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);
36 static struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
39 return per_cpu_ptr(q->queue_ctx, cpu);
43 * This assumes per-cpu software queueing queues. They could be per-node
44 * as well, for instance. For now this is hardcoded as-is. Note that we don't
45 * care about preemption, since we know the ctx's are persistent. This does
46 * mean that we can't rely on ctx always matching the currently running CPU.
48 static struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
50 return __blk_mq_get_ctx(q, get_cpu());
53 static void blk_mq_put_ctx(struct blk_mq_ctx *ctx)
59 * Check if any of the ctx's have pending work in this hardware queue
61 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
65 for (i = 0; i < hctx->ctx_map.map_size; i++)
66 if (hctx->ctx_map.map[i].word)
72 static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx,
73 struct blk_mq_ctx *ctx)
75 return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word];
78 #define CTX_TO_BIT(hctx, ctx) \
79 ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
82 * Mark this ctx as having pending work in this hardware queue
84 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
85 struct blk_mq_ctx *ctx)
87 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
89 if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word))
90 set_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
93 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
94 struct blk_mq_ctx *ctx)
96 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
98 clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
101 static int blk_mq_queue_enter(struct request_queue *q)
105 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
107 /* we have problems to freeze the queue if it's initializing */
108 if (!blk_queue_bypass(q) || !blk_queue_init_done(q))
111 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
113 spin_lock_irq(q->queue_lock);
114 ret = wait_event_interruptible_lock_irq(q->mq_freeze_wq,
115 !blk_queue_bypass(q) || blk_queue_dying(q),
117 /* inc usage with lock hold to avoid freeze_queue runs here */
118 if (!ret && !blk_queue_dying(q))
119 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
120 else if (blk_queue_dying(q))
122 spin_unlock_irq(q->queue_lock);
127 static void blk_mq_queue_exit(struct request_queue *q)
129 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
132 static void __blk_mq_drain_queue(struct request_queue *q)
137 spin_lock_irq(q->queue_lock);
138 count = percpu_counter_sum(&q->mq_usage_counter);
139 spin_unlock_irq(q->queue_lock);
143 blk_mq_run_queues(q, false);
149 * Guarantee no request is in use, so we can change any data structure of
150 * the queue afterward.
152 static void blk_mq_freeze_queue(struct request_queue *q)
156 spin_lock_irq(q->queue_lock);
157 drain = !q->bypass_depth++;
158 queue_flag_set(QUEUE_FLAG_BYPASS, q);
159 spin_unlock_irq(q->queue_lock);
162 __blk_mq_drain_queue(q);
165 void blk_mq_drain_queue(struct request_queue *q)
167 __blk_mq_drain_queue(q);
170 static void blk_mq_unfreeze_queue(struct request_queue *q)
174 spin_lock_irq(q->queue_lock);
175 if (!--q->bypass_depth) {
176 queue_flag_clear(QUEUE_FLAG_BYPASS, q);
179 WARN_ON_ONCE(q->bypass_depth < 0);
180 spin_unlock_irq(q->queue_lock);
182 wake_up_all(&q->mq_freeze_wq);
185 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
187 return blk_mq_has_free_tags(hctx->tags);
189 EXPORT_SYMBOL(blk_mq_can_queue);
191 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
192 struct request *rq, unsigned int rw_flags)
194 if (blk_queue_io_stat(q))
195 rw_flags |= REQ_IO_STAT;
197 INIT_LIST_HEAD(&rq->queuelist);
198 /* csd/requeue_work/fifo_time is initialized before use */
201 rq->cmd_flags |= rw_flags;
202 /* do not touch atomic flags, it needs atomic ops against the timer */
204 INIT_HLIST_NODE(&rq->hash);
205 RB_CLEAR_NODE(&rq->rb_node);
208 #ifdef CONFIG_BLK_CGROUP
210 set_start_time_ns(rq);
211 rq->io_start_time_ns = 0;
213 rq->nr_phys_segments = 0;
214 #if defined(CONFIG_BLK_DEV_INTEGRITY)
215 rq->nr_integrity_segments = 0;
218 /* tag was already set */
226 INIT_LIST_HEAD(&rq->timeout_list);
228 rq->end_io_data = NULL;
231 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
234 static struct request *
235 __blk_mq_alloc_request(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
236 struct blk_mq_ctx *ctx, int rw, gfp_t gfp, bool reserved)
241 tag = blk_mq_get_tag(hctx, &ctx->last_tag, gfp, reserved);
242 if (tag != BLK_MQ_TAG_FAIL) {
243 rq = hctx->tags->rqs[tag];
246 if (blk_mq_tag_busy(hctx)) {
247 rq->cmd_flags = REQ_MQ_INFLIGHT;
248 atomic_inc(&hctx->nr_active);
252 blk_mq_rq_ctx_init(q, ctx, rq, rw);
259 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
262 struct blk_mq_ctx *ctx;
263 struct blk_mq_hw_ctx *hctx;
266 if (blk_mq_queue_enter(q))
269 ctx = blk_mq_get_ctx(q);
270 hctx = q->mq_ops->map_queue(q, ctx->cpu);
272 rq = __blk_mq_alloc_request(q, hctx, ctx, rw, gfp & ~__GFP_WAIT,
274 if (!rq && (gfp & __GFP_WAIT)) {
275 __blk_mq_run_hw_queue(hctx);
278 ctx = blk_mq_get_ctx(q);
279 hctx = q->mq_ops->map_queue(q, ctx->cpu);
280 rq = __blk_mq_alloc_request(q, hctx, ctx, rw, gfp, reserved);
285 EXPORT_SYMBOL(blk_mq_alloc_request);
287 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
288 struct blk_mq_ctx *ctx, struct request *rq)
290 const int tag = rq->tag;
291 struct request_queue *q = rq->q;
293 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
294 atomic_dec(&hctx->nr_active);
296 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
297 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
298 blk_mq_queue_exit(q);
301 void blk_mq_free_request(struct request *rq)
303 struct blk_mq_ctx *ctx = rq->mq_ctx;
304 struct blk_mq_hw_ctx *hctx;
305 struct request_queue *q = rq->q;
307 ctx->rq_completed[rq_is_sync(rq)]++;
309 hctx = q->mq_ops->map_queue(q, ctx->cpu);
310 __blk_mq_free_request(hctx, ctx, rq);
314 * Clone all relevant state from a request that has been put on hold in
315 * the flush state machine into the preallocated flush request that hangs
316 * off the request queue.
318 * For a driver the flush request should be invisible, that's why we are
319 * impersonating the original request here.
321 void blk_mq_clone_flush_request(struct request *flush_rq,
322 struct request *orig_rq)
324 struct blk_mq_hw_ctx *hctx =
325 orig_rq->q->mq_ops->map_queue(orig_rq->q, orig_rq->mq_ctx->cpu);
327 flush_rq->mq_ctx = orig_rq->mq_ctx;
328 flush_rq->tag = orig_rq->tag;
329 memcpy(blk_mq_rq_to_pdu(flush_rq), blk_mq_rq_to_pdu(orig_rq),
333 inline void __blk_mq_end_io(struct request *rq, int error)
335 blk_account_io_done(rq);
338 rq->end_io(rq, error);
340 if (unlikely(blk_bidi_rq(rq)))
341 blk_mq_free_request(rq->next_rq);
342 blk_mq_free_request(rq);
345 EXPORT_SYMBOL(__blk_mq_end_io);
347 void blk_mq_end_io(struct request *rq, int error)
349 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
351 __blk_mq_end_io(rq, error);
353 EXPORT_SYMBOL(blk_mq_end_io);
355 static void __blk_mq_complete_request_remote(void *data)
357 struct request *rq = data;
359 rq->q->softirq_done_fn(rq);
362 static void blk_mq_ipi_complete_request(struct request *rq)
364 struct blk_mq_ctx *ctx = rq->mq_ctx;
368 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
369 rq->q->softirq_done_fn(rq);
374 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
375 shared = cpus_share_cache(cpu, ctx->cpu);
377 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
378 rq->csd.func = __blk_mq_complete_request_remote;
381 smp_call_function_single_async(ctx->cpu, &rq->csd);
383 rq->q->softirq_done_fn(rq);
388 void __blk_mq_complete_request(struct request *rq)
390 struct request_queue *q = rq->q;
392 if (!q->softirq_done_fn)
393 blk_mq_end_io(rq, rq->errors);
395 blk_mq_ipi_complete_request(rq);
399 * blk_mq_complete_request - end I/O on a request
400 * @rq: the request being processed
403 * Ends all I/O on a request. It does not handle partial completions.
404 * The actual completion happens out-of-order, through a IPI handler.
406 void blk_mq_complete_request(struct request *rq)
408 struct request_queue *q = rq->q;
410 if (unlikely(blk_should_fake_timeout(q)))
412 if (!blk_mark_rq_complete(rq))
413 __blk_mq_complete_request(rq);
415 EXPORT_SYMBOL(blk_mq_complete_request);
417 static void blk_mq_start_request(struct request *rq, bool last)
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);
428 * Just mark start time and set the started bit. Due to memory
429 * ordering, we know we'll see the correct deadline as long as
430 * REQ_ATOMIC_STARTED is seen. Use the default queue timeout,
431 * unless one has been set in the request.
434 rq->deadline = jiffies + q->rq_timeout;
436 rq->deadline = jiffies + rq->timeout;
439 * Mark us as started and clear complete. Complete might have been
440 * set if requeue raced with timeout, which then marked it as
441 * complete. So be sure to clear complete again when we start
442 * the request, otherwise we'll ignore the completion event.
444 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
445 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
446 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
447 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
449 if (q->dma_drain_size && blk_rq_bytes(rq)) {
451 * Make sure space for the drain appears. We know we can do
452 * this because max_hw_segments has been adjusted to be one
453 * fewer than the device can handle.
455 rq->nr_phys_segments++;
459 * Flag the last request in the series so that drivers know when IO
460 * should be kicked off, if they don't do it on a per-request basis.
462 * Note: the flag isn't the only condition drivers should do kick off.
463 * If drive is busy, the last request might not have the bit set.
466 rq->cmd_flags |= REQ_END;
469 static void __blk_mq_requeue_request(struct request *rq)
471 struct request_queue *q = rq->q;
473 trace_block_rq_requeue(q, rq);
474 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
476 rq->cmd_flags &= ~REQ_END;
478 if (q->dma_drain_size && blk_rq_bytes(rq))
479 rq->nr_phys_segments--;
482 void blk_mq_requeue_request(struct request *rq)
484 __blk_mq_requeue_request(rq);
485 blk_clear_rq_complete(rq);
487 BUG_ON(blk_queued_rq(rq));
488 blk_mq_add_to_requeue_list(rq, true);
490 EXPORT_SYMBOL(blk_mq_requeue_request);
492 static void blk_mq_requeue_work(struct work_struct *work)
494 struct request_queue *q =
495 container_of(work, struct request_queue, requeue_work);
497 struct request *rq, *next;
500 spin_lock_irqsave(&q->requeue_lock, flags);
501 list_splice_init(&q->requeue_list, &rq_list);
502 spin_unlock_irqrestore(&q->requeue_lock, flags);
504 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
505 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
508 rq->cmd_flags &= ~REQ_SOFTBARRIER;
509 list_del_init(&rq->queuelist);
510 blk_mq_insert_request(rq, true, false, false);
513 while (!list_empty(&rq_list)) {
514 rq = list_entry(rq_list.next, struct request, queuelist);
515 list_del_init(&rq->queuelist);
516 blk_mq_insert_request(rq, false, false, false);
519 blk_mq_run_queues(q, false);
522 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
524 struct request_queue *q = rq->q;
528 * We abuse this flag that is otherwise used by the I/O scheduler to
529 * request head insertation from the workqueue.
531 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
533 spin_lock_irqsave(&q->requeue_lock, flags);
535 rq->cmd_flags |= REQ_SOFTBARRIER;
536 list_add(&rq->queuelist, &q->requeue_list);
538 list_add_tail(&rq->queuelist, &q->requeue_list);
540 spin_unlock_irqrestore(&q->requeue_lock, flags);
542 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
544 void blk_mq_kick_requeue_list(struct request_queue *q)
546 kblockd_schedule_work(&q->requeue_work);
548 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
550 struct request *blk_mq_tag_to_rq(struct blk_mq_hw_ctx *hctx, unsigned int tag)
552 struct request_queue *q = hctx->queue;
554 if ((q->flush_rq->cmd_flags & REQ_FLUSH_SEQ) &&
555 q->flush_rq->tag == tag)
558 return hctx->tags->rqs[tag];
560 EXPORT_SYMBOL(blk_mq_tag_to_rq);
562 struct blk_mq_timeout_data {
563 struct blk_mq_hw_ctx *hctx;
565 unsigned int *next_set;
568 static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
570 struct blk_mq_timeout_data *data = __data;
571 struct blk_mq_hw_ctx *hctx = data->hctx;
574 /* It may not be in flight yet (this is where
575 * the REQ_ATOMIC_STARTED flag comes in). The requests are
576 * statically allocated, so we know it's always safe to access the
577 * memory associated with a bit offset into ->rqs[].
583 tag = find_next_zero_bit(free_tags, hctx->tags->nr_tags, tag);
584 if (tag >= hctx->tags->nr_tags)
587 rq = blk_mq_tag_to_rq(hctx, tag++);
588 if (rq->q != hctx->queue)
590 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
593 blk_rq_check_expired(rq, data->next, data->next_set);
597 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
599 unsigned int *next_set)
601 struct blk_mq_timeout_data data = {
604 .next_set = next_set,
608 * Ask the tagging code to iterate busy requests, so we can
609 * check them for timeout.
611 blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
614 static enum blk_eh_timer_return blk_mq_rq_timed_out(struct request *rq)
616 struct request_queue *q = rq->q;
619 * We know that complete is set at this point. If STARTED isn't set
620 * anymore, then the request isn't active and the "timeout" should
621 * just be ignored. This can happen due to the bitflag ordering.
622 * Timeout first checks if STARTED is set, and if it is, assumes
623 * the request is active. But if we race with completion, then
624 * we both flags will get cleared. So check here again, and ignore
625 * a timeout event with a request that isn't active.
627 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
628 return BLK_EH_NOT_HANDLED;
630 if (!q->mq_ops->timeout)
631 return BLK_EH_RESET_TIMER;
633 return q->mq_ops->timeout(rq);
636 static void blk_mq_rq_timer(unsigned long data)
638 struct request_queue *q = (struct request_queue *) data;
639 struct blk_mq_hw_ctx *hctx;
640 unsigned long next = 0;
643 queue_for_each_hw_ctx(q, hctx, i) {
645 * If not software queues are currently mapped to this
646 * hardware queue, there's nothing to check
648 if (!hctx->nr_ctx || !hctx->tags)
651 blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
655 next = blk_rq_timeout(round_jiffies_up(next));
656 mod_timer(&q->timeout, next);
658 queue_for_each_hw_ctx(q, hctx, i)
659 blk_mq_tag_idle(hctx);
664 * Reverse check our software queue for entries that we could potentially
665 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
666 * too much time checking for merges.
668 static bool blk_mq_attempt_merge(struct request_queue *q,
669 struct blk_mq_ctx *ctx, struct bio *bio)
674 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
680 if (!blk_rq_merge_ok(rq, bio))
683 el_ret = blk_try_merge(rq, bio);
684 if (el_ret == ELEVATOR_BACK_MERGE) {
685 if (bio_attempt_back_merge(q, rq, bio)) {
690 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
691 if (bio_attempt_front_merge(q, rq, bio)) {
703 * Process software queues that have been marked busy, splicing them
704 * to the for-dispatch
706 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
708 struct blk_mq_ctx *ctx;
711 for (i = 0; i < hctx->ctx_map.map_size; i++) {
712 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
713 unsigned int off, bit;
719 off = i * hctx->ctx_map.bits_per_word;
721 bit = find_next_bit(&bm->word, bm->depth, bit);
722 if (bit >= bm->depth)
725 ctx = hctx->ctxs[bit + off];
726 clear_bit(bit, &bm->word);
727 spin_lock(&ctx->lock);
728 list_splice_tail_init(&ctx->rq_list, list);
729 spin_unlock(&ctx->lock);
737 * Run this hardware queue, pulling any software queues mapped to it in.
738 * Note that this function currently has various problems around ordering
739 * of IO. In particular, we'd like FIFO behaviour on handling existing
740 * items on the hctx->dispatch list. Ignore that for now.
742 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
744 struct request_queue *q = hctx->queue;
749 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
751 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
757 * Touch any software queue that has pending entries.
759 flush_busy_ctxs(hctx, &rq_list);
762 * If we have previous entries on our dispatch list, grab them
763 * and stuff them at the front for more fair dispatch.
765 if (!list_empty_careful(&hctx->dispatch)) {
766 spin_lock(&hctx->lock);
767 if (!list_empty(&hctx->dispatch))
768 list_splice_init(&hctx->dispatch, &rq_list);
769 spin_unlock(&hctx->lock);
773 * Now process all the entries, sending them to the driver.
776 while (!list_empty(&rq_list)) {
779 rq = list_first_entry(&rq_list, struct request, queuelist);
780 list_del_init(&rq->queuelist);
782 blk_mq_start_request(rq, list_empty(&rq_list));
784 ret = q->mq_ops->queue_rq(hctx, rq);
786 case BLK_MQ_RQ_QUEUE_OK:
789 case BLK_MQ_RQ_QUEUE_BUSY:
790 list_add(&rq->queuelist, &rq_list);
791 __blk_mq_requeue_request(rq);
794 pr_err("blk-mq: bad return on queue: %d\n", ret);
795 case BLK_MQ_RQ_QUEUE_ERROR:
797 blk_mq_end_io(rq, rq->errors);
801 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
806 hctx->dispatched[0]++;
807 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
808 hctx->dispatched[ilog2(queued) + 1]++;
811 * Any items that need requeuing? Stuff them into hctx->dispatch,
812 * that is where we will continue on next queue run.
814 if (!list_empty(&rq_list)) {
815 spin_lock(&hctx->lock);
816 list_splice(&rq_list, &hctx->dispatch);
817 spin_unlock(&hctx->lock);
822 * It'd be great if the workqueue API had a way to pass
823 * in a mask and had some smarts for more clever placement.
824 * For now we just round-robin here, switching for every
825 * BLK_MQ_CPU_WORK_BATCH queued items.
827 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
829 int cpu = hctx->next_cpu;
831 if (--hctx->next_cpu_batch <= 0) {
834 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
835 if (next_cpu >= nr_cpu_ids)
836 next_cpu = cpumask_first(hctx->cpumask);
838 hctx->next_cpu = next_cpu;
839 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
845 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
847 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
850 if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask))
851 __blk_mq_run_hw_queue(hctx);
852 else if (hctx->queue->nr_hw_queues == 1)
853 kblockd_schedule_delayed_work(&hctx->run_work, 0);
857 cpu = blk_mq_hctx_next_cpu(hctx);
858 kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0);
862 void blk_mq_run_queues(struct request_queue *q, bool async)
864 struct blk_mq_hw_ctx *hctx;
867 queue_for_each_hw_ctx(q, hctx, i) {
868 if ((!blk_mq_hctx_has_pending(hctx) &&
869 list_empty_careful(&hctx->dispatch)) ||
870 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
874 blk_mq_run_hw_queue(hctx, async);
878 EXPORT_SYMBOL(blk_mq_run_queues);
880 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
882 cancel_delayed_work(&hctx->run_work);
883 cancel_delayed_work(&hctx->delay_work);
884 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
886 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
888 void blk_mq_stop_hw_queues(struct request_queue *q)
890 struct blk_mq_hw_ctx *hctx;
893 queue_for_each_hw_ctx(q, hctx, i)
894 blk_mq_stop_hw_queue(hctx);
896 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
898 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
900 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
903 __blk_mq_run_hw_queue(hctx);
906 EXPORT_SYMBOL(blk_mq_start_hw_queue);
908 void blk_mq_start_hw_queues(struct request_queue *q)
910 struct blk_mq_hw_ctx *hctx;
913 queue_for_each_hw_ctx(q, hctx, i)
914 blk_mq_start_hw_queue(hctx);
916 EXPORT_SYMBOL(blk_mq_start_hw_queues);
919 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
921 struct blk_mq_hw_ctx *hctx;
924 queue_for_each_hw_ctx(q, hctx, i) {
925 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
928 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
930 blk_mq_run_hw_queue(hctx, async);
934 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
936 static void blk_mq_run_work_fn(struct work_struct *work)
938 struct blk_mq_hw_ctx *hctx;
940 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
942 __blk_mq_run_hw_queue(hctx);
945 static void blk_mq_delay_work_fn(struct work_struct *work)
947 struct blk_mq_hw_ctx *hctx;
949 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
951 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
952 __blk_mq_run_hw_queue(hctx);
955 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
957 unsigned long tmo = msecs_to_jiffies(msecs);
959 if (hctx->queue->nr_hw_queues == 1)
960 kblockd_schedule_delayed_work(&hctx->delay_work, tmo);
964 cpu = blk_mq_hctx_next_cpu(hctx);
965 kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo);
968 EXPORT_SYMBOL(blk_mq_delay_queue);
970 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
971 struct request *rq, bool at_head)
973 struct blk_mq_ctx *ctx = rq->mq_ctx;
975 trace_block_rq_insert(hctx->queue, rq);
978 list_add(&rq->queuelist, &ctx->rq_list);
980 list_add_tail(&rq->queuelist, &ctx->rq_list);
982 blk_mq_hctx_mark_pending(hctx, ctx);
985 * We do this early, to ensure we are on the right CPU.
990 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
993 struct request_queue *q = rq->q;
994 struct blk_mq_hw_ctx *hctx;
995 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
997 current_ctx = blk_mq_get_ctx(q);
998 if (!cpu_online(ctx->cpu))
999 rq->mq_ctx = ctx = current_ctx;
1001 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1003 if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA) &&
1004 !(rq->cmd_flags & (REQ_FLUSH_SEQ))) {
1005 blk_insert_flush(rq);
1007 spin_lock(&ctx->lock);
1008 __blk_mq_insert_request(hctx, rq, at_head);
1009 spin_unlock(&ctx->lock);
1013 blk_mq_run_hw_queue(hctx, async);
1015 blk_mq_put_ctx(current_ctx);
1018 static void blk_mq_insert_requests(struct request_queue *q,
1019 struct blk_mq_ctx *ctx,
1020 struct list_head *list,
1025 struct blk_mq_hw_ctx *hctx;
1026 struct blk_mq_ctx *current_ctx;
1028 trace_block_unplug(q, depth, !from_schedule);
1030 current_ctx = blk_mq_get_ctx(q);
1032 if (!cpu_online(ctx->cpu))
1034 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1037 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1040 spin_lock(&ctx->lock);
1041 while (!list_empty(list)) {
1044 rq = list_first_entry(list, struct request, queuelist);
1045 list_del_init(&rq->queuelist);
1047 __blk_mq_insert_request(hctx, rq, false);
1049 spin_unlock(&ctx->lock);
1051 blk_mq_run_hw_queue(hctx, from_schedule);
1052 blk_mq_put_ctx(current_ctx);
1055 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1057 struct request *rqa = container_of(a, struct request, queuelist);
1058 struct request *rqb = container_of(b, struct request, queuelist);
1060 return !(rqa->mq_ctx < rqb->mq_ctx ||
1061 (rqa->mq_ctx == rqb->mq_ctx &&
1062 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1065 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1067 struct blk_mq_ctx *this_ctx;
1068 struct request_queue *this_q;
1071 LIST_HEAD(ctx_list);
1074 list_splice_init(&plug->mq_list, &list);
1076 list_sort(NULL, &list, plug_ctx_cmp);
1082 while (!list_empty(&list)) {
1083 rq = list_entry_rq(list.next);
1084 list_del_init(&rq->queuelist);
1086 if (rq->mq_ctx != this_ctx) {
1088 blk_mq_insert_requests(this_q, this_ctx,
1093 this_ctx = rq->mq_ctx;
1099 list_add_tail(&rq->queuelist, &ctx_list);
1103 * If 'this_ctx' is set, we know we have entries to complete
1104 * on 'ctx_list'. Do those.
1107 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1112 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1114 init_request_from_bio(rq, bio);
1116 if (blk_do_io_stat(rq)) {
1117 rq->start_time = jiffies;
1118 blk_account_io_start(rq, 1);
1122 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1123 struct blk_mq_ctx *ctx,
1124 struct request *rq, struct bio *bio)
1126 struct request_queue *q = hctx->queue;
1128 if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE)) {
1129 blk_mq_bio_to_request(rq, bio);
1130 spin_lock(&ctx->lock);
1132 __blk_mq_insert_request(hctx, rq, false);
1133 spin_unlock(&ctx->lock);
1136 spin_lock(&ctx->lock);
1137 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1138 blk_mq_bio_to_request(rq, bio);
1142 spin_unlock(&ctx->lock);
1143 __blk_mq_free_request(hctx, ctx, rq);
1148 struct blk_map_ctx {
1149 struct blk_mq_hw_ctx *hctx;
1150 struct blk_mq_ctx *ctx;
1153 static struct request *blk_mq_map_request(struct request_queue *q,
1155 struct blk_map_ctx *data)
1157 struct blk_mq_hw_ctx *hctx;
1158 struct blk_mq_ctx *ctx;
1160 int rw = bio_data_dir(bio);
1162 if (unlikely(blk_mq_queue_enter(q))) {
1163 bio_endio(bio, -EIO);
1167 ctx = blk_mq_get_ctx(q);
1168 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1170 if (rw_is_sync(bio->bi_rw))
1173 trace_block_getrq(q, bio, rw);
1174 rq = __blk_mq_alloc_request(q, hctx, ctx, rw, GFP_ATOMIC, false);
1175 if (unlikely(!rq)) {
1176 __blk_mq_run_hw_queue(hctx);
1177 blk_mq_put_ctx(ctx);
1178 trace_block_sleeprq(q, bio, rw);
1180 ctx = blk_mq_get_ctx(q);
1181 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1182 rq = __blk_mq_alloc_request(q, hctx, ctx, rw,
1183 __GFP_WAIT|GFP_ATOMIC, false);
1193 * Multiple hardware queue variant. This will not use per-process plugs,
1194 * but will attempt to bypass the hctx queueing if we can go straight to
1195 * hardware for SYNC IO.
1197 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1199 const int is_sync = rw_is_sync(bio->bi_rw);
1200 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1201 struct blk_map_ctx data;
1204 blk_queue_bounce(q, &bio);
1206 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1207 bio_endio(bio, -EIO);
1211 rq = blk_mq_map_request(q, bio, &data);
1215 if (unlikely(is_flush_fua)) {
1216 blk_mq_bio_to_request(rq, bio);
1217 blk_insert_flush(rq);
1224 blk_mq_bio_to_request(rq, bio);
1225 blk_mq_start_request(rq, true);
1229 * For OK queue, we are done. For error, kill it. Any other
1230 * error (busy), just add it to our list as we previously
1233 ret = q->mq_ops->queue_rq(data.hctx, rq);
1234 if (ret == BLK_MQ_RQ_QUEUE_OK)
1237 __blk_mq_requeue_request(rq);
1239 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1241 blk_mq_end_io(rq, rq->errors);
1247 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1249 * For a SYNC request, send it to the hardware immediately. For
1250 * an ASYNC request, just ensure that we run it later on. The
1251 * latter allows for merging opportunities and more efficient
1255 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1258 blk_mq_put_ctx(data.ctx);
1262 * Single hardware queue variant. This will attempt to use any per-process
1263 * plug for merging and IO deferral.
1265 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1267 const int is_sync = rw_is_sync(bio->bi_rw);
1268 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1269 unsigned int use_plug, request_count = 0;
1270 struct blk_map_ctx data;
1274 * If we have multiple hardware queues, just go directly to
1275 * one of those for sync IO.
1277 use_plug = !is_flush_fua && !is_sync;
1279 blk_queue_bounce(q, &bio);
1281 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1282 bio_endio(bio, -EIO);
1286 if (use_plug && !blk_queue_nomerges(q) &&
1287 blk_attempt_plug_merge(q, bio, &request_count))
1290 rq = blk_mq_map_request(q, bio, &data);
1292 if (unlikely(is_flush_fua)) {
1293 blk_mq_bio_to_request(rq, bio);
1294 blk_insert_flush(rq);
1299 * A task plug currently exists. Since this is completely lockless,
1300 * utilize that to temporarily store requests until the task is
1301 * either done or scheduled away.
1304 struct blk_plug *plug = current->plug;
1307 blk_mq_bio_to_request(rq, bio);
1308 if (list_empty(&plug->mq_list))
1309 trace_block_plug(q);
1310 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1311 blk_flush_plug_list(plug, false);
1312 trace_block_plug(q);
1314 list_add_tail(&rq->queuelist, &plug->mq_list);
1315 blk_mq_put_ctx(data.ctx);
1320 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1322 * For a SYNC request, send it to the hardware immediately. For
1323 * an ASYNC request, just ensure that we run it later on. The
1324 * latter allows for merging opportunities and more efficient
1328 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1331 blk_mq_put_ctx(data.ctx);
1335 * Default mapping to a software queue, since we use one per CPU.
1337 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1339 return q->queue_hw_ctx[q->mq_map[cpu]];
1341 EXPORT_SYMBOL(blk_mq_map_queue);
1343 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1344 struct blk_mq_tags *tags, unsigned int hctx_idx)
1348 if (tags->rqs && set->ops->exit_request) {
1351 for (i = 0; i < tags->nr_tags; i++) {
1354 set->ops->exit_request(set->driver_data, tags->rqs[i],
1359 while (!list_empty(&tags->page_list)) {
1360 page = list_first_entry(&tags->page_list, struct page, lru);
1361 list_del_init(&page->lru);
1362 __free_pages(page, page->private);
1367 blk_mq_free_tags(tags);
1370 static size_t order_to_size(unsigned int order)
1372 return (size_t)PAGE_SIZE << order;
1375 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1376 unsigned int hctx_idx)
1378 struct blk_mq_tags *tags;
1379 unsigned int i, j, entries_per_page, max_order = 4;
1380 size_t rq_size, left;
1382 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1387 INIT_LIST_HEAD(&tags->page_list);
1389 tags->rqs = kmalloc_node(set->queue_depth * sizeof(struct request *),
1390 GFP_KERNEL, set->numa_node);
1392 blk_mq_free_tags(tags);
1397 * rq_size is the size of the request plus driver payload, rounded
1398 * to the cacheline size
1400 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1402 left = rq_size * set->queue_depth;
1404 for (i = 0; i < set->queue_depth; ) {
1405 int this_order = max_order;
1410 while (left < order_to_size(this_order - 1) && this_order)
1414 page = alloc_pages_node(set->numa_node, GFP_KERNEL,
1420 if (order_to_size(this_order) < rq_size)
1427 page->private = this_order;
1428 list_add_tail(&page->lru, &tags->page_list);
1430 p = page_address(page);
1431 entries_per_page = order_to_size(this_order) / rq_size;
1432 to_do = min(entries_per_page, set->queue_depth - i);
1433 left -= to_do * rq_size;
1434 for (j = 0; j < to_do; j++) {
1436 if (set->ops->init_request) {
1437 if (set->ops->init_request(set->driver_data,
1438 tags->rqs[i], hctx_idx, i,
1451 pr_warn("%s: failed to allocate requests\n", __func__);
1452 blk_mq_free_rq_map(set, tags, hctx_idx);
1456 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1461 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1463 unsigned int bpw = 8, total, num_maps, i;
1465 bitmap->bits_per_word = bpw;
1467 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1468 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1473 bitmap->map_size = num_maps;
1476 for (i = 0; i < num_maps; i++) {
1477 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1478 total -= bitmap->map[i].depth;
1484 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1486 struct request_queue *q = hctx->queue;
1487 struct blk_mq_ctx *ctx;
1491 * Move ctx entries to new CPU, if this one is going away.
1493 ctx = __blk_mq_get_ctx(q, cpu);
1495 spin_lock(&ctx->lock);
1496 if (!list_empty(&ctx->rq_list)) {
1497 list_splice_init(&ctx->rq_list, &tmp);
1498 blk_mq_hctx_clear_pending(hctx, ctx);
1500 spin_unlock(&ctx->lock);
1502 if (list_empty(&tmp))
1505 ctx = blk_mq_get_ctx(q);
1506 spin_lock(&ctx->lock);
1508 while (!list_empty(&tmp)) {
1511 rq = list_first_entry(&tmp, struct request, queuelist);
1513 list_move_tail(&rq->queuelist, &ctx->rq_list);
1516 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1517 blk_mq_hctx_mark_pending(hctx, ctx);
1519 spin_unlock(&ctx->lock);
1521 blk_mq_run_hw_queue(hctx, true);
1522 blk_mq_put_ctx(ctx);
1526 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1528 struct request_queue *q = hctx->queue;
1529 struct blk_mq_tag_set *set = q->tag_set;
1531 if (set->tags[hctx->queue_num])
1534 set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1535 if (!set->tags[hctx->queue_num])
1538 hctx->tags = set->tags[hctx->queue_num];
1542 static int blk_mq_hctx_notify(void *data, unsigned long action,
1545 struct blk_mq_hw_ctx *hctx = data;
1547 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1548 return blk_mq_hctx_cpu_offline(hctx, cpu);
1549 else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1550 return blk_mq_hctx_cpu_online(hctx, cpu);
1555 static void blk_mq_exit_hw_queues(struct request_queue *q,
1556 struct blk_mq_tag_set *set, int nr_queue)
1558 struct blk_mq_hw_ctx *hctx;
1561 queue_for_each_hw_ctx(q, hctx, i) {
1565 if (set->ops->exit_hctx)
1566 set->ops->exit_hctx(hctx, i);
1568 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1570 blk_mq_free_bitmap(&hctx->ctx_map);
1575 static void blk_mq_free_hw_queues(struct request_queue *q,
1576 struct blk_mq_tag_set *set)
1578 struct blk_mq_hw_ctx *hctx;
1581 queue_for_each_hw_ctx(q, hctx, i) {
1582 free_cpumask_var(hctx->cpumask);
1587 static int blk_mq_init_hw_queues(struct request_queue *q,
1588 struct blk_mq_tag_set *set)
1590 struct blk_mq_hw_ctx *hctx;
1594 * Initialize hardware queues
1596 queue_for_each_hw_ctx(q, hctx, i) {
1599 node = hctx->numa_node;
1600 if (node == NUMA_NO_NODE)
1601 node = hctx->numa_node = set->numa_node;
1603 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1604 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1605 spin_lock_init(&hctx->lock);
1606 INIT_LIST_HEAD(&hctx->dispatch);
1608 hctx->queue_num = i;
1609 hctx->flags = set->flags;
1610 hctx->cmd_size = set->cmd_size;
1612 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1613 blk_mq_hctx_notify, hctx);
1614 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1616 hctx->tags = set->tags[i];
1619 * Allocate space for all possible cpus to avoid allocation in
1622 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1627 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1632 if (set->ops->init_hctx &&
1633 set->ops->init_hctx(hctx, set->driver_data, i))
1637 if (i == q->nr_hw_queues)
1643 blk_mq_exit_hw_queues(q, set, i);
1648 static void blk_mq_init_cpu_queues(struct request_queue *q,
1649 unsigned int nr_hw_queues)
1653 for_each_possible_cpu(i) {
1654 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1655 struct blk_mq_hw_ctx *hctx;
1657 memset(__ctx, 0, sizeof(*__ctx));
1659 spin_lock_init(&__ctx->lock);
1660 INIT_LIST_HEAD(&__ctx->rq_list);
1663 /* If the cpu isn't online, the cpu is mapped to first hctx */
1667 hctx = q->mq_ops->map_queue(q, i);
1668 cpumask_set_cpu(i, hctx->cpumask);
1672 * Set local node, IFF we have more than one hw queue. If
1673 * not, we remain on the home node of the device
1675 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1676 hctx->numa_node = cpu_to_node(i);
1680 static void blk_mq_map_swqueue(struct request_queue *q)
1683 struct blk_mq_hw_ctx *hctx;
1684 struct blk_mq_ctx *ctx;
1686 queue_for_each_hw_ctx(q, hctx, i) {
1687 cpumask_clear(hctx->cpumask);
1692 * Map software to hardware queues
1694 queue_for_each_ctx(q, ctx, i) {
1695 /* If the cpu isn't online, the cpu is mapped to first hctx */
1699 hctx = q->mq_ops->map_queue(q, i);
1700 cpumask_set_cpu(i, hctx->cpumask);
1701 ctx->index_hw = hctx->nr_ctx;
1702 hctx->ctxs[hctx->nr_ctx++] = ctx;
1705 queue_for_each_hw_ctx(q, hctx, i) {
1707 * If not software queues are mapped to this hardware queue,
1708 * disable it and free the request entries
1710 if (!hctx->nr_ctx) {
1711 struct blk_mq_tag_set *set = q->tag_set;
1714 blk_mq_free_rq_map(set, set->tags[i], i);
1715 set->tags[i] = NULL;
1722 * Initialize batch roundrobin counts
1724 hctx->next_cpu = cpumask_first(hctx->cpumask);
1725 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1729 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1731 struct blk_mq_hw_ctx *hctx;
1732 struct request_queue *q;
1736 if (set->tag_list.next == set->tag_list.prev)
1741 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1742 blk_mq_freeze_queue(q);
1744 queue_for_each_hw_ctx(q, hctx, i) {
1746 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1748 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1750 blk_mq_unfreeze_queue(q);
1754 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1756 struct blk_mq_tag_set *set = q->tag_set;
1758 blk_mq_freeze_queue(q);
1760 mutex_lock(&set->tag_list_lock);
1761 list_del_init(&q->tag_set_list);
1762 blk_mq_update_tag_set_depth(set);
1763 mutex_unlock(&set->tag_list_lock);
1765 blk_mq_unfreeze_queue(q);
1768 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1769 struct request_queue *q)
1773 mutex_lock(&set->tag_list_lock);
1774 list_add_tail(&q->tag_set_list, &set->tag_list);
1775 blk_mq_update_tag_set_depth(set);
1776 mutex_unlock(&set->tag_list_lock);
1779 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1781 struct blk_mq_hw_ctx **hctxs;
1782 struct blk_mq_ctx *ctx;
1783 struct request_queue *q;
1787 ctx = alloc_percpu(struct blk_mq_ctx);
1789 return ERR_PTR(-ENOMEM);
1791 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1797 map = blk_mq_make_queue_map(set);
1801 for (i = 0; i < set->nr_hw_queues; i++) {
1802 int node = blk_mq_hw_queue_to_node(map, i);
1804 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1809 if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL))
1812 atomic_set(&hctxs[i]->nr_active, 0);
1813 hctxs[i]->numa_node = node;
1814 hctxs[i]->queue_num = i;
1817 q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1821 if (percpu_counter_init(&q->mq_usage_counter, 0))
1824 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1825 blk_queue_rq_timeout(q, 30000);
1827 q->nr_queues = nr_cpu_ids;
1828 q->nr_hw_queues = set->nr_hw_queues;
1832 q->queue_hw_ctx = hctxs;
1834 q->mq_ops = set->ops;
1835 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1837 if (!(set->flags & BLK_MQ_F_SG_MERGE))
1838 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1840 q->sg_reserved_size = INT_MAX;
1842 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1843 INIT_LIST_HEAD(&q->requeue_list);
1844 spin_lock_init(&q->requeue_lock);
1846 if (q->nr_hw_queues > 1)
1847 blk_queue_make_request(q, blk_mq_make_request);
1849 blk_queue_make_request(q, blk_sq_make_request);
1851 blk_queue_rq_timed_out(q, blk_mq_rq_timed_out);
1853 blk_queue_rq_timeout(q, set->timeout);
1856 * Do this after blk_queue_make_request() overrides it...
1858 q->nr_requests = set->queue_depth;
1860 if (set->ops->complete)
1861 blk_queue_softirq_done(q, set->ops->complete);
1863 blk_mq_init_flush(q);
1864 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1866 q->flush_rq = kzalloc(round_up(sizeof(struct request) +
1867 set->cmd_size, cache_line_size()),
1872 if (blk_mq_init_hw_queues(q, set))
1875 mutex_lock(&all_q_mutex);
1876 list_add_tail(&q->all_q_node, &all_q_list);
1877 mutex_unlock(&all_q_mutex);
1879 blk_mq_add_queue_tag_set(set, q);
1881 blk_mq_map_swqueue(q);
1888 blk_cleanup_queue(q);
1891 for (i = 0; i < set->nr_hw_queues; i++) {
1894 free_cpumask_var(hctxs[i]->cpumask);
1901 return ERR_PTR(-ENOMEM);
1903 EXPORT_SYMBOL(blk_mq_init_queue);
1905 void blk_mq_free_queue(struct request_queue *q)
1907 struct blk_mq_tag_set *set = q->tag_set;
1909 blk_mq_del_queue_tag_set(q);
1911 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
1912 blk_mq_free_hw_queues(q, set);
1914 percpu_counter_destroy(&q->mq_usage_counter);
1916 free_percpu(q->queue_ctx);
1917 kfree(q->queue_hw_ctx);
1920 q->queue_ctx = NULL;
1921 q->queue_hw_ctx = NULL;
1924 mutex_lock(&all_q_mutex);
1925 list_del_init(&q->all_q_node);
1926 mutex_unlock(&all_q_mutex);
1929 /* Basically redo blk_mq_init_queue with queue frozen */
1930 static void blk_mq_queue_reinit(struct request_queue *q)
1932 blk_mq_freeze_queue(q);
1934 blk_mq_sysfs_unregister(q);
1936 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1939 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1940 * we should change hctx numa_node according to new topology (this
1941 * involves free and re-allocate memory, worthy doing?)
1944 blk_mq_map_swqueue(q);
1946 blk_mq_sysfs_register(q);
1948 blk_mq_unfreeze_queue(q);
1951 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1952 unsigned long action, void *hcpu)
1954 struct request_queue *q;
1957 * Before new mappings are established, hotadded cpu might already
1958 * start handling requests. This doesn't break anything as we map
1959 * offline CPUs to first hardware queue. We will re-init the queue
1960 * below to get optimal settings.
1962 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1963 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1966 mutex_lock(&all_q_mutex);
1967 list_for_each_entry(q, &all_q_list, all_q_node)
1968 blk_mq_queue_reinit(q);
1969 mutex_unlock(&all_q_mutex);
1973 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
1977 if (!set->nr_hw_queues)
1979 if (!set->queue_depth || set->queue_depth > BLK_MQ_MAX_DEPTH)
1981 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
1984 if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue)
1988 set->tags = kmalloc_node(set->nr_hw_queues *
1989 sizeof(struct blk_mq_tags *),
1990 GFP_KERNEL, set->numa_node);
1994 for (i = 0; i < set->nr_hw_queues; i++) {
1995 set->tags[i] = blk_mq_init_rq_map(set, i);
2000 mutex_init(&set->tag_list_lock);
2001 INIT_LIST_HEAD(&set->tag_list);
2007 blk_mq_free_rq_map(set, set->tags[i], i);
2011 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2013 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2017 for (i = 0; i < set->nr_hw_queues; i++) {
2019 blk_mq_free_rq_map(set, set->tags[i], i);
2024 EXPORT_SYMBOL(blk_mq_free_tag_set);
2026 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2028 struct blk_mq_tag_set *set = q->tag_set;
2029 struct blk_mq_hw_ctx *hctx;
2032 if (!set || nr > set->queue_depth)
2036 queue_for_each_hw_ctx(q, hctx, i) {
2037 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2043 q->nr_requests = nr;
2048 void blk_mq_disable_hotplug(void)
2050 mutex_lock(&all_q_mutex);
2053 void blk_mq_enable_hotplug(void)
2055 mutex_unlock(&all_q_mutex);
2058 static int __init blk_mq_init(void)
2062 /* Must be called after percpu_counter_hotcpu_callback() */
2063 hotcpu_notifier(blk_mq_queue_reinit_notify, -10);
2067 subsys_initcall(blk_mq_init);