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 void __blk_mq_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);
389 * blk_mq_complete_request - end I/O on a request
390 * @rq: the request being processed
393 * Ends all I/O on a request. It does not handle partial completions.
394 * The actual completion happens out-of-order, through a IPI handler.
396 void blk_mq_complete_request(struct request *rq)
398 struct request_queue *q = rq->q;
400 if (unlikely(blk_should_fake_timeout(q)))
402 if (!blk_mark_rq_complete(rq)) {
403 if (q->softirq_done_fn)
404 __blk_mq_complete_request(rq);
406 blk_mq_end_io(rq, rq->errors);
409 EXPORT_SYMBOL(blk_mq_complete_request);
411 static void blk_mq_start_request(struct request *rq, bool last)
413 struct request_queue *q = rq->q;
415 trace_block_rq_issue(q, rq);
417 rq->resid_len = blk_rq_bytes(rq);
418 if (unlikely(blk_bidi_rq(rq)))
419 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
422 * Just mark start time and set the started bit. Due to memory
423 * ordering, we know we'll see the correct deadline as long as
424 * REQ_ATOMIC_STARTED is seen. Use the default queue timeout,
425 * unless one has been set in the request.
428 rq->deadline = jiffies + q->rq_timeout;
430 rq->deadline = jiffies + rq->timeout;
433 * Mark us as started and clear complete. Complete might have been
434 * set if requeue raced with timeout, which then marked it as
435 * complete. So be sure to clear complete again when we start
436 * the request, otherwise we'll ignore the completion event.
438 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
439 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
440 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
441 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
443 if (q->dma_drain_size && blk_rq_bytes(rq)) {
445 * Make sure space for the drain appears. We know we can do
446 * this because max_hw_segments has been adjusted to be one
447 * fewer than the device can handle.
449 rq->nr_phys_segments++;
453 * Flag the last request in the series so that drivers know when IO
454 * should be kicked off, if they don't do it on a per-request basis.
456 * Note: the flag isn't the only condition drivers should do kick off.
457 * If drive is busy, the last request might not have the bit set.
460 rq->cmd_flags |= REQ_END;
463 static void __blk_mq_requeue_request(struct request *rq)
465 struct request_queue *q = rq->q;
467 trace_block_rq_requeue(q, rq);
468 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
470 rq->cmd_flags &= ~REQ_END;
472 if (q->dma_drain_size && blk_rq_bytes(rq))
473 rq->nr_phys_segments--;
476 void blk_mq_requeue_request(struct request *rq)
478 __blk_mq_requeue_request(rq);
479 blk_clear_rq_complete(rq);
481 BUG_ON(blk_queued_rq(rq));
482 blk_mq_add_to_requeue_list(rq, true);
484 EXPORT_SYMBOL(blk_mq_requeue_request);
486 static void blk_mq_requeue_work(struct work_struct *work)
488 struct request_queue *q =
489 container_of(work, struct request_queue, requeue_work);
491 struct request *rq, *next;
494 spin_lock_irqsave(&q->requeue_lock, flags);
495 list_splice_init(&q->requeue_list, &rq_list);
496 spin_unlock_irqrestore(&q->requeue_lock, flags);
498 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
499 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
502 rq->cmd_flags &= ~REQ_SOFTBARRIER;
503 list_del_init(&rq->queuelist);
504 blk_mq_insert_request(rq, true, false, false);
507 while (!list_empty(&rq_list)) {
508 rq = list_entry(rq_list.next, struct request, queuelist);
509 list_del_init(&rq->queuelist);
510 blk_mq_insert_request(rq, false, false, false);
513 blk_mq_run_queues(q, false);
516 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
518 struct request_queue *q = rq->q;
522 * We abuse this flag that is otherwise used by the I/O scheduler to
523 * request head insertation from the workqueue.
525 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
527 spin_lock_irqsave(&q->requeue_lock, flags);
529 rq->cmd_flags |= REQ_SOFTBARRIER;
530 list_add(&rq->queuelist, &q->requeue_list);
532 list_add_tail(&rq->queuelist, &q->requeue_list);
534 spin_unlock_irqrestore(&q->requeue_lock, flags);
536 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
538 void blk_mq_kick_requeue_list(struct request_queue *q)
540 kblockd_schedule_work(&q->requeue_work);
542 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
544 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
546 return tags->rqs[tag];
548 EXPORT_SYMBOL(blk_mq_tag_to_rq);
550 struct blk_mq_timeout_data {
551 struct blk_mq_hw_ctx *hctx;
553 unsigned int *next_set;
556 static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
558 struct blk_mq_timeout_data *data = __data;
559 struct blk_mq_hw_ctx *hctx = data->hctx;
562 /* It may not be in flight yet (this is where
563 * the REQ_ATOMIC_STARTED flag comes in). The requests are
564 * statically allocated, so we know it's always safe to access the
565 * memory associated with a bit offset into ->rqs[].
571 tag = find_next_zero_bit(free_tags, hctx->tags->nr_tags, tag);
572 if (tag >= hctx->tags->nr_tags)
575 rq = blk_mq_tag_to_rq(hctx->tags, tag++);
576 if (rq->q != hctx->queue)
578 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
581 blk_rq_check_expired(rq, data->next, data->next_set);
585 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
587 unsigned int *next_set)
589 struct blk_mq_timeout_data data = {
592 .next_set = next_set,
596 * Ask the tagging code to iterate busy requests, so we can
597 * check them for timeout.
599 blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
602 static enum blk_eh_timer_return blk_mq_rq_timed_out(struct request *rq)
604 struct request_queue *q = rq->q;
607 * We know that complete is set at this point. If STARTED isn't set
608 * anymore, then the request isn't active and the "timeout" should
609 * just be ignored. This can happen due to the bitflag ordering.
610 * Timeout first checks if STARTED is set, and if it is, assumes
611 * the request is active. But if we race with completion, then
612 * we both flags will get cleared. So check here again, and ignore
613 * a timeout event with a request that isn't active.
615 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
616 return BLK_EH_NOT_HANDLED;
618 if (!q->mq_ops->timeout)
619 return BLK_EH_RESET_TIMER;
621 return q->mq_ops->timeout(rq);
624 static void blk_mq_rq_timer(unsigned long data)
626 struct request_queue *q = (struct request_queue *) data;
627 struct blk_mq_hw_ctx *hctx;
628 unsigned long next = 0;
631 queue_for_each_hw_ctx(q, hctx, i) {
633 * If not software queues are currently mapped to this
634 * hardware queue, there's nothing to check
636 if (!hctx->nr_ctx || !hctx->tags)
639 blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
643 next = blk_rq_timeout(round_jiffies_up(next));
644 mod_timer(&q->timeout, next);
646 queue_for_each_hw_ctx(q, hctx, i)
647 blk_mq_tag_idle(hctx);
652 * Reverse check our software queue for entries that we could potentially
653 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
654 * too much time checking for merges.
656 static bool blk_mq_attempt_merge(struct request_queue *q,
657 struct blk_mq_ctx *ctx, struct bio *bio)
662 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
668 if (!blk_rq_merge_ok(rq, bio))
671 el_ret = blk_try_merge(rq, bio);
672 if (el_ret == ELEVATOR_BACK_MERGE) {
673 if (bio_attempt_back_merge(q, rq, bio)) {
678 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
679 if (bio_attempt_front_merge(q, rq, bio)) {
691 * Process software queues that have been marked busy, splicing them
692 * to the for-dispatch
694 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
696 struct blk_mq_ctx *ctx;
699 for (i = 0; i < hctx->ctx_map.map_size; i++) {
700 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
701 unsigned int off, bit;
707 off = i * hctx->ctx_map.bits_per_word;
709 bit = find_next_bit(&bm->word, bm->depth, bit);
710 if (bit >= bm->depth)
713 ctx = hctx->ctxs[bit + off];
714 clear_bit(bit, &bm->word);
715 spin_lock(&ctx->lock);
716 list_splice_tail_init(&ctx->rq_list, list);
717 spin_unlock(&ctx->lock);
725 * Run this hardware queue, pulling any software queues mapped to it in.
726 * Note that this function currently has various problems around ordering
727 * of IO. In particular, we'd like FIFO behaviour on handling existing
728 * items on the hctx->dispatch list. Ignore that for now.
730 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
732 struct request_queue *q = hctx->queue;
737 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
739 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
745 * Touch any software queue that has pending entries.
747 flush_busy_ctxs(hctx, &rq_list);
750 * If we have previous entries on our dispatch list, grab them
751 * and stuff them at the front for more fair dispatch.
753 if (!list_empty_careful(&hctx->dispatch)) {
754 spin_lock(&hctx->lock);
755 if (!list_empty(&hctx->dispatch))
756 list_splice_init(&hctx->dispatch, &rq_list);
757 spin_unlock(&hctx->lock);
761 * Now process all the entries, sending them to the driver.
764 while (!list_empty(&rq_list)) {
767 rq = list_first_entry(&rq_list, struct request, queuelist);
768 list_del_init(&rq->queuelist);
770 blk_mq_start_request(rq, list_empty(&rq_list));
772 ret = q->mq_ops->queue_rq(hctx, rq);
774 case BLK_MQ_RQ_QUEUE_OK:
777 case BLK_MQ_RQ_QUEUE_BUSY:
778 list_add(&rq->queuelist, &rq_list);
779 __blk_mq_requeue_request(rq);
782 pr_err("blk-mq: bad return on queue: %d\n", ret);
783 case BLK_MQ_RQ_QUEUE_ERROR:
785 blk_mq_end_io(rq, rq->errors);
789 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
794 hctx->dispatched[0]++;
795 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
796 hctx->dispatched[ilog2(queued) + 1]++;
799 * Any items that need requeuing? Stuff them into hctx->dispatch,
800 * that is where we will continue on next queue run.
802 if (!list_empty(&rq_list)) {
803 spin_lock(&hctx->lock);
804 list_splice(&rq_list, &hctx->dispatch);
805 spin_unlock(&hctx->lock);
810 * It'd be great if the workqueue API had a way to pass
811 * in a mask and had some smarts for more clever placement.
812 * For now we just round-robin here, switching for every
813 * BLK_MQ_CPU_WORK_BATCH queued items.
815 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
817 int cpu = hctx->next_cpu;
819 if (--hctx->next_cpu_batch <= 0) {
822 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
823 if (next_cpu >= nr_cpu_ids)
824 next_cpu = cpumask_first(hctx->cpumask);
826 hctx->next_cpu = next_cpu;
827 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
833 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
835 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
838 if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask))
839 __blk_mq_run_hw_queue(hctx);
840 else if (hctx->queue->nr_hw_queues == 1)
841 kblockd_schedule_delayed_work(&hctx->run_work, 0);
845 cpu = blk_mq_hctx_next_cpu(hctx);
846 kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0);
850 void blk_mq_run_queues(struct request_queue *q, bool async)
852 struct blk_mq_hw_ctx *hctx;
855 queue_for_each_hw_ctx(q, hctx, i) {
856 if ((!blk_mq_hctx_has_pending(hctx) &&
857 list_empty_careful(&hctx->dispatch)) ||
858 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
862 blk_mq_run_hw_queue(hctx, async);
866 EXPORT_SYMBOL(blk_mq_run_queues);
868 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
870 cancel_delayed_work(&hctx->run_work);
871 cancel_delayed_work(&hctx->delay_work);
872 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
874 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
876 void blk_mq_stop_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_stop_hw_queue(hctx);
884 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
886 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
888 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
891 __blk_mq_run_hw_queue(hctx);
894 EXPORT_SYMBOL(blk_mq_start_hw_queue);
896 void blk_mq_start_hw_queues(struct request_queue *q)
898 struct blk_mq_hw_ctx *hctx;
901 queue_for_each_hw_ctx(q, hctx, i)
902 blk_mq_start_hw_queue(hctx);
904 EXPORT_SYMBOL(blk_mq_start_hw_queues);
907 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
909 struct blk_mq_hw_ctx *hctx;
912 queue_for_each_hw_ctx(q, hctx, i) {
913 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
916 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
918 blk_mq_run_hw_queue(hctx, async);
922 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
924 static void blk_mq_run_work_fn(struct work_struct *work)
926 struct blk_mq_hw_ctx *hctx;
928 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
930 __blk_mq_run_hw_queue(hctx);
933 static void blk_mq_delay_work_fn(struct work_struct *work)
935 struct blk_mq_hw_ctx *hctx;
937 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
939 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
940 __blk_mq_run_hw_queue(hctx);
943 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
945 unsigned long tmo = msecs_to_jiffies(msecs);
947 if (hctx->queue->nr_hw_queues == 1)
948 kblockd_schedule_delayed_work(&hctx->delay_work, tmo);
952 cpu = blk_mq_hctx_next_cpu(hctx);
953 kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo);
956 EXPORT_SYMBOL(blk_mq_delay_queue);
958 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
959 struct request *rq, bool at_head)
961 struct blk_mq_ctx *ctx = rq->mq_ctx;
963 trace_block_rq_insert(hctx->queue, rq);
966 list_add(&rq->queuelist, &ctx->rq_list);
968 list_add_tail(&rq->queuelist, &ctx->rq_list);
970 blk_mq_hctx_mark_pending(hctx, ctx);
973 * We do this early, to ensure we are on the right CPU.
978 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
981 struct request_queue *q = rq->q;
982 struct blk_mq_hw_ctx *hctx;
983 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
985 current_ctx = blk_mq_get_ctx(q);
986 if (!cpu_online(ctx->cpu))
987 rq->mq_ctx = ctx = current_ctx;
989 hctx = q->mq_ops->map_queue(q, ctx->cpu);
991 if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA) &&
992 !(rq->cmd_flags & (REQ_FLUSH_SEQ))) {
993 blk_insert_flush(rq);
995 spin_lock(&ctx->lock);
996 __blk_mq_insert_request(hctx, rq, at_head);
997 spin_unlock(&ctx->lock);
1001 blk_mq_run_hw_queue(hctx, async);
1003 blk_mq_put_ctx(current_ctx);
1006 static void blk_mq_insert_requests(struct request_queue *q,
1007 struct blk_mq_ctx *ctx,
1008 struct list_head *list,
1013 struct blk_mq_hw_ctx *hctx;
1014 struct blk_mq_ctx *current_ctx;
1016 trace_block_unplug(q, depth, !from_schedule);
1018 current_ctx = blk_mq_get_ctx(q);
1020 if (!cpu_online(ctx->cpu))
1022 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1025 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1028 spin_lock(&ctx->lock);
1029 while (!list_empty(list)) {
1032 rq = list_first_entry(list, struct request, queuelist);
1033 list_del_init(&rq->queuelist);
1035 __blk_mq_insert_request(hctx, rq, false);
1037 spin_unlock(&ctx->lock);
1039 blk_mq_run_hw_queue(hctx, from_schedule);
1040 blk_mq_put_ctx(current_ctx);
1043 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1045 struct request *rqa = container_of(a, struct request, queuelist);
1046 struct request *rqb = container_of(b, struct request, queuelist);
1048 return !(rqa->mq_ctx < rqb->mq_ctx ||
1049 (rqa->mq_ctx == rqb->mq_ctx &&
1050 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1053 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1055 struct blk_mq_ctx *this_ctx;
1056 struct request_queue *this_q;
1059 LIST_HEAD(ctx_list);
1062 list_splice_init(&plug->mq_list, &list);
1064 list_sort(NULL, &list, plug_ctx_cmp);
1070 while (!list_empty(&list)) {
1071 rq = list_entry_rq(list.next);
1072 list_del_init(&rq->queuelist);
1074 if (rq->mq_ctx != this_ctx) {
1076 blk_mq_insert_requests(this_q, this_ctx,
1081 this_ctx = rq->mq_ctx;
1087 list_add_tail(&rq->queuelist, &ctx_list);
1091 * If 'this_ctx' is set, we know we have entries to complete
1092 * on 'ctx_list'. Do those.
1095 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1100 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1102 init_request_from_bio(rq, bio);
1104 if (blk_do_io_stat(rq)) {
1105 rq->start_time = jiffies;
1106 blk_account_io_start(rq, 1);
1110 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1111 struct blk_mq_ctx *ctx,
1112 struct request *rq, struct bio *bio)
1114 struct request_queue *q = hctx->queue;
1116 if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE)) {
1117 blk_mq_bio_to_request(rq, bio);
1118 spin_lock(&ctx->lock);
1120 __blk_mq_insert_request(hctx, rq, false);
1121 spin_unlock(&ctx->lock);
1124 spin_lock(&ctx->lock);
1125 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1126 blk_mq_bio_to_request(rq, bio);
1130 spin_unlock(&ctx->lock);
1131 __blk_mq_free_request(hctx, ctx, rq);
1136 struct blk_map_ctx {
1137 struct blk_mq_hw_ctx *hctx;
1138 struct blk_mq_ctx *ctx;
1141 static struct request *blk_mq_map_request(struct request_queue *q,
1143 struct blk_map_ctx *data)
1145 struct blk_mq_hw_ctx *hctx;
1146 struct blk_mq_ctx *ctx;
1148 int rw = bio_data_dir(bio);
1150 if (unlikely(blk_mq_queue_enter(q))) {
1151 bio_endio(bio, -EIO);
1155 ctx = blk_mq_get_ctx(q);
1156 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1158 if (rw_is_sync(bio->bi_rw))
1161 trace_block_getrq(q, bio, rw);
1162 rq = __blk_mq_alloc_request(q, hctx, ctx, rw, GFP_ATOMIC, false);
1163 if (unlikely(!rq)) {
1164 __blk_mq_run_hw_queue(hctx);
1165 blk_mq_put_ctx(ctx);
1166 trace_block_sleeprq(q, bio, rw);
1168 ctx = blk_mq_get_ctx(q);
1169 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1170 rq = __blk_mq_alloc_request(q, hctx, ctx, rw,
1171 __GFP_WAIT|GFP_ATOMIC, false);
1181 * Multiple hardware queue variant. This will not use per-process plugs,
1182 * but will attempt to bypass the hctx queueing if we can go straight to
1183 * hardware for SYNC IO.
1185 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1187 const int is_sync = rw_is_sync(bio->bi_rw);
1188 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1189 struct blk_map_ctx data;
1192 blk_queue_bounce(q, &bio);
1194 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1195 bio_endio(bio, -EIO);
1199 rq = blk_mq_map_request(q, bio, &data);
1203 if (unlikely(is_flush_fua)) {
1204 blk_mq_bio_to_request(rq, bio);
1205 blk_insert_flush(rq);
1212 blk_mq_bio_to_request(rq, bio);
1213 blk_mq_start_request(rq, true);
1216 * For OK queue, we are done. For error, kill it. Any other
1217 * error (busy), just add it to our list as we previously
1220 ret = q->mq_ops->queue_rq(data.hctx, rq);
1221 if (ret == BLK_MQ_RQ_QUEUE_OK)
1224 __blk_mq_requeue_request(rq);
1226 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1228 blk_mq_end_io(rq, rq->errors);
1234 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1236 * For a SYNC request, send it to the hardware immediately. For
1237 * an ASYNC request, just ensure that we run it later on. The
1238 * latter allows for merging opportunities and more efficient
1242 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1245 blk_mq_put_ctx(data.ctx);
1249 * Single hardware queue variant. This will attempt to use any per-process
1250 * plug for merging and IO deferral.
1252 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1254 const int is_sync = rw_is_sync(bio->bi_rw);
1255 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1256 unsigned int use_plug, request_count = 0;
1257 struct blk_map_ctx data;
1261 * If we have multiple hardware queues, just go directly to
1262 * one of those for sync IO.
1264 use_plug = !is_flush_fua && !is_sync;
1266 blk_queue_bounce(q, &bio);
1268 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1269 bio_endio(bio, -EIO);
1273 if (use_plug && !blk_queue_nomerges(q) &&
1274 blk_attempt_plug_merge(q, bio, &request_count))
1277 rq = blk_mq_map_request(q, bio, &data);
1279 if (unlikely(is_flush_fua)) {
1280 blk_mq_bio_to_request(rq, bio);
1281 blk_insert_flush(rq);
1286 * A task plug currently exists. Since this is completely lockless,
1287 * utilize that to temporarily store requests until the task is
1288 * either done or scheduled away.
1291 struct blk_plug *plug = current->plug;
1294 blk_mq_bio_to_request(rq, bio);
1295 if (list_empty(&plug->mq_list))
1296 trace_block_plug(q);
1297 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1298 blk_flush_plug_list(plug, false);
1299 trace_block_plug(q);
1301 list_add_tail(&rq->queuelist, &plug->mq_list);
1302 blk_mq_put_ctx(data.ctx);
1307 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1309 * For a SYNC request, send it to the hardware immediately. For
1310 * an ASYNC request, just ensure that we run it later on. The
1311 * latter allows for merging opportunities and more efficient
1315 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1318 blk_mq_put_ctx(data.ctx);
1322 * Default mapping to a software queue, since we use one per CPU.
1324 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1326 return q->queue_hw_ctx[q->mq_map[cpu]];
1328 EXPORT_SYMBOL(blk_mq_map_queue);
1330 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1331 struct blk_mq_tags *tags, unsigned int hctx_idx)
1335 if (tags->rqs && set->ops->exit_request) {
1338 for (i = 0; i < tags->nr_tags; i++) {
1341 set->ops->exit_request(set->driver_data, tags->rqs[i],
1346 while (!list_empty(&tags->page_list)) {
1347 page = list_first_entry(&tags->page_list, struct page, lru);
1348 list_del_init(&page->lru);
1349 __free_pages(page, page->private);
1354 blk_mq_free_tags(tags);
1357 static size_t order_to_size(unsigned int order)
1359 return (size_t)PAGE_SIZE << order;
1362 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1363 unsigned int hctx_idx)
1365 struct blk_mq_tags *tags;
1366 unsigned int i, j, entries_per_page, max_order = 4;
1367 size_t rq_size, left;
1369 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1374 INIT_LIST_HEAD(&tags->page_list);
1376 tags->rqs = kmalloc_node(set->queue_depth * sizeof(struct request *),
1377 GFP_KERNEL, set->numa_node);
1379 blk_mq_free_tags(tags);
1384 * rq_size is the size of the request plus driver payload, rounded
1385 * to the cacheline size
1387 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1389 left = rq_size * set->queue_depth;
1391 for (i = 0; i < set->queue_depth; ) {
1392 int this_order = max_order;
1397 while (left < order_to_size(this_order - 1) && this_order)
1401 page = alloc_pages_node(set->numa_node, GFP_KERNEL,
1407 if (order_to_size(this_order) < rq_size)
1414 page->private = this_order;
1415 list_add_tail(&page->lru, &tags->page_list);
1417 p = page_address(page);
1418 entries_per_page = order_to_size(this_order) / rq_size;
1419 to_do = min(entries_per_page, set->queue_depth - i);
1420 left -= to_do * rq_size;
1421 for (j = 0; j < to_do; j++) {
1423 if (set->ops->init_request) {
1424 if (set->ops->init_request(set->driver_data,
1425 tags->rqs[i], hctx_idx, i,
1438 pr_warn("%s: failed to allocate requests\n", __func__);
1439 blk_mq_free_rq_map(set, tags, hctx_idx);
1443 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1448 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1450 unsigned int bpw = 8, total, num_maps, i;
1452 bitmap->bits_per_word = bpw;
1454 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1455 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1460 bitmap->map_size = num_maps;
1463 for (i = 0; i < num_maps; i++) {
1464 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1465 total -= bitmap->map[i].depth;
1471 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1473 struct request_queue *q = hctx->queue;
1474 struct blk_mq_ctx *ctx;
1478 * Move ctx entries to new CPU, if this one is going away.
1480 ctx = __blk_mq_get_ctx(q, cpu);
1482 spin_lock(&ctx->lock);
1483 if (!list_empty(&ctx->rq_list)) {
1484 list_splice_init(&ctx->rq_list, &tmp);
1485 blk_mq_hctx_clear_pending(hctx, ctx);
1487 spin_unlock(&ctx->lock);
1489 if (list_empty(&tmp))
1492 ctx = blk_mq_get_ctx(q);
1493 spin_lock(&ctx->lock);
1495 while (!list_empty(&tmp)) {
1498 rq = list_first_entry(&tmp, struct request, queuelist);
1500 list_move_tail(&rq->queuelist, &ctx->rq_list);
1503 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1504 blk_mq_hctx_mark_pending(hctx, ctx);
1506 spin_unlock(&ctx->lock);
1508 blk_mq_run_hw_queue(hctx, true);
1509 blk_mq_put_ctx(ctx);
1513 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1515 struct request_queue *q = hctx->queue;
1516 struct blk_mq_tag_set *set = q->tag_set;
1518 if (set->tags[hctx->queue_num])
1521 set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1522 if (!set->tags[hctx->queue_num])
1525 hctx->tags = set->tags[hctx->queue_num];
1529 static int blk_mq_hctx_notify(void *data, unsigned long action,
1532 struct blk_mq_hw_ctx *hctx = data;
1534 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1535 return blk_mq_hctx_cpu_offline(hctx, cpu);
1536 else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1537 return blk_mq_hctx_cpu_online(hctx, cpu);
1542 static void blk_mq_exit_hw_queues(struct request_queue *q,
1543 struct blk_mq_tag_set *set, int nr_queue)
1545 struct blk_mq_hw_ctx *hctx;
1548 queue_for_each_hw_ctx(q, hctx, i) {
1552 if (set->ops->exit_hctx)
1553 set->ops->exit_hctx(hctx, i);
1555 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1557 blk_mq_free_bitmap(&hctx->ctx_map);
1562 static void blk_mq_free_hw_queues(struct request_queue *q,
1563 struct blk_mq_tag_set *set)
1565 struct blk_mq_hw_ctx *hctx;
1568 queue_for_each_hw_ctx(q, hctx, i) {
1569 free_cpumask_var(hctx->cpumask);
1574 static int blk_mq_init_hw_queues(struct request_queue *q,
1575 struct blk_mq_tag_set *set)
1577 struct blk_mq_hw_ctx *hctx;
1581 * Initialize hardware queues
1583 queue_for_each_hw_ctx(q, hctx, i) {
1586 node = hctx->numa_node;
1587 if (node == NUMA_NO_NODE)
1588 node = hctx->numa_node = set->numa_node;
1590 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1591 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1592 spin_lock_init(&hctx->lock);
1593 INIT_LIST_HEAD(&hctx->dispatch);
1595 hctx->queue_num = i;
1596 hctx->flags = set->flags;
1597 hctx->cmd_size = set->cmd_size;
1599 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1600 blk_mq_hctx_notify, hctx);
1601 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1603 hctx->tags = set->tags[i];
1606 * Allocate space for all possible cpus to avoid allocation in
1609 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1614 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1619 if (set->ops->init_hctx &&
1620 set->ops->init_hctx(hctx, set->driver_data, i))
1624 if (i == q->nr_hw_queues)
1630 blk_mq_exit_hw_queues(q, set, i);
1635 static void blk_mq_init_cpu_queues(struct request_queue *q,
1636 unsigned int nr_hw_queues)
1640 for_each_possible_cpu(i) {
1641 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1642 struct blk_mq_hw_ctx *hctx;
1644 memset(__ctx, 0, sizeof(*__ctx));
1646 spin_lock_init(&__ctx->lock);
1647 INIT_LIST_HEAD(&__ctx->rq_list);
1650 /* If the cpu isn't online, the cpu is mapped to first hctx */
1654 hctx = q->mq_ops->map_queue(q, i);
1655 cpumask_set_cpu(i, hctx->cpumask);
1659 * Set local node, IFF we have more than one hw queue. If
1660 * not, we remain on the home node of the device
1662 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1663 hctx->numa_node = cpu_to_node(i);
1667 static void blk_mq_map_swqueue(struct request_queue *q)
1670 struct blk_mq_hw_ctx *hctx;
1671 struct blk_mq_ctx *ctx;
1673 queue_for_each_hw_ctx(q, hctx, i) {
1674 cpumask_clear(hctx->cpumask);
1679 * Map software to hardware queues
1681 queue_for_each_ctx(q, ctx, i) {
1682 /* If the cpu isn't online, the cpu is mapped to first hctx */
1686 hctx = q->mq_ops->map_queue(q, i);
1687 cpumask_set_cpu(i, hctx->cpumask);
1688 ctx->index_hw = hctx->nr_ctx;
1689 hctx->ctxs[hctx->nr_ctx++] = ctx;
1692 queue_for_each_hw_ctx(q, hctx, i) {
1694 * If not software queues are mapped to this hardware queue,
1695 * disable it and free the request entries
1697 if (!hctx->nr_ctx) {
1698 struct blk_mq_tag_set *set = q->tag_set;
1701 blk_mq_free_rq_map(set, set->tags[i], i);
1702 set->tags[i] = NULL;
1709 * Initialize batch roundrobin counts
1711 hctx->next_cpu = cpumask_first(hctx->cpumask);
1712 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1716 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1718 struct blk_mq_hw_ctx *hctx;
1719 struct request_queue *q;
1723 if (set->tag_list.next == set->tag_list.prev)
1728 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1729 blk_mq_freeze_queue(q);
1731 queue_for_each_hw_ctx(q, hctx, i) {
1733 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1735 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1737 blk_mq_unfreeze_queue(q);
1741 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1743 struct blk_mq_tag_set *set = q->tag_set;
1745 blk_mq_freeze_queue(q);
1747 mutex_lock(&set->tag_list_lock);
1748 list_del_init(&q->tag_set_list);
1749 blk_mq_update_tag_set_depth(set);
1750 mutex_unlock(&set->tag_list_lock);
1752 blk_mq_unfreeze_queue(q);
1755 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1756 struct request_queue *q)
1760 mutex_lock(&set->tag_list_lock);
1761 list_add_tail(&q->tag_set_list, &set->tag_list);
1762 blk_mq_update_tag_set_depth(set);
1763 mutex_unlock(&set->tag_list_lock);
1766 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1768 struct blk_mq_hw_ctx **hctxs;
1769 struct blk_mq_ctx *ctx;
1770 struct request_queue *q;
1774 ctx = alloc_percpu(struct blk_mq_ctx);
1776 return ERR_PTR(-ENOMEM);
1778 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1784 map = blk_mq_make_queue_map(set);
1788 for (i = 0; i < set->nr_hw_queues; i++) {
1789 int node = blk_mq_hw_queue_to_node(map, i);
1791 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1796 if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL))
1799 atomic_set(&hctxs[i]->nr_active, 0);
1800 hctxs[i]->numa_node = node;
1801 hctxs[i]->queue_num = i;
1804 q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1808 if (percpu_counter_init(&q->mq_usage_counter, 0))
1811 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1812 blk_queue_rq_timeout(q, 30000);
1814 q->nr_queues = nr_cpu_ids;
1815 q->nr_hw_queues = set->nr_hw_queues;
1819 q->queue_hw_ctx = hctxs;
1821 q->mq_ops = set->ops;
1822 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1824 if (!(set->flags & BLK_MQ_F_SG_MERGE))
1825 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1827 q->sg_reserved_size = INT_MAX;
1829 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1830 INIT_LIST_HEAD(&q->requeue_list);
1831 spin_lock_init(&q->requeue_lock);
1833 if (q->nr_hw_queues > 1)
1834 blk_queue_make_request(q, blk_mq_make_request);
1836 blk_queue_make_request(q, blk_sq_make_request);
1838 blk_queue_rq_timed_out(q, blk_mq_rq_timed_out);
1840 blk_queue_rq_timeout(q, set->timeout);
1843 * Do this after blk_queue_make_request() overrides it...
1845 q->nr_requests = set->queue_depth;
1847 if (set->ops->complete)
1848 blk_queue_softirq_done(q, set->ops->complete);
1850 blk_mq_init_flush(q);
1851 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1853 q->flush_rq = kzalloc(round_up(sizeof(struct request) +
1854 set->cmd_size, cache_line_size()),
1859 if (blk_mq_init_hw_queues(q, set))
1862 mutex_lock(&all_q_mutex);
1863 list_add_tail(&q->all_q_node, &all_q_list);
1864 mutex_unlock(&all_q_mutex);
1866 blk_mq_add_queue_tag_set(set, q);
1868 blk_mq_map_swqueue(q);
1875 blk_cleanup_queue(q);
1878 for (i = 0; i < set->nr_hw_queues; i++) {
1881 free_cpumask_var(hctxs[i]->cpumask);
1888 return ERR_PTR(-ENOMEM);
1890 EXPORT_SYMBOL(blk_mq_init_queue);
1892 void blk_mq_free_queue(struct request_queue *q)
1894 struct blk_mq_tag_set *set = q->tag_set;
1896 blk_mq_del_queue_tag_set(q);
1898 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
1899 blk_mq_free_hw_queues(q, set);
1901 percpu_counter_destroy(&q->mq_usage_counter);
1903 free_percpu(q->queue_ctx);
1904 kfree(q->queue_hw_ctx);
1907 q->queue_ctx = NULL;
1908 q->queue_hw_ctx = NULL;
1911 mutex_lock(&all_q_mutex);
1912 list_del_init(&q->all_q_node);
1913 mutex_unlock(&all_q_mutex);
1916 /* Basically redo blk_mq_init_queue with queue frozen */
1917 static void blk_mq_queue_reinit(struct request_queue *q)
1919 blk_mq_freeze_queue(q);
1921 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1924 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1925 * we should change hctx numa_node according to new topology (this
1926 * involves free and re-allocate memory, worthy doing?)
1929 blk_mq_map_swqueue(q);
1931 blk_mq_unfreeze_queue(q);
1934 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1935 unsigned long action, void *hcpu)
1937 struct request_queue *q;
1940 * Before new mappings are established, hotadded cpu might already
1941 * start handling requests. This doesn't break anything as we map
1942 * offline CPUs to first hardware queue. We will re-init the queue
1943 * below to get optimal settings.
1945 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1946 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1949 mutex_lock(&all_q_mutex);
1950 list_for_each_entry(q, &all_q_list, all_q_node)
1951 blk_mq_queue_reinit(q);
1952 mutex_unlock(&all_q_mutex);
1956 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
1960 if (!set->nr_hw_queues)
1962 if (!set->queue_depth || set->queue_depth > BLK_MQ_MAX_DEPTH)
1964 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
1967 if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue)
1971 set->tags = kmalloc_node(set->nr_hw_queues *
1972 sizeof(struct blk_mq_tags *),
1973 GFP_KERNEL, set->numa_node);
1977 for (i = 0; i < set->nr_hw_queues; i++) {
1978 set->tags[i] = blk_mq_init_rq_map(set, i);
1983 mutex_init(&set->tag_list_lock);
1984 INIT_LIST_HEAD(&set->tag_list);
1990 blk_mq_free_rq_map(set, set->tags[i], i);
1994 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
1996 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2000 for (i = 0; i < set->nr_hw_queues; i++) {
2002 blk_mq_free_rq_map(set, set->tags[i], i);
2007 EXPORT_SYMBOL(blk_mq_free_tag_set);
2009 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2011 struct blk_mq_tag_set *set = q->tag_set;
2012 struct blk_mq_hw_ctx *hctx;
2015 if (!set || nr > set->queue_depth)
2019 queue_for_each_hw_ctx(q, hctx, i) {
2020 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2026 q->nr_requests = nr;
2031 void blk_mq_disable_hotplug(void)
2033 mutex_lock(&all_q_mutex);
2036 void blk_mq_enable_hotplug(void)
2038 mutex_unlock(&all_q_mutex);
2041 static int __init blk_mq_init(void)
2045 /* Must be called after percpu_counter_hotcpu_callback() */
2046 hotcpu_notifier(blk_mq_queue_reinit_notify, -10);
2050 subsys_initcall(blk_mq_init);