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)
83 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
85 /* we have problems to freeze the queue if it's initializing */
86 if (!blk_queue_bypass(q) || !blk_queue_init_done(q))
89 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
91 spin_lock_irq(q->queue_lock);
92 ret = wait_event_interruptible_lock_irq(q->mq_freeze_wq,
93 !blk_queue_bypass(q) || blk_queue_dying(q),
95 /* inc usage with lock hold to avoid freeze_queue runs here */
96 if (!ret && !blk_queue_dying(q))
97 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
98 else if (blk_queue_dying(q))
100 spin_unlock_irq(q->queue_lock);
105 static void blk_mq_queue_exit(struct request_queue *q)
107 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
110 static void __blk_mq_drain_queue(struct request_queue *q)
115 spin_lock_irq(q->queue_lock);
116 count = percpu_counter_sum(&q->mq_usage_counter);
117 spin_unlock_irq(q->queue_lock);
121 blk_mq_run_queues(q, false);
127 * Guarantee no request is in use, so we can change any data structure of
128 * the queue afterward.
130 static void blk_mq_freeze_queue(struct request_queue *q)
134 spin_lock_irq(q->queue_lock);
135 drain = !q->bypass_depth++;
136 queue_flag_set(QUEUE_FLAG_BYPASS, q);
137 spin_unlock_irq(q->queue_lock);
140 __blk_mq_drain_queue(q);
143 void blk_mq_drain_queue(struct request_queue *q)
145 __blk_mq_drain_queue(q);
148 static void blk_mq_unfreeze_queue(struct request_queue *q)
152 spin_lock_irq(q->queue_lock);
153 if (!--q->bypass_depth) {
154 queue_flag_clear(QUEUE_FLAG_BYPASS, q);
157 WARN_ON_ONCE(q->bypass_depth < 0);
158 spin_unlock_irq(q->queue_lock);
160 wake_up_all(&q->mq_freeze_wq);
163 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
165 return blk_mq_has_free_tags(hctx->tags);
167 EXPORT_SYMBOL(blk_mq_can_queue);
169 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
170 struct request *rq, unsigned int rw_flags)
172 if (blk_queue_io_stat(q))
173 rw_flags |= REQ_IO_STAT;
175 INIT_LIST_HEAD(&rq->queuelist);
176 /* csd/requeue_work/fifo_time is initialized before use */
179 rq->cmd_flags |= rw_flags;
180 /* do not touch atomic flags, it needs atomic ops against the timer */
182 INIT_HLIST_NODE(&rq->hash);
183 RB_CLEAR_NODE(&rq->rb_node);
186 #ifdef CONFIG_BLK_CGROUP
188 set_start_time_ns(rq);
189 rq->io_start_time_ns = 0;
191 rq->nr_phys_segments = 0;
192 #if defined(CONFIG_BLK_DEV_INTEGRITY)
193 rq->nr_integrity_segments = 0;
196 /* tag was already set */
204 INIT_LIST_HEAD(&rq->timeout_list);
206 rq->end_io_data = NULL;
209 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
212 static struct request *
213 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
218 tag = blk_mq_get_tag(data);
219 if (tag != BLK_MQ_TAG_FAIL) {
220 rq = data->hctx->tags->rqs[tag];
223 if (blk_mq_tag_busy(data->hctx)) {
224 rq->cmd_flags = REQ_MQ_INFLIGHT;
225 atomic_inc(&data->hctx->nr_active);
229 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
236 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
239 struct blk_mq_ctx *ctx;
240 struct blk_mq_hw_ctx *hctx;
242 struct blk_mq_alloc_data alloc_data;
244 if (blk_mq_queue_enter(q))
247 ctx = blk_mq_get_ctx(q);
248 hctx = q->mq_ops->map_queue(q, ctx->cpu);
249 blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
250 reserved, ctx, hctx);
252 rq = __blk_mq_alloc_request(&alloc_data, rw);
253 if (!rq && (gfp & __GFP_WAIT)) {
254 __blk_mq_run_hw_queue(hctx);
257 ctx = blk_mq_get_ctx(q);
258 hctx = q->mq_ops->map_queue(q, ctx->cpu);
259 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
261 rq = __blk_mq_alloc_request(&alloc_data, rw);
262 ctx = alloc_data.ctx;
267 EXPORT_SYMBOL(blk_mq_alloc_request);
269 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
270 struct blk_mq_ctx *ctx, struct request *rq)
272 const int tag = rq->tag;
273 struct request_queue *q = rq->q;
275 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
276 atomic_dec(&hctx->nr_active);
278 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
279 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
280 blk_mq_queue_exit(q);
283 void blk_mq_free_request(struct request *rq)
285 struct blk_mq_ctx *ctx = rq->mq_ctx;
286 struct blk_mq_hw_ctx *hctx;
287 struct request_queue *q = rq->q;
289 ctx->rq_completed[rq_is_sync(rq)]++;
291 hctx = q->mq_ops->map_queue(q, ctx->cpu);
292 __blk_mq_free_request(hctx, ctx, rq);
296 * Clone all relevant state from a request that has been put on hold in
297 * the flush state machine into the preallocated flush request that hangs
298 * off the request queue.
300 * For a driver the flush request should be invisible, that's why we are
301 * impersonating the original request here.
303 void blk_mq_clone_flush_request(struct request *flush_rq,
304 struct request *orig_rq)
306 struct blk_mq_hw_ctx *hctx =
307 orig_rq->q->mq_ops->map_queue(orig_rq->q, orig_rq->mq_ctx->cpu);
309 flush_rq->mq_ctx = orig_rq->mq_ctx;
310 flush_rq->tag = orig_rq->tag;
311 memcpy(blk_mq_rq_to_pdu(flush_rq), blk_mq_rq_to_pdu(orig_rq),
315 inline void __blk_mq_end_io(struct request *rq, int error)
317 blk_account_io_done(rq);
320 rq->end_io(rq, error);
322 if (unlikely(blk_bidi_rq(rq)))
323 blk_mq_free_request(rq->next_rq);
324 blk_mq_free_request(rq);
327 EXPORT_SYMBOL(__blk_mq_end_io);
329 void blk_mq_end_io(struct request *rq, int error)
331 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
333 __blk_mq_end_io(rq, error);
335 EXPORT_SYMBOL(blk_mq_end_io);
337 static void __blk_mq_complete_request_remote(void *data)
339 struct request *rq = data;
341 rq->q->softirq_done_fn(rq);
344 static void blk_mq_ipi_complete_request(struct request *rq)
346 struct blk_mq_ctx *ctx = rq->mq_ctx;
350 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
351 rq->q->softirq_done_fn(rq);
356 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
357 shared = cpus_share_cache(cpu, ctx->cpu);
359 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
360 rq->csd.func = __blk_mq_complete_request_remote;
363 smp_call_function_single_async(ctx->cpu, &rq->csd);
365 rq->q->softirq_done_fn(rq);
370 void __blk_mq_complete_request(struct request *rq)
372 struct request_queue *q = rq->q;
374 if (!q->softirq_done_fn)
375 blk_mq_end_io(rq, rq->errors);
377 blk_mq_ipi_complete_request(rq);
381 * blk_mq_complete_request - end I/O on a request
382 * @rq: the request being processed
385 * Ends all I/O on a request. It does not handle partial completions.
386 * The actual completion happens out-of-order, through a IPI handler.
388 void blk_mq_complete_request(struct request *rq)
390 struct request_queue *q = rq->q;
392 if (unlikely(blk_should_fake_timeout(q)))
394 if (!blk_mark_rq_complete(rq))
395 __blk_mq_complete_request(rq);
397 EXPORT_SYMBOL(blk_mq_complete_request);
399 static void blk_mq_start_request(struct request *rq, bool last)
401 struct request_queue *q = rq->q;
403 trace_block_rq_issue(q, rq);
405 rq->resid_len = blk_rq_bytes(rq);
406 if (unlikely(blk_bidi_rq(rq)))
407 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
410 * Just mark start time and set the started bit. Due to memory
411 * ordering, we know we'll see the correct deadline as long as
412 * REQ_ATOMIC_STARTED is seen. Use the default queue timeout,
413 * unless one has been set in the request.
416 rq->deadline = jiffies + q->rq_timeout;
418 rq->deadline = jiffies + rq->timeout;
421 * Mark us as started and clear complete. Complete might have been
422 * set if requeue raced with timeout, which then marked it as
423 * complete. So be sure to clear complete again when we start
424 * the request, otherwise we'll ignore the completion event.
426 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
427 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
428 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
429 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
431 if (q->dma_drain_size && blk_rq_bytes(rq)) {
433 * Make sure space for the drain appears. We know we can do
434 * this because max_hw_segments has been adjusted to be one
435 * fewer than the device can handle.
437 rq->nr_phys_segments++;
441 * Flag the last request in the series so that drivers know when IO
442 * should be kicked off, if they don't do it on a per-request basis.
444 * Note: the flag isn't the only condition drivers should do kick off.
445 * If drive is busy, the last request might not have the bit set.
448 rq->cmd_flags |= REQ_END;
451 static void __blk_mq_requeue_request(struct request *rq)
453 struct request_queue *q = rq->q;
455 trace_block_rq_requeue(q, rq);
456 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
458 rq->cmd_flags &= ~REQ_END;
460 if (q->dma_drain_size && blk_rq_bytes(rq))
461 rq->nr_phys_segments--;
464 void blk_mq_requeue_request(struct request *rq)
466 __blk_mq_requeue_request(rq);
467 blk_clear_rq_complete(rq);
469 BUG_ON(blk_queued_rq(rq));
470 blk_mq_add_to_requeue_list(rq, true);
472 EXPORT_SYMBOL(blk_mq_requeue_request);
474 static void blk_mq_requeue_work(struct work_struct *work)
476 struct request_queue *q =
477 container_of(work, struct request_queue, requeue_work);
479 struct request *rq, *next;
482 spin_lock_irqsave(&q->requeue_lock, flags);
483 list_splice_init(&q->requeue_list, &rq_list);
484 spin_unlock_irqrestore(&q->requeue_lock, flags);
486 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
487 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
490 rq->cmd_flags &= ~REQ_SOFTBARRIER;
491 list_del_init(&rq->queuelist);
492 blk_mq_insert_request(rq, true, false, false);
495 while (!list_empty(&rq_list)) {
496 rq = list_entry(rq_list.next, struct request, queuelist);
497 list_del_init(&rq->queuelist);
498 blk_mq_insert_request(rq, false, false, false);
501 blk_mq_run_queues(q, false);
504 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
506 struct request_queue *q = rq->q;
510 * We abuse this flag that is otherwise used by the I/O scheduler to
511 * request head insertation from the workqueue.
513 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
515 spin_lock_irqsave(&q->requeue_lock, flags);
517 rq->cmd_flags |= REQ_SOFTBARRIER;
518 list_add(&rq->queuelist, &q->requeue_list);
520 list_add_tail(&rq->queuelist, &q->requeue_list);
522 spin_unlock_irqrestore(&q->requeue_lock, flags);
524 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
526 void blk_mq_kick_requeue_list(struct request_queue *q)
528 kblockd_schedule_work(&q->requeue_work);
530 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
532 struct request *blk_mq_tag_to_rq(struct blk_mq_hw_ctx *hctx, unsigned int tag)
534 struct request_queue *q = hctx->queue;
536 if ((q->flush_rq->cmd_flags & REQ_FLUSH_SEQ) &&
537 q->flush_rq->tag == tag)
540 return hctx->tags->rqs[tag];
542 EXPORT_SYMBOL(blk_mq_tag_to_rq);
544 struct blk_mq_timeout_data {
545 struct blk_mq_hw_ctx *hctx;
547 unsigned int *next_set;
550 static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
552 struct blk_mq_timeout_data *data = __data;
553 struct blk_mq_hw_ctx *hctx = data->hctx;
556 /* It may not be in flight yet (this is where
557 * the REQ_ATOMIC_STARTED flag comes in). The requests are
558 * statically allocated, so we know it's always safe to access the
559 * memory associated with a bit offset into ->rqs[].
565 tag = find_next_zero_bit(free_tags, hctx->tags->nr_tags, tag);
566 if (tag >= hctx->tags->nr_tags)
569 rq = blk_mq_tag_to_rq(hctx, tag++);
570 if (rq->q != hctx->queue)
572 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
575 blk_rq_check_expired(rq, data->next, data->next_set);
579 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
581 unsigned int *next_set)
583 struct blk_mq_timeout_data data = {
586 .next_set = next_set,
590 * Ask the tagging code to iterate busy requests, so we can
591 * check them for timeout.
593 blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
596 static enum blk_eh_timer_return blk_mq_rq_timed_out(struct request *rq)
598 struct request_queue *q = rq->q;
601 * We know that complete is set at this point. If STARTED isn't set
602 * anymore, then the request isn't active and the "timeout" should
603 * just be ignored. This can happen due to the bitflag ordering.
604 * Timeout first checks if STARTED is set, and if it is, assumes
605 * the request is active. But if we race with completion, then
606 * we both flags will get cleared. So check here again, and ignore
607 * a timeout event with a request that isn't active.
609 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
610 return BLK_EH_NOT_HANDLED;
612 if (!q->mq_ops->timeout)
613 return BLK_EH_RESET_TIMER;
615 return q->mq_ops->timeout(rq);
618 static void blk_mq_rq_timer(unsigned long data)
620 struct request_queue *q = (struct request_queue *) data;
621 struct blk_mq_hw_ctx *hctx;
622 unsigned long next = 0;
625 queue_for_each_hw_ctx(q, hctx, i) {
627 * If not software queues are currently mapped to this
628 * hardware queue, there's nothing to check
630 if (!hctx->nr_ctx || !hctx->tags)
633 blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
637 next = blk_rq_timeout(round_jiffies_up(next));
638 mod_timer(&q->timeout, next);
640 queue_for_each_hw_ctx(q, hctx, i)
641 blk_mq_tag_idle(hctx);
646 * Reverse check our software queue for entries that we could potentially
647 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
648 * too much time checking for merges.
650 static bool blk_mq_attempt_merge(struct request_queue *q,
651 struct blk_mq_ctx *ctx, struct bio *bio)
656 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
662 if (!blk_rq_merge_ok(rq, bio))
665 el_ret = blk_try_merge(rq, bio);
666 if (el_ret == ELEVATOR_BACK_MERGE) {
667 if (bio_attempt_back_merge(q, rq, bio)) {
672 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
673 if (bio_attempt_front_merge(q, rq, bio)) {
685 * Process software queues that have been marked busy, splicing them
686 * to the for-dispatch
688 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
690 struct blk_mq_ctx *ctx;
693 for (i = 0; i < hctx->ctx_map.map_size; i++) {
694 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
695 unsigned int off, bit;
701 off = i * hctx->ctx_map.bits_per_word;
703 bit = find_next_bit(&bm->word, bm->depth, bit);
704 if (bit >= bm->depth)
707 ctx = hctx->ctxs[bit + off];
708 clear_bit(bit, &bm->word);
709 spin_lock(&ctx->lock);
710 list_splice_tail_init(&ctx->rq_list, list);
711 spin_unlock(&ctx->lock);
719 * Run this hardware queue, pulling any software queues mapped to it in.
720 * Note that this function currently has various problems around ordering
721 * of IO. In particular, we'd like FIFO behaviour on handling existing
722 * items on the hctx->dispatch list. Ignore that for now.
724 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
726 struct request_queue *q = hctx->queue;
731 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
733 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
739 * Touch any software queue that has pending entries.
741 flush_busy_ctxs(hctx, &rq_list);
744 * If we have previous entries on our dispatch list, grab them
745 * and stuff them at the front for more fair dispatch.
747 if (!list_empty_careful(&hctx->dispatch)) {
748 spin_lock(&hctx->lock);
749 if (!list_empty(&hctx->dispatch))
750 list_splice_init(&hctx->dispatch, &rq_list);
751 spin_unlock(&hctx->lock);
755 * Now process all the entries, sending them to the driver.
758 while (!list_empty(&rq_list)) {
761 rq = list_first_entry(&rq_list, struct request, queuelist);
762 list_del_init(&rq->queuelist);
764 blk_mq_start_request(rq, list_empty(&rq_list));
766 ret = q->mq_ops->queue_rq(hctx, rq);
768 case BLK_MQ_RQ_QUEUE_OK:
771 case BLK_MQ_RQ_QUEUE_BUSY:
772 list_add(&rq->queuelist, &rq_list);
773 __blk_mq_requeue_request(rq);
776 pr_err("blk-mq: bad return on queue: %d\n", ret);
777 case BLK_MQ_RQ_QUEUE_ERROR:
779 blk_mq_end_io(rq, rq->errors);
783 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
788 hctx->dispatched[0]++;
789 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
790 hctx->dispatched[ilog2(queued) + 1]++;
793 * Any items that need requeuing? Stuff them into hctx->dispatch,
794 * that is where we will continue on next queue run.
796 if (!list_empty(&rq_list)) {
797 spin_lock(&hctx->lock);
798 list_splice(&rq_list, &hctx->dispatch);
799 spin_unlock(&hctx->lock);
804 * It'd be great if the workqueue API had a way to pass
805 * in a mask and had some smarts for more clever placement.
806 * For now we just round-robin here, switching for every
807 * BLK_MQ_CPU_WORK_BATCH queued items.
809 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
811 int cpu = hctx->next_cpu;
813 if (--hctx->next_cpu_batch <= 0) {
816 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
817 if (next_cpu >= nr_cpu_ids)
818 next_cpu = cpumask_first(hctx->cpumask);
820 hctx->next_cpu = next_cpu;
821 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
827 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
829 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
832 if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask))
833 __blk_mq_run_hw_queue(hctx);
834 else if (hctx->queue->nr_hw_queues == 1)
835 kblockd_schedule_delayed_work(&hctx->run_work, 0);
839 cpu = blk_mq_hctx_next_cpu(hctx);
840 kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0);
844 void blk_mq_run_queues(struct request_queue *q, bool async)
846 struct blk_mq_hw_ctx *hctx;
849 queue_for_each_hw_ctx(q, hctx, i) {
850 if ((!blk_mq_hctx_has_pending(hctx) &&
851 list_empty_careful(&hctx->dispatch)) ||
852 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
856 blk_mq_run_hw_queue(hctx, async);
860 EXPORT_SYMBOL(blk_mq_run_queues);
862 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
864 cancel_delayed_work(&hctx->run_work);
865 cancel_delayed_work(&hctx->delay_work);
866 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
868 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
870 void blk_mq_stop_hw_queues(struct request_queue *q)
872 struct blk_mq_hw_ctx *hctx;
875 queue_for_each_hw_ctx(q, hctx, i)
876 blk_mq_stop_hw_queue(hctx);
878 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
880 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
882 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
885 __blk_mq_run_hw_queue(hctx);
888 EXPORT_SYMBOL(blk_mq_start_hw_queue);
890 void blk_mq_start_hw_queues(struct request_queue *q)
892 struct blk_mq_hw_ctx *hctx;
895 queue_for_each_hw_ctx(q, hctx, i)
896 blk_mq_start_hw_queue(hctx);
898 EXPORT_SYMBOL(blk_mq_start_hw_queues);
901 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
903 struct blk_mq_hw_ctx *hctx;
906 queue_for_each_hw_ctx(q, hctx, i) {
907 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
910 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
912 blk_mq_run_hw_queue(hctx, async);
916 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
918 static void blk_mq_run_work_fn(struct work_struct *work)
920 struct blk_mq_hw_ctx *hctx;
922 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
924 __blk_mq_run_hw_queue(hctx);
927 static void blk_mq_delay_work_fn(struct work_struct *work)
929 struct blk_mq_hw_ctx *hctx;
931 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
933 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
934 __blk_mq_run_hw_queue(hctx);
937 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
939 unsigned long tmo = msecs_to_jiffies(msecs);
941 if (hctx->queue->nr_hw_queues == 1)
942 kblockd_schedule_delayed_work(&hctx->delay_work, tmo);
946 cpu = blk_mq_hctx_next_cpu(hctx);
947 kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo);
950 EXPORT_SYMBOL(blk_mq_delay_queue);
952 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
953 struct request *rq, bool at_head)
955 struct blk_mq_ctx *ctx = rq->mq_ctx;
957 trace_block_rq_insert(hctx->queue, rq);
960 list_add(&rq->queuelist, &ctx->rq_list);
962 list_add_tail(&rq->queuelist, &ctx->rq_list);
964 blk_mq_hctx_mark_pending(hctx, ctx);
967 * We do this early, to ensure we are on the right CPU.
972 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
975 struct request_queue *q = rq->q;
976 struct blk_mq_hw_ctx *hctx;
977 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
979 current_ctx = blk_mq_get_ctx(q);
980 if (!cpu_online(ctx->cpu))
981 rq->mq_ctx = ctx = current_ctx;
983 hctx = q->mq_ops->map_queue(q, ctx->cpu);
985 if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA) &&
986 !(rq->cmd_flags & (REQ_FLUSH_SEQ))) {
987 blk_insert_flush(rq);
989 spin_lock(&ctx->lock);
990 __blk_mq_insert_request(hctx, rq, at_head);
991 spin_unlock(&ctx->lock);
995 blk_mq_run_hw_queue(hctx, async);
997 blk_mq_put_ctx(current_ctx);
1000 static void blk_mq_insert_requests(struct request_queue *q,
1001 struct blk_mq_ctx *ctx,
1002 struct list_head *list,
1007 struct blk_mq_hw_ctx *hctx;
1008 struct blk_mq_ctx *current_ctx;
1010 trace_block_unplug(q, depth, !from_schedule);
1012 current_ctx = blk_mq_get_ctx(q);
1014 if (!cpu_online(ctx->cpu))
1016 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1019 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1022 spin_lock(&ctx->lock);
1023 while (!list_empty(list)) {
1026 rq = list_first_entry(list, struct request, queuelist);
1027 list_del_init(&rq->queuelist);
1029 __blk_mq_insert_request(hctx, rq, false);
1031 spin_unlock(&ctx->lock);
1033 blk_mq_run_hw_queue(hctx, from_schedule);
1034 blk_mq_put_ctx(current_ctx);
1037 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1039 struct request *rqa = container_of(a, struct request, queuelist);
1040 struct request *rqb = container_of(b, struct request, queuelist);
1042 return !(rqa->mq_ctx < rqb->mq_ctx ||
1043 (rqa->mq_ctx == rqb->mq_ctx &&
1044 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1047 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1049 struct blk_mq_ctx *this_ctx;
1050 struct request_queue *this_q;
1053 LIST_HEAD(ctx_list);
1056 list_splice_init(&plug->mq_list, &list);
1058 list_sort(NULL, &list, plug_ctx_cmp);
1064 while (!list_empty(&list)) {
1065 rq = list_entry_rq(list.next);
1066 list_del_init(&rq->queuelist);
1068 if (rq->mq_ctx != this_ctx) {
1070 blk_mq_insert_requests(this_q, this_ctx,
1075 this_ctx = rq->mq_ctx;
1081 list_add_tail(&rq->queuelist, &ctx_list);
1085 * If 'this_ctx' is set, we know we have entries to complete
1086 * on 'ctx_list'. Do those.
1089 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1094 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1096 init_request_from_bio(rq, bio);
1098 if (blk_do_io_stat(rq)) {
1099 rq->start_time = jiffies;
1100 blk_account_io_start(rq, 1);
1104 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1105 struct blk_mq_ctx *ctx,
1106 struct request *rq, struct bio *bio)
1108 struct request_queue *q = hctx->queue;
1110 if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE)) {
1111 blk_mq_bio_to_request(rq, bio);
1112 spin_lock(&ctx->lock);
1114 __blk_mq_insert_request(hctx, rq, false);
1115 spin_unlock(&ctx->lock);
1118 spin_lock(&ctx->lock);
1119 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1120 blk_mq_bio_to_request(rq, bio);
1124 spin_unlock(&ctx->lock);
1125 __blk_mq_free_request(hctx, ctx, rq);
1130 struct blk_map_ctx {
1131 struct blk_mq_hw_ctx *hctx;
1132 struct blk_mq_ctx *ctx;
1135 static struct request *blk_mq_map_request(struct request_queue *q,
1137 struct blk_map_ctx *data)
1139 struct blk_mq_hw_ctx *hctx;
1140 struct blk_mq_ctx *ctx;
1142 int rw = bio_data_dir(bio);
1143 struct blk_mq_alloc_data alloc_data;
1145 if (unlikely(blk_mq_queue_enter(q))) {
1146 bio_endio(bio, -EIO);
1150 ctx = blk_mq_get_ctx(q);
1151 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1153 if (rw_is_sync(bio->bi_rw))
1156 trace_block_getrq(q, bio, rw);
1157 blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1159 rq = __blk_mq_alloc_request(&alloc_data, rw);
1160 if (unlikely(!rq)) {
1161 __blk_mq_run_hw_queue(hctx);
1162 blk_mq_put_ctx(ctx);
1163 trace_block_sleeprq(q, bio, rw);
1165 ctx = blk_mq_get_ctx(q);
1166 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1167 blk_mq_set_alloc_data(&alloc_data, q,
1168 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1169 rq = __blk_mq_alloc_request(&alloc_data, rw);
1170 ctx = alloc_data.ctx;
1171 hctx = alloc_data.hctx;
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);
1217 * For OK queue, we are done. For error, kill it. Any other
1218 * error (busy), just add it to our list as we previously
1221 ret = q->mq_ops->queue_rq(data.hctx, rq);
1222 if (ret == BLK_MQ_RQ_QUEUE_OK)
1225 __blk_mq_requeue_request(rq);
1227 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1229 blk_mq_end_io(rq, rq->errors);
1235 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1237 * For a SYNC request, send it to the hardware immediately. For
1238 * an ASYNC request, just ensure that we run it later on. The
1239 * latter allows for merging opportunities and more efficient
1243 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1246 blk_mq_put_ctx(data.ctx);
1250 * Single hardware queue variant. This will attempt to use any per-process
1251 * plug for merging and IO deferral.
1253 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1255 const int is_sync = rw_is_sync(bio->bi_rw);
1256 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1257 unsigned int use_plug, request_count = 0;
1258 struct blk_map_ctx data;
1262 * If we have multiple hardware queues, just go directly to
1263 * one of those for sync IO.
1265 use_plug = !is_flush_fua && !is_sync;
1267 blk_queue_bounce(q, &bio);
1269 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1270 bio_endio(bio, -EIO);
1274 if (use_plug && !blk_queue_nomerges(q) &&
1275 blk_attempt_plug_merge(q, bio, &request_count))
1278 rq = blk_mq_map_request(q, bio, &data);
1282 if (unlikely(is_flush_fua)) {
1283 blk_mq_bio_to_request(rq, bio);
1284 blk_insert_flush(rq);
1289 * A task plug currently exists. Since this is completely lockless,
1290 * utilize that to temporarily store requests until the task is
1291 * either done or scheduled away.
1294 struct blk_plug *plug = current->plug;
1297 blk_mq_bio_to_request(rq, bio);
1298 if (list_empty(&plug->mq_list))
1299 trace_block_plug(q);
1300 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1301 blk_flush_plug_list(plug, false);
1302 trace_block_plug(q);
1304 list_add_tail(&rq->queuelist, &plug->mq_list);
1305 blk_mq_put_ctx(data.ctx);
1310 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1312 * For a SYNC request, send it to the hardware immediately. For
1313 * an ASYNC request, just ensure that we run it later on. The
1314 * latter allows for merging opportunities and more efficient
1318 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1321 blk_mq_put_ctx(data.ctx);
1325 * Default mapping to a software queue, since we use one per CPU.
1327 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1329 return q->queue_hw_ctx[q->mq_map[cpu]];
1331 EXPORT_SYMBOL(blk_mq_map_queue);
1333 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1334 struct blk_mq_tags *tags, unsigned int hctx_idx)
1338 if (tags->rqs && set->ops->exit_request) {
1341 for (i = 0; i < tags->nr_tags; i++) {
1344 set->ops->exit_request(set->driver_data, tags->rqs[i],
1349 while (!list_empty(&tags->page_list)) {
1350 page = list_first_entry(&tags->page_list, struct page, lru);
1351 list_del_init(&page->lru);
1352 __free_pages(page, page->private);
1357 blk_mq_free_tags(tags);
1360 static size_t order_to_size(unsigned int order)
1362 return (size_t)PAGE_SIZE << order;
1365 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1366 unsigned int hctx_idx)
1368 struct blk_mq_tags *tags;
1369 unsigned int i, j, entries_per_page, max_order = 4;
1370 size_t rq_size, left;
1372 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1377 INIT_LIST_HEAD(&tags->page_list);
1379 tags->rqs = kmalloc_node(set->queue_depth * sizeof(struct request *),
1380 GFP_KERNEL, set->numa_node);
1382 blk_mq_free_tags(tags);
1387 * rq_size is the size of the request plus driver payload, rounded
1388 * to the cacheline size
1390 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1392 left = rq_size * set->queue_depth;
1394 for (i = 0; i < set->queue_depth; ) {
1395 int this_order = max_order;
1400 while (left < order_to_size(this_order - 1) && this_order)
1404 page = alloc_pages_node(set->numa_node, GFP_KERNEL,
1410 if (order_to_size(this_order) < rq_size)
1417 page->private = this_order;
1418 list_add_tail(&page->lru, &tags->page_list);
1420 p = page_address(page);
1421 entries_per_page = order_to_size(this_order) / rq_size;
1422 to_do = min(entries_per_page, set->queue_depth - i);
1423 left -= to_do * rq_size;
1424 for (j = 0; j < to_do; j++) {
1426 if (set->ops->init_request) {
1427 if (set->ops->init_request(set->driver_data,
1428 tags->rqs[i], hctx_idx, i,
1441 pr_warn("%s: failed to allocate requests\n", __func__);
1442 blk_mq_free_rq_map(set, tags, hctx_idx);
1446 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1451 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1453 unsigned int bpw = 8, total, num_maps, i;
1455 bitmap->bits_per_word = bpw;
1457 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1458 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1463 bitmap->map_size = num_maps;
1466 for (i = 0; i < num_maps; i++) {
1467 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1468 total -= bitmap->map[i].depth;
1474 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1476 struct request_queue *q = hctx->queue;
1477 struct blk_mq_ctx *ctx;
1481 * Move ctx entries to new CPU, if this one is going away.
1483 ctx = __blk_mq_get_ctx(q, cpu);
1485 spin_lock(&ctx->lock);
1486 if (!list_empty(&ctx->rq_list)) {
1487 list_splice_init(&ctx->rq_list, &tmp);
1488 blk_mq_hctx_clear_pending(hctx, ctx);
1490 spin_unlock(&ctx->lock);
1492 if (list_empty(&tmp))
1495 ctx = blk_mq_get_ctx(q);
1496 spin_lock(&ctx->lock);
1498 while (!list_empty(&tmp)) {
1501 rq = list_first_entry(&tmp, struct request, queuelist);
1503 list_move_tail(&rq->queuelist, &ctx->rq_list);
1506 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1507 blk_mq_hctx_mark_pending(hctx, ctx);
1509 spin_unlock(&ctx->lock);
1511 blk_mq_run_hw_queue(hctx, true);
1512 blk_mq_put_ctx(ctx);
1516 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1518 struct request_queue *q = hctx->queue;
1519 struct blk_mq_tag_set *set = q->tag_set;
1521 if (set->tags[hctx->queue_num])
1524 set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1525 if (!set->tags[hctx->queue_num])
1528 hctx->tags = set->tags[hctx->queue_num];
1532 static int blk_mq_hctx_notify(void *data, unsigned long action,
1535 struct blk_mq_hw_ctx *hctx = data;
1537 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1538 return blk_mq_hctx_cpu_offline(hctx, cpu);
1539 else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1540 return blk_mq_hctx_cpu_online(hctx, cpu);
1545 static void blk_mq_exit_hw_queues(struct request_queue *q,
1546 struct blk_mq_tag_set *set, int nr_queue)
1548 struct blk_mq_hw_ctx *hctx;
1551 queue_for_each_hw_ctx(q, hctx, i) {
1555 if (set->ops->exit_hctx)
1556 set->ops->exit_hctx(hctx, i);
1558 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1560 blk_mq_free_bitmap(&hctx->ctx_map);
1565 static void blk_mq_free_hw_queues(struct request_queue *q,
1566 struct blk_mq_tag_set *set)
1568 struct blk_mq_hw_ctx *hctx;
1571 queue_for_each_hw_ctx(q, hctx, i) {
1572 free_cpumask_var(hctx->cpumask);
1577 static int blk_mq_init_hw_queues(struct request_queue *q,
1578 struct blk_mq_tag_set *set)
1580 struct blk_mq_hw_ctx *hctx;
1584 * Initialize hardware queues
1586 queue_for_each_hw_ctx(q, hctx, i) {
1589 node = hctx->numa_node;
1590 if (node == NUMA_NO_NODE)
1591 node = hctx->numa_node = set->numa_node;
1593 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1594 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1595 spin_lock_init(&hctx->lock);
1596 INIT_LIST_HEAD(&hctx->dispatch);
1598 hctx->queue_num = i;
1599 hctx->flags = set->flags;
1600 hctx->cmd_size = set->cmd_size;
1602 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1603 blk_mq_hctx_notify, hctx);
1604 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1606 hctx->tags = set->tags[i];
1609 * Allocate space for all possible cpus to avoid allocation in
1612 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1617 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1622 if (set->ops->init_hctx &&
1623 set->ops->init_hctx(hctx, set->driver_data, i))
1627 if (i == q->nr_hw_queues)
1633 blk_mq_exit_hw_queues(q, set, i);
1638 static void blk_mq_init_cpu_queues(struct request_queue *q,
1639 unsigned int nr_hw_queues)
1643 for_each_possible_cpu(i) {
1644 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1645 struct blk_mq_hw_ctx *hctx;
1647 memset(__ctx, 0, sizeof(*__ctx));
1649 spin_lock_init(&__ctx->lock);
1650 INIT_LIST_HEAD(&__ctx->rq_list);
1653 /* If the cpu isn't online, the cpu is mapped to first hctx */
1657 hctx = q->mq_ops->map_queue(q, i);
1658 cpumask_set_cpu(i, hctx->cpumask);
1662 * Set local node, IFF we have more than one hw queue. If
1663 * not, we remain on the home node of the device
1665 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1666 hctx->numa_node = cpu_to_node(i);
1670 static void blk_mq_map_swqueue(struct request_queue *q)
1673 struct blk_mq_hw_ctx *hctx;
1674 struct blk_mq_ctx *ctx;
1676 queue_for_each_hw_ctx(q, hctx, i) {
1677 cpumask_clear(hctx->cpumask);
1682 * Map software to hardware queues
1684 queue_for_each_ctx(q, ctx, i) {
1685 /* If the cpu isn't online, the cpu is mapped to first hctx */
1689 hctx = q->mq_ops->map_queue(q, i);
1690 cpumask_set_cpu(i, hctx->cpumask);
1691 ctx->index_hw = hctx->nr_ctx;
1692 hctx->ctxs[hctx->nr_ctx++] = ctx;
1695 queue_for_each_hw_ctx(q, hctx, i) {
1697 * If not software queues are mapped to this hardware queue,
1698 * disable it and free the request entries
1700 if (!hctx->nr_ctx) {
1701 struct blk_mq_tag_set *set = q->tag_set;
1704 blk_mq_free_rq_map(set, set->tags[i], i);
1705 set->tags[i] = NULL;
1712 * Initialize batch roundrobin counts
1714 hctx->next_cpu = cpumask_first(hctx->cpumask);
1715 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1719 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1721 struct blk_mq_hw_ctx *hctx;
1722 struct request_queue *q;
1726 if (set->tag_list.next == set->tag_list.prev)
1731 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1732 blk_mq_freeze_queue(q);
1734 queue_for_each_hw_ctx(q, hctx, i) {
1736 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1738 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1740 blk_mq_unfreeze_queue(q);
1744 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1746 struct blk_mq_tag_set *set = q->tag_set;
1748 blk_mq_freeze_queue(q);
1750 mutex_lock(&set->tag_list_lock);
1751 list_del_init(&q->tag_set_list);
1752 blk_mq_update_tag_set_depth(set);
1753 mutex_unlock(&set->tag_list_lock);
1755 blk_mq_unfreeze_queue(q);
1758 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1759 struct request_queue *q)
1763 mutex_lock(&set->tag_list_lock);
1764 list_add_tail(&q->tag_set_list, &set->tag_list);
1765 blk_mq_update_tag_set_depth(set);
1766 mutex_unlock(&set->tag_list_lock);
1769 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1771 struct blk_mq_hw_ctx **hctxs;
1772 struct blk_mq_ctx __percpu *ctx;
1773 struct request_queue *q;
1777 ctx = alloc_percpu(struct blk_mq_ctx);
1779 return ERR_PTR(-ENOMEM);
1781 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1787 map = blk_mq_make_queue_map(set);
1791 for (i = 0; i < set->nr_hw_queues; i++) {
1792 int node = blk_mq_hw_queue_to_node(map, i);
1794 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1799 if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL))
1802 atomic_set(&hctxs[i]->nr_active, 0);
1803 hctxs[i]->numa_node = node;
1804 hctxs[i]->queue_num = i;
1807 q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1811 if (percpu_counter_init(&q->mq_usage_counter, 0))
1814 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1815 blk_queue_rq_timeout(q, 30000);
1817 q->nr_queues = nr_cpu_ids;
1818 q->nr_hw_queues = set->nr_hw_queues;
1822 q->queue_hw_ctx = hctxs;
1824 q->mq_ops = set->ops;
1825 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1827 if (!(set->flags & BLK_MQ_F_SG_MERGE))
1828 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1830 q->sg_reserved_size = INT_MAX;
1832 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1833 INIT_LIST_HEAD(&q->requeue_list);
1834 spin_lock_init(&q->requeue_lock);
1836 if (q->nr_hw_queues > 1)
1837 blk_queue_make_request(q, blk_mq_make_request);
1839 blk_queue_make_request(q, blk_sq_make_request);
1841 blk_queue_rq_timed_out(q, blk_mq_rq_timed_out);
1843 blk_queue_rq_timeout(q, set->timeout);
1846 * Do this after blk_queue_make_request() overrides it...
1848 q->nr_requests = set->queue_depth;
1850 if (set->ops->complete)
1851 blk_queue_softirq_done(q, set->ops->complete);
1853 blk_mq_init_flush(q);
1854 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1856 q->flush_rq = kzalloc(round_up(sizeof(struct request) +
1857 set->cmd_size, cache_line_size()),
1862 if (blk_mq_init_hw_queues(q, set))
1865 mutex_lock(&all_q_mutex);
1866 list_add_tail(&q->all_q_node, &all_q_list);
1867 mutex_unlock(&all_q_mutex);
1869 blk_mq_add_queue_tag_set(set, q);
1871 blk_mq_map_swqueue(q);
1878 blk_cleanup_queue(q);
1881 for (i = 0; i < set->nr_hw_queues; i++) {
1884 free_cpumask_var(hctxs[i]->cpumask);
1891 return ERR_PTR(-ENOMEM);
1893 EXPORT_SYMBOL(blk_mq_init_queue);
1895 void blk_mq_free_queue(struct request_queue *q)
1897 struct blk_mq_tag_set *set = q->tag_set;
1899 blk_mq_del_queue_tag_set(q);
1901 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
1902 blk_mq_free_hw_queues(q, set);
1904 percpu_counter_destroy(&q->mq_usage_counter);
1906 free_percpu(q->queue_ctx);
1907 kfree(q->queue_hw_ctx);
1910 q->queue_ctx = NULL;
1911 q->queue_hw_ctx = NULL;
1914 mutex_lock(&all_q_mutex);
1915 list_del_init(&q->all_q_node);
1916 mutex_unlock(&all_q_mutex);
1919 /* Basically redo blk_mq_init_queue with queue frozen */
1920 static void blk_mq_queue_reinit(struct request_queue *q)
1922 blk_mq_freeze_queue(q);
1924 blk_mq_sysfs_unregister(q);
1926 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1929 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1930 * we should change hctx numa_node according to new topology (this
1931 * involves free and re-allocate memory, worthy doing?)
1934 blk_mq_map_swqueue(q);
1936 blk_mq_sysfs_register(q);
1938 blk_mq_unfreeze_queue(q);
1941 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1942 unsigned long action, void *hcpu)
1944 struct request_queue *q;
1947 * Before new mappings are established, hotadded cpu might already
1948 * start handling requests. This doesn't break anything as we map
1949 * offline CPUs to first hardware queue. We will re-init the queue
1950 * below to get optimal settings.
1952 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1953 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1956 mutex_lock(&all_q_mutex);
1957 list_for_each_entry(q, &all_q_list, all_q_node)
1958 blk_mq_queue_reinit(q);
1959 mutex_unlock(&all_q_mutex);
1963 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
1967 if (!set->nr_hw_queues)
1969 if (!set->queue_depth || set->queue_depth > BLK_MQ_MAX_DEPTH)
1971 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
1974 if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue)
1978 set->tags = kmalloc_node(set->nr_hw_queues *
1979 sizeof(struct blk_mq_tags *),
1980 GFP_KERNEL, set->numa_node);
1984 for (i = 0; i < set->nr_hw_queues; i++) {
1985 set->tags[i] = blk_mq_init_rq_map(set, i);
1990 mutex_init(&set->tag_list_lock);
1991 INIT_LIST_HEAD(&set->tag_list);
1997 blk_mq_free_rq_map(set, set->tags[i], i);
2001 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2003 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2007 for (i = 0; i < set->nr_hw_queues; i++) {
2009 blk_mq_free_rq_map(set, set->tags[i], i);
2014 EXPORT_SYMBOL(blk_mq_free_tag_set);
2016 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2018 struct blk_mq_tag_set *set = q->tag_set;
2019 struct blk_mq_hw_ctx *hctx;
2022 if (!set || nr > set->queue_depth)
2026 queue_for_each_hw_ctx(q, hctx, i) {
2027 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2033 q->nr_requests = nr;
2038 void blk_mq_disable_hotplug(void)
2040 mutex_lock(&all_q_mutex);
2043 void blk_mq_enable_hotplug(void)
2045 mutex_unlock(&all_q_mutex);
2048 static int __init blk_mq_init(void)
2052 /* Must be called after percpu_counter_hotcpu_callback() */
2053 hotcpu_notifier(blk_mq_queue_reinit_notify, -10);
2057 subsys_initcall(blk_mq_init);