1 #include <linux/kernel.h>
2 #include <linux/module.h>
3 #include <linux/backing-dev.h>
5 #include <linux/blkdev.h>
7 #include <linux/init.h>
8 #include <linux/slab.h>
9 #include <linux/workqueue.h>
10 #include <linux/smp.h>
11 #include <linux/llist.h>
12 #include <linux/list_sort.h>
13 #include <linux/cpu.h>
14 #include <linux/cache.h>
15 #include <linux/sched/sysctl.h>
16 #include <linux/delay.h>
18 #include <trace/events/block.h>
20 #include <linux/blk-mq.h>
23 #include "blk-mq-tag.h"
25 static DEFINE_MUTEX(all_q_mutex);
26 static LIST_HEAD(all_q_list);
28 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
30 static struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
33 return per_cpu_ptr(q->queue_ctx, cpu);
37 * This assumes per-cpu software queueing queues. They could be per-node
38 * as well, for instance. For now this is hardcoded as-is. Note that we don't
39 * care about preemption, since we know the ctx's are persistent. This does
40 * mean that we can't rely on ctx always matching the currently running CPU.
42 static struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
44 return __blk_mq_get_ctx(q, get_cpu());
47 static void blk_mq_put_ctx(struct blk_mq_ctx *ctx)
53 * Check if any of the ctx's have pending work in this hardware queue
55 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
59 for (i = 0; i < hctx->ctx_map.map_size; i++)
60 if (hctx->ctx_map.map[i].word)
66 static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx,
67 struct blk_mq_ctx *ctx)
69 return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word];
72 #define CTX_TO_BIT(hctx, ctx) \
73 ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
76 * Mark this ctx as having pending work in this hardware queue
78 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
79 struct blk_mq_ctx *ctx)
81 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
83 if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word))
84 set_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
87 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
88 struct blk_mq_ctx *ctx)
90 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
92 clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
95 static int blk_mq_queue_enter(struct request_queue *q)
99 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
101 /* we have problems to freeze the queue if it's initializing */
102 if (!blk_queue_bypass(q) || !blk_queue_init_done(q))
105 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
107 spin_lock_irq(q->queue_lock);
108 ret = wait_event_interruptible_lock_irq(q->mq_freeze_wq,
109 !blk_queue_bypass(q) || blk_queue_dying(q),
111 /* inc usage with lock hold to avoid freeze_queue runs here */
112 if (!ret && !blk_queue_dying(q))
113 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
114 else if (blk_queue_dying(q))
116 spin_unlock_irq(q->queue_lock);
121 static void blk_mq_queue_exit(struct request_queue *q)
123 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
126 static void __blk_mq_drain_queue(struct request_queue *q)
131 spin_lock_irq(q->queue_lock);
132 count = percpu_counter_sum(&q->mq_usage_counter);
133 spin_unlock_irq(q->queue_lock);
137 blk_mq_run_queues(q, false);
143 * Guarantee no request is in use, so we can change any data structure of
144 * the queue afterward.
146 static void blk_mq_freeze_queue(struct request_queue *q)
150 spin_lock_irq(q->queue_lock);
151 drain = !q->bypass_depth++;
152 queue_flag_set(QUEUE_FLAG_BYPASS, q);
153 spin_unlock_irq(q->queue_lock);
156 __blk_mq_drain_queue(q);
159 void blk_mq_drain_queue(struct request_queue *q)
161 __blk_mq_drain_queue(q);
164 static void blk_mq_unfreeze_queue(struct request_queue *q)
168 spin_lock_irq(q->queue_lock);
169 if (!--q->bypass_depth) {
170 queue_flag_clear(QUEUE_FLAG_BYPASS, q);
173 WARN_ON_ONCE(q->bypass_depth < 0);
174 spin_unlock_irq(q->queue_lock);
176 wake_up_all(&q->mq_freeze_wq);
179 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
181 return blk_mq_has_free_tags(hctx->tags);
183 EXPORT_SYMBOL(blk_mq_can_queue);
185 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
186 struct request *rq, unsigned int rw_flags)
188 if (blk_queue_io_stat(q))
189 rw_flags |= REQ_IO_STAT;
191 INIT_LIST_HEAD(&rq->queuelist);
192 /* csd/requeue_work/fifo_time is initialized before use */
195 rq->cmd_flags |= rw_flags;
197 /* do not touch atomic flags, it needs atomic ops against the timer */
200 rq->__sector = (sector_t) -1;
203 INIT_HLIST_NODE(&rq->hash);
204 RB_CLEAR_NODE(&rq->rb_node);
205 memset(&rq->flush, 0, max(sizeof(rq->flush), sizeof(rq->elv)));
208 rq->start_time = jiffies;
209 #ifdef CONFIG_BLK_CGROUP
211 set_start_time_ns(rq);
212 rq->io_start_time_ns = 0;
214 rq->nr_phys_segments = 0;
215 #if defined(CONFIG_BLK_DEV_INTEGRITY)
216 rq->nr_integrity_segments = 0;
220 /* tag was already set */
222 memset(rq->__cmd, 0, sizeof(rq->__cmd));
224 rq->cmd_len = BLK_MAX_CDB;
232 INIT_LIST_HEAD(&rq->timeout_list);
236 rq->end_io_data = NULL;
239 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
242 static struct request *
243 __blk_mq_alloc_request(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
244 struct blk_mq_ctx *ctx, int rw, gfp_t gfp, bool reserved)
249 tag = blk_mq_get_tag(hctx, &ctx->last_tag, gfp, reserved);
250 if (tag != BLK_MQ_TAG_FAIL) {
251 rq = hctx->tags->rqs[tag];
254 if (blk_mq_tag_busy(hctx)) {
255 rq->cmd_flags = REQ_MQ_INFLIGHT;
256 atomic_inc(&hctx->nr_active);
260 blk_mq_rq_ctx_init(q, ctx, rq, rw);
267 static struct request *blk_mq_alloc_request_pinned(struct request_queue *q,
271 bool gfp_mask = gfp & ~__GFP_WAIT;
275 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
276 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q, ctx->cpu);
278 rq = __blk_mq_alloc_request(q, hctx, ctx, rw, gfp_mask,
283 if (!(gfp & __GFP_WAIT)) {
288 __blk_mq_run_hw_queue(hctx);
296 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
301 if (blk_mq_queue_enter(q))
304 rq = blk_mq_alloc_request_pinned(q, rw, gfp, reserved);
306 blk_mq_put_ctx(rq->mq_ctx);
309 EXPORT_SYMBOL(blk_mq_alloc_request);
311 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
312 struct blk_mq_ctx *ctx, struct request *rq)
314 const int tag = rq->tag;
315 struct request_queue *q = rq->q;
317 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
318 atomic_dec(&hctx->nr_active);
320 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
321 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
322 blk_mq_queue_exit(q);
325 void blk_mq_free_request(struct request *rq)
327 struct blk_mq_ctx *ctx = rq->mq_ctx;
328 struct blk_mq_hw_ctx *hctx;
329 struct request_queue *q = rq->q;
331 ctx->rq_completed[rq_is_sync(rq)]++;
333 hctx = q->mq_ops->map_queue(q, ctx->cpu);
334 __blk_mq_free_request(hctx, ctx, rq);
338 * Clone all relevant state from a request that has been put on hold in
339 * the flush state machine into the preallocated flush request that hangs
340 * off the request queue.
342 * For a driver the flush request should be invisible, that's why we are
343 * impersonating the original request here.
345 void blk_mq_clone_flush_request(struct request *flush_rq,
346 struct request *orig_rq)
348 struct blk_mq_hw_ctx *hctx =
349 orig_rq->q->mq_ops->map_queue(orig_rq->q, orig_rq->mq_ctx->cpu);
351 flush_rq->mq_ctx = orig_rq->mq_ctx;
352 flush_rq->tag = orig_rq->tag;
353 memcpy(blk_mq_rq_to_pdu(flush_rq), blk_mq_rq_to_pdu(orig_rq),
357 inline void __blk_mq_end_io(struct request *rq, int error)
359 blk_account_io_done(rq);
362 rq->end_io(rq, error);
364 if (unlikely(blk_bidi_rq(rq)))
365 blk_mq_free_request(rq->next_rq);
366 blk_mq_free_request(rq);
369 EXPORT_SYMBOL(__blk_mq_end_io);
371 void blk_mq_end_io(struct request *rq, int error)
373 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
375 __blk_mq_end_io(rq, error);
377 EXPORT_SYMBOL(blk_mq_end_io);
379 static void __blk_mq_complete_request_remote(void *data)
381 struct request *rq = data;
383 rq->q->softirq_done_fn(rq);
386 void __blk_mq_complete_request(struct request *rq)
388 struct blk_mq_ctx *ctx = rq->mq_ctx;
392 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
393 rq->q->softirq_done_fn(rq);
398 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
399 shared = cpus_share_cache(cpu, ctx->cpu);
401 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
402 rq->csd.func = __blk_mq_complete_request_remote;
405 smp_call_function_single_async(ctx->cpu, &rq->csd);
407 rq->q->softirq_done_fn(rq);
413 * blk_mq_complete_request - end I/O on a request
414 * @rq: the request being processed
417 * Ends all I/O on a request. It does not handle partial completions.
418 * The actual completion happens out-of-order, through a IPI handler.
420 void blk_mq_complete_request(struct request *rq)
422 struct request_queue *q = rq->q;
424 if (unlikely(blk_should_fake_timeout(q)))
426 if (!blk_mark_rq_complete(rq)) {
427 if (q->softirq_done_fn)
428 __blk_mq_complete_request(rq);
430 blk_mq_end_io(rq, rq->errors);
433 EXPORT_SYMBOL(blk_mq_complete_request);
435 static void blk_mq_start_request(struct request *rq, bool last)
437 struct request_queue *q = rq->q;
439 trace_block_rq_issue(q, rq);
441 rq->resid_len = blk_rq_bytes(rq);
442 if (unlikely(blk_bidi_rq(rq)))
443 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
446 * Just mark start time and set the started bit. Due to memory
447 * ordering, we know we'll see the correct deadline as long as
448 * REQ_ATOMIC_STARTED is seen. Use the default queue timeout,
449 * unless one has been set in the request.
452 rq->deadline = jiffies + q->rq_timeout;
454 rq->deadline = jiffies + rq->timeout;
457 * Mark us as started and clear complete. Complete might have been
458 * set if requeue raced with timeout, which then marked it as
459 * complete. So be sure to clear complete again when we start
460 * the request, otherwise we'll ignore the completion event.
462 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
463 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
465 if (q->dma_drain_size && blk_rq_bytes(rq)) {
467 * Make sure space for the drain appears. We know we can do
468 * this because max_hw_segments has been adjusted to be one
469 * fewer than the device can handle.
471 rq->nr_phys_segments++;
475 * Flag the last request in the series so that drivers know when IO
476 * should be kicked off, if they don't do it on a per-request basis.
478 * Note: the flag isn't the only condition drivers should do kick off.
479 * If drive is busy, the last request might not have the bit set.
482 rq->cmd_flags |= REQ_END;
485 static void __blk_mq_requeue_request(struct request *rq)
487 struct request_queue *q = rq->q;
489 trace_block_rq_requeue(q, rq);
490 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
492 rq->cmd_flags &= ~REQ_END;
494 if (q->dma_drain_size && blk_rq_bytes(rq))
495 rq->nr_phys_segments--;
498 void blk_mq_requeue_request(struct request *rq)
500 __blk_mq_requeue_request(rq);
501 blk_clear_rq_complete(rq);
503 BUG_ON(blk_queued_rq(rq));
504 blk_mq_add_to_requeue_list(rq, true);
506 EXPORT_SYMBOL(blk_mq_requeue_request);
508 static void blk_mq_requeue_work(struct work_struct *work)
510 struct request_queue *q =
511 container_of(work, struct request_queue, requeue_work);
513 struct request *rq, *next;
516 spin_lock_irqsave(&q->requeue_lock, flags);
517 list_splice_init(&q->requeue_list, &rq_list);
518 spin_unlock_irqrestore(&q->requeue_lock, flags);
520 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
521 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
524 rq->cmd_flags &= ~REQ_SOFTBARRIER;
525 list_del_init(&rq->queuelist);
526 blk_mq_insert_request(rq, true, false, false);
529 while (!list_empty(&rq_list)) {
530 rq = list_entry(rq_list.next, struct request, queuelist);
531 list_del_init(&rq->queuelist);
532 blk_mq_insert_request(rq, false, false, false);
535 blk_mq_run_queues(q, false);
538 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
540 struct request_queue *q = rq->q;
544 * We abuse this flag that is otherwise used by the I/O scheduler to
545 * request head insertation from the workqueue.
547 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
549 spin_lock_irqsave(&q->requeue_lock, flags);
551 rq->cmd_flags |= REQ_SOFTBARRIER;
552 list_add(&rq->queuelist, &q->requeue_list);
554 list_add_tail(&rq->queuelist, &q->requeue_list);
556 spin_unlock_irqrestore(&q->requeue_lock, flags);
558 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
560 void blk_mq_kick_requeue_list(struct request_queue *q)
562 kblockd_schedule_work(&q->requeue_work);
564 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
566 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
568 return tags->rqs[tag];
570 EXPORT_SYMBOL(blk_mq_tag_to_rq);
572 struct blk_mq_timeout_data {
573 struct blk_mq_hw_ctx *hctx;
575 unsigned int *next_set;
578 static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
580 struct blk_mq_timeout_data *data = __data;
581 struct blk_mq_hw_ctx *hctx = data->hctx;
584 /* It may not be in flight yet (this is where
585 * the REQ_ATOMIC_STARTED flag comes in). The requests are
586 * statically allocated, so we know it's always safe to access the
587 * memory associated with a bit offset into ->rqs[].
593 tag = find_next_zero_bit(free_tags, hctx->tags->nr_tags, tag);
594 if (tag >= hctx->tags->nr_tags)
597 rq = blk_mq_tag_to_rq(hctx->tags, tag++);
598 if (rq->q != hctx->queue)
600 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
603 blk_rq_check_expired(rq, data->next, data->next_set);
607 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
609 unsigned int *next_set)
611 struct blk_mq_timeout_data data = {
614 .next_set = next_set,
618 * Ask the tagging code to iterate busy requests, so we can
619 * check them for timeout.
621 blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
624 static enum blk_eh_timer_return blk_mq_rq_timed_out(struct request *rq)
626 struct request_queue *q = rq->q;
629 * We know that complete is set at this point. If STARTED isn't set
630 * anymore, then the request isn't active and the "timeout" should
631 * just be ignored. This can happen due to the bitflag ordering.
632 * Timeout first checks if STARTED is set, and if it is, assumes
633 * the request is active. But if we race with completion, then
634 * we both flags will get cleared. So check here again, and ignore
635 * a timeout event with a request that isn't active.
637 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
638 return BLK_EH_NOT_HANDLED;
640 if (!q->mq_ops->timeout)
641 return BLK_EH_RESET_TIMER;
643 return q->mq_ops->timeout(rq);
646 static void blk_mq_rq_timer(unsigned long data)
648 struct request_queue *q = (struct request_queue *) data;
649 struct blk_mq_hw_ctx *hctx;
650 unsigned long next = 0;
653 queue_for_each_hw_ctx(q, hctx, i) {
655 * If not software queues are currently mapped to this
656 * hardware queue, there's nothing to check
658 if (!hctx->nr_ctx || !hctx->tags)
661 blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
665 next = blk_rq_timeout(round_jiffies_up(next));
666 mod_timer(&q->timeout, next);
668 queue_for_each_hw_ctx(q, hctx, i)
669 blk_mq_tag_idle(hctx);
674 * Reverse check our software queue for entries that we could potentially
675 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
676 * too much time checking for merges.
678 static bool blk_mq_attempt_merge(struct request_queue *q,
679 struct blk_mq_ctx *ctx, struct bio *bio)
684 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
690 if (!blk_rq_merge_ok(rq, bio))
693 el_ret = blk_try_merge(rq, bio);
694 if (el_ret == ELEVATOR_BACK_MERGE) {
695 if (bio_attempt_back_merge(q, rq, bio)) {
700 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
701 if (bio_attempt_front_merge(q, rq, bio)) {
713 * Process software queues that have been marked busy, splicing them
714 * to the for-dispatch
716 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
718 struct blk_mq_ctx *ctx;
721 for (i = 0; i < hctx->ctx_map.map_size; i++) {
722 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
723 unsigned int off, bit;
729 off = i * hctx->ctx_map.bits_per_word;
731 bit = find_next_bit(&bm->word, bm->depth, bit);
732 if (bit >= bm->depth)
735 ctx = hctx->ctxs[bit + off];
736 clear_bit(bit, &bm->word);
737 spin_lock(&ctx->lock);
738 list_splice_tail_init(&ctx->rq_list, list);
739 spin_unlock(&ctx->lock);
747 * Run this hardware queue, pulling any software queues mapped to it in.
748 * Note that this function currently has various problems around ordering
749 * of IO. In particular, we'd like FIFO behaviour on handling existing
750 * items on the hctx->dispatch list. Ignore that for now.
752 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
754 struct request_queue *q = hctx->queue;
759 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
761 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
767 * Touch any software queue that has pending entries.
769 flush_busy_ctxs(hctx, &rq_list);
772 * If we have previous entries on our dispatch list, grab them
773 * and stuff them at the front for more fair dispatch.
775 if (!list_empty_careful(&hctx->dispatch)) {
776 spin_lock(&hctx->lock);
777 if (!list_empty(&hctx->dispatch))
778 list_splice_init(&hctx->dispatch, &rq_list);
779 spin_unlock(&hctx->lock);
783 * Now process all the entries, sending them to the driver.
786 while (!list_empty(&rq_list)) {
789 rq = list_first_entry(&rq_list, struct request, queuelist);
790 list_del_init(&rq->queuelist);
792 blk_mq_start_request(rq, list_empty(&rq_list));
794 ret = q->mq_ops->queue_rq(hctx, rq);
796 case BLK_MQ_RQ_QUEUE_OK:
799 case BLK_MQ_RQ_QUEUE_BUSY:
800 list_add(&rq->queuelist, &rq_list);
801 __blk_mq_requeue_request(rq);
804 pr_err("blk-mq: bad return on queue: %d\n", ret);
805 case BLK_MQ_RQ_QUEUE_ERROR:
807 blk_mq_end_io(rq, rq->errors);
811 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
816 hctx->dispatched[0]++;
817 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
818 hctx->dispatched[ilog2(queued) + 1]++;
821 * Any items that need requeuing? Stuff them into hctx->dispatch,
822 * that is where we will continue on next queue run.
824 if (!list_empty(&rq_list)) {
825 spin_lock(&hctx->lock);
826 list_splice(&rq_list, &hctx->dispatch);
827 spin_unlock(&hctx->lock);
832 * It'd be great if the workqueue API had a way to pass
833 * in a mask and had some smarts for more clever placement.
834 * For now we just round-robin here, switching for every
835 * BLK_MQ_CPU_WORK_BATCH queued items.
837 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
839 int cpu = hctx->next_cpu;
841 if (--hctx->next_cpu_batch <= 0) {
844 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
845 if (next_cpu >= nr_cpu_ids)
846 next_cpu = cpumask_first(hctx->cpumask);
848 hctx->next_cpu = next_cpu;
849 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
855 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
857 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
860 if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask))
861 __blk_mq_run_hw_queue(hctx);
862 else if (hctx->queue->nr_hw_queues == 1)
863 kblockd_schedule_delayed_work(&hctx->run_work, 0);
867 cpu = blk_mq_hctx_next_cpu(hctx);
868 kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0);
872 void blk_mq_run_queues(struct request_queue *q, bool async)
874 struct blk_mq_hw_ctx *hctx;
877 queue_for_each_hw_ctx(q, hctx, i) {
878 if ((!blk_mq_hctx_has_pending(hctx) &&
879 list_empty_careful(&hctx->dispatch)) ||
880 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
884 blk_mq_run_hw_queue(hctx, async);
888 EXPORT_SYMBOL(blk_mq_run_queues);
890 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
892 cancel_delayed_work(&hctx->run_work);
893 cancel_delayed_work(&hctx->delay_work);
894 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
896 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
898 void blk_mq_stop_hw_queues(struct request_queue *q)
900 struct blk_mq_hw_ctx *hctx;
903 queue_for_each_hw_ctx(q, hctx, i)
904 blk_mq_stop_hw_queue(hctx);
906 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
908 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
910 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
913 __blk_mq_run_hw_queue(hctx);
916 EXPORT_SYMBOL(blk_mq_start_hw_queue);
918 void blk_mq_start_hw_queues(struct request_queue *q)
920 struct blk_mq_hw_ctx *hctx;
923 queue_for_each_hw_ctx(q, hctx, i)
924 blk_mq_start_hw_queue(hctx);
926 EXPORT_SYMBOL(blk_mq_start_hw_queues);
929 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
931 struct blk_mq_hw_ctx *hctx;
934 queue_for_each_hw_ctx(q, hctx, i) {
935 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
938 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
940 blk_mq_run_hw_queue(hctx, async);
944 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
946 static void blk_mq_run_work_fn(struct work_struct *work)
948 struct blk_mq_hw_ctx *hctx;
950 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
952 __blk_mq_run_hw_queue(hctx);
955 static void blk_mq_delay_work_fn(struct work_struct *work)
957 struct blk_mq_hw_ctx *hctx;
959 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
961 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
962 __blk_mq_run_hw_queue(hctx);
965 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
967 unsigned long tmo = msecs_to_jiffies(msecs);
969 if (hctx->queue->nr_hw_queues == 1)
970 kblockd_schedule_delayed_work(&hctx->delay_work, tmo);
974 cpu = blk_mq_hctx_next_cpu(hctx);
975 kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo);
978 EXPORT_SYMBOL(blk_mq_delay_queue);
980 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
981 struct request *rq, bool at_head)
983 struct blk_mq_ctx *ctx = rq->mq_ctx;
985 trace_block_rq_insert(hctx->queue, rq);
988 list_add(&rq->queuelist, &ctx->rq_list);
990 list_add_tail(&rq->queuelist, &ctx->rq_list);
992 blk_mq_hctx_mark_pending(hctx, ctx);
995 * We do this early, to ensure we are on the right CPU.
1000 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1003 struct request_queue *q = rq->q;
1004 struct blk_mq_hw_ctx *hctx;
1005 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
1007 current_ctx = blk_mq_get_ctx(q);
1008 if (!cpu_online(ctx->cpu))
1009 rq->mq_ctx = ctx = current_ctx;
1011 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1013 if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA) &&
1014 !(rq->cmd_flags & (REQ_FLUSH_SEQ))) {
1015 blk_insert_flush(rq);
1017 spin_lock(&ctx->lock);
1018 __blk_mq_insert_request(hctx, rq, at_head);
1019 spin_unlock(&ctx->lock);
1023 blk_mq_run_hw_queue(hctx, async);
1025 blk_mq_put_ctx(current_ctx);
1028 static void blk_mq_insert_requests(struct request_queue *q,
1029 struct blk_mq_ctx *ctx,
1030 struct list_head *list,
1035 struct blk_mq_hw_ctx *hctx;
1036 struct blk_mq_ctx *current_ctx;
1038 trace_block_unplug(q, depth, !from_schedule);
1040 current_ctx = blk_mq_get_ctx(q);
1042 if (!cpu_online(ctx->cpu))
1044 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1047 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1050 spin_lock(&ctx->lock);
1051 while (!list_empty(list)) {
1054 rq = list_first_entry(list, struct request, queuelist);
1055 list_del_init(&rq->queuelist);
1057 __blk_mq_insert_request(hctx, rq, false);
1059 spin_unlock(&ctx->lock);
1061 blk_mq_run_hw_queue(hctx, from_schedule);
1062 blk_mq_put_ctx(current_ctx);
1065 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1067 struct request *rqa = container_of(a, struct request, queuelist);
1068 struct request *rqb = container_of(b, struct request, queuelist);
1070 return !(rqa->mq_ctx < rqb->mq_ctx ||
1071 (rqa->mq_ctx == rqb->mq_ctx &&
1072 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1075 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1077 struct blk_mq_ctx *this_ctx;
1078 struct request_queue *this_q;
1081 LIST_HEAD(ctx_list);
1084 list_splice_init(&plug->mq_list, &list);
1086 list_sort(NULL, &list, plug_ctx_cmp);
1092 while (!list_empty(&list)) {
1093 rq = list_entry_rq(list.next);
1094 list_del_init(&rq->queuelist);
1096 if (rq->mq_ctx != this_ctx) {
1098 blk_mq_insert_requests(this_q, this_ctx,
1103 this_ctx = rq->mq_ctx;
1109 list_add_tail(&rq->queuelist, &ctx_list);
1113 * If 'this_ctx' is set, we know we have entries to complete
1114 * on 'ctx_list'. Do those.
1117 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1122 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1124 init_request_from_bio(rq, bio);
1125 blk_account_io_start(rq, 1);
1128 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1129 struct blk_mq_ctx *ctx,
1130 struct request *rq, struct bio *bio)
1132 struct request_queue *q = hctx->queue;
1134 if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE)) {
1135 blk_mq_bio_to_request(rq, bio);
1136 spin_lock(&ctx->lock);
1138 __blk_mq_insert_request(hctx, rq, false);
1139 spin_unlock(&ctx->lock);
1142 spin_lock(&ctx->lock);
1143 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1144 blk_mq_bio_to_request(rq, bio);
1148 spin_unlock(&ctx->lock);
1149 __blk_mq_free_request(hctx, ctx, rq);
1154 struct blk_map_ctx {
1155 struct blk_mq_hw_ctx *hctx;
1156 struct blk_mq_ctx *ctx;
1159 static struct request *blk_mq_map_request(struct request_queue *q,
1161 struct blk_map_ctx *data)
1163 struct blk_mq_hw_ctx *hctx;
1164 struct blk_mq_ctx *ctx;
1166 int rw = bio_data_dir(bio);
1168 if (unlikely(blk_mq_queue_enter(q))) {
1169 bio_endio(bio, -EIO);
1173 ctx = blk_mq_get_ctx(q);
1174 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1176 if (rw_is_sync(bio->bi_rw))
1179 trace_block_getrq(q, bio, rw);
1180 rq = __blk_mq_alloc_request(q, hctx, ctx, rw, GFP_ATOMIC, false);
1181 if (unlikely(!rq)) {
1182 __blk_mq_run_hw_queue(hctx);
1183 blk_mq_put_ctx(ctx);
1184 trace_block_sleeprq(q, bio, rw);
1186 ctx = blk_mq_get_ctx(q);
1187 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1188 rq = __blk_mq_alloc_request(q, hctx, ctx, rw,
1189 __GFP_WAIT|GFP_ATOMIC, false);
1199 * Multiple hardware queue variant. This will not use per-process plugs,
1200 * but will attempt to bypass the hctx queueing if we can go straight to
1201 * hardware for SYNC IO.
1203 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1205 const int is_sync = rw_is_sync(bio->bi_rw);
1206 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1207 struct blk_map_ctx data;
1210 blk_queue_bounce(q, &bio);
1212 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1213 bio_endio(bio, -EIO);
1217 rq = blk_mq_map_request(q, bio, &data);
1221 if (unlikely(is_flush_fua)) {
1222 blk_mq_bio_to_request(rq, bio);
1223 blk_insert_flush(rq);
1230 blk_mq_bio_to_request(rq, bio);
1231 blk_mq_start_request(rq, true);
1234 * For OK queue, we are done. For error, kill it. Any other
1235 * error (busy), just add it to our list as we previously
1238 ret = q->mq_ops->queue_rq(data.hctx, rq);
1239 if (ret == BLK_MQ_RQ_QUEUE_OK)
1242 __blk_mq_requeue_request(rq);
1244 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1246 blk_mq_end_io(rq, rq->errors);
1252 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1254 * For a SYNC request, send it to the hardware immediately. For
1255 * an ASYNC request, just ensure that we run it later on. The
1256 * latter allows for merging opportunities and more efficient
1260 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1263 blk_mq_put_ctx(data.ctx);
1267 * Single hardware queue variant. This will attempt to use any per-process
1268 * plug for merging and IO deferral.
1270 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1272 const int is_sync = rw_is_sync(bio->bi_rw);
1273 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1274 unsigned int use_plug, request_count = 0;
1275 struct blk_map_ctx data;
1279 * If we have multiple hardware queues, just go directly to
1280 * one of those for sync IO.
1282 use_plug = !is_flush_fua && !is_sync;
1284 blk_queue_bounce(q, &bio);
1286 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1287 bio_endio(bio, -EIO);
1291 if (use_plug && !blk_queue_nomerges(q) &&
1292 blk_attempt_plug_merge(q, bio, &request_count))
1295 rq = blk_mq_map_request(q, bio, &data);
1297 if (unlikely(is_flush_fua)) {
1298 blk_mq_bio_to_request(rq, bio);
1299 blk_insert_flush(rq);
1304 * A task plug currently exists. Since this is completely lockless,
1305 * utilize that to temporarily store requests until the task is
1306 * either done or scheduled away.
1309 struct blk_plug *plug = current->plug;
1312 blk_mq_bio_to_request(rq, bio);
1313 if (list_empty(&plug->mq_list))
1314 trace_block_plug(q);
1315 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1316 blk_flush_plug_list(plug, false);
1317 trace_block_plug(q);
1319 list_add_tail(&rq->queuelist, &plug->mq_list);
1320 blk_mq_put_ctx(data.ctx);
1325 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1327 * For a SYNC request, send it to the hardware immediately. For
1328 * an ASYNC request, just ensure that we run it later on. The
1329 * latter allows for merging opportunities and more efficient
1333 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1336 blk_mq_put_ctx(data.ctx);
1340 * Default mapping to a software queue, since we use one per CPU.
1342 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1344 return q->queue_hw_ctx[q->mq_map[cpu]];
1346 EXPORT_SYMBOL(blk_mq_map_queue);
1348 struct blk_mq_hw_ctx *blk_mq_alloc_single_hw_queue(struct blk_mq_tag_set *set,
1349 unsigned int hctx_index,
1352 return kzalloc_node(sizeof(struct blk_mq_hw_ctx), GFP_KERNEL, node);
1354 EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue);
1356 void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx *hctx,
1357 unsigned int hctx_index)
1361 EXPORT_SYMBOL(blk_mq_free_single_hw_queue);
1363 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1364 struct blk_mq_tags *tags, unsigned int hctx_idx)
1368 if (tags->rqs && set->ops->exit_request) {
1371 for (i = 0; i < tags->nr_tags; i++) {
1374 set->ops->exit_request(set->driver_data, tags->rqs[i],
1379 while (!list_empty(&tags->page_list)) {
1380 page = list_first_entry(&tags->page_list, struct page, lru);
1381 list_del_init(&page->lru);
1382 __free_pages(page, page->private);
1387 blk_mq_free_tags(tags);
1390 static size_t order_to_size(unsigned int order)
1392 return (size_t)PAGE_SIZE << order;
1395 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1396 unsigned int hctx_idx)
1398 struct blk_mq_tags *tags;
1399 unsigned int i, j, entries_per_page, max_order = 4;
1400 size_t rq_size, left;
1402 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1407 INIT_LIST_HEAD(&tags->page_list);
1409 tags->rqs = kmalloc_node(set->queue_depth * sizeof(struct request *),
1410 GFP_KERNEL, set->numa_node);
1412 blk_mq_free_tags(tags);
1417 * rq_size is the size of the request plus driver payload, rounded
1418 * to the cacheline size
1420 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1422 left = rq_size * set->queue_depth;
1424 for (i = 0; i < set->queue_depth; ) {
1425 int this_order = max_order;
1430 while (left < order_to_size(this_order - 1) && this_order)
1434 page = alloc_pages_node(set->numa_node, GFP_KERNEL,
1440 if (order_to_size(this_order) < rq_size)
1447 page->private = this_order;
1448 list_add_tail(&page->lru, &tags->page_list);
1450 p = page_address(page);
1451 entries_per_page = order_to_size(this_order) / rq_size;
1452 to_do = min(entries_per_page, set->queue_depth - i);
1453 left -= to_do * rq_size;
1454 for (j = 0; j < to_do; j++) {
1456 if (set->ops->init_request) {
1457 if (set->ops->init_request(set->driver_data,
1458 tags->rqs[i], hctx_idx, i,
1471 pr_warn("%s: failed to allocate requests\n", __func__);
1472 blk_mq_free_rq_map(set, tags, hctx_idx);
1476 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1481 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1483 unsigned int bpw = 8, total, num_maps, i;
1485 bitmap->bits_per_word = bpw;
1487 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1488 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1493 bitmap->map_size = num_maps;
1496 for (i = 0; i < num_maps; i++) {
1497 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1498 total -= bitmap->map[i].depth;
1504 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1506 struct request_queue *q = hctx->queue;
1507 struct blk_mq_ctx *ctx;
1511 * Move ctx entries to new CPU, if this one is going away.
1513 ctx = __blk_mq_get_ctx(q, cpu);
1515 spin_lock(&ctx->lock);
1516 if (!list_empty(&ctx->rq_list)) {
1517 list_splice_init(&ctx->rq_list, &tmp);
1518 blk_mq_hctx_clear_pending(hctx, ctx);
1520 spin_unlock(&ctx->lock);
1522 if (list_empty(&tmp))
1525 ctx = blk_mq_get_ctx(q);
1526 spin_lock(&ctx->lock);
1528 while (!list_empty(&tmp)) {
1531 rq = list_first_entry(&tmp, struct request, queuelist);
1533 list_move_tail(&rq->queuelist, &ctx->rq_list);
1536 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1537 blk_mq_hctx_mark_pending(hctx, ctx);
1539 spin_unlock(&ctx->lock);
1541 blk_mq_run_hw_queue(hctx, true);
1542 blk_mq_put_ctx(ctx);
1546 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1548 struct request_queue *q = hctx->queue;
1549 struct blk_mq_tag_set *set = q->tag_set;
1551 if (set->tags[hctx->queue_num])
1554 set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1555 if (!set->tags[hctx->queue_num])
1558 hctx->tags = set->tags[hctx->queue_num];
1562 static int blk_mq_hctx_notify(void *data, unsigned long action,
1565 struct blk_mq_hw_ctx *hctx = data;
1567 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1568 return blk_mq_hctx_cpu_offline(hctx, cpu);
1569 else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1570 return blk_mq_hctx_cpu_online(hctx, cpu);
1575 static void blk_mq_exit_hw_queues(struct request_queue *q,
1576 struct blk_mq_tag_set *set, int nr_queue)
1578 struct blk_mq_hw_ctx *hctx;
1581 queue_for_each_hw_ctx(q, hctx, i) {
1585 if (set->ops->exit_hctx)
1586 set->ops->exit_hctx(hctx, i);
1588 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1590 blk_mq_free_bitmap(&hctx->ctx_map);
1595 static void blk_mq_free_hw_queues(struct request_queue *q,
1596 struct blk_mq_tag_set *set)
1598 struct blk_mq_hw_ctx *hctx;
1601 queue_for_each_hw_ctx(q, hctx, i) {
1602 free_cpumask_var(hctx->cpumask);
1603 set->ops->free_hctx(hctx, i);
1607 static int blk_mq_init_hw_queues(struct request_queue *q,
1608 struct blk_mq_tag_set *set)
1610 struct blk_mq_hw_ctx *hctx;
1614 * Initialize hardware queues
1616 queue_for_each_hw_ctx(q, hctx, i) {
1619 node = hctx->numa_node;
1620 if (node == NUMA_NO_NODE)
1621 node = hctx->numa_node = set->numa_node;
1623 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1624 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1625 spin_lock_init(&hctx->lock);
1626 INIT_LIST_HEAD(&hctx->dispatch);
1628 hctx->queue_num = i;
1629 hctx->flags = set->flags;
1630 hctx->cmd_size = set->cmd_size;
1632 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1633 blk_mq_hctx_notify, hctx);
1634 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1636 hctx->tags = set->tags[i];
1639 * Allocate space for all possible cpus to avoid allocation in
1642 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1647 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1652 if (set->ops->init_hctx &&
1653 set->ops->init_hctx(hctx, set->driver_data, i))
1657 if (i == q->nr_hw_queues)
1663 blk_mq_exit_hw_queues(q, set, i);
1668 static void blk_mq_init_cpu_queues(struct request_queue *q,
1669 unsigned int nr_hw_queues)
1673 for_each_possible_cpu(i) {
1674 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1675 struct blk_mq_hw_ctx *hctx;
1677 memset(__ctx, 0, sizeof(*__ctx));
1679 spin_lock_init(&__ctx->lock);
1680 INIT_LIST_HEAD(&__ctx->rq_list);
1683 /* If the cpu isn't online, the cpu is mapped to first hctx */
1687 hctx = q->mq_ops->map_queue(q, i);
1688 cpumask_set_cpu(i, hctx->cpumask);
1692 * Set local node, IFF we have more than one hw queue. If
1693 * not, we remain on the home node of the device
1695 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1696 hctx->numa_node = cpu_to_node(i);
1700 static void blk_mq_map_swqueue(struct request_queue *q)
1703 struct blk_mq_hw_ctx *hctx;
1704 struct blk_mq_ctx *ctx;
1706 queue_for_each_hw_ctx(q, hctx, i) {
1707 cpumask_clear(hctx->cpumask);
1712 * Map software to hardware queues
1714 queue_for_each_ctx(q, ctx, i) {
1715 /* If the cpu isn't online, the cpu is mapped to first hctx */
1719 hctx = q->mq_ops->map_queue(q, i);
1720 cpumask_set_cpu(i, hctx->cpumask);
1721 ctx->index_hw = hctx->nr_ctx;
1722 hctx->ctxs[hctx->nr_ctx++] = ctx;
1725 queue_for_each_hw_ctx(q, hctx, i) {
1727 * If not software queues are mapped to this hardware queue,
1728 * disable it and free the request entries
1730 if (!hctx->nr_ctx) {
1731 struct blk_mq_tag_set *set = q->tag_set;
1734 blk_mq_free_rq_map(set, set->tags[i], i);
1735 set->tags[i] = NULL;
1742 * Initialize batch roundrobin counts
1744 hctx->next_cpu = cpumask_first(hctx->cpumask);
1745 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1749 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1751 struct blk_mq_hw_ctx *hctx;
1752 struct request_queue *q;
1756 if (set->tag_list.next == set->tag_list.prev)
1761 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1762 blk_mq_freeze_queue(q);
1764 queue_for_each_hw_ctx(q, hctx, i) {
1766 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1768 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1770 blk_mq_unfreeze_queue(q);
1774 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1776 struct blk_mq_tag_set *set = q->tag_set;
1778 blk_mq_freeze_queue(q);
1780 mutex_lock(&set->tag_list_lock);
1781 list_del_init(&q->tag_set_list);
1782 blk_mq_update_tag_set_depth(set);
1783 mutex_unlock(&set->tag_list_lock);
1785 blk_mq_unfreeze_queue(q);
1788 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1789 struct request_queue *q)
1793 mutex_lock(&set->tag_list_lock);
1794 list_add_tail(&q->tag_set_list, &set->tag_list);
1795 blk_mq_update_tag_set_depth(set);
1796 mutex_unlock(&set->tag_list_lock);
1799 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1801 struct blk_mq_hw_ctx **hctxs;
1802 struct blk_mq_ctx *ctx;
1803 struct request_queue *q;
1807 ctx = alloc_percpu(struct blk_mq_ctx);
1809 return ERR_PTR(-ENOMEM);
1811 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1817 map = blk_mq_make_queue_map(set);
1821 for (i = 0; i < set->nr_hw_queues; i++) {
1822 int node = blk_mq_hw_queue_to_node(map, i);
1824 hctxs[i] = set->ops->alloc_hctx(set, i, node);
1828 if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL))
1831 atomic_set(&hctxs[i]->nr_active, 0);
1832 hctxs[i]->numa_node = node;
1833 hctxs[i]->queue_num = i;
1836 q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1840 if (percpu_counter_init(&q->mq_usage_counter, 0))
1843 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1844 blk_queue_rq_timeout(q, 30000);
1846 q->nr_queues = nr_cpu_ids;
1847 q->nr_hw_queues = set->nr_hw_queues;
1851 q->queue_hw_ctx = hctxs;
1853 q->mq_ops = set->ops;
1854 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1856 q->sg_reserved_size = INT_MAX;
1858 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1859 INIT_LIST_HEAD(&q->requeue_list);
1860 spin_lock_init(&q->requeue_lock);
1862 if (q->nr_hw_queues > 1)
1863 blk_queue_make_request(q, blk_mq_make_request);
1865 blk_queue_make_request(q, blk_sq_make_request);
1867 blk_queue_rq_timed_out(q, blk_mq_rq_timed_out);
1869 blk_queue_rq_timeout(q, set->timeout);
1872 * Do this after blk_queue_make_request() overrides it...
1874 q->nr_requests = set->queue_depth;
1876 if (set->ops->complete)
1877 blk_queue_softirq_done(q, set->ops->complete);
1879 blk_mq_init_flush(q);
1880 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1882 q->flush_rq = kzalloc(round_up(sizeof(struct request) +
1883 set->cmd_size, cache_line_size()),
1888 if (blk_mq_init_hw_queues(q, set))
1891 mutex_lock(&all_q_mutex);
1892 list_add_tail(&q->all_q_node, &all_q_list);
1893 mutex_unlock(&all_q_mutex);
1895 blk_mq_add_queue_tag_set(set, q);
1897 blk_mq_map_swqueue(q);
1904 blk_cleanup_queue(q);
1907 for (i = 0; i < set->nr_hw_queues; i++) {
1910 free_cpumask_var(hctxs[i]->cpumask);
1911 set->ops->free_hctx(hctxs[i], i);
1917 return ERR_PTR(-ENOMEM);
1919 EXPORT_SYMBOL(blk_mq_init_queue);
1921 void blk_mq_free_queue(struct request_queue *q)
1923 struct blk_mq_tag_set *set = q->tag_set;
1925 blk_mq_del_queue_tag_set(q);
1927 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
1928 blk_mq_free_hw_queues(q, set);
1930 percpu_counter_destroy(&q->mq_usage_counter);
1932 free_percpu(q->queue_ctx);
1933 kfree(q->queue_hw_ctx);
1936 q->queue_ctx = NULL;
1937 q->queue_hw_ctx = NULL;
1940 mutex_lock(&all_q_mutex);
1941 list_del_init(&q->all_q_node);
1942 mutex_unlock(&all_q_mutex);
1945 /* Basically redo blk_mq_init_queue with queue frozen */
1946 static void blk_mq_queue_reinit(struct request_queue *q)
1948 blk_mq_freeze_queue(q);
1950 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1953 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1954 * we should change hctx numa_node according to new topology (this
1955 * involves free and re-allocate memory, worthy doing?)
1958 blk_mq_map_swqueue(q);
1960 blk_mq_unfreeze_queue(q);
1963 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1964 unsigned long action, void *hcpu)
1966 struct request_queue *q;
1969 * Before new mappings are established, hotadded cpu might already
1970 * start handling requests. This doesn't break anything as we map
1971 * offline CPUs to first hardware queue. We will re-init the queue
1972 * below to get optimal settings.
1974 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1975 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1978 mutex_lock(&all_q_mutex);
1979 list_for_each_entry(q, &all_q_list, all_q_node)
1980 blk_mq_queue_reinit(q);
1981 mutex_unlock(&all_q_mutex);
1985 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
1989 if (!set->nr_hw_queues)
1991 if (!set->queue_depth || set->queue_depth > BLK_MQ_MAX_DEPTH)
1993 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
1996 if (!set->nr_hw_queues ||
1997 !set->ops->queue_rq || !set->ops->map_queue ||
1998 !set->ops->alloc_hctx || !set->ops->free_hctx)
2002 set->tags = kmalloc_node(set->nr_hw_queues *
2003 sizeof(struct blk_mq_tags *),
2004 GFP_KERNEL, set->numa_node);
2008 for (i = 0; i < set->nr_hw_queues; i++) {
2009 set->tags[i] = blk_mq_init_rq_map(set, i);
2014 mutex_init(&set->tag_list_lock);
2015 INIT_LIST_HEAD(&set->tag_list);
2021 blk_mq_free_rq_map(set, set->tags[i], i);
2025 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2027 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2031 for (i = 0; i < set->nr_hw_queues; i++) {
2033 blk_mq_free_rq_map(set, set->tags[i], i);
2038 EXPORT_SYMBOL(blk_mq_free_tag_set);
2040 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2042 struct blk_mq_tag_set *set = q->tag_set;
2043 struct blk_mq_hw_ctx *hctx;
2046 if (!set || nr > set->queue_depth)
2050 queue_for_each_hw_ctx(q, hctx, i) {
2051 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2057 q->nr_requests = nr;
2062 void blk_mq_disable_hotplug(void)
2064 mutex_lock(&all_q_mutex);
2067 void blk_mq_enable_hotplug(void)
2069 mutex_unlock(&all_q_mutex);
2072 static int __init blk_mq_init(void)
2076 /* Must be called after percpu_counter_hotcpu_callback() */
2077 hotcpu_notifier(blk_mq_queue_reinit_notify, -10);
2081 subsys_initcall(blk_mq_init);