpus_share_cache(cpu, ctx->cpu);
343 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
344 rq->csd.func = __blk_mq_complete_request_remote;
347 smp_call_function_single_async(ctx->cpu, &rq->csd);
349 rq->q->softirq_done_fn(rq);
354 void __blk_mq_complete_request(struct request *rq)
356 struct request_queue *q = rq->q;
358 if (!q->softirq_done_fn)
359 blk_mq_end_io(rq, rq->errors);
361 blk_mq_ipi_complete_request(rq);
365 * blk_mq_complete_request - end I/O on a request
366 * @rq: the request being processed
369 * Ends all I/O on a request. It does not handle partial completions.
370 * The actual completion happens out-of-order, through a IPI handler.
372 void blk_mq_complete_request(struct request *rq)
374 struct request_queue *q = rq->q;
376 if (unlikely(blk_should_fake_timeout(q)))
378 if (!blk_mark_rq_complete(rq))
379 __blk_mq_complete_request(rq);
381 EXPORT_SYMBOL(blk_mq_complete_request);
383 static void blk_mq_start_request(struct request *rq, bool last)
385 struct request_queue *q = rq->q;
387 trace_block_rq_issue(q, rq);
389 rq->resid_len = blk_rq_bytes(rq);
390 if (unlikely(blk_bidi_rq(rq)))
391 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
396 * Ensure that ->deadline is visible before set the started
397 * flag and clear the completed flag.
399 smp_mb__before_atomic();
402 * Mark us as started and clear complete. Complete might have been
403 * set if requeue raced with timeout, which then marked it as
404 * complete. So be sure to clear complete again when we start
405 * the request, otherwise we'll ignore the completion event.
407 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
408 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
409 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
410 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
412 if (q->dma_drain_size && blk_rq_bytes(rq)) {
414 * Make sure space for the drain appears. We know we can do
415 * this because max_hw_segments has been adjusted to be one
416 * fewer than the device can handle.
418 rq->nr_phys_segments++;
422 * Flag the last request in the series so that drivers know when IO
423 * should be kicked off, if they don't do it on a per-request basis.
425 * Note: the flag isn't the only condition drivers should do kick off.
426 * If drive is busy, the last request might not have the bit set.
429 rq->cmd_flags |= REQ_END;
432 static void __blk_mq_requeue_request(struct request *rq)
434 struct request_queue *q = rq->q;
436 trace_block_rq_requeue(q, rq);
437 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
439 rq->cmd_flags &= ~REQ_END;
441 if (q->dma_drain_size && blk_rq_bytes(rq))
442 rq->nr_phys_segments--;
445 void blk_mq_requeue_request(struct request *rq)
447 __blk_mq_requeue_request(rq);
448 blk_clear_rq_complete(rq);
450 BUG_ON(blk_queued_rq(rq));
451 blk_mq_add_to_requeue_list(rq, true);
453 EXPORT_SYMBOL(blk_mq_requeue_request);
455 static void blk_mq_requeue_work(struct work_struct *work)
457 struct request_queue *q =
458 container_of(work, struct request_queue, requeue_work);
460 struct request *rq, *next;
463 spin_lock_irqsave(&q->requeue_lock, flags);
464 list_splice_init(&q->requeue_list, &rq_list);
465 spin_unlock_irqrestore(&q->requeue_lock, flags);
467 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
468 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
471 rq->cmd_flags &= ~REQ_SOFTBARRIER;
472 list_del_init(&rq->queuelist);
473 blk_mq_insert_request(rq, true, false, false);
476 while (!list_empty(&rq_list)) {
477 rq = list_entry(rq_list.next, struct request, queuelist);
478 list_del_init(&rq->queuelist);
479 blk_mq_insert_request(rq, false, false, false);
483 * Use the start variant of queue running here, so that running
484 * the requeue work will kick stopped queues.
486 blk_mq_start_hw_queues(q);
489 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
491 struct request_queue *q = rq->q;
495 * We abuse this flag that is otherwise used by the I/O scheduler to
496 * request head insertation from the workqueue.
498 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
500 spin_lock_irqsave(&q->requeue_lock, flags);
502 rq->cmd_flags |= REQ_SOFTBARRIER;
503 list_add(&rq->queuelist, &q->requeue_list);
505 list_add_tail(&rq->queuelist, &q->requeue_list);
507 spin_unlock_irqrestore(&q->requeue_lock, flags);
509 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
511 void blk_mq_kick_requeue_list(struct request_queue *q)
513 kblockd_schedule_work(&q->requeue_work);
515 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
517 static inline bool is_flush_request(struct request *rq, unsigned int tag)
519 return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
520 rq->q->flush_rq->tag == tag);
523 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
525 struct request *rq = tags->rqs[tag];
527 if (!is_flush_request(rq, tag))
530 return rq->q->flush_rq;
532 EXPORT_SYMBOL(blk_mq_tag_to_rq);
534 struct blk_mq_timeout_data {
535 struct blk_mq_hw_ctx *hctx;
537 unsigned int *next_set;
540 static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
542 struct blk_mq_timeout_data *data = __data;
543 struct blk_mq_hw_ctx *hctx = data->hctx;
546 /* It may not be in flight yet (this is where
547 * the REQ_ATOMIC_STARTED flag comes in). The requests are
548 * statically allocated, so we know it's always safe to access the
549 * memory associated with a bit offset into ->rqs[].
555 tag = find_next_zero_bit(free_tags, hctx->tags->nr_tags, tag);
556 if (tag >= hctx->tags->nr_tags)
559 rq = blk_mq_tag_to_rq(hctx->tags, tag++);
560 if (rq->q != hctx->queue)
562 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
565 blk_rq_check_expired(rq, data->next, data->next_set);
569 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
571 unsigned int *next_set)
573 struct blk_mq_timeout_data data = {
576 .next_set = next_set,
580 * Ask the tagging code to iterate busy requests, so we can
581 * check them for timeout.
583 blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
586 static enum blk_eh_timer_return blk_mq_rq_timed_out(struct request *rq)
588 struct request_queue *q = rq->q;
591 * We know that complete is set at this point. If STARTED isn't set
592 * anymore, then the request isn't active and the "timeout" should
593 * just be ignored. This can happen due to the bitflag ordering.
594 * Timeout first checks if STARTED is set, and if it is, assumes
595 * the request is active. But if we race with completion, then
596 * we both flags will get cleared. So check here again, and ignore
597 * a timeout event with a request that isn't active.
599 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
600 return BLK_EH_NOT_HANDLED;
602 if (!q->mq_ops->timeout)
603 return BLK_EH_RESET_TIMER;
605 return q->mq_ops->timeout(rq);
608 static void blk_mq_rq_timer(unsigned long data)
610 struct request_queue *q = (struct request_queue *) data;
611 struct blk_mq_hw_ctx *hctx;
612 unsigned long next = 0;
615 queue_for_each_hw_ctx(q, hctx, i) {
617 * If not software queues are currently mapped to this
618 * hardware queue, there's nothing to check
620 if (!hctx->nr_ctx || !hctx->tags)
623 blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
627 next = blk_rq_timeout(round_jiffies_up(next));
628 mod_timer(&q->timeout, next);
630 queue_for_each_hw_ctx(q, hctx, i)
631 blk_mq_tag_idle(hctx);
636 * Reverse check our software queue for entries that we could potentially
637 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
638 * too much time checking for merges.
640 static bool blk_mq_attempt_merge(struct request_queue *q,
641 struct blk_mq_ctx *ctx, struct bio *bio)
646 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
652 if (!blk_rq_merge_ok(rq, bio))
655 el_ret = blk_try_merge(rq, bio);
656 if (el_ret == ELEVATOR_BACK_MERGE) {
657 if (bio_attempt_back_merge(q, rq, bio)) {
662 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
663 if (bio_attempt_front_merge(q, rq, bio)) {
675 * Process software queues that have been marked busy, splicing them
676 * to the for-dispatch
678 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
680 struct blk_mq_ctx *ctx;
683 for (i = 0; i < hctx->ctx_map.map_size; i++) {
684 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
685 unsigned int off, bit;
691 off = i * hctx->ctx_map.bits_per_word;
693 bit = find_next_bit(&bm->word, bm->depth, bit);
694 if (bit >= bm->depth)
697 ctx = hctx->ctxs[bit + off];
698 clear_bit(bit, &bm->word);
699 spin_lock(&ctx->lock);
700 list_splice_tail_init(&ctx->rq_list, list);
701 spin_unlock(&ctx->lock);
709 * Run this hardware queue, pulling any software queues mapped to it in.
710 * Note that this function currently has various problems around ordering
711 * of IO. In particular, we'd like FIFO behaviour on handling existing
712 * items on the hctx->dispatch list. Ignore that for now.
714 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
716 struct request_queue *q = hctx->queue;
721 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
723 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
729 * Touch any software queue that has pending entries.
731 flush_busy_ctxs(hctx, &rq_list);
734 * If we have previous entries on our dispatch list, grab them
735 * and stuff them at the front for more fair dispatch.
737 if (!list_empty_careful(&hctx->dispatch)) {
738 spin_lock(&hctx->lock);
739 if (!list_empty(&hctx->dispatch))
740 list_splice_init(&hctx->dispatch, &rq_list);
741 spin_unlock(&hctx->lock);
745 * Now process all the entries, sending them to the driver.
748 while (!list_empty(&rq_list)) {
751 rq = list_first_entry(&rq_list, struct request, queuelist);
752 list_del_init(&rq->queuelist);
754 blk_mq_start_request(rq, list_empty(&rq_list));
756 ret = q->mq_ops->queue_rq(hctx, rq);
758 case BLK_MQ_RQ_QUEUE_OK:
761 case BLK_MQ_RQ_QUEUE_BUSY:
762 list_add(&rq->queuelist, &rq_list);
763 __blk_mq_requeue_request(rq);
766 pr_err("blk-mq: bad return on queue: %d\n", ret);
767 case BLK_MQ_RQ_QUEUE_ERROR:
769 blk_mq_end_io(rq, rq->errors);
773 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
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 int cpu = hctx->next_cpu;
803 if (--hctx->next_cpu_batch <= 0) {
806 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
807 if (next_cpu >= nr_cpu_ids)
808 next_cpu = cpumask_first(hctx->cpumask);
810 hctx->next_cpu = next_cpu;
811 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
817 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
819 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
822 if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask))
823 __blk_mq_run_hw_queue(hctx);
824 else if (hctx->queue->nr_hw_queues == 1)
825 kblockd_schedule_delayed_work(&hctx->run_work, 0);
829 cpu = blk_mq_hctx_next_cpu(hctx);
830 kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0);
834 void blk_mq_run_queues(struct request_queue *q, bool async)
836 struct blk_mq_hw_ctx *hctx;
839 queue_for_each_hw_ctx(q, hctx, i) {
840 if ((!blk_mq_hctx_has_pending(hctx) &&
841 list_empty_careful(&hctx->dispatch)) ||
842 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
846 blk_mq_run_hw_queue(hctx, async);
850 EXPORT_SYMBOL(blk_mq_run_queues);
852 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
854 cancel_delayed_work(&hctx->run_work);
855 cancel_delayed_work(&hctx->delay_work);
856 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
858 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
860 void blk_mq_stop_hw_queues(struct request_queue *q)
862 struct blk_mq_hw_ctx *hctx;
865 queue_for_each_hw_ctx(q, hctx, i)
866 blk_mq_stop_hw_queue(hctx);
868 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
870 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
872 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
875 blk_mq_run_hw_queue(hctx, false);
878 EXPORT_SYMBOL(blk_mq_start_hw_queue);
880 void blk_mq_start_hw_queues(struct request_queue *q)
882 struct blk_mq_hw_ctx *hctx;
885 queue_for_each_hw_ctx(q, hctx, i)
886 blk_mq_start_hw_queue(hctx);
888 EXPORT_SYMBOL(blk_mq_start_hw_queues);
891 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
893 struct blk_mq_hw_ctx *hctx;
896 queue_for_each_hw_ctx(q, hctx, i) {
897 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
900 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
902 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 unsigned long tmo = msecs_to_jiffies(msecs);
931 if (hctx->queue->nr_hw_queues == 1)
932 kblockd_schedule_delayed_work(&hctx->delay_work, tmo);
936 cpu = blk_mq_hctx_next_cpu(hctx);
937 kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo);
940 EXPORT_SYMBOL(blk_mq_delay_queue);
942 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
943 struct request *rq, bool at_head)
945 struct blk_mq_ctx *ctx = rq->mq_ctx;
947 trace_block_rq_insert(hctx->queue, rq);
950 list_add(&rq->queuelist, &ctx->rq_list);
952 list_add_tail(&rq->queuelist, &ctx->rq_list);
954 blk_mq_hctx_mark_pending(hctx, ctx);
957 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
960 struct request_queue *q = rq->q;
961 struct blk_mq_hw_ctx *hctx;
962 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
964 current_ctx = blk_mq_get_ctx(q);
965 if (!cpu_online(ctx->cpu))
966 rq->mq_ctx = ctx = current_ctx;
968 hctx = q->mq_ops->map_queue(q, ctx->cpu);
970 spin_lock(&ctx->lock);
971 __blk_mq_insert_request(hctx, rq, at_head);
972 spin_unlock(&ctx->lock);
975 blk_mq_run_hw_queue(hctx, async);
977 blk_mq_put_ctx(current_ctx);
980 static void blk_mq_insert_requests(struct request_queue *q,
981 struct blk_mq_ctx *ctx,
982 struct list_head *list,
987 struct blk_mq_hw_ctx *hctx;
988 struct blk_mq_ctx *current_ctx;
990 trace_block_unplug(q, depth, !from_schedule);
992 current_ctx = blk_mq_get_ctx(q);
994 if (!cpu_online(ctx->cpu))
996 hctx = q->mq_ops->map_queue(q, ctx->cpu);
999 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1002 spin_lock(&ctx->lock);
1003 while (!list_empty(list)) {
1006 rq = list_first_entry(list, struct request, queuelist);
1007 list_del_init(&rq->queuelist);
1009 __blk_mq_insert_request(hctx, rq, false);
1011 spin_unlock(&ctx->lock);
1013 blk_mq_run_hw_queue(hctx, from_schedule);
1014 blk_mq_put_ctx(current_ctx);
1017 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1019 struct request *rqa = container_of(a, struct request, queuelist);
1020 struct request *rqb = container_of(b, struct request, queuelist);
1022 return !(rqa->mq_ctx < rqb->mq_ctx ||
1023 (rqa->mq_ctx == rqb->mq_ctx &&
1024 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1027 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1029 struct blk_mq_ctx *this_ctx;
1030 struct request_queue *this_q;
1033 LIST_HEAD(ctx_list);
1036 list_splice_init(&plug->mq_list, &list);
1038 list_sort(NULL, &list, plug_ctx_cmp);
1044 while (!list_empty(&list)) {
1045 rq = list_entry_rq(list.next);
1046 list_del_init(&rq->queuelist);
1048 if (rq->mq_ctx != this_ctx) {
1050 blk_mq_insert_requests(this_q, this_ctx,
1055 this_ctx = rq->mq_ctx;
1061 list_add_tail(&rq->queuelist, &ctx_list);
1065 * If 'this_ctx' is set, we know we have entries to complete
1066 * on 'ctx_list'. Do those.
1069 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1074 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1076 init_request_from_bio(rq, bio);
1078 if (blk_do_io_stat(rq))
1079 blk_account_io_start(rq, 1);
1082 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1084 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1085 !blk_queue_nomerges(hctx->queue);
1088 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1089 struct blk_mq_ctx *ctx,
1090 struct request *rq, struct bio *bio)
1092 if (!hctx_allow_merges(hctx)) {
1093 blk_mq_bio_to_request(rq, bio);
1094 spin_lock(&ctx->lock);
1096 __blk_mq_insert_request(hctx, rq, false);
1097 spin_unlock(&ctx->lock);
1100 struct request_queue *q = hctx->queue;
1102 spin_lock(&ctx->lock);
1103 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1104 blk_mq_bio_to_request(rq, bio);
1108 spin_unlock(&ctx->lock);
1109 __blk_mq_free_request(hctx, ctx, rq);
1114 struct blk_map_ctx {
1115 struct blk_mq_hw_ctx *hctx;
1116 struct blk_mq_ctx *ctx;
1119 static struct request *blk_mq_map_request(struct request_queue *q,
1121 struct blk_map_ctx *data)
1123 struct blk_mq_hw_ctx *hctx;
1124 struct blk_mq_ctx *ctx;
1126 int rw = bio_data_dir(bio);
1127 struct blk_mq_alloc_data alloc_data;
1129 if (unlikely(blk_mq_queue_enter(q))) {
1130 bio_endio(bio, -EIO);
1134 ctx = blk_mq_get_ctx(q);
1135 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1137 if (rw_is_sync(bio->bi_rw))
1140 trace_block_getrq(q, bio, rw);
1141 blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1143 rq = __blk_mq_alloc_request(&alloc_data, rw);
1144 if (unlikely(!rq)) {
1145 __blk_mq_run_hw_queue(hctx);
1146 blk_mq_put_ctx(ctx);
1147 trace_block_sleeprq(q, bio, rw);
1149 ctx = blk_mq_get_ctx(q);
1150 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1151 blk_mq_set_alloc_data(&alloc_data, q,
1152 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1153 rq = __blk_mq_alloc_request(&alloc_data, rw);
1154 ctx = alloc_data.ctx;
1155 hctx = alloc_data.hctx;
1165 * Multiple hardware queue variant. This will not use per-process plugs,
1166 * but will attempt to bypass the hctx queueing if we can go straight to
1167 * hardware for SYNC IO.
1169 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1171 const int is_sync = rw_is_sync(bio->bi_rw);
1172 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1173 struct blk_map_ctx data;
1176 blk_queue_bounce(q, &bio);
1178 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1179 bio_endio(bio, -EIO);
1183 rq = blk_mq_map_request(q, bio, &data);
1187 if (unlikely(is_flush_fua)) {
1188 blk_mq_bio_to_request(rq, bio);
1189 blk_insert_flush(rq);
1196 blk_mq_bio_to_request(rq, bio);
1197 blk_mq_start_request(rq, true);
1200 * For OK queue, we are done. For error, kill it. Any other
1201 * error (busy), just add it to our list as we previously
1204 ret = q->mq_ops->queue_rq(data.hctx, rq);
1205 if (ret == BLK_MQ_RQ_QUEUE_OK)
1208 __blk_mq_requeue_request(rq);
1210 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1212 blk_mq_end_io(rq, rq->errors);
1218 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1220 * For a SYNC request, send it to the hardware immediately. For
1221 * an ASYNC request, just ensure that we run it later on. The
1222 * latter allows for merging opportunities and more efficient
1226 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1229 blk_mq_put_ctx(data.ctx);
1233 * Single hardware queue variant. This will attempt to use any per-process
1234 * plug for merging and IO deferral.
1236 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1238 const int is_sync = rw_is_sync(bio->bi_rw);
1239 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1240 unsigned int use_plug, request_count = 0;
1241 struct blk_map_ctx data;
1245 * If we have multiple hardware queues, just go directly to
1246 * one of those for sync IO.
1248 use_plug = !is_flush_fua && !is_sync;
1250 blk_queue_bounce(q, &bio);
1252 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1253 bio_endio(bio, -EIO);
1257 if (use_plug && !blk_queue_nomerges(q) &&
1258 blk_attempt_plug_merge(q, bio, &request_count))
1261 rq = blk_mq_map_request(q, bio, &data);
1265 if (unlikely(is_flush_fua)) {
1266 blk_mq_bio_to_request(rq, bio);
1267 blk_insert_flush(rq);
1272 * A task plug currently exists. Since this is completely lockless,
1273 * utilize that to temporarily store requests until the task is
1274 * either done or scheduled away.
1277 struct blk_plug *plug = current->plug;
1280 blk_mq_bio_to_request(rq, bio);
1281 if (list_empty(&plug->mq_list))
1282 trace_block_plug(q);
1283 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1284 blk_flush_plug_list(plug, false);
1285 trace_block_plug(q);
1287 list_add_tail(&rq->queuelist, &plug->mq_list);
1288 blk_mq_put_ctx(data.ctx);
1293 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1295 * For a SYNC request, send it to the hardware immediately. For
1296 * an ASYNC request, just ensure that we run it later on. The
1297 * latter allows for merging opportunities and more efficient
1301 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1304 blk_mq_put_ctx(data.ctx);
1308 * Default mapping to a software queue, since we use one per CPU.
1310 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1312 return q->queue_hw_ctx[q->mq_map[cpu]];
1314 EXPORT_SYMBOL(blk_mq_map_queue);
1316 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1317 struct blk_mq_tags *tags, unsigned int hctx_idx)
1321 if (tags->rqs && set->ops->exit_request) {
1324 for (i = 0; i < tags->nr_tags; i++) {
1327 set->ops->exit_request(set->driver_data, tags->rqs[i],
1329 tags->rqs[i] = NULL;
1333 while (!list_empty(&tags->page_list)) {
1334 page = list_first_entry(&tags->page_list, struct page, lru);
1335 list_del_init(&page->lru);
1336 __free_pages(page, page->private);
1341 blk_mq_free_tags(tags);
1344 static size_t order_to_size(unsigned int order)
1346 return (size_t)PAGE_SIZE << order;
1349 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1350 unsigned int hctx_idx)
1352 struct blk_mq_tags *tags;
1353 unsigned int i, j, entries_per_page, max_order = 4;
1354 size_t rq_size, left;
1356 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1361 INIT_LIST_HEAD(&tags->page_list);
1363 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1364 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1367 blk_mq_free_tags(tags);
1372 * rq_size is the size of the request plus driver payload, rounded
1373 * to the cacheline size
1375 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1377 left = rq_size * set->queue_depth;
1379 for (i = 0; i < set->queue_depth; ) {
1380 int this_order = max_order;
1385 while (left < order_to_size(this_order - 1) && this_order)
1389 page = alloc_pages_node(set->numa_node,
1390 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1396 if (order_to_size(this_order) < rq_size)
1403 page->private = this_order;
1404 list_add_tail(&page->lru, &tags->page_list);
1406 p = page_address(page);
1407 entries_per_page = order_to_size(this_order) / rq_size;
1408 to_do = min(entries_per_page, set->queue_depth - i);
1409 left -= to_do * rq_size;
1410 for (j = 0; j < to_do; j++) {
1412 tags->rqs[i]->atomic_flags = 0;
1413 tags->rqs[i]->cmd_flags = 0;
1414 if (set->ops->init_request) {
1415 if (set->ops->init_request(set->driver_data,
1416 tags->rqs[i], hctx_idx, i,
1418 tags->rqs[i] = NULL;
1431 blk_mq_free_rq_map(set, tags, hctx_idx);
1435 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1440 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1442 unsigned int bpw = 8, total, num_maps, i;
1444 bitmap->bits_per_word = bpw;
1446 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1447 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1452 bitmap->map_size = num_maps;
1455 for (i = 0; i < num_maps; i++) {
1456 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1457 total -= bitmap->map[i].depth;
1463 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1465 struct request_queue *q = hctx->queue;
1466 struct blk_mq_ctx *ctx;
1470 * Move ctx entries to new CPU, if this one is going away.
1472 ctx = __blk_mq_get_ctx(q, cpu);
1474 spin_lock(&ctx->lock);
1475 if (!list_empty(&ctx->rq_list)) {
1476 list_splice_init(&ctx->rq_list, &tmp);
1477 blk_mq_hctx_clear_pending(hctx, ctx);
1479 spin_unlock(&ctx->lock);
1481 if (list_empty(&tmp))
1484 ctx = blk_mq_get_ctx(q);
1485 spin_lock(&ctx->lock);
1487 while (!list_empty(&tmp)) {
1490 rq = list_first_entry(&tmp, struct request, queuelist);
1492 list_move_tail(&rq->queuelist, &ctx->rq_list);
1495 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1496 blk_mq_hctx_mark_pending(hctx, ctx);
1498 spin_unlock(&ctx->lock);
1500 blk_mq_run_hw_queue(hctx, true);
1501 blk_mq_put_ctx(ctx);
1505 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1507 struct request_queue *q = hctx->queue;
1508 struct blk_mq_tag_set *set = q->tag_set;
1510 if (set->tags[hctx->queue_num])
1513 set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1514 if (!set->tags[hctx->queue_num])
1517 hctx->tags = set->tags[hctx->queue_num];
1521 static int blk_mq_hctx_notify(void *data, unsigned long action,
1524 struct blk_mq_hw_ctx *hctx = data;
1526 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1527 return blk_mq_hctx_cpu_offline(hctx, cpu);
1528 else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1529 return blk_mq_hctx_cpu_online(hctx, cpu);
1534 static void blk_mq_exit_hw_queues(struct request_queue *q,
1535 struct blk_mq_tag_set *set, int nr_queue)
1537 struct blk_mq_hw_ctx *hctx;
1540 queue_for_each_hw_ctx(q, hctx, i) {
1544 blk_mq_tag_idle(hctx);
1546 if (set->ops->exit_hctx)
1547 set->ops->exit_hctx(hctx, i);
1549 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1551 blk_mq_free_bitmap(&hctx->ctx_map);
1556 static void blk_mq_free_hw_queues(struct request_queue *q,
1557 struct blk_mq_tag_set *set)
1559 struct blk_mq_hw_ctx *hctx;
1562 queue_for_each_hw_ctx(q, hctx, i) {
1563 free_cpumask_var(hctx->cpumask);
1568 static int blk_mq_init_hw_queues(struct request_queue *q,
1569 struct blk_mq_tag_set *set)
1571 struct blk_mq_hw_ctx *hctx;
1575 * Initialize hardware queues
1577 queue_for_each_hw_ctx(q, hctx, i) {
1580 node = hctx->numa_node;
1581 if (node == NUMA_NO_NODE)
1582 node = hctx->numa_node = set->numa_node;
1584 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1585 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1586 spin_lock_init(&hctx->lock);
1587 INIT_LIST_HEAD(&hctx->dispatch);
1589 hctx->queue_num = i;
1590 hctx->flags = set->flags;
1591 hctx->cmd_size = set->cmd_size;
1593 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1594 blk_mq_hctx_notify, hctx);
1595 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1597 hctx->tags = set->tags[i];
1600 * Allocate space for all possible cpus to avoid allocation at
1603 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1608 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1613 if (set->ops->init_hctx &&
1614 set->ops->init_hctx(hctx, set->driver_data, i))
1618 if (i == q->nr_hw_queues)
1624 blk_mq_exit_hw_queues(q, set, i);
1629 static void blk_mq_init_cpu_queues(struct request_queue *q,
1630 unsigned int nr_hw_queues)
1634 for_each_possible_cpu(i) {
1635 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1636 struct blk_mq_hw_ctx *hctx;
1638 memset(__ctx, 0, sizeof(*__ctx));
1640 spin_lock_init(&__ctx->lock);
1641 INIT_LIST_HEAD(&__ctx->rq_list);
1644 /* If the cpu isn't online, the cpu is mapped to first hctx */
1648 hctx = q->mq_ops->map_queue(q, i);
1649 cpumask_set_cpu(i, hctx->cpumask);
1653 * Set local node, IFF we have more than one hw queue. If
1654 * not, we remain on the home node of the device
1656 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1657 hctx->numa_node = cpu_to_node(i);
1661 static void blk_mq_map_swqueue(struct request_queue *q)
1664 struct blk_mq_hw_ctx *hctx;
1665 struct blk_mq_ctx *ctx;
1667 queue_for_each_hw_ctx(q, hctx, i) {
1668 cpumask_clear(hctx->cpumask);
1673 * Map software to hardware queues
1675 queue_for_each_ctx(q, ctx, i) {
1676 /* If the cpu isn't online, the cpu is mapped to first hctx */
1680 hctx = q->mq_ops->map_queue(q, i);
1681 cpumask_set_cpu(i, hctx->cpumask);
1682 ctx->index_hw = hctx->nr_ctx;
1683 hctx->ctxs[hctx->nr_ctx++] = ctx;
1686 queue_for_each_hw_ctx(q, hctx, i) {
1688 * If no software queues are mapped to this hardware queue,
1689 * disable it and free the request entries.
1691 if (!hctx->nr_ctx) {
1692 struct blk_mq_tag_set *set = q->tag_set;
1695 blk_mq_free_rq_map(set, set->tags[i], i);
1696 set->tags[i] = NULL;
1703 * Initialize batch roundrobin counts
1705 hctx->next_cpu = cpumask_first(hctx->cpumask);
1706 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1710 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1712 struct blk_mq_hw_ctx *hctx;
1713 struct request_queue *q;
1717 if (set->tag_list.next == set->tag_list.prev)
1722 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1723 blk_mq_freeze_queue(q);
1725 queue_for_each_hw_ctx(q, hctx, i) {
1727 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1729 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1731 blk_mq_unfreeze_queue(q);
1735 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1737 struct blk_mq_tag_set *set = q->tag_set;
1739 mutex_lock(&set->tag_list_lock);
1740 list_del_init(&q->tag_set_list);
1741 blk_mq_update_tag_set_depth(set);
1742 mutex_unlock(&set->tag_list_lock);
1745 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1746 struct request_queue *q)
1750 mutex_lock(&set->tag_list_lock);
1751 list_add_tail(&q->tag_set_list, &set->tag_list);
1752 blk_mq_update_tag_set_depth(set);
1753 mutex_unlock(&set->tag_list_lock);
1756 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1758 struct blk_mq_hw_ctx **hctxs;
1759 struct blk_mq_ctx __percpu *ctx;
1760 struct request_queue *q;
1764 ctx = alloc_percpu(struct blk_mq_ctx);
1766 return ERR_PTR(-ENOMEM);
1768 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1774 map = blk_mq_make_queue_map(set);
1778 for (i = 0; i < set->nr_hw_queues; i++) {
1779 int node = blk_mq_hw_queue_to_node(map, i);
1781 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1786 if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL))
1789 atomic_set(&hctxs[i]->nr_active, 0);
1790 hctxs[i]->numa_node = node;
1791 hctxs[i]->queue_num = i;
1794 q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1799 * Init percpu_ref in atomic mode so that it's faster to shutdown.
1800 * See blk_register_queue() for details.
1802 if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release,
1803 PERCPU_REF_INIT_ATOMIC, GFP_KERNEL))
1806 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1807 blk_queue_rq_timeout(q, 30000);
1809 q->nr_queues = nr_cpu_ids;
1810 q->nr_hw_queues = set->nr_hw_queues;
1814 q->queue_hw_ctx = hctxs;
1816 q->mq_ops = set->ops;
1817 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1819 if (!(set->flags & BLK_MQ_F_SG_MERGE))
1820 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1822 q->sg_reserved_size = INT_MAX;
1824 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1825 INIT_LIST_HEAD(&q->requeue_list);
1826 spin_lock_init(&q->requeue_lock);
1828 if (q->nr_hw_queues > 1)
1829 blk_queue_make_request(q, blk_mq_make_request);
1831 blk_queue_make_request(q, blk_sq_make_request);
1833 blk_queue_rq_timed_out(q, blk_mq_rq_timed_out);
1835 blk_queue_rq_timeout(q, set->timeout);
1838 * Do this after blk_queue_make_request() overrides it...
1840 q->nr_requests = set->queue_depth;
1842 if (set->ops->complete)
1843 blk_queue_softirq_done(q, set->ops->complete);
1845 blk_mq_init_flush(q);
1846 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1848 q->flush_rq = kzalloc(round_up(sizeof(struct request) +
1849 set->cmd_size, cache_line_size()),
1854 if (blk_mq_init_hw_queues(q, set))
1857 mutex_lock(&all_q_mutex);
1858 list_add_tail(&q->all_q_node, &all_q_list);
1859 mutex_unlock(&all_q_mutex);
1861 blk_mq_add_queue_tag_set(set, q);
1863 blk_mq_map_swqueue(q);
1870 blk_cleanup_queue(q);
1873 for (i = 0; i < set->nr_hw_queues; i++) {
1876 free_cpumask_var(hctxs[i]->cpumask);
1883 return ERR_PTR(-ENOMEM);
1885 EXPORT_SYMBOL(blk_mq_init_queue);
1887 void blk_mq_free_queue(struct request_queue *q)
1889 struct blk_mq_tag_set *set = q->tag_set;
1891 blk_mq_del_queue_tag_set(q);
1893 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
1894 blk_mq_free_hw_queues(q, set);
1896 percpu_ref_exit(&q->mq_usage_counter);
1898 free_percpu(q->queue_ctx);
1899 kfree(q->queue_hw_ctx);
1902 q->queue_ctx = NULL;
1903 q->queue_hw_ctx = NULL;
1906 mutex_lock(&all_q_mutex);
1907 list_del_init(&q->all_q_node);
1908 mutex_unlock(&all_q_mutex);
1911 /* Basically redo blk_mq_init_queue with queue frozen */
1912 static void blk_mq_queue_reinit(struct request_queue *q)
1914 blk_mq_freeze_queue(q);
1916 blk_mq_sysfs_unregister(q);
1918 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1921 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1922 * we should change hctx numa_node according to new topology (this
1923 * involves free and re-allocate memory, worthy doing?)
1926 blk_mq_map_swqueue(q);
1928 blk_mq_sysfs_register(q);
1930 blk_mq_unfreeze_queue(q);
1933 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1934 unsigned long action, void *hcpu)
1936 struct request_queue *q;
1939 * Before new mappings are established, hotadded cpu might already
1940 * start handling requests. This doesn't break anything as we map
1941 * offline CPUs to first hardware queue. We will re-init the queue
1942 * below to get optimal settings.
1944 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1945 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1948 mutex_lock(&all_q_mutex);
1949 list_for_each_entry(q, &all_q_list, all_q_node)
1950 blk_mq_queue_reinit(q);
1951 mutex_unlock(&all_q_mutex);
1955 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
1959 for (i = 0; i < set->nr_hw_queues; i++) {
1960 set->tags[i] = blk_mq_init_rq_map(set, i);
1969 blk_mq_free_rq_map(set, set->tags[i], i);
1975 * Allocate the request maps associated with this tag_set. Note that this
1976 * may reduce the depth asked for, if memory is tight. set->queue_depth
1977 * will be updated to reflect the allocated depth.
1979 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
1984 depth = set->queue_depth;
1986 err = __blk_mq_alloc_rq_maps(set);
1990 set->queue_depth >>= 1;
1991 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
1995 } while (set->queue_depth);
1997 if (!set->queue_depth || err) {
1998 pr_err("blk-mq: failed to allocate request map\n");
2002 if (depth != set->queue_depth)
2003 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2004 depth, set->queue_depth);
2010 * Alloc a tag set to be associated with one or more request queues.
2011 * May fail with EINVAL for various error conditions. May adjust the
2012 * requested depth down, if if it too large. In that case, the set
2013 * value will be stored in set->queue_depth.
2015 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2017 if (!set->nr_hw_queues)
2019 if (!set->queue_depth)
2021 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2024 if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue)
2027 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2028 pr_info("blk-mq: reduced tag depth to %u\n",
2030 set->queue_depth = BLK_MQ_MAX_DEPTH;
2033 set->tags = kmalloc_node(set->nr_hw_queues *
2034 sizeof(struct blk_mq_tags *),
2035 GFP_KERNEL, set->numa_node);
2039 if (blk_mq_alloc_rq_maps(set))
2042 mutex_init(&set->tag_list_lock);
2043 INIT_LIST_HEAD(&set->tag_list);
2051 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2053 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2057 for (i = 0; i < set->nr_hw_queues; i++) {
2059 blk_mq_free_rq_map(set, set->tags[i], i);
2065 EXPORT_SYMBOL(blk_mq_free_tag_set);
2067 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2069 struct blk_mq_tag_set *set = q->tag_set;
2070 struct blk_mq_hw_ctx *hctx;
2073 if (!set || nr > set->queue_depth)
2077 queue_for_each_hw_ctx(q, hctx, i) {
2078 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2084 q->nr_requests = nr;
2089 void blk_mq_disable_hotplug(void)
2091 mutex_lock(&all_q_mutex);
2094 void blk_mq_enable_hotplug(void)
2096 mutex_unlock(&all_q_mutex);
2099 static int __init blk_mq_init(void)
2103 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2107 subsys_initcall(blk_mq_init);