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_hw_ctx *hctx, unsigned int tag)
546 struct request_queue *q = hctx->queue;
548 if ((q->flush_rq->cmd_flags & REQ_FLUSH_SEQ) &&
549 q->flush_rq->tag == tag)
552 return hctx->tags->rqs[tag];
554 EXPORT_SYMBOL(blk_mq_tag_to_rq);
556 struct blk_mq_timeout_data {
557 struct blk_mq_hw_ctx *hctx;
559 unsigned int *next_set;
562 static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
564 struct blk_mq_timeout_data *data = __data;
565 struct blk_mq_hw_ctx *hctx = data->hctx;
568 /* It may not be in flight yet (this is where
569 * the REQ_ATOMIC_STARTED flag comes in). The requests are
570 * statically allocated, so we know it's always safe to access the
571 * memory associated with a bit offset into ->rqs[].
577 tag = find_next_zero_bit(free_tags, hctx->tags->nr_tags, tag);
578 if (tag >= hctx->tags->nr_tags)
581 rq = blk_mq_tag_to_rq(hctx, tag++);
582 if (rq->q != hctx->queue)
584 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
587 blk_rq_check_expired(rq, data->next, data->next_set);
591 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
593 unsigned int *next_set)
595 struct blk_mq_timeout_data data = {
598 .next_set = next_set,
602 * Ask the tagging code to iterate busy requests, so we can
603 * check them for timeout.
605 blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
608 static enum blk_eh_timer_return blk_mq_rq_timed_out(struct request *rq)
610 struct request_queue *q = rq->q;
613 * We know that complete is set at this point. If STARTED isn't set
614 * anymore, then the request isn't active and the "timeout" should
615 * just be ignored. This can happen due to the bitflag ordering.
616 * Timeout first checks if STARTED is set, and if it is, assumes
617 * the request is active. But if we race with completion, then
618 * we both flags will get cleared. So check here again, and ignore
619 * a timeout event with a request that isn't active.
621 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
622 return BLK_EH_NOT_HANDLED;
624 if (!q->mq_ops->timeout)
625 return BLK_EH_RESET_TIMER;
627 return q->mq_ops->timeout(rq);
630 static void blk_mq_rq_timer(unsigned long data)
632 struct request_queue *q = (struct request_queue *) data;
633 struct blk_mq_hw_ctx *hctx;
634 unsigned long next = 0;
637 queue_for_each_hw_ctx(q, hctx, i) {
639 * If not software queues are currently mapped to this
640 * hardware queue, there's nothing to check
642 if (!hctx->nr_ctx || !hctx->tags)
645 blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
649 next = blk_rq_timeout(round_jiffies_up(next));
650 mod_timer(&q->timeout, next);
652 queue_for_each_hw_ctx(q, hctx, i)
653 blk_mq_tag_idle(hctx);
658 * Reverse check our software queue for entries that we could potentially
659 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
660 * too much time checking for merges.
662 static bool blk_mq_attempt_merge(struct request_queue *q,
663 struct blk_mq_ctx *ctx, struct bio *bio)
668 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
674 if (!blk_rq_merge_ok(rq, bio))
677 el_ret = blk_try_merge(rq, bio);
678 if (el_ret == ELEVATOR_BACK_MERGE) {
679 if (bio_attempt_back_merge(q, rq, bio)) {
684 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
685 if (bio_attempt_front_merge(q, rq, bio)) {
697 * Process software queues that have been marked busy, splicing them
698 * to the for-dispatch
700 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
702 struct blk_mq_ctx *ctx;
705 for (i = 0; i < hctx->ctx_map.map_size; i++) {
706 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
707 unsigned int off, bit;
713 off = i * hctx->ctx_map.bits_per_word;
715 bit = find_next_bit(&bm->word, bm->depth, bit);
716 if (bit >= bm->depth)
719 ctx = hctx->ctxs[bit + off];
720 clear_bit(bit, &bm->word);
721 spin_lock(&ctx->lock);
722 list_splice_tail_init(&ctx->rq_list, list);
723 spin_unlock(&ctx->lock);
731 * Run this hardware queue, pulling any software queues mapped to it in.
732 * Note that this function currently has various problems around ordering
733 * of IO. In particular, we'd like FIFO behaviour on handling existing
734 * items on the hctx->dispatch list. Ignore that for now.
736 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
738 struct request_queue *q = hctx->queue;
743 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
745 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
751 * Touch any software queue that has pending entries.
753 flush_busy_ctxs(hctx, &rq_list);
756 * If we have previous entries on our dispatch list, grab them
757 * and stuff them at the front for more fair dispatch.
759 if (!list_empty_careful(&hctx->dispatch)) {
760 spin_lock(&hctx->lock);
761 if (!list_empty(&hctx->dispatch))
762 list_splice_init(&hctx->dispatch, &rq_list);
763 spin_unlock(&hctx->lock);
767 * Now process all the entries, sending them to the driver.
770 while (!list_empty(&rq_list)) {
773 rq = list_first_entry(&rq_list, struct request, queuelist);
774 list_del_init(&rq->queuelist);
776 blk_mq_start_request(rq, list_empty(&rq_list));
778 ret = q->mq_ops->queue_rq(hctx, rq);
780 case BLK_MQ_RQ_QUEUE_OK:
783 case BLK_MQ_RQ_QUEUE_BUSY:
784 list_add(&rq->queuelist, &rq_list);
785 __blk_mq_requeue_request(rq);
788 pr_err("blk-mq: bad return on queue: %d\n", ret);
789 case BLK_MQ_RQ_QUEUE_ERROR:
791 blk_mq_end_io(rq, rq->errors);
795 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
800 hctx->dispatched[0]++;
801 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
802 hctx->dispatched[ilog2(queued) + 1]++;
805 * Any items that need requeuing? Stuff them into hctx->dispatch,
806 * that is where we will continue on next queue run.
808 if (!list_empty(&rq_list)) {
809 spin_lock(&hctx->lock);
810 list_splice(&rq_list, &hctx->dispatch);
811 spin_unlock(&hctx->lock);
816 * It'd be great if the workqueue API had a way to pass
817 * in a mask and had some smarts for more clever placement.
818 * For now we just round-robin here, switching for every
819 * BLK_MQ_CPU_WORK_BATCH queued items.
821 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
823 int cpu = hctx->next_cpu;
825 if (--hctx->next_cpu_batch <= 0) {
828 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
829 if (next_cpu >= nr_cpu_ids)
830 next_cpu = cpumask_first(hctx->cpumask);
832 hctx->next_cpu = next_cpu;
833 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
839 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
841 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
844 if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask))
845 __blk_mq_run_hw_queue(hctx);
846 else if (hctx->queue->nr_hw_queues == 1)
847 kblockd_schedule_delayed_work(&hctx->run_work, 0);
851 cpu = blk_mq_hctx_next_cpu(hctx);
852 kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0);
856 void blk_mq_run_queues(struct request_queue *q, bool async)
858 struct blk_mq_hw_ctx *hctx;
861 queue_for_each_hw_ctx(q, hctx, i) {
862 if ((!blk_mq_hctx_has_pending(hctx) &&
863 list_empty_careful(&hctx->dispatch)) ||
864 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
868 blk_mq_run_hw_queue(hctx, async);
872 EXPORT_SYMBOL(blk_mq_run_queues);
874 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
876 cancel_delayed_work(&hctx->run_work);
877 cancel_delayed_work(&hctx->delay_work);
878 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
880 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
882 void blk_mq_stop_hw_queues(struct request_queue *q)
884 struct blk_mq_hw_ctx *hctx;
887 queue_for_each_hw_ctx(q, hctx, i)
888 blk_mq_stop_hw_queue(hctx);
890 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
892 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
894 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
897 __blk_mq_run_hw_queue(hctx);
900 EXPORT_SYMBOL(blk_mq_start_hw_queue);
902 void blk_mq_start_hw_queues(struct request_queue *q)
904 struct blk_mq_hw_ctx *hctx;
907 queue_for_each_hw_ctx(q, hctx, i)
908 blk_mq_start_hw_queue(hctx);
910 EXPORT_SYMBOL(blk_mq_start_hw_queues);
913 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
915 struct blk_mq_hw_ctx *hctx;
918 queue_for_each_hw_ctx(q, hctx, i) {
919 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
922 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
924 blk_mq_run_hw_queue(hctx, async);
928 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
930 static void blk_mq_run_work_fn(struct work_struct *work)
932 struct blk_mq_hw_ctx *hctx;
934 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
936 __blk_mq_run_hw_queue(hctx);
939 static void blk_mq_delay_work_fn(struct work_struct *work)
941 struct blk_mq_hw_ctx *hctx;
943 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
945 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
946 __blk_mq_run_hw_queue(hctx);
949 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
951 unsigned long tmo = msecs_to_jiffies(msecs);
953 if (hctx->queue->nr_hw_queues == 1)
954 kblockd_schedule_delayed_work(&hctx->delay_work, tmo);
958 cpu = blk_mq_hctx_next_cpu(hctx);
959 kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo);
962 EXPORT_SYMBOL(blk_mq_delay_queue);
964 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
965 struct request *rq, bool at_head)
967 struct blk_mq_ctx *ctx = rq->mq_ctx;
969 trace_block_rq_insert(hctx->queue, rq);
972 list_add(&rq->queuelist, &ctx->rq_list);
974 list_add_tail(&rq->queuelist, &ctx->rq_list);
976 blk_mq_hctx_mark_pending(hctx, ctx);
979 * We do this early, to ensure we are on the right CPU.
984 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
987 struct request_queue *q = rq->q;
988 struct blk_mq_hw_ctx *hctx;
989 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
991 current_ctx = blk_mq_get_ctx(q);
992 if (!cpu_online(ctx->cpu))
993 rq->mq_ctx = ctx = current_ctx;
995 hctx = q->mq_ops->map_queue(q, ctx->cpu);
997 if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA) &&
998 !(rq->cmd_flags & (REQ_FLUSH_SEQ))) {
999 blk_insert_flush(rq);
1001 spin_lock(&ctx->lock);
1002 __blk_mq_insert_request(hctx, rq, at_head);
1003 spin_unlock(&ctx->lock);
1007 blk_mq_run_hw_queue(hctx, async);
1009 blk_mq_put_ctx(current_ctx);
1012 static void blk_mq_insert_requests(struct request_queue *q,
1013 struct blk_mq_ctx *ctx,
1014 struct list_head *list,
1019 struct blk_mq_hw_ctx *hctx;
1020 struct blk_mq_ctx *current_ctx;
1022 trace_block_unplug(q, depth, !from_schedule);
1024 current_ctx = blk_mq_get_ctx(q);
1026 if (!cpu_online(ctx->cpu))
1028 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1031 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1034 spin_lock(&ctx->lock);
1035 while (!list_empty(list)) {
1038 rq = list_first_entry(list, struct request, queuelist);
1039 list_del_init(&rq->queuelist);
1041 __blk_mq_insert_request(hctx, rq, false);
1043 spin_unlock(&ctx->lock);
1045 blk_mq_run_hw_queue(hctx, from_schedule);
1046 blk_mq_put_ctx(current_ctx);
1049 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1051 struct request *rqa = container_of(a, struct request, queuelist);
1052 struct request *rqb = container_of(b, struct request, queuelist);
1054 return !(rqa->mq_ctx < rqb->mq_ctx ||
1055 (rqa->mq_ctx == rqb->mq_ctx &&
1056 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1059 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1061 struct blk_mq_ctx *this_ctx;
1062 struct request_queue *this_q;
1065 LIST_HEAD(ctx_list);
1068 list_splice_init(&plug->mq_list, &list);
1070 list_sort(NULL, &list, plug_ctx_cmp);
1076 while (!list_empty(&list)) {
1077 rq = list_entry_rq(list.next);
1078 list_del_init(&rq->queuelist);
1080 if (rq->mq_ctx != this_ctx) {
1082 blk_mq_insert_requests(this_q, this_ctx,
1087 this_ctx = rq->mq_ctx;
1093 list_add_tail(&rq->queuelist, &ctx_list);
1097 * If 'this_ctx' is set, we know we have entries to complete
1098 * on 'ctx_list'. Do those.
1101 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1106 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1108 init_request_from_bio(rq, bio);
1110 if (blk_do_io_stat(rq)) {
1111 rq->start_time = jiffies;
1112 blk_account_io_start(rq, 1);
1116 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1117 struct blk_mq_ctx *ctx,
1118 struct request *rq, struct bio *bio)
1120 struct request_queue *q = hctx->queue;
1122 if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE)) {
1123 blk_mq_bio_to_request(rq, bio);
1124 spin_lock(&ctx->lock);
1126 __blk_mq_insert_request(hctx, rq, false);
1127 spin_unlock(&ctx->lock);
1130 spin_lock(&ctx->lock);
1131 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1132 blk_mq_bio_to_request(rq, bio);
1136 spin_unlock(&ctx->lock);
1137 __blk_mq_free_request(hctx, ctx, rq);
1142 struct blk_map_ctx {
1143 struct blk_mq_hw_ctx *hctx;
1144 struct blk_mq_ctx *ctx;
1147 static struct request *blk_mq_map_request(struct request_queue *q,
1149 struct blk_map_ctx *data)
1151 struct blk_mq_hw_ctx *hctx;
1152 struct blk_mq_ctx *ctx;
1154 int rw = bio_data_dir(bio);
1156 if (unlikely(blk_mq_queue_enter(q))) {
1157 bio_endio(bio, -EIO);
1161 ctx = blk_mq_get_ctx(q);
1162 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1164 if (rw_is_sync(bio->bi_rw))
1167 trace_block_getrq(q, bio, rw);
1168 rq = __blk_mq_alloc_request(q, hctx, ctx, rw, GFP_ATOMIC, false);
1169 if (unlikely(!rq)) {
1170 __blk_mq_run_hw_queue(hctx);
1171 blk_mq_put_ctx(ctx);
1172 trace_block_sleeprq(q, bio, rw);
1174 ctx = blk_mq_get_ctx(q);
1175 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1176 rq = __blk_mq_alloc_request(q, hctx, ctx, rw,
1177 __GFP_WAIT|GFP_ATOMIC, false);
1187 * Multiple hardware queue variant. This will not use per-process plugs,
1188 * but will attempt to bypass the hctx queueing if we can go straight to
1189 * hardware for SYNC IO.
1191 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1193 const int is_sync = rw_is_sync(bio->bi_rw);
1194 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1195 struct blk_map_ctx data;
1198 blk_queue_bounce(q, &bio);
1200 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1201 bio_endio(bio, -EIO);
1205 rq = blk_mq_map_request(q, bio, &data);
1209 if (unlikely(is_flush_fua)) {
1210 blk_mq_bio_to_request(rq, bio);
1211 blk_insert_flush(rq);
1218 blk_mq_bio_to_request(rq, bio);
1219 blk_mq_start_request(rq, true);
1222 * For OK queue, we are done. For error, kill it. Any other
1223 * error (busy), just add it to our list as we previously
1226 ret = q->mq_ops->queue_rq(data.hctx, rq);
1227 if (ret == BLK_MQ_RQ_QUEUE_OK)
1230 __blk_mq_requeue_request(rq);
1232 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1234 blk_mq_end_io(rq, rq->errors);
1240 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1242 * For a SYNC request, send it to the hardware immediately. For
1243 * an ASYNC request, just ensure that we run it later on. The
1244 * latter allows for merging opportunities and more efficient
1248 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1251 blk_mq_put_ctx(data.ctx);
1255 * Single hardware queue variant. This will attempt to use any per-process
1256 * plug for merging and IO deferral.
1258 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1260 const int is_sync = rw_is_sync(bio->bi_rw);
1261 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1262 unsigned int use_plug, request_count = 0;
1263 struct blk_map_ctx data;
1267 * If we have multiple hardware queues, just go directly to
1268 * one of those for sync IO.
1270 use_plug = !is_flush_fua && !is_sync;
1272 blk_queue_bounce(q, &bio);
1274 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1275 bio_endio(bio, -EIO);
1279 if (use_plug && !blk_queue_nomerges(q) &&
1280 blk_attempt_plug_merge(q, bio, &request_count))
1283 rq = blk_mq_map_request(q, bio, &data);
1285 if (unlikely(is_flush_fua)) {
1286 blk_mq_bio_to_request(rq, bio);
1287 blk_insert_flush(rq);
1292 * A task plug currently exists. Since this is completely lockless,
1293 * utilize that to temporarily store requests until the task is
1294 * either done or scheduled away.
1297 struct blk_plug *plug = current->plug;
1300 blk_mq_bio_to_request(rq, bio);
1301 if (list_empty(&plug->mq_list))
1302 trace_block_plug(q);
1303 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1304 blk_flush_plug_list(plug, false);
1305 trace_block_plug(q);
1307 list_add_tail(&rq->queuelist, &plug->mq_list);
1308 blk_mq_put_ctx(data.ctx);
1313 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1315 * For a SYNC request, send it to the hardware immediately. For
1316 * an ASYNC request, just ensure that we run it later on. The
1317 * latter allows for merging opportunities and more efficient
1321 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1324 blk_mq_put_ctx(data.ctx);
1328 * Default mapping to a software queue, since we use one per CPU.
1330 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1332 return q->queue_hw_ctx[q->mq_map[cpu]];
1334 EXPORT_SYMBOL(blk_mq_map_queue);
1336 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1337 struct blk_mq_tags *tags, unsigned int hctx_idx)
1341 if (tags->rqs && set->ops->exit_request) {
1344 for (i = 0; i < tags->nr_tags; i++) {
1347 set->ops->exit_request(set->driver_data, tags->rqs[i],
1352 while (!list_empty(&tags->page_list)) {
1353 page = list_first_entry(&tags->page_list, struct page, lru);
1354 list_del_init(&page->lru);
1355 __free_pages(page, page->private);
1360 blk_mq_free_tags(tags);
1363 static size_t order_to_size(unsigned int order)
1365 return (size_t)PAGE_SIZE << order;
1368 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1369 unsigned int hctx_idx)
1371 struct blk_mq_tags *tags;
1372 unsigned int i, j, entries_per_page, max_order = 4;
1373 size_t rq_size, left;
1375 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1380 INIT_LIST_HEAD(&tags->page_list);
1382 tags->rqs = kmalloc_node(set->queue_depth * sizeof(struct request *),
1383 GFP_KERNEL, set->numa_node);
1385 blk_mq_free_tags(tags);
1390 * rq_size is the size of the request plus driver payload, rounded
1391 * to the cacheline size
1393 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1395 left = rq_size * set->queue_depth;
1397 for (i = 0; i < set->queue_depth; ) {
1398 int this_order = max_order;
1403 while (left < order_to_size(this_order - 1) && this_order)
1407 page = alloc_pages_node(set->numa_node, GFP_KERNEL,
1413 if (order_to_size(this_order) < rq_size)
1420 page->private = this_order;
1421 list_add_tail(&page->lru, &tags->page_list);
1423 p = page_address(page);
1424 entries_per_page = order_to_size(this_order) / rq_size;
1425 to_do = min(entries_per_page, set->queue_depth - i);
1426 left -= to_do * rq_size;
1427 for (j = 0; j < to_do; j++) {
1429 if (set->ops->init_request) {
1430 if (set->ops->init_request(set->driver_data,
1431 tags->rqs[i], hctx_idx, i,
1444 pr_warn("%s: failed to allocate requests\n", __func__);
1445 blk_mq_free_rq_map(set, tags, hctx_idx);
1449 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1454 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1456 unsigned int bpw = 8, total, num_maps, i;
1458 bitmap->bits_per_word = bpw;
1460 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1461 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1466 bitmap->map_size = num_maps;
1469 for (i = 0; i < num_maps; i++) {
1470 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1471 total -= bitmap->map[i].depth;
1477 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1479 struct request_queue *q = hctx->queue;
1480 struct blk_mq_ctx *ctx;
1484 * Move ctx entries to new CPU, if this one is going away.
1486 ctx = __blk_mq_get_ctx(q, cpu);
1488 spin_lock(&ctx->lock);
1489 if (!list_empty(&ctx->rq_list)) {
1490 list_splice_init(&ctx->rq_list, &tmp);
1491 blk_mq_hctx_clear_pending(hctx, ctx);
1493 spin_unlock(&ctx->lock);
1495 if (list_empty(&tmp))
1498 ctx = blk_mq_get_ctx(q);
1499 spin_lock(&ctx->lock);
1501 while (!list_empty(&tmp)) {
1504 rq = list_first_entry(&tmp, struct request, queuelist);
1506 list_move_tail(&rq->queuelist, &ctx->rq_list);
1509 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1510 blk_mq_hctx_mark_pending(hctx, ctx);
1512 spin_unlock(&ctx->lock);
1514 blk_mq_run_hw_queue(hctx, true);
1515 blk_mq_put_ctx(ctx);
1519 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1521 struct request_queue *q = hctx->queue;
1522 struct blk_mq_tag_set *set = q->tag_set;
1524 if (set->tags[hctx->queue_num])
1527 set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1528 if (!set->tags[hctx->queue_num])
1531 hctx->tags = set->tags[hctx->queue_num];
1535 static int blk_mq_hctx_notify(void *data, unsigned long action,
1538 struct blk_mq_hw_ctx *hctx = data;
1540 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1541 return blk_mq_hctx_cpu_offline(hctx, cpu);
1542 else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1543 return blk_mq_hctx_cpu_online(hctx, cpu);
1548 static void blk_mq_exit_hw_queues(struct request_queue *q,
1549 struct blk_mq_tag_set *set, int nr_queue)
1551 struct blk_mq_hw_ctx *hctx;
1554 queue_for_each_hw_ctx(q, hctx, i) {
1558 if (set->ops->exit_hctx)
1559 set->ops->exit_hctx(hctx, i);
1561 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1563 blk_mq_free_bitmap(&hctx->ctx_map);
1568 static void blk_mq_free_hw_queues(struct request_queue *q,
1569 struct blk_mq_tag_set *set)
1571 struct blk_mq_hw_ctx *hctx;
1574 queue_for_each_hw_ctx(q, hctx, i) {
1575 free_cpumask_var(hctx->cpumask);
1580 static int blk_mq_init_hw_queues(struct request_queue *q,
1581 struct blk_mq_tag_set *set)
1583 struct blk_mq_hw_ctx *hctx;
1587 * Initialize hardware queues
1589 queue_for_each_hw_ctx(q, hctx, i) {
1592 node = hctx->numa_node;
1593 if (node == NUMA_NO_NODE)
1594 node = hctx->numa_node = set->numa_node;
1596 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1597 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1598 spin_lock_init(&hctx->lock);
1599 INIT_LIST_HEAD(&hctx->dispatch);
1601 hctx->queue_num = i;
1602 hctx->flags = set->flags;
1603 hctx->cmd_size = set->cmd_size;
1605 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1606 blk_mq_hctx_notify, hctx);
1607 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1609 hctx->tags = set->tags[i];
1612 * Allocate space for all possible cpus to avoid allocation in
1615 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1620 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1625 if (set->ops->init_hctx &&
1626 set->ops->init_hctx(hctx, set->driver_data, i))
1630 if (i == q->nr_hw_queues)
1636 blk_mq_exit_hw_queues(q, set, i);
1641 static void blk_mq_init_cpu_queues(struct request_queue *q,
1642 unsigned int nr_hw_queues)
1646 for_each_possible_cpu(i) {
1647 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1648 struct blk_mq_hw_ctx *hctx;
1650 memset(__ctx, 0, sizeof(*__ctx));
1652 spin_lock_init(&__ctx->lock);
1653 INIT_LIST_HEAD(&__ctx->rq_list);
1656 /* If the cpu isn't online, the cpu is mapped to first hctx */
1660 hctx = q->mq_ops->map_queue(q, i);
1661 cpumask_set_cpu(i, hctx->cpumask);
1665 * Set local node, IFF we have more than one hw queue. If
1666 * not, we remain on the home node of the device
1668 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1669 hctx->numa_node = cpu_to_node(i);
1673 static void blk_mq_map_swqueue(struct request_queue *q)
1676 struct blk_mq_hw_ctx *hctx;
1677 struct blk_mq_ctx *ctx;
1679 queue_for_each_hw_ctx(q, hctx, i) {
1680 cpumask_clear(hctx->cpumask);
1685 * Map software to hardware queues
1687 queue_for_each_ctx(q, ctx, i) {
1688 /* If the cpu isn't online, the cpu is mapped to first hctx */
1692 hctx = q->mq_ops->map_queue(q, i);
1693 cpumask_set_cpu(i, hctx->cpumask);
1694 ctx->index_hw = hctx->nr_ctx;
1695 hctx->ctxs[hctx->nr_ctx++] = ctx;
1698 queue_for_each_hw_ctx(q, hctx, i) {
1700 * If not software queues are mapped to this hardware queue,
1701 * disable it and free the request entries
1703 if (!hctx->nr_ctx) {
1704 struct blk_mq_tag_set *set = q->tag_set;
1707 blk_mq_free_rq_map(set, set->tags[i], i);
1708 set->tags[i] = NULL;
1715 * Initialize batch roundrobin counts
1717 hctx->next_cpu = cpumask_first(hctx->cpumask);
1718 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1722 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1724 struct blk_mq_hw_ctx *hctx;
1725 struct request_queue *q;
1729 if (set->tag_list.next == set->tag_list.prev)
1734 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1735 blk_mq_freeze_queue(q);
1737 queue_for_each_hw_ctx(q, hctx, i) {
1739 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1741 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1743 blk_mq_unfreeze_queue(q);
1747 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1749 struct blk_mq_tag_set *set = q->tag_set;
1751 blk_mq_freeze_queue(q);
1753 mutex_lock(&set->tag_list_lock);
1754 list_del_init(&q->tag_set_list);
1755 blk_mq_update_tag_set_depth(set);
1756 mutex_unlock(&set->tag_list_lock);
1758 blk_mq_unfreeze_queue(q);
1761 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1762 struct request_queue *q)
1766 mutex_lock(&set->tag_list_lock);
1767 list_add_tail(&q->tag_set_list, &set->tag_list);
1768 blk_mq_update_tag_set_depth(set);
1769 mutex_unlock(&set->tag_list_lock);
1772 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1774 struct blk_mq_hw_ctx **hctxs;
1775 struct blk_mq_ctx *ctx;
1776 struct request_queue *q;
1780 ctx = alloc_percpu(struct blk_mq_ctx);
1782 return ERR_PTR(-ENOMEM);
1784 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1790 map = blk_mq_make_queue_map(set);
1794 for (i = 0; i < set->nr_hw_queues; i++) {
1795 int node = blk_mq_hw_queue_to_node(map, i);
1797 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1802 if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL))
1805 atomic_set(&hctxs[i]->nr_active, 0);
1806 hctxs[i]->numa_node = node;
1807 hctxs[i]->queue_num = i;
1810 q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1814 if (percpu_counter_init(&q->mq_usage_counter, 0))
1817 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1818 blk_queue_rq_timeout(q, 30000);
1820 q->nr_queues = nr_cpu_ids;
1821 q->nr_hw_queues = set->nr_hw_queues;
1825 q->queue_hw_ctx = hctxs;
1827 q->mq_ops = set->ops;
1828 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1830 if (!(set->flags & BLK_MQ_F_SG_MERGE))
1831 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1833 q->sg_reserved_size = INT_MAX;
1835 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1836 INIT_LIST_HEAD(&q->requeue_list);
1837 spin_lock_init(&q->requeue_lock);
1839 if (q->nr_hw_queues > 1)
1840 blk_queue_make_request(q, blk_mq_make_request);
1842 blk_queue_make_request(q, blk_sq_make_request);
1844 blk_queue_rq_timed_out(q, blk_mq_rq_timed_out);
1846 blk_queue_rq_timeout(q, set->timeout);
1849 * Do this after blk_queue_make_request() overrides it...
1851 q->nr_requests = set->queue_depth;
1853 if (set->ops->complete)
1854 blk_queue_softirq_done(q, set->ops->complete);
1856 blk_mq_init_flush(q);
1857 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1859 q->flush_rq = kzalloc(round_up(sizeof(struct request) +
1860 set->cmd_size, cache_line_size()),
1865 if (blk_mq_init_hw_queues(q, set))
1868 mutex_lock(&all_q_mutex);
1869 list_add_tail(&q->all_q_node, &all_q_list);
1870 mutex_unlock(&all_q_mutex);
1872 blk_mq_add_queue_tag_set(set, q);
1874 blk_mq_map_swqueue(q);
1881 blk_cleanup_queue(q);
1884 for (i = 0; i < set->nr_hw_queues; i++) {
1887 free_cpumask_var(hctxs[i]->cpumask);
1894 return ERR_PTR(-ENOMEM);
1896 EXPORT_SYMBOL(blk_mq_init_queue);
1898 void blk_mq_free_queue(struct request_queue *q)
1900 struct blk_mq_tag_set *set = q->tag_set;
1902 blk_mq_del_queue_tag_set(q);
1904 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
1905 blk_mq_free_hw_queues(q, set);
1907 percpu_counter_destroy(&q->mq_usage_counter);
1909 free_percpu(q->queue_ctx);
1910 kfree(q->queue_hw_ctx);
1913 q->queue_ctx = NULL;
1914 q->queue_hw_ctx = NULL;
1917 mutex_lock(&all_q_mutex);
1918 list_del_init(&q->all_q_node);
1919 mutex_unlock(&all_q_mutex);
1922 /* Basically redo blk_mq_init_queue with queue frozen */
1923 static void blk_mq_queue_reinit(struct request_queue *q)
1925 blk_mq_freeze_queue(q);
1927 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1930 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1931 * we should change hctx numa_node according to new topology (this
1932 * involves free and re-allocate memory, worthy doing?)
1935 blk_mq_map_swqueue(q);
1937 blk_mq_unfreeze_queue(q);
1940 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1941 unsigned long action, void *hcpu)
1943 struct request_queue *q;
1946 * Before new mappings are established, hotadded cpu might already
1947 * start handling requests. This doesn't break anything as we map
1948 * offline CPUs to first hardware queue. We will re-init the queue
1949 * below to get optimal settings.
1951 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1952 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1955 mutex_lock(&all_q_mutex);
1956 list_for_each_entry(q, &all_q_list, all_q_node)
1957 blk_mq_queue_reinit(q);
1958 mutex_unlock(&all_q_mutex);
1962 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
1966 if (!set->nr_hw_queues)
1968 if (!set->queue_depth || set->queue_depth > BLK_MQ_MAX_DEPTH)
1970 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
1973 if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue)
1977 set->tags = kmalloc_node(set->nr_hw_queues *
1978 sizeof(struct blk_mq_tags *),
1979 GFP_KERNEL, set->numa_node);
1983 for (i = 0; i < set->nr_hw_queues; i++) {
1984 set->tags[i] = blk_mq_init_rq_map(set, i);
1989 mutex_init(&set->tag_list_lock);
1990 INIT_LIST_HEAD(&set->tag_list);
1996 blk_mq_free_rq_map(set, set->tags[i], i);
2000 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2002 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2006 for (i = 0; i < set->nr_hw_queues; i++) {
2008 blk_mq_free_rq_map(set, set->tags[i], i);
2013 EXPORT_SYMBOL(blk_mq_free_tag_set);
2015 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2017 struct blk_mq_tag_set *set = q->tag_set;
2018 struct blk_mq_hw_ctx *hctx;
2021 if (!set || nr > set->queue_depth)
2025 queue_for_each_hw_ctx(q, hctx, i) {
2026 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2032 q->nr_requests = nr;
2037 void blk_mq_disable_hotplug(void)
2039 mutex_lock(&all_q_mutex);
2042 void blk_mq_enable_hotplug(void)
2044 mutex_unlock(&all_q_mutex);
2047 static int __init blk_mq_init(void)
2051 /* Must be called after percpu_counter_hotcpu_callback() */
2052 hotcpu_notifier(blk_mq_queue_reinit_notify, -10);
2056 subsys_initcall(blk_mq_init);