HID: hiddev: Fix slab-out-of-bounds write in hiddev_ioctl_usage()
[linux/fpc-iii.git] / block / blk-mq.c
blobe027b8ed603085f1ce87263d4967e41f3f979b99
1 /*
2 * Block multiqueue core code
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
6 */
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>
12 #include <linux/kmemleak.h>
13 #include <linux/mm.h>
14 #include <linux/init.h>
15 #include <linux/slab.h>
16 #include <linux/workqueue.h>
17 #include <linux/smp.h>
18 #include <linux/llist.h>
19 #include <linux/list_sort.h>
20 #include <linux/cpu.h>
21 #include <linux/cache.h>
22 #include <linux/sched/sysctl.h>
23 #include <linux/delay.h>
24 #include <linux/crash_dump.h>
26 #include <trace/events/block.h>
28 #include <linux/blk-mq.h>
29 #include "blk.h"
30 #include "blk-mq.h"
31 #include "blk-mq-tag.h"
33 static DEFINE_MUTEX(all_q_mutex);
34 static LIST_HEAD(all_q_list);
36 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
39 * Check if any of the ctx's have pending work in this hardware queue
41 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
43 unsigned int i;
45 for (i = 0; i < hctx->ctx_map.size; i++)
46 if (hctx->ctx_map.map[i].word)
47 return true;
49 return false;
52 static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx,
53 struct blk_mq_ctx *ctx)
55 return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word];
58 #define CTX_TO_BIT(hctx, ctx) \
59 ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
62 * Mark this ctx as having pending work in this hardware queue
64 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
65 struct blk_mq_ctx *ctx)
67 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
69 if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word))
70 set_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
73 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
74 struct blk_mq_ctx *ctx)
76 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
78 clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
81 void blk_mq_freeze_queue_start(struct request_queue *q)
83 int freeze_depth;
85 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
86 if (freeze_depth == 1) {
87 percpu_ref_kill(&q->q_usage_counter);
88 blk_mq_run_hw_queues(q, false);
91 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
93 static void blk_mq_freeze_queue_wait(struct request_queue *q)
95 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
99 * Guarantee no request is in use, so we can change any data structure of
100 * the queue afterward.
102 void blk_freeze_queue(struct request_queue *q)
105 * In the !blk_mq case we are only calling this to kill the
106 * q_usage_counter, otherwise this increases the freeze depth
107 * and waits for it to return to zero. For this reason there is
108 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
109 * exported to drivers as the only user for unfreeze is blk_mq.
111 blk_mq_freeze_queue_start(q);
112 blk_mq_freeze_queue_wait(q);
115 void blk_mq_freeze_queue(struct request_queue *q)
118 * ...just an alias to keep freeze and unfreeze actions balanced
119 * in the blk_mq_* namespace
121 blk_freeze_queue(q);
123 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
125 void blk_mq_unfreeze_queue(struct request_queue *q)
127 int freeze_depth;
129 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
130 WARN_ON_ONCE(freeze_depth < 0);
131 if (!freeze_depth) {
132 percpu_ref_reinit(&q->q_usage_counter);
133 wake_up_all(&q->mq_freeze_wq);
136 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
138 void blk_mq_wake_waiters(struct request_queue *q)
140 struct blk_mq_hw_ctx *hctx;
141 unsigned int i;
143 queue_for_each_hw_ctx(q, hctx, i)
144 if (blk_mq_hw_queue_mapped(hctx))
145 blk_mq_tag_wakeup_all(hctx->tags, true);
148 * If we are called because the queue has now been marked as
149 * dying, we need to ensure that processes currently waiting on
150 * the queue are notified as well.
152 wake_up_all(&q->mq_freeze_wq);
155 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
157 return blk_mq_has_free_tags(hctx->tags);
159 EXPORT_SYMBOL(blk_mq_can_queue);
161 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
162 struct request *rq, unsigned int rw_flags)
164 if (blk_queue_io_stat(q))
165 rw_flags |= REQ_IO_STAT;
167 INIT_LIST_HEAD(&rq->queuelist);
168 /* csd/requeue_work/fifo_time is initialized before use */
169 rq->q = q;
170 rq->mq_ctx = ctx;
171 rq->cmd_flags |= rw_flags;
172 /* do not touch atomic flags, it needs atomic ops against the timer */
173 rq->cpu = -1;
174 INIT_HLIST_NODE(&rq->hash);
175 RB_CLEAR_NODE(&rq->rb_node);
176 rq->rq_disk = NULL;
177 rq->part = NULL;
178 rq->start_time = jiffies;
179 #ifdef CONFIG_BLK_CGROUP
180 rq->rl = NULL;
181 set_start_time_ns(rq);
182 rq->io_start_time_ns = 0;
183 #endif
184 rq->nr_phys_segments = 0;
185 #if defined(CONFIG_BLK_DEV_INTEGRITY)
186 rq->nr_integrity_segments = 0;
187 #endif
188 rq->special = NULL;
189 /* tag was already set */
190 rq->errors = 0;
192 rq->cmd = rq->__cmd;
194 rq->extra_len = 0;
195 rq->sense_len = 0;
196 rq->resid_len = 0;
197 rq->sense = NULL;
199 INIT_LIST_HEAD(&rq->timeout_list);
200 rq->timeout = 0;
202 rq->end_io = NULL;
203 rq->end_io_data = NULL;
204 rq->next_rq = NULL;
206 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
209 static struct request *
210 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
212 struct request *rq;
213 unsigned int tag;
215 tag = blk_mq_get_tag(data);
216 if (tag != BLK_MQ_TAG_FAIL) {
217 rq = data->hctx->tags->rqs[tag];
219 if (blk_mq_tag_busy(data->hctx)) {
220 rq->cmd_flags = REQ_MQ_INFLIGHT;
221 atomic_inc(&data->hctx->nr_active);
224 rq->tag = tag;
225 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
226 return rq;
229 return NULL;
232 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
233 bool reserved)
235 struct blk_mq_ctx *ctx;
236 struct blk_mq_hw_ctx *hctx;
237 struct request *rq;
238 struct blk_mq_alloc_data alloc_data;
239 int ret;
241 ret = blk_queue_enter(q, gfp);
242 if (ret)
243 return ERR_PTR(ret);
245 ctx = blk_mq_get_ctx(q);
246 hctx = q->mq_ops->map_queue(q, ctx->cpu);
247 blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_DIRECT_RECLAIM,
248 reserved, ctx, hctx);
250 rq = __blk_mq_alloc_request(&alloc_data, rw);
251 if (!rq && (gfp & __GFP_DIRECT_RECLAIM)) {
252 __blk_mq_run_hw_queue(hctx);
253 blk_mq_put_ctx(ctx);
255 ctx = blk_mq_get_ctx(q);
256 hctx = q->mq_ops->map_queue(q, ctx->cpu);
257 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
258 hctx);
259 rq = __blk_mq_alloc_request(&alloc_data, rw);
260 ctx = alloc_data.ctx;
262 blk_mq_put_ctx(ctx);
263 if (!rq) {
264 blk_queue_exit(q);
265 return ERR_PTR(-EWOULDBLOCK);
267 return rq;
269 EXPORT_SYMBOL(blk_mq_alloc_request);
271 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
272 struct blk_mq_ctx *ctx, struct request *rq)
274 const int tag = rq->tag;
275 struct request_queue *q = rq->q;
277 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
278 atomic_dec(&hctx->nr_active);
279 rq->cmd_flags = 0;
281 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
282 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
283 blk_queue_exit(q);
286 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
288 struct blk_mq_ctx *ctx = rq->mq_ctx;
290 ctx->rq_completed[rq_is_sync(rq)]++;
291 __blk_mq_free_request(hctx, ctx, rq);
294 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
296 void blk_mq_free_request(struct request *rq)
298 struct blk_mq_hw_ctx *hctx;
299 struct request_queue *q = rq->q;
301 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
302 blk_mq_free_hctx_request(hctx, rq);
304 EXPORT_SYMBOL_GPL(blk_mq_free_request);
306 inline void __blk_mq_end_request(struct request *rq, int error)
308 blk_account_io_done(rq);
310 if (rq->end_io) {
311 rq->end_io(rq, error);
312 } else {
313 if (unlikely(blk_bidi_rq(rq)))
314 blk_mq_free_request(rq->next_rq);
315 blk_mq_free_request(rq);
318 EXPORT_SYMBOL(__blk_mq_end_request);
320 void blk_mq_end_request(struct request *rq, int error)
322 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
323 BUG();
324 __blk_mq_end_request(rq, error);
326 EXPORT_SYMBOL(blk_mq_end_request);
328 static void __blk_mq_complete_request_remote(void *data)
330 struct request *rq = data;
332 rq->q->softirq_done_fn(rq);
335 static void blk_mq_ipi_complete_request(struct request *rq)
337 struct blk_mq_ctx *ctx = rq->mq_ctx;
338 bool shared = false;
339 int cpu;
341 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
342 rq->q->softirq_done_fn(rq);
343 return;
346 cpu = get_cpu();
347 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
348 shared = cpus_share_cache(cpu, ctx->cpu);
350 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
351 rq->csd.func = __blk_mq_complete_request_remote;
352 rq->csd.info = rq;
353 rq->csd.flags = 0;
354 smp_call_function_single_async(ctx->cpu, &rq->csd);
355 } else {
356 rq->q->softirq_done_fn(rq);
358 put_cpu();
361 static void __blk_mq_complete_request(struct request *rq)
363 struct request_queue *q = rq->q;
365 if (!q->softirq_done_fn)
366 blk_mq_end_request(rq, rq->errors);
367 else
368 blk_mq_ipi_complete_request(rq);
372 * blk_mq_complete_request - end I/O on a request
373 * @rq: the request being processed
375 * Description:
376 * Ends all I/O on a request. It does not handle partial completions.
377 * The actual completion happens out-of-order, through a IPI handler.
379 void blk_mq_complete_request(struct request *rq, int error)
381 struct request_queue *q = rq->q;
383 if (unlikely(blk_should_fake_timeout(q)))
384 return;
385 if (!blk_mark_rq_complete(rq)) {
386 rq->errors = error;
387 __blk_mq_complete_request(rq);
390 EXPORT_SYMBOL(blk_mq_complete_request);
392 int blk_mq_request_started(struct request *rq)
394 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
396 EXPORT_SYMBOL_GPL(blk_mq_request_started);
398 void blk_mq_start_request(struct request *rq)
400 struct request_queue *q = rq->q;
402 trace_block_rq_issue(q, rq);
404 rq->resid_len = blk_rq_bytes(rq);
405 if (unlikely(blk_bidi_rq(rq)))
406 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
408 blk_add_timer(rq);
411 * Ensure that ->deadline is visible before set the started
412 * flag and clear the completed flag.
414 smp_mb__before_atomic();
417 * Mark us as started and clear complete. Complete might have been
418 * set if requeue raced with timeout, which then marked it as
419 * complete. So be sure to clear complete again when we start
420 * the request, otherwise we'll ignore the completion event.
422 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
423 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
424 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
425 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
427 if (q->dma_drain_size && blk_rq_bytes(rq)) {
429 * Make sure space for the drain appears. We know we can do
430 * this because max_hw_segments has been adjusted to be one
431 * fewer than the device can handle.
433 rq->nr_phys_segments++;
436 EXPORT_SYMBOL(blk_mq_start_request);
438 static void __blk_mq_requeue_request(struct request *rq)
440 struct request_queue *q = rq->q;
442 trace_block_rq_requeue(q, rq);
444 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
445 if (q->dma_drain_size && blk_rq_bytes(rq))
446 rq->nr_phys_segments--;
450 void blk_mq_requeue_request(struct request *rq)
452 __blk_mq_requeue_request(rq);
454 BUG_ON(blk_queued_rq(rq));
455 blk_mq_add_to_requeue_list(rq, true);
457 EXPORT_SYMBOL(blk_mq_requeue_request);
459 static void blk_mq_requeue_work(struct work_struct *work)
461 struct request_queue *q =
462 container_of(work, struct request_queue, requeue_work);
463 LIST_HEAD(rq_list);
464 struct request *rq, *next;
465 unsigned long flags;
467 spin_lock_irqsave(&q->requeue_lock, flags);
468 list_splice_init(&q->requeue_list, &rq_list);
469 spin_unlock_irqrestore(&q->requeue_lock, flags);
471 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
472 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
473 continue;
475 rq->cmd_flags &= ~REQ_SOFTBARRIER;
476 list_del_init(&rq->queuelist);
477 blk_mq_insert_request(rq, true, false, false);
480 while (!list_empty(&rq_list)) {
481 rq = list_entry(rq_list.next, struct request, queuelist);
482 list_del_init(&rq->queuelist);
483 blk_mq_insert_request(rq, false, false, false);
487 * Use the start variant of queue running here, so that running
488 * the requeue work will kick stopped queues.
490 blk_mq_start_hw_queues(q);
493 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
495 struct request_queue *q = rq->q;
496 unsigned long flags;
499 * We abuse this flag that is otherwise used by the I/O scheduler to
500 * request head insertation from the workqueue.
502 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
504 spin_lock_irqsave(&q->requeue_lock, flags);
505 if (at_head) {
506 rq->cmd_flags |= REQ_SOFTBARRIER;
507 list_add(&rq->queuelist, &q->requeue_list);
508 } else {
509 list_add_tail(&rq->queuelist, &q->requeue_list);
511 spin_unlock_irqrestore(&q->requeue_lock, flags);
513 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
515 void blk_mq_cancel_requeue_work(struct request_queue *q)
517 cancel_work_sync(&q->requeue_work);
519 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
521 void blk_mq_kick_requeue_list(struct request_queue *q)
523 kblockd_schedule_work(&q->requeue_work);
525 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
527 void blk_mq_abort_requeue_list(struct request_queue *q)
529 unsigned long flags;
530 LIST_HEAD(rq_list);
532 spin_lock_irqsave(&q->requeue_lock, flags);
533 list_splice_init(&q->requeue_list, &rq_list);
534 spin_unlock_irqrestore(&q->requeue_lock, flags);
536 while (!list_empty(&rq_list)) {
537 struct request *rq;
539 rq = list_first_entry(&rq_list, struct request, queuelist);
540 list_del_init(&rq->queuelist);
541 rq->errors = -EIO;
542 blk_mq_end_request(rq, rq->errors);
545 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
547 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
549 return tags->rqs[tag];
551 EXPORT_SYMBOL(blk_mq_tag_to_rq);
553 struct blk_mq_timeout_data {
554 unsigned long next;
555 unsigned int next_set;
558 void blk_mq_rq_timed_out(struct request *req, bool reserved)
560 struct blk_mq_ops *ops = req->q->mq_ops;
561 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
564 * We know that complete is set at this point. If STARTED isn't set
565 * anymore, then the request isn't active and the "timeout" should
566 * just be ignored. This can happen due to the bitflag ordering.
567 * Timeout first checks if STARTED is set, and if it is, assumes
568 * the request is active. But if we race with completion, then
569 * we both flags will get cleared. So check here again, and ignore
570 * a timeout event with a request that isn't active.
572 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
573 return;
575 if (ops->timeout)
576 ret = ops->timeout(req, reserved);
578 switch (ret) {
579 case BLK_EH_HANDLED:
580 __blk_mq_complete_request(req);
581 break;
582 case BLK_EH_RESET_TIMER:
583 blk_add_timer(req);
584 blk_clear_rq_complete(req);
585 break;
586 case BLK_EH_NOT_HANDLED:
587 break;
588 default:
589 printk(KERN_ERR "block: bad eh return: %d\n", ret);
590 break;
594 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
595 struct request *rq, void *priv, bool reserved)
597 struct blk_mq_timeout_data *data = priv;
599 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
601 * If a request wasn't started before the queue was
602 * marked dying, kill it here or it'll go unnoticed.
604 if (unlikely(blk_queue_dying(rq->q))) {
605 rq->errors = -EIO;
606 blk_mq_end_request(rq, rq->errors);
608 return;
610 if (rq->cmd_flags & REQ_NO_TIMEOUT)
611 return;
613 if (time_after_eq(jiffies, rq->deadline)) {
614 if (!blk_mark_rq_complete(rq))
615 blk_mq_rq_timed_out(rq, reserved);
616 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
617 data->next = rq->deadline;
618 data->next_set = 1;
622 static void blk_mq_rq_timer(unsigned long priv)
624 struct request_queue *q = (struct request_queue *)priv;
625 struct blk_mq_timeout_data data = {
626 .next = 0,
627 .next_set = 0,
629 int i;
631 /* A deadlock might occur if a request is stuck requiring a
632 * timeout at the same time a queue freeze is waiting
633 * completion, since the timeout code would not be able to
634 * acquire the queue reference here.
636 * That's why we don't use blk_queue_enter here; instead, we use
637 * percpu_ref_tryget directly, because we need to be able to
638 * obtain a reference even in the short window between the queue
639 * starting to freeze, by dropping the first reference in
640 * blk_mq_freeze_queue_start, and the moment the last request is
641 * consumed, marked by the instant q_usage_counter reaches
642 * zero.
644 if (!percpu_ref_tryget(&q->q_usage_counter))
645 return;
647 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
649 if (data.next_set) {
650 data.next = blk_rq_timeout(round_jiffies_up(data.next));
651 mod_timer(&q->timeout, data.next);
652 } else {
653 struct blk_mq_hw_ctx *hctx;
655 queue_for_each_hw_ctx(q, hctx, i) {
656 /* the hctx may be unmapped, so check it here */
657 if (blk_mq_hw_queue_mapped(hctx))
658 blk_mq_tag_idle(hctx);
661 blk_queue_exit(q);
665 * Reverse check our software queue for entries that we could potentially
666 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
667 * too much time checking for merges.
669 static bool blk_mq_attempt_merge(struct request_queue *q,
670 struct blk_mq_ctx *ctx, struct bio *bio)
672 struct request *rq;
673 int checked = 8;
675 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
676 int el_ret;
678 if (!checked--)
679 break;
681 if (!blk_rq_merge_ok(rq, bio))
682 continue;
684 el_ret = blk_try_merge(rq, bio);
685 if (el_ret == ELEVATOR_BACK_MERGE) {
686 if (bio_attempt_back_merge(q, rq, bio)) {
687 ctx->rq_merged++;
688 return true;
690 break;
691 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
692 if (bio_attempt_front_merge(q, rq, bio)) {
693 ctx->rq_merged++;
694 return true;
696 break;
700 return false;
704 * Process software queues that have been marked busy, splicing them
705 * to the for-dispatch
707 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
709 struct blk_mq_ctx *ctx;
710 int i;
712 for (i = 0; i < hctx->ctx_map.size; i++) {
713 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
714 unsigned int off, bit;
716 if (!bm->word)
717 continue;
719 bit = 0;
720 off = i * hctx->ctx_map.bits_per_word;
721 do {
722 bit = find_next_bit(&bm->word, bm->depth, bit);
723 if (bit >= bm->depth)
724 break;
726 ctx = hctx->ctxs[bit + off];
727 clear_bit(bit, &bm->word);
728 spin_lock(&ctx->lock);
729 list_splice_tail_init(&ctx->rq_list, list);
730 spin_unlock(&ctx->lock);
732 bit++;
733 } while (1);
738 * Run this hardware queue, pulling any software queues mapped to it in.
739 * Note that this function currently has various problems around ordering
740 * of IO. In particular, we'd like FIFO behaviour on handling existing
741 * items on the hctx->dispatch list. Ignore that for now.
743 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
745 struct request_queue *q = hctx->queue;
746 struct request *rq;
747 LIST_HEAD(rq_list);
748 LIST_HEAD(driver_list);
749 struct list_head *dptr;
750 int queued;
752 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
754 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
755 return;
757 hctx->run++;
760 * Touch any software queue that has pending entries.
762 flush_busy_ctxs(hctx, &rq_list);
765 * If we have previous entries on our dispatch list, grab them
766 * and stuff them at the front for more fair dispatch.
768 if (!list_empty_careful(&hctx->dispatch)) {
769 spin_lock(&hctx->lock);
770 if (!list_empty(&hctx->dispatch))
771 list_splice_init(&hctx->dispatch, &rq_list);
772 spin_unlock(&hctx->lock);
776 * Start off with dptr being NULL, so we start the first request
777 * immediately, even if we have more pending.
779 dptr = NULL;
782 * Now process all the entries, sending them to the driver.
784 queued = 0;
785 while (!list_empty(&rq_list)) {
786 struct blk_mq_queue_data bd;
787 int ret;
789 rq = list_first_entry(&rq_list, struct request, queuelist);
790 list_del_init(&rq->queuelist);
792 bd.rq = rq;
793 bd.list = dptr;
794 bd.last = list_empty(&rq_list);
796 ret = q->mq_ops->queue_rq(hctx, &bd);
797 switch (ret) {
798 case BLK_MQ_RQ_QUEUE_OK:
799 queued++;
800 break;
801 case BLK_MQ_RQ_QUEUE_BUSY:
802 list_add(&rq->queuelist, &rq_list);
803 __blk_mq_requeue_request(rq);
804 break;
805 default:
806 pr_err("blk-mq: bad return on queue: %d\n", ret);
807 case BLK_MQ_RQ_QUEUE_ERROR:
808 rq->errors = -EIO;
809 blk_mq_end_request(rq, rq->errors);
810 break;
813 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
814 break;
817 * We've done the first request. If we have more than 1
818 * left in the list, set dptr to defer issue.
820 if (!dptr && rq_list.next != rq_list.prev)
821 dptr = &driver_list;
824 if (!queued)
825 hctx->dispatched[0]++;
826 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
827 hctx->dispatched[ilog2(queued) + 1]++;
830 * Any items that need requeuing? Stuff them into hctx->dispatch,
831 * that is where we will continue on next queue run.
833 if (!list_empty(&rq_list)) {
834 spin_lock(&hctx->lock);
835 list_splice(&rq_list, &hctx->dispatch);
836 spin_unlock(&hctx->lock);
838 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
839 * it's possible the queue is stopped and restarted again
840 * before this. Queue restart will dispatch requests. And since
841 * requests in rq_list aren't added into hctx->dispatch yet,
842 * the requests in rq_list might get lost.
844 * blk_mq_run_hw_queue() already checks the STOPPED bit
846 blk_mq_run_hw_queue(hctx, true);
851 * It'd be great if the workqueue API had a way to pass
852 * in a mask and had some smarts for more clever placement.
853 * For now we just round-robin here, switching for every
854 * BLK_MQ_CPU_WORK_BATCH queued items.
856 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
858 if (hctx->queue->nr_hw_queues == 1)
859 return WORK_CPU_UNBOUND;
861 if (--hctx->next_cpu_batch <= 0) {
862 int next_cpu;
864 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
865 if (next_cpu >= nr_cpu_ids)
866 next_cpu = cpumask_first(hctx->cpumask);
868 hctx->next_cpu = next_cpu;
869 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
872 return hctx->next_cpu;
875 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
877 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
878 !blk_mq_hw_queue_mapped(hctx)))
879 return;
881 if (!async) {
882 int cpu = get_cpu();
883 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
884 __blk_mq_run_hw_queue(hctx);
885 put_cpu();
886 return;
889 put_cpu();
892 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
893 &hctx->run_work, 0);
896 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
898 struct blk_mq_hw_ctx *hctx;
899 int i;
901 queue_for_each_hw_ctx(q, hctx, i) {
902 if ((!blk_mq_hctx_has_pending(hctx) &&
903 list_empty_careful(&hctx->dispatch)) ||
904 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
905 continue;
907 blk_mq_run_hw_queue(hctx, async);
910 EXPORT_SYMBOL(blk_mq_run_hw_queues);
912 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
914 cancel_delayed_work(&hctx->run_work);
915 cancel_delayed_work(&hctx->delay_work);
916 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
918 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
920 void blk_mq_stop_hw_queues(struct request_queue *q)
922 struct blk_mq_hw_ctx *hctx;
923 int i;
925 queue_for_each_hw_ctx(q, hctx, i)
926 blk_mq_stop_hw_queue(hctx);
928 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
930 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
932 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
934 blk_mq_run_hw_queue(hctx, false);
936 EXPORT_SYMBOL(blk_mq_start_hw_queue);
938 void blk_mq_start_hw_queues(struct request_queue *q)
940 struct blk_mq_hw_ctx *hctx;
941 int i;
943 queue_for_each_hw_ctx(q, hctx, i)
944 blk_mq_start_hw_queue(hctx);
946 EXPORT_SYMBOL(blk_mq_start_hw_queues);
948 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
950 struct blk_mq_hw_ctx *hctx;
951 int i;
953 queue_for_each_hw_ctx(q, hctx, i) {
954 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
955 continue;
957 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
958 blk_mq_run_hw_queue(hctx, async);
961 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
963 static void blk_mq_run_work_fn(struct work_struct *work)
965 struct blk_mq_hw_ctx *hctx;
967 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
969 __blk_mq_run_hw_queue(hctx);
972 static void blk_mq_delay_work_fn(struct work_struct *work)
974 struct blk_mq_hw_ctx *hctx;
976 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
978 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
979 __blk_mq_run_hw_queue(hctx);
982 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
984 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
985 return;
987 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
988 &hctx->delay_work, msecs_to_jiffies(msecs));
990 EXPORT_SYMBOL(blk_mq_delay_queue);
992 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
993 struct blk_mq_ctx *ctx,
994 struct request *rq,
995 bool at_head)
997 trace_block_rq_insert(hctx->queue, rq);
999 if (at_head)
1000 list_add(&rq->queuelist, &ctx->rq_list);
1001 else
1002 list_add_tail(&rq->queuelist, &ctx->rq_list);
1005 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
1006 struct request *rq, bool at_head)
1008 struct blk_mq_ctx *ctx = rq->mq_ctx;
1010 __blk_mq_insert_req_list(hctx, ctx, rq, at_head);
1011 blk_mq_hctx_mark_pending(hctx, ctx);
1014 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1015 bool async)
1017 struct request_queue *q = rq->q;
1018 struct blk_mq_hw_ctx *hctx;
1019 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
1021 current_ctx = blk_mq_get_ctx(q);
1022 if (!cpu_online(ctx->cpu))
1023 rq->mq_ctx = ctx = current_ctx;
1025 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1027 spin_lock(&ctx->lock);
1028 __blk_mq_insert_request(hctx, rq, at_head);
1029 spin_unlock(&ctx->lock);
1031 if (run_queue)
1032 blk_mq_run_hw_queue(hctx, async);
1034 blk_mq_put_ctx(current_ctx);
1037 static void blk_mq_insert_requests(struct request_queue *q,
1038 struct blk_mq_ctx *ctx,
1039 struct list_head *list,
1040 int depth,
1041 bool from_schedule)
1044 struct blk_mq_hw_ctx *hctx;
1045 struct blk_mq_ctx *current_ctx;
1047 trace_block_unplug(q, depth, !from_schedule);
1049 current_ctx = blk_mq_get_ctx(q);
1051 if (!cpu_online(ctx->cpu))
1052 ctx = current_ctx;
1053 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1056 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1057 * offline now
1059 spin_lock(&ctx->lock);
1060 while (!list_empty(list)) {
1061 struct request *rq;
1063 rq = list_first_entry(list, struct request, queuelist);
1064 list_del_init(&rq->queuelist);
1065 rq->mq_ctx = ctx;
1066 __blk_mq_insert_req_list(hctx, ctx, rq, false);
1068 blk_mq_hctx_mark_pending(hctx, ctx);
1069 spin_unlock(&ctx->lock);
1071 blk_mq_run_hw_queue(hctx, from_schedule);
1072 blk_mq_put_ctx(current_ctx);
1075 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1077 struct request *rqa = container_of(a, struct request, queuelist);
1078 struct request *rqb = container_of(b, struct request, queuelist);
1080 return !(rqa->mq_ctx < rqb->mq_ctx ||
1081 (rqa->mq_ctx == rqb->mq_ctx &&
1082 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1085 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1087 struct blk_mq_ctx *this_ctx;
1088 struct request_queue *this_q;
1089 struct request *rq;
1090 LIST_HEAD(list);
1091 LIST_HEAD(ctx_list);
1092 unsigned int depth;
1094 list_splice_init(&plug->mq_list, &list);
1096 list_sort(NULL, &list, plug_ctx_cmp);
1098 this_q = NULL;
1099 this_ctx = NULL;
1100 depth = 0;
1102 while (!list_empty(&list)) {
1103 rq = list_entry_rq(list.next);
1104 list_del_init(&rq->queuelist);
1105 BUG_ON(!rq->q);
1106 if (rq->mq_ctx != this_ctx) {
1107 if (this_ctx) {
1108 blk_mq_insert_requests(this_q, this_ctx,
1109 &ctx_list, depth,
1110 from_schedule);
1113 this_ctx = rq->mq_ctx;
1114 this_q = rq->q;
1115 depth = 0;
1118 depth++;
1119 list_add_tail(&rq->queuelist, &ctx_list);
1123 * If 'this_ctx' is set, we know we have entries to complete
1124 * on 'ctx_list'. Do those.
1126 if (this_ctx) {
1127 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1128 from_schedule);
1132 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1134 init_request_from_bio(rq, bio);
1136 if (blk_do_io_stat(rq))
1137 blk_account_io_start(rq, 1);
1140 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1142 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1143 !blk_queue_nomerges(hctx->queue);
1146 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1147 struct blk_mq_ctx *ctx,
1148 struct request *rq, struct bio *bio)
1150 if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
1151 blk_mq_bio_to_request(rq, bio);
1152 spin_lock(&ctx->lock);
1153 insert_rq:
1154 __blk_mq_insert_request(hctx, rq, false);
1155 spin_unlock(&ctx->lock);
1156 return false;
1157 } else {
1158 struct request_queue *q = hctx->queue;
1160 spin_lock(&ctx->lock);
1161 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1162 blk_mq_bio_to_request(rq, bio);
1163 goto insert_rq;
1166 spin_unlock(&ctx->lock);
1167 __blk_mq_free_request(hctx, ctx, rq);
1168 return true;
1172 struct blk_map_ctx {
1173 struct blk_mq_hw_ctx *hctx;
1174 struct blk_mq_ctx *ctx;
1177 static struct request *blk_mq_map_request(struct request_queue *q,
1178 struct bio *bio,
1179 struct blk_map_ctx *data)
1181 struct blk_mq_hw_ctx *hctx;
1182 struct blk_mq_ctx *ctx;
1183 struct request *rq;
1184 int rw = bio_data_dir(bio);
1185 struct blk_mq_alloc_data alloc_data;
1187 blk_queue_enter_live(q);
1188 ctx = blk_mq_get_ctx(q);
1189 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1191 if (rw_is_sync(bio->bi_rw))
1192 rw |= REQ_SYNC;
1194 trace_block_getrq(q, bio, rw);
1195 blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1196 hctx);
1197 rq = __blk_mq_alloc_request(&alloc_data, rw);
1198 if (unlikely(!rq)) {
1199 __blk_mq_run_hw_queue(hctx);
1200 blk_mq_put_ctx(ctx);
1201 trace_block_sleeprq(q, bio, rw);
1203 ctx = blk_mq_get_ctx(q);
1204 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1205 blk_mq_set_alloc_data(&alloc_data, q,
1206 __GFP_RECLAIM|__GFP_HIGH, false, ctx, hctx);
1207 rq = __blk_mq_alloc_request(&alloc_data, rw);
1208 ctx = alloc_data.ctx;
1209 hctx = alloc_data.hctx;
1212 hctx->queued++;
1213 data->hctx = hctx;
1214 data->ctx = ctx;
1215 return rq;
1218 static int blk_mq_direct_issue_request(struct request *rq, blk_qc_t *cookie)
1220 int ret;
1221 struct request_queue *q = rq->q;
1222 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q,
1223 rq->mq_ctx->cpu);
1224 struct blk_mq_queue_data bd = {
1225 .rq = rq,
1226 .list = NULL,
1227 .last = 1
1229 blk_qc_t new_cookie = blk_tag_to_qc_t(rq->tag, hctx->queue_num);
1232 * For OK queue, we are done. For error, kill it. Any other
1233 * error (busy), just add it to our list as we previously
1234 * would have done
1236 ret = q->mq_ops->queue_rq(hctx, &bd);
1237 if (ret == BLK_MQ_RQ_QUEUE_OK) {
1238 *cookie = new_cookie;
1239 return 0;
1242 __blk_mq_requeue_request(rq);
1244 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1245 *cookie = BLK_QC_T_NONE;
1246 rq->errors = -EIO;
1247 blk_mq_end_request(rq, rq->errors);
1248 return 0;
1251 return -1;
1255 * Multiple hardware queue variant. This will not use per-process plugs,
1256 * but will attempt to bypass the hctx queueing if we can go straight to
1257 * hardware for SYNC IO.
1259 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1261 const int is_sync = rw_is_sync(bio->bi_rw);
1262 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1263 struct blk_map_ctx data;
1264 struct request *rq;
1265 unsigned int request_count = 0;
1266 struct blk_plug *plug;
1267 struct request *same_queue_rq = NULL;
1268 blk_qc_t cookie;
1270 blk_queue_bounce(q, &bio);
1272 blk_queue_split(q, &bio, q->bio_split);
1274 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1275 bio_io_error(bio);
1276 return BLK_QC_T_NONE;
1279 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1280 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1281 return BLK_QC_T_NONE;
1283 rq = blk_mq_map_request(q, bio, &data);
1284 if (unlikely(!rq))
1285 return BLK_QC_T_NONE;
1287 cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
1289 if (unlikely(is_flush_fua)) {
1290 blk_mq_bio_to_request(rq, bio);
1291 blk_insert_flush(rq);
1292 goto run_queue;
1295 plug = current->plug;
1297 * If the driver supports defer issued based on 'last', then
1298 * queue it up like normal since we can potentially save some
1299 * CPU this way.
1301 if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1302 !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1303 struct request *old_rq = NULL;
1305 blk_mq_bio_to_request(rq, bio);
1308 * We do limited pluging. If the bio can be merged, do that.
1309 * Otherwise the existing request in the plug list will be
1310 * issued. So the plug list will have one request at most
1312 if (plug) {
1314 * The plug list might get flushed before this. If that
1315 * happens, same_queue_rq is invalid and plug list is
1316 * empty
1318 if (same_queue_rq && !list_empty(&plug->mq_list)) {
1319 old_rq = same_queue_rq;
1320 list_del_init(&old_rq->queuelist);
1322 list_add_tail(&rq->queuelist, &plug->mq_list);
1323 } else /* is_sync */
1324 old_rq = rq;
1325 blk_mq_put_ctx(data.ctx);
1326 if (!old_rq)
1327 goto done;
1328 if (test_bit(BLK_MQ_S_STOPPED, &data.hctx->state) ||
1329 blk_mq_direct_issue_request(old_rq, &cookie) != 0)
1330 blk_mq_insert_request(old_rq, false, true, true);
1331 goto done;
1334 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1336 * For a SYNC request, send it to the hardware immediately. For
1337 * an ASYNC request, just ensure that we run it later on. The
1338 * latter allows for merging opportunities and more efficient
1339 * dispatching.
1341 run_queue:
1342 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1344 blk_mq_put_ctx(data.ctx);
1345 done:
1346 return cookie;
1350 * Single hardware queue variant. This will attempt to use any per-process
1351 * plug for merging and IO deferral.
1353 static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio)
1355 const int is_sync = rw_is_sync(bio->bi_rw);
1356 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1357 struct blk_plug *plug;
1358 unsigned int request_count = 0;
1359 struct blk_map_ctx data;
1360 struct request *rq;
1361 blk_qc_t cookie;
1363 blk_queue_bounce(q, &bio);
1365 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1366 bio_io_error(bio);
1367 return BLK_QC_T_NONE;
1370 blk_queue_split(q, &bio, q->bio_split);
1372 if (!is_flush_fua && !blk_queue_nomerges(q)) {
1373 if (blk_attempt_plug_merge(q, bio, &request_count, NULL))
1374 return BLK_QC_T_NONE;
1375 } else
1376 request_count = blk_plug_queued_count(q);
1378 rq = blk_mq_map_request(q, bio, &data);
1379 if (unlikely(!rq))
1380 return BLK_QC_T_NONE;
1382 cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
1384 if (unlikely(is_flush_fua)) {
1385 blk_mq_bio_to_request(rq, bio);
1386 blk_insert_flush(rq);
1387 goto run_queue;
1391 * A task plug currently exists. Since this is completely lockless,
1392 * utilize that to temporarily store requests until the task is
1393 * either done or scheduled away.
1395 plug = current->plug;
1396 if (plug) {
1397 blk_mq_bio_to_request(rq, bio);
1398 if (!request_count)
1399 trace_block_plug(q);
1401 blk_mq_put_ctx(data.ctx);
1403 if (request_count >= BLK_MAX_REQUEST_COUNT) {
1404 blk_flush_plug_list(plug, false);
1405 trace_block_plug(q);
1408 list_add_tail(&rq->queuelist, &plug->mq_list);
1409 return cookie;
1412 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1414 * For a SYNC request, send it to the hardware immediately. For
1415 * an ASYNC request, just ensure that we run it later on. The
1416 * latter allows for merging opportunities and more efficient
1417 * dispatching.
1419 run_queue:
1420 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1423 blk_mq_put_ctx(data.ctx);
1424 return cookie;
1428 * Default mapping to a software queue, since we use one per CPU.
1430 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1432 return q->queue_hw_ctx[q->mq_map[cpu]];
1434 EXPORT_SYMBOL(blk_mq_map_queue);
1436 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1437 struct blk_mq_tags *tags, unsigned int hctx_idx)
1439 struct page *page;
1441 if (tags->rqs && set->ops->exit_request) {
1442 int i;
1444 for (i = 0; i < tags->nr_tags; i++) {
1445 if (!tags->rqs[i])
1446 continue;
1447 set->ops->exit_request(set->driver_data, tags->rqs[i],
1448 hctx_idx, i);
1449 tags->rqs[i] = NULL;
1453 while (!list_empty(&tags->page_list)) {
1454 page = list_first_entry(&tags->page_list, struct page, lru);
1455 list_del_init(&page->lru);
1457 * Remove kmemleak object previously allocated in
1458 * blk_mq_init_rq_map().
1460 kmemleak_free(page_address(page));
1461 __free_pages(page, page->private);
1464 kfree(tags->rqs);
1466 blk_mq_free_tags(tags);
1469 static size_t order_to_size(unsigned int order)
1471 return (size_t)PAGE_SIZE << order;
1474 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1475 unsigned int hctx_idx)
1477 struct blk_mq_tags *tags;
1478 unsigned int i, j, entries_per_page, max_order = 4;
1479 size_t rq_size, left;
1481 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1482 set->numa_node,
1483 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1484 if (!tags)
1485 return NULL;
1487 INIT_LIST_HEAD(&tags->page_list);
1489 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1490 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1491 set->numa_node);
1492 if (!tags->rqs) {
1493 blk_mq_free_tags(tags);
1494 return NULL;
1498 * rq_size is the size of the request plus driver payload, rounded
1499 * to the cacheline size
1501 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1502 cache_line_size());
1503 left = rq_size * set->queue_depth;
1505 for (i = 0; i < set->queue_depth; ) {
1506 int this_order = max_order;
1507 struct page *page;
1508 int to_do;
1509 void *p;
1511 while (this_order && left < order_to_size(this_order - 1))
1512 this_order--;
1514 do {
1515 page = alloc_pages_node(set->numa_node,
1516 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1517 this_order);
1518 if (page)
1519 break;
1520 if (!this_order--)
1521 break;
1522 if (order_to_size(this_order) < rq_size)
1523 break;
1524 } while (1);
1526 if (!page)
1527 goto fail;
1529 page->private = this_order;
1530 list_add_tail(&page->lru, &tags->page_list);
1532 p = page_address(page);
1534 * Allow kmemleak to scan these pages as they contain pointers
1535 * to additional allocations like via ops->init_request().
1537 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1538 entries_per_page = order_to_size(this_order) / rq_size;
1539 to_do = min(entries_per_page, set->queue_depth - i);
1540 left -= to_do * rq_size;
1541 for (j = 0; j < to_do; j++) {
1542 tags->rqs[i] = p;
1543 if (set->ops->init_request) {
1544 if (set->ops->init_request(set->driver_data,
1545 tags->rqs[i], hctx_idx, i,
1546 set->numa_node)) {
1547 tags->rqs[i] = NULL;
1548 goto fail;
1552 p += rq_size;
1553 i++;
1556 return tags;
1558 fail:
1559 blk_mq_free_rq_map(set, tags, hctx_idx);
1560 return NULL;
1563 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1565 kfree(bitmap->map);
1568 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1570 unsigned int bpw = 8, total, num_maps, i;
1572 bitmap->bits_per_word = bpw;
1574 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1575 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1576 GFP_KERNEL, node);
1577 if (!bitmap->map)
1578 return -ENOMEM;
1580 total = nr_cpu_ids;
1581 for (i = 0; i < num_maps; i++) {
1582 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1583 total -= bitmap->map[i].depth;
1586 return 0;
1589 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1591 struct request_queue *q = hctx->queue;
1592 struct blk_mq_ctx *ctx;
1593 LIST_HEAD(tmp);
1596 * Move ctx entries to new CPU, if this one is going away.
1598 ctx = __blk_mq_get_ctx(q, cpu);
1600 spin_lock(&ctx->lock);
1601 if (!list_empty(&ctx->rq_list)) {
1602 list_splice_init(&ctx->rq_list, &tmp);
1603 blk_mq_hctx_clear_pending(hctx, ctx);
1605 spin_unlock(&ctx->lock);
1607 if (list_empty(&tmp))
1608 return NOTIFY_OK;
1610 ctx = blk_mq_get_ctx(q);
1611 spin_lock(&ctx->lock);
1613 while (!list_empty(&tmp)) {
1614 struct request *rq;
1616 rq = list_first_entry(&tmp, struct request, queuelist);
1617 rq->mq_ctx = ctx;
1618 list_move_tail(&rq->queuelist, &ctx->rq_list);
1621 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1622 blk_mq_hctx_mark_pending(hctx, ctx);
1624 spin_unlock(&ctx->lock);
1626 blk_mq_run_hw_queue(hctx, true);
1627 blk_mq_put_ctx(ctx);
1628 return NOTIFY_OK;
1631 static int blk_mq_hctx_notify(void *data, unsigned long action,
1632 unsigned int cpu)
1634 struct blk_mq_hw_ctx *hctx = data;
1636 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1637 return blk_mq_hctx_cpu_offline(hctx, cpu);
1640 * In case of CPU online, tags may be reallocated
1641 * in blk_mq_map_swqueue() after mapping is updated.
1644 return NOTIFY_OK;
1647 /* hctx->ctxs will be freed in queue's release handler */
1648 static void blk_mq_exit_hctx(struct request_queue *q,
1649 struct blk_mq_tag_set *set,
1650 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1652 unsigned flush_start_tag = set->queue_depth;
1654 if (blk_mq_hw_queue_mapped(hctx))
1655 blk_mq_tag_idle(hctx);
1657 if (set->ops->exit_request)
1658 set->ops->exit_request(set->driver_data,
1659 hctx->fq->flush_rq, hctx_idx,
1660 flush_start_tag + hctx_idx);
1662 if (set->ops->exit_hctx)
1663 set->ops->exit_hctx(hctx, hctx_idx);
1665 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1666 blk_free_flush_queue(hctx->fq);
1667 blk_mq_free_bitmap(&hctx->ctx_map);
1670 static void blk_mq_exit_hw_queues(struct request_queue *q,
1671 struct blk_mq_tag_set *set, int nr_queue)
1673 struct blk_mq_hw_ctx *hctx;
1674 unsigned int i;
1676 queue_for_each_hw_ctx(q, hctx, i) {
1677 if (i == nr_queue)
1678 break;
1679 blk_mq_exit_hctx(q, set, hctx, i);
1683 static void blk_mq_free_hw_queues(struct request_queue *q,
1684 struct blk_mq_tag_set *set)
1686 struct blk_mq_hw_ctx *hctx;
1687 unsigned int i;
1689 queue_for_each_hw_ctx(q, hctx, i)
1690 free_cpumask_var(hctx->cpumask);
1693 static int blk_mq_init_hctx(struct request_queue *q,
1694 struct blk_mq_tag_set *set,
1695 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1697 int node;
1698 unsigned flush_start_tag = set->queue_depth;
1700 node = hctx->numa_node;
1701 if (node == NUMA_NO_NODE)
1702 node = hctx->numa_node = set->numa_node;
1704 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1705 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1706 spin_lock_init(&hctx->lock);
1707 INIT_LIST_HEAD(&hctx->dispatch);
1708 hctx->queue = q;
1709 hctx->queue_num = hctx_idx;
1710 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1712 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1713 blk_mq_hctx_notify, hctx);
1714 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1716 hctx->tags = set->tags[hctx_idx];
1719 * Allocate space for all possible cpus to avoid allocation at
1720 * runtime
1722 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1723 GFP_KERNEL, node);
1724 if (!hctx->ctxs)
1725 goto unregister_cpu_notifier;
1727 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1728 goto free_ctxs;
1730 hctx->nr_ctx = 0;
1732 if (set->ops->init_hctx &&
1733 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1734 goto free_bitmap;
1736 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1737 if (!hctx->fq)
1738 goto exit_hctx;
1740 if (set->ops->init_request &&
1741 set->ops->init_request(set->driver_data,
1742 hctx->fq->flush_rq, hctx_idx,
1743 flush_start_tag + hctx_idx, node))
1744 goto free_fq;
1746 return 0;
1748 free_fq:
1749 kfree(hctx->fq);
1750 exit_hctx:
1751 if (set->ops->exit_hctx)
1752 set->ops->exit_hctx(hctx, hctx_idx);
1753 free_bitmap:
1754 blk_mq_free_bitmap(&hctx->ctx_map);
1755 free_ctxs:
1756 kfree(hctx->ctxs);
1757 unregister_cpu_notifier:
1758 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1760 return -1;
1763 static int blk_mq_init_hw_queues(struct request_queue *q,
1764 struct blk_mq_tag_set *set)
1766 struct blk_mq_hw_ctx *hctx;
1767 unsigned int i;
1770 * Initialize hardware queues
1772 queue_for_each_hw_ctx(q, hctx, i) {
1773 if (blk_mq_init_hctx(q, set, hctx, i))
1774 break;
1777 if (i == q->nr_hw_queues)
1778 return 0;
1781 * Init failed
1783 blk_mq_exit_hw_queues(q, set, i);
1785 return 1;
1788 static void blk_mq_init_cpu_queues(struct request_queue *q,
1789 unsigned int nr_hw_queues)
1791 unsigned int i;
1793 for_each_possible_cpu(i) {
1794 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1795 struct blk_mq_hw_ctx *hctx;
1797 memset(__ctx, 0, sizeof(*__ctx));
1798 __ctx->cpu = i;
1799 spin_lock_init(&__ctx->lock);
1800 INIT_LIST_HEAD(&__ctx->rq_list);
1801 __ctx->queue = q;
1803 /* If the cpu isn't online, the cpu is mapped to first hctx */
1804 if (!cpu_online(i))
1805 continue;
1807 hctx = q->mq_ops->map_queue(q, i);
1810 * Set local node, IFF we have more than one hw queue. If
1811 * not, we remain on the home node of the device
1813 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1814 hctx->numa_node = cpu_to_node(i);
1818 static void blk_mq_map_swqueue(struct request_queue *q,
1819 const struct cpumask *online_mask)
1821 unsigned int i;
1822 struct blk_mq_hw_ctx *hctx;
1823 struct blk_mq_ctx *ctx;
1824 struct blk_mq_tag_set *set = q->tag_set;
1827 * Avoid others reading imcomplete hctx->cpumask through sysfs
1829 mutex_lock(&q->sysfs_lock);
1831 queue_for_each_hw_ctx(q, hctx, i) {
1832 cpumask_clear(hctx->cpumask);
1833 hctx->nr_ctx = 0;
1837 * Map software to hardware queues
1839 queue_for_each_ctx(q, ctx, i) {
1840 /* If the cpu isn't online, the cpu is mapped to first hctx */
1841 if (!cpumask_test_cpu(i, online_mask))
1842 continue;
1844 hctx = q->mq_ops->map_queue(q, i);
1845 cpumask_set_cpu(i, hctx->cpumask);
1846 ctx->index_hw = hctx->nr_ctx;
1847 hctx->ctxs[hctx->nr_ctx++] = ctx;
1850 mutex_unlock(&q->sysfs_lock);
1852 queue_for_each_hw_ctx(q, hctx, i) {
1853 struct blk_mq_ctxmap *map = &hctx->ctx_map;
1856 * If no software queues are mapped to this hardware queue,
1857 * disable it and free the request entries.
1859 if (!hctx->nr_ctx) {
1860 if (set->tags[i]) {
1861 blk_mq_free_rq_map(set, set->tags[i], i);
1862 set->tags[i] = NULL;
1864 hctx->tags = NULL;
1865 continue;
1868 /* unmapped hw queue can be remapped after CPU topo changed */
1869 if (!set->tags[i])
1870 set->tags[i] = blk_mq_init_rq_map(set, i);
1871 hctx->tags = set->tags[i];
1872 WARN_ON(!hctx->tags);
1875 * Set the map size to the number of mapped software queues.
1876 * This is more accurate and more efficient than looping
1877 * over all possibly mapped software queues.
1879 map->size = DIV_ROUND_UP(hctx->nr_ctx, map->bits_per_word);
1882 * Initialize batch roundrobin counts
1884 hctx->next_cpu = cpumask_first(hctx->cpumask);
1885 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1888 queue_for_each_ctx(q, ctx, i) {
1889 if (!cpumask_test_cpu(i, online_mask))
1890 continue;
1892 hctx = q->mq_ops->map_queue(q, i);
1893 cpumask_set_cpu(i, hctx->tags->cpumask);
1897 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
1899 struct blk_mq_hw_ctx *hctx;
1900 int i;
1902 queue_for_each_hw_ctx(q, hctx, i) {
1903 if (shared)
1904 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1905 else
1906 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1910 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
1912 struct request_queue *q;
1914 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1915 blk_mq_freeze_queue(q);
1916 queue_set_hctx_shared(q, shared);
1917 blk_mq_unfreeze_queue(q);
1921 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1923 struct blk_mq_tag_set *set = q->tag_set;
1925 mutex_lock(&set->tag_list_lock);
1926 list_del_init(&q->tag_set_list);
1927 if (list_is_singular(&set->tag_list)) {
1928 /* just transitioned to unshared */
1929 set->flags &= ~BLK_MQ_F_TAG_SHARED;
1930 /* update existing queue */
1931 blk_mq_update_tag_set_depth(set, false);
1933 mutex_unlock(&set->tag_list_lock);
1936 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1937 struct request_queue *q)
1939 q->tag_set = set;
1941 mutex_lock(&set->tag_list_lock);
1943 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
1944 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
1945 set->flags |= BLK_MQ_F_TAG_SHARED;
1946 /* update existing queue */
1947 blk_mq_update_tag_set_depth(set, true);
1949 if (set->flags & BLK_MQ_F_TAG_SHARED)
1950 queue_set_hctx_shared(q, true);
1951 list_add_tail(&q->tag_set_list, &set->tag_list);
1953 mutex_unlock(&set->tag_list_lock);
1957 * It is the actual release handler for mq, but we do it from
1958 * request queue's release handler for avoiding use-after-free
1959 * and headache because q->mq_kobj shouldn't have been introduced,
1960 * but we can't group ctx/kctx kobj without it.
1962 void blk_mq_release(struct request_queue *q)
1964 struct blk_mq_hw_ctx *hctx;
1965 unsigned int i;
1967 /* hctx kobj stays in hctx */
1968 queue_for_each_hw_ctx(q, hctx, i) {
1969 if (!hctx)
1970 continue;
1971 kfree(hctx->ctxs);
1972 kfree(hctx);
1975 kfree(q->mq_map);
1976 q->mq_map = NULL;
1978 kfree(q->queue_hw_ctx);
1980 /* ctx kobj stays in queue_ctx */
1981 free_percpu(q->queue_ctx);
1984 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1986 struct request_queue *uninit_q, *q;
1988 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1989 if (!uninit_q)
1990 return ERR_PTR(-ENOMEM);
1992 q = blk_mq_init_allocated_queue(set, uninit_q);
1993 if (IS_ERR(q))
1994 blk_cleanup_queue(uninit_q);
1996 return q;
1998 EXPORT_SYMBOL(blk_mq_init_queue);
2000 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2001 struct request_queue *q)
2003 struct blk_mq_hw_ctx **hctxs;
2004 struct blk_mq_ctx __percpu *ctx;
2005 unsigned int *map;
2006 int i;
2008 ctx = alloc_percpu(struct blk_mq_ctx);
2009 if (!ctx)
2010 return ERR_PTR(-ENOMEM);
2012 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
2013 set->numa_node);
2015 if (!hctxs)
2016 goto err_percpu;
2018 map = blk_mq_make_queue_map(set);
2019 if (!map)
2020 goto err_map;
2022 for (i = 0; i < set->nr_hw_queues; i++) {
2023 int node = blk_mq_hw_queue_to_node(map, i);
2025 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
2026 GFP_KERNEL, node);
2027 if (!hctxs[i])
2028 goto err_hctxs;
2030 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2031 node))
2032 goto err_hctxs;
2034 atomic_set(&hctxs[i]->nr_active, 0);
2035 hctxs[i]->numa_node = node;
2036 hctxs[i]->queue_num = i;
2039 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
2040 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2042 q->nr_queues = nr_cpu_ids;
2043 q->nr_hw_queues = set->nr_hw_queues;
2044 q->mq_map = map;
2046 q->queue_ctx = ctx;
2047 q->queue_hw_ctx = hctxs;
2049 q->mq_ops = set->ops;
2050 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2052 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2053 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2055 q->sg_reserved_size = INT_MAX;
2057 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
2058 INIT_LIST_HEAD(&q->requeue_list);
2059 spin_lock_init(&q->requeue_lock);
2061 if (q->nr_hw_queues > 1)
2062 blk_queue_make_request(q, blk_mq_make_request);
2063 else
2064 blk_queue_make_request(q, blk_sq_make_request);
2067 * Do this after blk_queue_make_request() overrides it...
2069 q->nr_requests = set->queue_depth;
2071 if (set->ops->complete)
2072 blk_queue_softirq_done(q, set->ops->complete);
2074 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2076 if (blk_mq_init_hw_queues(q, set))
2077 goto err_hctxs;
2079 get_online_cpus();
2080 mutex_lock(&all_q_mutex);
2082 list_add_tail(&q->all_q_node, &all_q_list);
2083 blk_mq_add_queue_tag_set(set, q);
2084 blk_mq_map_swqueue(q, cpu_online_mask);
2086 mutex_unlock(&all_q_mutex);
2087 put_online_cpus();
2089 return q;
2091 err_hctxs:
2092 kfree(map);
2093 for (i = 0; i < set->nr_hw_queues; i++) {
2094 if (!hctxs[i])
2095 break;
2096 free_cpumask_var(hctxs[i]->cpumask);
2097 kfree(hctxs[i]);
2099 err_map:
2100 kfree(hctxs);
2101 err_percpu:
2102 free_percpu(ctx);
2103 return ERR_PTR(-ENOMEM);
2105 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2107 void blk_mq_free_queue(struct request_queue *q)
2109 struct blk_mq_tag_set *set = q->tag_set;
2111 mutex_lock(&all_q_mutex);
2112 list_del_init(&q->all_q_node);
2113 mutex_unlock(&all_q_mutex);
2115 blk_mq_del_queue_tag_set(q);
2117 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2118 blk_mq_free_hw_queues(q, set);
2121 /* Basically redo blk_mq_init_queue with queue frozen */
2122 static void blk_mq_queue_reinit(struct request_queue *q,
2123 const struct cpumask *online_mask)
2125 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2127 blk_mq_sysfs_unregister(q);
2129 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues, online_mask);
2132 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2133 * we should change hctx numa_node according to new topology (this
2134 * involves free and re-allocate memory, worthy doing?)
2137 blk_mq_map_swqueue(q, online_mask);
2139 blk_mq_sysfs_register(q);
2142 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2143 unsigned long action, void *hcpu)
2145 struct request_queue *q;
2146 int cpu = (unsigned long)hcpu;
2148 * New online cpumask which is going to be set in this hotplug event.
2149 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2150 * one-by-one and dynamically allocating this could result in a failure.
2152 static struct cpumask online_new;
2155 * Before hotadded cpu starts handling requests, new mappings must
2156 * be established. Otherwise, these requests in hw queue might
2157 * never be dispatched.
2159 * For example, there is a single hw queue (hctx) and two CPU queues
2160 * (ctx0 for CPU0, and ctx1 for CPU1).
2162 * Now CPU1 is just onlined and a request is inserted into
2163 * ctx1->rq_list and set bit0 in pending bitmap as ctx1->index_hw is
2164 * still zero.
2166 * And then while running hw queue, flush_busy_ctxs() finds bit0 is
2167 * set in pending bitmap and tries to retrieve requests in
2168 * hctx->ctxs[0]->rq_list. But htx->ctxs[0] is a pointer to ctx0,
2169 * so the request in ctx1->rq_list is ignored.
2171 switch (action & ~CPU_TASKS_FROZEN) {
2172 case CPU_DEAD:
2173 case CPU_UP_CANCELED:
2174 cpumask_copy(&online_new, cpu_online_mask);
2175 break;
2176 case CPU_UP_PREPARE:
2177 cpumask_copy(&online_new, cpu_online_mask);
2178 cpumask_set_cpu(cpu, &online_new);
2179 break;
2180 default:
2181 return NOTIFY_OK;
2184 mutex_lock(&all_q_mutex);
2187 * We need to freeze and reinit all existing queues. Freezing
2188 * involves synchronous wait for an RCU grace period and doing it
2189 * one by one may take a long time. Start freezing all queues in
2190 * one swoop and then wait for the completions so that freezing can
2191 * take place in parallel.
2193 list_for_each_entry(q, &all_q_list, all_q_node)
2194 blk_mq_freeze_queue_start(q);
2195 list_for_each_entry(q, &all_q_list, all_q_node) {
2196 blk_mq_freeze_queue_wait(q);
2199 * timeout handler can't touch hw queue during the
2200 * reinitialization
2202 del_timer_sync(&q->timeout);
2205 list_for_each_entry(q, &all_q_list, all_q_node)
2206 blk_mq_queue_reinit(q, &online_new);
2208 list_for_each_entry(q, &all_q_list, all_q_node)
2209 blk_mq_unfreeze_queue(q);
2211 mutex_unlock(&all_q_mutex);
2212 return NOTIFY_OK;
2215 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2217 int i;
2219 for (i = 0; i < set->nr_hw_queues; i++) {
2220 set->tags[i] = blk_mq_init_rq_map(set, i);
2221 if (!set->tags[i])
2222 goto out_unwind;
2225 return 0;
2227 out_unwind:
2228 while (--i >= 0)
2229 blk_mq_free_rq_map(set, set->tags[i], i);
2231 return -ENOMEM;
2235 * Allocate the request maps associated with this tag_set. Note that this
2236 * may reduce the depth asked for, if memory is tight. set->queue_depth
2237 * will be updated to reflect the allocated depth.
2239 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2241 unsigned int depth;
2242 int err;
2244 depth = set->queue_depth;
2245 do {
2246 err = __blk_mq_alloc_rq_maps(set);
2247 if (!err)
2248 break;
2250 set->queue_depth >>= 1;
2251 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2252 err = -ENOMEM;
2253 break;
2255 } while (set->queue_depth);
2257 if (!set->queue_depth || err) {
2258 pr_err("blk-mq: failed to allocate request map\n");
2259 return -ENOMEM;
2262 if (depth != set->queue_depth)
2263 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2264 depth, set->queue_depth);
2266 return 0;
2269 struct cpumask *blk_mq_tags_cpumask(struct blk_mq_tags *tags)
2271 return tags->cpumask;
2273 EXPORT_SYMBOL_GPL(blk_mq_tags_cpumask);
2276 * Alloc a tag set to be associated with one or more request queues.
2277 * May fail with EINVAL for various error conditions. May adjust the
2278 * requested depth down, if if it too large. In that case, the set
2279 * value will be stored in set->queue_depth.
2281 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2283 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2285 if (!set->nr_hw_queues)
2286 return -EINVAL;
2287 if (!set->queue_depth)
2288 return -EINVAL;
2289 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2290 return -EINVAL;
2292 if (!set->ops->queue_rq || !set->ops->map_queue)
2293 return -EINVAL;
2295 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2296 pr_info("blk-mq: reduced tag depth to %u\n",
2297 BLK_MQ_MAX_DEPTH);
2298 set->queue_depth = BLK_MQ_MAX_DEPTH;
2302 * If a crashdump is active, then we are potentially in a very
2303 * memory constrained environment. Limit us to 1 queue and
2304 * 64 tags to prevent using too much memory.
2306 if (is_kdump_kernel()) {
2307 set->nr_hw_queues = 1;
2308 set->queue_depth = min(64U, set->queue_depth);
2311 set->tags = kmalloc_node(set->nr_hw_queues *
2312 sizeof(struct blk_mq_tags *),
2313 GFP_KERNEL, set->numa_node);
2314 if (!set->tags)
2315 return -ENOMEM;
2317 if (blk_mq_alloc_rq_maps(set))
2318 goto enomem;
2320 mutex_init(&set->tag_list_lock);
2321 INIT_LIST_HEAD(&set->tag_list);
2323 return 0;
2324 enomem:
2325 kfree(set->tags);
2326 set->tags = NULL;
2327 return -ENOMEM;
2329 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2331 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2333 int i;
2335 for (i = 0; i < set->nr_hw_queues; i++) {
2336 if (set->tags[i])
2337 blk_mq_free_rq_map(set, set->tags[i], i);
2340 kfree(set->tags);
2341 set->tags = NULL;
2343 EXPORT_SYMBOL(blk_mq_free_tag_set);
2345 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2347 struct blk_mq_tag_set *set = q->tag_set;
2348 struct blk_mq_hw_ctx *hctx;
2349 int i, ret;
2351 if (!set || nr > set->queue_depth)
2352 return -EINVAL;
2354 ret = 0;
2355 queue_for_each_hw_ctx(q, hctx, i) {
2356 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2357 if (ret)
2358 break;
2361 if (!ret)
2362 q->nr_requests = nr;
2364 return ret;
2367 void blk_mq_disable_hotplug(void)
2369 mutex_lock(&all_q_mutex);
2372 void blk_mq_enable_hotplug(void)
2374 mutex_unlock(&all_q_mutex);
2377 static int __init blk_mq_init(void)
2379 blk_mq_cpu_init();
2381 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2383 return 0;
2385 subsys_initcall(blk_mq_init);