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);
482 blk_mq_run_queues(q, false);
485 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
487 struct request_queue *q = rq->q;
491 * We abuse this flag that is otherwise used by the I/O scheduler to
492 * request head insertation from the workqueue.
494 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
496 spin_lock_irqsave(&q->requeue_lock, flags);
498 rq->cmd_flags |= REQ_SOFTBARRIER;
499 list_add(&rq->queuelist, &q->requeue_list);
501 list_add_tail(&rq->queuelist, &q->requeue_list);
503 spin_unlock_irqrestore(&q->requeue_lock, flags);
505 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
507 void blk_mq_kick_requeue_list(struct request_queue *q)
509 kblockd_schedule_work(&q->requeue_work);
511 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
513 static inline bool is_flush_request(struct request *rq, unsigned int tag)
515 return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
516 rq->q->flush_rq->tag == tag);
519 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
521 struct request *rq = tags->rqs[tag];
523 if (!is_flush_request(rq, tag))
526 return rq->q->flush_rq;
528 EXPORT_SYMBOL(blk_mq_tag_to_rq);
530 struct blk_mq_timeout_data {
531 struct blk_mq_hw_ctx *hctx;
533 unsigned int *next_set;
536 static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
538 struct blk_mq_timeout_data *data = __data;
539 struct blk_mq_hw_ctx *hctx = data->hctx;
542 /* It may not be in flight yet (this is where
543 * the REQ_ATOMIC_STARTED flag comes in). The requests are
544 * statically allocated, so we know it's always safe to access the
545 * memory associated with a bit offset into ->rqs[].
551 tag = find_next_zero_bit(free_tags, hctx->tags->nr_tags, tag);
552 if (tag >= hctx->tags->nr_tags)
555 rq = blk_mq_tag_to_rq(hctx->tags, tag++);
556 if (rq->q != hctx->queue)
558 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
561 blk_rq_check_expired(rq, data->next, data->next_set);
565 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
567 unsigned int *next_set)
569 struct blk_mq_timeout_data data = {
572 .next_set = next_set,
576 * Ask the tagging code to iterate busy requests, so we can
577 * check them for timeout.
579 blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
582 static enum blk_eh_timer_return blk_mq_rq_timed_out(struct request *rq)
584 struct request_queue *q = rq->q;
587 * We know that complete is set at this point. If STARTED isn't set
588 * anymore, then the request isn't active and the "timeout" should
589 * just be ignored. This can happen due to the bitflag ordering.
590 * Timeout first checks if STARTED is set, and if it is, assumes
591 * the request is active. But if we race with completion, then
592 * we both flags will get cleared. So check here again, and ignore
593 * a timeout event with a request that isn't active.
595 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
596 return BLK_EH_NOT_HANDLED;
598 if (!q->mq_ops->timeout)
599 return BLK_EH_RESET_TIMER;
601 return q->mq_ops->timeout(rq);
604 static void blk_mq_rq_timer(unsigned long data)
606 struct request_queue *q = (struct request_queue *) data;
607 struct blk_mq_hw_ctx *hctx;
608 unsigned long next = 0;
611 queue_for_each_hw_ctx(q, hctx, i) {
613 * If not software queues are currently mapped to this
614 * hardware queue, there's nothing to check
616 if (!hctx->nr_ctx || !hctx->tags)
619 blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
623 next = blk_rq_timeout(round_jiffies_up(next));
624 mod_timer(&q->timeout, next);
626 queue_for_each_hw_ctx(q, hctx, i)
627 blk_mq_tag_idle(hctx);
632 * Reverse check our software queue for entries that we could potentially
633 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
634 * too much time checking for merges.
636 static bool blk_mq_attempt_merge(struct request_queue *q,
637 struct blk_mq_ctx *ctx, struct bio *bio)
642 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
648 if (!blk_rq_merge_ok(rq, bio))
651 el_ret = blk_try_merge(rq, bio);
652 if (el_ret == ELEVATOR_BACK_MERGE) {
653 if (bio_attempt_back_merge(q, rq, bio)) {
658 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
659 if (bio_attempt_front_merge(q, rq, bio)) {
671 * Process software queues that have been marked busy, splicing them
672 * to the for-dispatch
674 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
676 struct blk_mq_ctx *ctx;
679 for (i = 0; i < hctx->ctx_map.map_size; i++) {
680 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
681 unsigned int off, bit;
687 off = i * hctx->ctx_map.bits_per_word;
689 bit = find_next_bit(&bm->word, bm->depth, bit);
690 if (bit >= bm->depth)
693 ctx = hctx->ctxs[bit + off];
694 clear_bit(bit, &bm->word);
695 spin_lock(&ctx->lock);
696 list_splice_tail_init(&ctx->rq_list, list);
697 spin_unlock(&ctx->lock);
705 * Run this hardware queue, pulling any software queues mapped to it in.
706 * Note that this function currently has various problems around ordering
707 * of IO. In particular, we'd like FIFO behaviour on handling existing
708 * items on the hctx->dispatch list. Ignore that for now.
710 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
712 struct request_queue *q = hctx->queue;
717 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
719 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
725 * Touch any software queue that has pending entries.
727 flush_busy_ctxs(hctx, &rq_list);
730 * If we have previous entries on our dispatch list, grab them
731 * and stuff them at the front for more fair dispatch.
733 if (!list_empty_careful(&hctx->dispatch)) {
734 spin_lock(&hctx->lock);
735 if (!list_empty(&hctx->dispatch))
736 list_splice_init(&hctx->dispatch, &rq_list);
737 spin_unlock(&hctx->lock);
741 * Now process all the entries, sending them to the driver.
744 while (!list_empty(&rq_list)) {
747 rq = list_first_entry(&rq_list, struct request, queuelist);
748 list_del_init(&rq->queuelist);
750 blk_mq_start_request(rq, list_empty(&rq_list));
752 ret = q->mq_ops->queue_rq(hctx, rq);
754 case BLK_MQ_RQ_QUEUE_OK:
757 case BLK_MQ_RQ_QUEUE_BUSY:
758 list_add(&rq->queuelist, &rq_list);
759 __blk_mq_requeue_request(rq);
762 pr_err("blk-mq: bad return on queue: %d\n", ret);
763 case BLK_MQ_RQ_QUEUE_ERROR:
765 blk_mq_end_io(rq, rq->errors);
769 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
774 hctx->dispatched[0]++;
775 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
776 hctx->dispatched[ilog2(queued) + 1]++;
779 * Any items that need requeuing? Stuff them into hctx->dispatch,
780 * that is where we will continue on next queue run.
782 if (!list_empty(&rq_list)) {
783 spin_lock(&hctx->lock);
784 list_splice(&rq_list, &hctx->dispatch);
785 spin_unlock(&hctx->lock);
790 * It'd be great if the workqueue API had a way to pass
791 * in a mask and had some smarts for more clever placement.
792 * For now we just round-robin here, switching for every
793 * BLK_MQ_CPU_WORK_BATCH queued items.
795 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
797 int cpu = hctx->next_cpu;
799 if (--hctx->next_cpu_batch <= 0) {
802 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
803 if (next_cpu >= nr_cpu_ids)
804 next_cpu = cpumask_first(hctx->cpumask);
806 hctx->next_cpu = next_cpu;
807 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
813 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
815 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
818 if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask))
819 __blk_mq_run_hw_queue(hctx);
820 else if (hctx->queue->nr_hw_queues == 1)
821 kblockd_schedule_delayed_work(&hctx->run_work, 0);
825 cpu = blk_mq_hctx_next_cpu(hctx);
826 kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0);
830 void blk_mq_run_queues(struct request_queue *q, bool async)
832 struct blk_mq_hw_ctx *hctx;
835 queue_for_each_hw_ctx(q, hctx, i) {
836 if ((!blk_mq_hctx_has_pending(hctx) &&
837 list_empty_careful(&hctx->dispatch)) ||
838 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
842 blk_mq_run_hw_queue(hctx, async);
846 EXPORT_SYMBOL(blk_mq_run_queues);
848 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
850 cancel_delayed_work(&hctx->run_work);
851 cancel_delayed_work(&hctx->delay_work);
852 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
854 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
856 void blk_mq_stop_hw_queues(struct request_queue *q)
858 struct blk_mq_hw_ctx *hctx;
861 queue_for_each_hw_ctx(q, hctx, i)
862 blk_mq_stop_hw_queue(hctx);
864 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
866 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
868 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
871 blk_mq_run_hw_queue(hctx, false);
874 EXPORT_SYMBOL(blk_mq_start_hw_queue);
876 void blk_mq_start_hw_queues(struct request_queue *q)
878 struct blk_mq_hw_ctx *hctx;
881 queue_for_each_hw_ctx(q, hctx, i)
882 blk_mq_start_hw_queue(hctx);
884 EXPORT_SYMBOL(blk_mq_start_hw_queues);
887 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
889 struct blk_mq_hw_ctx *hctx;
892 queue_for_each_hw_ctx(q, hctx, i) {
893 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
896 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
898 blk_mq_run_hw_queue(hctx, async);
902 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
904 static void blk_mq_run_work_fn(struct work_struct *work)
906 struct blk_mq_hw_ctx *hctx;
908 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
910 __blk_mq_run_hw_queue(hctx);
913 static void blk_mq_delay_work_fn(struct work_struct *work)
915 struct blk_mq_hw_ctx *hctx;
917 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
919 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
920 __blk_mq_run_hw_queue(hctx);
923 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
925 unsigned long tmo = msecs_to_jiffies(msecs);
927 if (hctx->queue->nr_hw_queues == 1)
928 kblockd_schedule_delayed_work(&hctx->delay_work, tmo);
932 cpu = blk_mq_hctx_next_cpu(hctx);
933 kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo);
936 EXPORT_SYMBOL(blk_mq_delay_queue);
938 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
939 struct request *rq, bool at_head)
941 struct blk_mq_ctx *ctx = rq->mq_ctx;
943 trace_block_rq_insert(hctx->queue, rq);
946 list_add(&rq->queuelist, &ctx->rq_list);
948 list_add_tail(&rq->queuelist, &ctx->rq_list);
950 blk_mq_hctx_mark_pending(hctx, ctx);
953 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
956 struct request_queue *q = rq->q;
957 struct blk_mq_hw_ctx *hctx;
958 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
960 current_ctx = blk_mq_get_ctx(q);
961 if (!cpu_online(ctx->cpu))
962 rq->mq_ctx = ctx = current_ctx;
964 hctx = q->mq_ops->map_queue(q, ctx->cpu);
966 spin_lock(&ctx->lock);
967 __blk_mq_insert_request(hctx, rq, at_head);
968 spin_unlock(&ctx->lock);
971 blk_mq_run_hw_queue(hctx, async);
973 blk_mq_put_ctx(current_ctx);
976 static void blk_mq_insert_requests(struct request_queue *q,
977 struct blk_mq_ctx *ctx,
978 struct list_head *list,
983 struct blk_mq_hw_ctx *hctx;
984 struct blk_mq_ctx *current_ctx;
986 trace_block_unplug(q, depth, !from_schedule);
988 current_ctx = blk_mq_get_ctx(q);
990 if (!cpu_online(ctx->cpu))
992 hctx = q->mq_ops->map_queue(q, ctx->cpu);
995 * preemption doesn't flush plug list, so it's possible ctx->cpu is
998 spin_lock(&ctx->lock);
999 while (!list_empty(list)) {
1002 rq = list_first_entry(list, struct request, queuelist);
1003 list_del_init(&rq->queuelist);
1005 __blk_mq_insert_request(hctx, rq, false);
1007 spin_unlock(&ctx->lock);
1009 blk_mq_run_hw_queue(hctx, from_schedule);
1010 blk_mq_put_ctx(current_ctx);
1013 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1015 struct request *rqa = container_of(a, struct request, queuelist);
1016 struct request *rqb = container_of(b, struct request, queuelist);
1018 return !(rqa->mq_ctx < rqb->mq_ctx ||
1019 (rqa->mq_ctx == rqb->mq_ctx &&
1020 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1023 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1025 struct blk_mq_ctx *this_ctx;
1026 struct request_queue *this_q;
1029 LIST_HEAD(ctx_list);
1032 list_splice_init(&plug->mq_list, &list);
1034 list_sort(NULL, &list, plug_ctx_cmp);
1040 while (!list_empty(&list)) {
1041 rq = list_entry_rq(list.next);
1042 list_del_init(&rq->queuelist);
1044 if (rq->mq_ctx != this_ctx) {
1046 blk_mq_insert_requests(this_q, this_ctx,
1051 this_ctx = rq->mq_ctx;
1057 list_add_tail(&rq->queuelist, &ctx_list);
1061 * If 'this_ctx' is set, we know we have entries to complete
1062 * on 'ctx_list'. Do those.
1065 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1070 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1072 init_request_from_bio(rq, bio);
1074 if (blk_do_io_stat(rq))
1075 blk_account_io_start(rq, 1);
1078 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1080 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1081 !blk_queue_nomerges(hctx->queue);
1084 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1085 struct blk_mq_ctx *ctx,
1086 struct request *rq, struct bio *bio)
1088 if (!hctx_allow_merges(hctx)) {
1089 blk_mq_bio_to_request(rq, bio);
1090 spin_lock(&ctx->lock);
1092 __blk_mq_insert_request(hctx, rq, false);
1093 spin_unlock(&ctx->lock);
1096 struct request_queue *q = hctx->queue;
1098 spin_lock(&ctx->lock);
1099 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1100 blk_mq_bio_to_request(rq, bio);
1104 spin_unlock(&ctx->lock);
1105 __blk_mq_free_request(hctx, ctx, rq);
1110 struct blk_map_ctx {
1111 struct blk_mq_hw_ctx *hctx;
1112 struct blk_mq_ctx *ctx;
1115 static struct request *blk_mq_map_request(struct request_queue *q,
1117 struct blk_map_ctx *data)
1119 struct blk_mq_hw_ctx *hctx;
1120 struct blk_mq_ctx *ctx;
1122 int rw = bio_data_dir(bio);
1123 struct blk_mq_alloc_data alloc_data;
1125 if (unlikely(blk_mq_queue_enter(q))) {
1126 bio_endio(bio, -EIO);
1130 ctx = blk_mq_get_ctx(q);
1131 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1133 if (rw_is_sync(bio->bi_rw))
1136 trace_block_getrq(q, bio, rw);
1137 blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1139 rq = __blk_mq_alloc_request(&alloc_data, rw);
1140 if (unlikely(!rq)) {
1141 __blk_mq_run_hw_queue(hctx);
1142 blk_mq_put_ctx(ctx);
1143 trace_block_sleeprq(q, bio, rw);
1145 ctx = blk_mq_get_ctx(q);
1146 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1147 blk_mq_set_alloc_data(&alloc_data, q,
1148 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1149 rq = __blk_mq_alloc_request(&alloc_data, rw);
1150 ctx = alloc_data.ctx;
1151 hctx = alloc_data.hctx;
1161 * Multiple hardware queue variant. This will not use per-process plugs,
1162 * but will attempt to bypass the hctx queueing if we can go straight to
1163 * hardware for SYNC IO.
1165 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1167 const int is_sync = rw_is_sync(bio->bi_rw);
1168 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1169 struct blk_map_ctx data;
1172 blk_queue_bounce(q, &bio);
1174 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1175 bio_endio(bio, -EIO);
1179 rq = blk_mq_map_request(q, bio, &data);
1183 if (unlikely(is_flush_fua)) {
1184 blk_mq_bio_to_request(rq, bio);
1185 blk_insert_flush(rq);
1192 blk_mq_bio_to_request(rq, bio);
1193 blk_mq_start_request(rq, true);
1196 * For OK queue, we are done. For error, kill it. Any other
1197 * error (busy), just add it to our list as we previously
1200 ret = q->mq_ops->queue_rq(data.hctx, rq);
1201 if (ret == BLK_MQ_RQ_QUEUE_OK)
1204 __blk_mq_requeue_request(rq);
1206 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1208 blk_mq_end_io(rq, rq->errors);
1214 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1216 * For a SYNC request, send it to the hardware immediately. For
1217 * an ASYNC request, just ensure that we run it later on. The
1218 * latter allows for merging opportunities and more efficient
1222 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1225 blk_mq_put_ctx(data.ctx);
1229 * Single hardware queue variant. This will attempt to use any per-process
1230 * plug for merging and IO deferral.
1232 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1234 const int is_sync = rw_is_sync(bio->bi_rw);
1235 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1236 unsigned int use_plug, request_count = 0;
1237 struct blk_map_ctx data;
1241 * If we have multiple hardware queues, just go directly to
1242 * one of those for sync IO.
1244 use_plug = !is_flush_fua && !is_sync;
1246 blk_queue_bounce(q, &bio);
1248 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1249 bio_endio(bio, -EIO);
1253 if (use_plug && !blk_queue_nomerges(q) &&
1254 blk_attempt_plug_merge(q, bio, &request_count))
1257 rq = blk_mq_map_request(q, bio, &data);
1261 if (unlikely(is_flush_fua)) {
1262 blk_mq_bio_to_request(rq, bio);
1263 blk_insert_flush(rq);
1268 * A task plug currently exists. Since this is completely lockless,
1269 * utilize that to temporarily store requests until the task is
1270 * either done or scheduled away.
1273 struct blk_plug *plug = current->plug;
1276 blk_mq_bio_to_request(rq, bio);
1277 if (list_empty(&plug->mq_list))
1278 trace_block_plug(q);
1279 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1280 blk_flush_plug_list(plug, false);
1281 trace_block_plug(q);
1283 list_add_tail(&rq->queuelist, &plug->mq_list);
1284 blk_mq_put_ctx(data.ctx);
1289 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1291 * For a SYNC request, send it to the hardware immediately. For
1292 * an ASYNC request, just ensure that we run it later on. The
1293 * latter allows for merging opportunities and more efficient
1297 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1300 blk_mq_put_ctx(data.ctx);
1304 * Default mapping to a software queue, since we use one per CPU.
1306 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1308 return q->queue_hw_ctx[q->mq_map[cpu]];
1310 EXPORT_SYMBOL(blk_mq_map_queue);
1312 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1313 struct blk_mq_tags *tags, unsigned int hctx_idx)
1317 if (tags->rqs && set->ops->exit_request) {
1320 for (i = 0; i < tags->nr_tags; i++) {
1323 set->ops->exit_request(set->driver_data, tags->rqs[i],
1325 tags->rqs[i] = NULL;
1329 while (!list_empty(&tags->page_list)) {
1330 page = list_first_entry(&tags->page_list, struct page, lru);
1331 list_del_init(&page->lru);
1332 __free_pages(page, page->private);
1337 blk_mq_free_tags(tags);
1340 static size_t order_to_size(unsigned int order)
1342 return (size_t)PAGE_SIZE << order;
1345 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1346 unsigned int hctx_idx)
1348 struct blk_mq_tags *tags;
1349 unsigned int i, j, entries_per_page, max_order = 4;
1350 size_t rq_size, left;
1352 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1357 INIT_LIST_HEAD(&tags->page_list);
1359 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1360 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1363 blk_mq_free_tags(tags);
1368 * rq_size is the size of the request plus driver payload, rounded
1369 * to the cacheline size
1371 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1373 left = rq_size * set->queue_depth;
1375 for (i = 0; i < set->queue_depth; ) {
1376 int this_order = max_order;
1381 while (left < order_to_size(this_order - 1) && this_order)
1385 page = alloc_pages_node(set->numa_node,
1386 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1392 if (order_to_size(this_order) < rq_size)
1399 page->private = this_order;
1400 list_add_tail(&page->lru, &tags->page_list);
1402 p = page_address(page);
1403 entries_per_page = order_to_size(this_order) / rq_size;
1404 to_do = min(entries_per_page, set->queue_depth - i);
1405 left -= to_do * rq_size;
1406 for (j = 0; j < to_do; j++) {
1408 tags->rqs[i]->atomic_flags = 0;
1409 tags->rqs[i]->cmd_flags = 0;
1410 if (set->ops->init_request) {
1411 if (set->ops->init_request(set->driver_data,
1412 tags->rqs[i], hctx_idx, i,
1414 tags->rqs[i] = NULL;
1427 blk_mq_free_rq_map(set, tags, hctx_idx);
1431 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1436 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1438 unsigned int bpw = 8, total, num_maps, i;
1440 bitmap->bits_per_word = bpw;
1442 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1443 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1448 bitmap->map_size = num_maps;
1451 for (i = 0; i < num_maps; i++) {
1452 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1453 total -= bitmap->map[i].depth;
1459 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1461 struct request_queue *q = hctx->queue;
1462 struct blk_mq_ctx *ctx;
1466 * Move ctx entries to new CPU, if this one is going away.
1468 ctx = __blk_mq_get_ctx(q, cpu);
1470 spin_lock(&ctx->lock);
1471 if (!list_empty(&ctx->rq_list)) {
1472 list_splice_init(&ctx->rq_list, &tmp);
1473 blk_mq_hctx_clear_pending(hctx, ctx);
1475 spin_unlock(&ctx->lock);
1477 if (list_empty(&tmp))
1480 ctx = blk_mq_get_ctx(q);
1481 spin_lock(&ctx->lock);
1483 while (!list_empty(&tmp)) {
1486 rq = list_first_entry(&tmp, struct request, queuelist);
1488 list_move_tail(&rq->queuelist, &ctx->rq_list);
1491 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1492 blk_mq_hctx_mark_pending(hctx, ctx);
1494 spin_unlock(&ctx->lock);
1496 blk_mq_run_hw_queue(hctx, true);
1497 blk_mq_put_ctx(ctx);
1501 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1503 struct request_queue *q = hctx->queue;
1504 struct blk_mq_tag_set *set = q->tag_set;
1506 if (set->tags[hctx->queue_num])
1509 set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1510 if (!set->tags[hctx->queue_num])
1513 hctx->tags = set->tags[hctx->queue_num];
1517 static int blk_mq_hctx_notify(void *data, unsigned long action,
1520 struct blk_mq_hw_ctx *hctx = data;
1522 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1523 return blk_mq_hctx_cpu_offline(hctx, cpu);
1524 else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1525 return blk_mq_hctx_cpu_online(hctx, cpu);
1530 static void blk_mq_exit_hw_queues(struct request_queue *q,
1531 struct blk_mq_tag_set *set, int nr_queue)
1533 struct blk_mq_hw_ctx *hctx;
1536 queue_for_each_hw_ctx(q, hctx, i) {
1540 blk_mq_tag_idle(hctx);
1542 if (set->ops->exit_hctx)
1543 set->ops->exit_hctx(hctx, i);
1545 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1547 blk_mq_free_bitmap(&hctx->ctx_map);
1552 static void blk_mq_free_hw_queues(struct request_queue *q,
1553 struct blk_mq_tag_set *set)
1555 struct blk_mq_hw_ctx *hctx;
1558 queue_for_each_hw_ctx(q, hctx, i) {
1559 free_cpumask_var(hctx->cpumask);
1564 static int blk_mq_init_hw_queues(struct request_queue *q,
1565 struct blk_mq_tag_set *set)
1567 struct blk_mq_hw_ctx *hctx;
1571 * Initialize hardware queues
1573 queue_for_each_hw_ctx(q, hctx, i) {
1576 node = hctx->numa_node;
1577 if (node == NUMA_NO_NODE)
1578 node = hctx->numa_node = set->numa_node;
1580 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1581 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1582 spin_lock_init(&hctx->lock);
1583 INIT_LIST_HEAD(&hctx->dispatch);
1585 hctx->queue_num = i;
1586 hctx->flags = set->flags;
1587 hctx->cmd_size = set->cmd_size;
1589 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1590 blk_mq_hctx_notify, hctx);
1591 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1593 hctx->tags = set->tags[i];
1596 * Allocate space for all possible cpus to avoid allocation at
1599 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1604 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1609 if (set->ops->init_hctx &&
1610 set->ops->init_hctx(hctx, set->driver_data, i))
1614 if (i == q->nr_hw_queues)
1620 blk_mq_exit_hw_queues(q, set, i);
1625 static void blk_mq_init_cpu_queues(struct request_queue *q,
1626 unsigned int nr_hw_queues)
1630 for_each_possible_cpu(i) {
1631 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1632 struct blk_mq_hw_ctx *hctx;
1634 memset(__ctx, 0, sizeof(*__ctx));
1636 spin_lock_init(&__ctx->lock);
1637 INIT_LIST_HEAD(&__ctx->rq_list);
1640 /* If the cpu isn't online, the cpu is mapped to first hctx */
1644 hctx = q->mq_ops->map_queue(q, i);
1645 cpumask_set_cpu(i, hctx->cpumask);
1649 * Set local node, IFF we have more than one hw queue. If
1650 * not, we remain on the home node of the device
1652 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1653 hctx->numa_node = cpu_to_node(i);
1657 static void blk_mq_map_swqueue(struct request_queue *q)
1660 struct blk_mq_hw_ctx *hctx;
1661 struct blk_mq_ctx *ctx;
1663 queue_for_each_hw_ctx(q, hctx, i) {
1664 cpumask_clear(hctx->cpumask);
1669 * Map software to hardware queues
1671 queue_for_each_ctx(q, ctx, i) {
1672 /* If the cpu isn't online, the cpu is mapped to first hctx */
1676 hctx = q->mq_ops->map_queue(q, i);
1677 cpumask_set_cpu(i, hctx->cpumask);
1678 ctx->index_hw = hctx->nr_ctx;
1679 hctx->ctxs[hctx->nr_ctx++] = ctx;
1682 queue_for_each_hw_ctx(q, hctx, i) {
1684 * If no software queues are mapped to this hardware queue,
1685 * disable it and free the request entries.
1687 if (!hctx->nr_ctx) {
1688 struct blk_mq_tag_set *set = q->tag_set;
1691 blk_mq_free_rq_map(set, set->tags[i], i);
1692 set->tags[i] = NULL;
1699 * Initialize batch roundrobin counts
1701 hctx->next_cpu = cpumask_first(hctx->cpumask);
1702 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1706 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1708 struct blk_mq_hw_ctx *hctx;
1709 struct request_queue *q;
1713 if (set->tag_list.next == set->tag_list.prev)
1718 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1719 blk_mq_freeze_queue(q);
1721 queue_for_each_hw_ctx(q, hctx, i) {
1723 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1725 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1727 blk_mq_unfreeze_queue(q);
1731 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1733 struct blk_mq_tag_set *set = q->tag_set;
1735 mutex_lock(&set->tag_list_lock);
1736 list_del_init(&q->tag_set_list);
1737 blk_mq_update_tag_set_depth(set);
1738 mutex_unlock(&set->tag_list_lock);
1741 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1742 struct request_queue *q)
1746 mutex_lock(&set->tag_list_lock);
1747 list_add_tail(&q->tag_set_list, &set->tag_list);
1748 blk_mq_update_tag_set_depth(set);
1749 mutex_unlock(&set->tag_list_lock);
1752 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1754 struct blk_mq_hw_ctx **hctxs;
1755 struct blk_mq_ctx __percpu *ctx;
1756 struct request_queue *q;
1760 ctx = alloc_percpu(struct blk_mq_ctx);
1762 return ERR_PTR(-ENOMEM);
1764 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1770 map = blk_mq_make_queue_map(set);
1774 for (i = 0; i < set->nr_hw_queues; i++) {
1775 int node = blk_mq_hw_queue_to_node(map, i);
1777 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1782 if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL))
1785 atomic_set(&hctxs[i]->nr_active, 0);
1786 hctxs[i]->numa_node = node;
1787 hctxs[i]->queue_num = i;
1790 q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1794 if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release))
1797 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1798 blk_queue_rq_timeout(q, 30000);
1800 q->nr_queues = nr_cpu_ids;
1801 q->nr_hw_queues = set->nr_hw_queues;
1805 q->queue_hw_ctx = hctxs;
1807 q->mq_ops = set->ops;
1808 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1810 if (!(set->flags & BLK_MQ_F_SG_MERGE))
1811 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1813 q->sg_reserved_size = INT_MAX;
1815 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1816 INIT_LIST_HEAD(&q->requeue_list);
1817 spin_lock_init(&q->requeue_lock);
1819 if (q->nr_hw_queues > 1)
1820 blk_queue_make_request(q, blk_mq_make_request);
1822 blk_queue_make_request(q, blk_sq_make_request);
1824 blk_queue_rq_timed_out(q, blk_mq_rq_timed_out);
1826 blk_queue_rq_timeout(q, set->timeout);
1829 * Do this after blk_queue_make_request() overrides it...
1831 q->nr_requests = set->queue_depth;
1833 if (set->ops->complete)
1834 blk_queue_softirq_done(q, set->ops->complete);
1836 blk_mq_init_flush(q);
1837 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1839 q->flush_rq = kzalloc(round_up(sizeof(struct request) +
1840 set->cmd_size, cache_line_size()),
1845 if (blk_mq_init_hw_queues(q, set))
1848 mutex_lock(&all_q_mutex);
1849 list_add_tail(&q->all_q_node, &all_q_list);
1850 mutex_unlock(&all_q_mutex);
1852 blk_mq_add_queue_tag_set(set, q);
1854 blk_mq_map_swqueue(q);
1861 blk_cleanup_queue(q);
1864 for (i = 0; i < set->nr_hw_queues; i++) {
1867 free_cpumask_var(hctxs[i]->cpumask);
1874 return ERR_PTR(-ENOMEM);
1876 EXPORT_SYMBOL(blk_mq_init_queue);
1878 void blk_mq_free_queue(struct request_queue *q)
1880 struct blk_mq_tag_set *set = q->tag_set;
1882 blk_mq_del_queue_tag_set(q);
1884 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
1885 blk_mq_free_hw_queues(q, set);
1887 percpu_ref_exit(&q->mq_usage_counter);
1889 free_percpu(q->queue_ctx);
1890 kfree(q->queue_hw_ctx);
1893 q->queue_ctx = NULL;
1894 q->queue_hw_ctx = NULL;
1897 mutex_lock(&all_q_mutex);
1898 list_del_init(&q->all_q_node);
1899 mutex_unlock(&all_q_mutex);
1902 /* Basically redo blk_mq_init_queue with queue frozen */
1903 static void blk_mq_queue_reinit(struct request_queue *q)
1905 blk_mq_freeze_queue(q);
1907 blk_mq_sysfs_unregister(q);
1909 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1912 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1913 * we should change hctx numa_node according to new topology (this
1914 * involves free and re-allocate memory, worthy doing?)
1917 blk_mq_map_swqueue(q);
1919 blk_mq_sysfs_register(q);
1921 blk_mq_unfreeze_queue(q);
1924 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1925 unsigned long action, void *hcpu)
1927 struct request_queue *q;
1930 * Before new mappings are established, hotadded cpu might already
1931 * start handling requests. This doesn't break anything as we map
1932 * offline CPUs to first hardware queue. We will re-init the queue
1933 * below to get optimal settings.
1935 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1936 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1939 mutex_lock(&all_q_mutex);
1940 list_for_each_entry(q, &all_q_list, all_q_node)
1941 blk_mq_queue_reinit(q);
1942 mutex_unlock(&all_q_mutex);
1946 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
1950 for (i = 0; i < set->nr_hw_queues; i++) {
1951 set->tags[i] = blk_mq_init_rq_map(set, i);
1960 blk_mq_free_rq_map(set, set->tags[i], i);
1967 * Allocate the request maps associated with this tag_set. Note that this
1968 * may reduce the depth asked for, if memory is tight. set->queue_depth
1969 * will be updated to reflect the allocated depth.
1971 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
1976 depth = set->queue_depth;
1978 err = __blk_mq_alloc_rq_maps(set);
1982 set->queue_depth >>= 1;
1983 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
1987 } while (set->queue_depth);
1989 if (!set->queue_depth || err) {
1990 pr_err("blk-mq: failed to allocate request map\n");
1994 if (depth != set->queue_depth)
1995 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
1996 depth, set->queue_depth);
2002 * Alloc a tag set to be associated with one or more request queues.
2003 * May fail with EINVAL for various error conditions. May adjust the
2004 * requested depth down, if if it too large. In that case, the set
2005 * value will be stored in set->queue_depth.
2007 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2009 if (!set->nr_hw_queues)
2011 if (!set->queue_depth)
2013 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2016 if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue)
2019 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2020 pr_info("blk-mq: reduced tag depth to %u\n",
2022 set->queue_depth = BLK_MQ_MAX_DEPTH;
2025 set->tags = kmalloc_node(set->nr_hw_queues *
2026 sizeof(struct blk_mq_tags *),
2027 GFP_KERNEL, set->numa_node);
2031 if (blk_mq_alloc_rq_maps(set))
2034 mutex_init(&set->tag_list_lock);
2035 INIT_LIST_HEAD(&set->tag_list);
2043 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2045 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2049 for (i = 0; i < set->nr_hw_queues; i++) {
2051 blk_mq_free_rq_map(set, set->tags[i], i);
2057 EXPORT_SYMBOL(blk_mq_free_tag_set);
2059 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2061 struct blk_mq_tag_set *set = q->tag_set;
2062 struct blk_mq_hw_ctx *hctx;
2065 if (!set || nr > set->queue_depth)
2069 queue_for_each_hw_ctx(q, hctx, i) {
2070 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2076 q->nr_requests = nr;
2081 void blk_mq_disable_hotplug(void)
2083 mutex_lock(&all_q_mutex);
2086 void blk_mq_enable_hotplug(void)
2088 mutex_unlock(&all_q_mutex);
2091 static int __init blk_mq_init(void)
2095 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2099 subsys_initcall(blk_mq_init);