xref: /linux/block/blk-mq.c (revision d91517839e5d95adc0cf4b28caa7af62a71de526)
1 #include <linux/kernel.h>
2 #include <linux/module.h>
3 #include <linux/backing-dev.h>
4 #include <linux/bio.h>
5 #include <linux/blkdev.h>
6 #include <linux/mm.h>
7 #include <linux/init.h>
8 #include <linux/slab.h>
9 #include <linux/workqueue.h>
10 #include <linux/smp.h>
11 #include <linux/llist.h>
12 #include <linux/list_sort.h>
13 #include <linux/cpu.h>
14 #include <linux/cache.h>
15 #include <linux/sched/sysctl.h>
16 #include <linux/delay.h>
17 
18 #include <trace/events/block.h>
19 
20 #include <linux/blk-mq.h>
21 #include "blk.h"
22 #include "blk-mq.h"
23 #include "blk-mq-tag.h"
24 
25 static DEFINE_MUTEX(all_q_mutex);
26 static LIST_HEAD(all_q_list);
27 
28 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
29 
30 static struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
31 					   unsigned int cpu)
32 {
33 	return per_cpu_ptr(q->queue_ctx, cpu);
34 }
35 
36 /*
37  * This assumes per-cpu software queueing queues. They could be per-node
38  * as well, for instance. For now this is hardcoded as-is. Note that we don't
39  * care about preemption, since we know the ctx's are persistent. This does
40  * mean that we can't rely on ctx always matching the currently running CPU.
41  */
42 static struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
43 {
44 	return __blk_mq_get_ctx(q, get_cpu());
45 }
46 
47 static void blk_mq_put_ctx(struct blk_mq_ctx *ctx)
48 {
49 	put_cpu();
50 }
51 
52 /*
53  * Check if any of the ctx's have pending work in this hardware queue
54  */
55 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
56 {
57 	unsigned int i;
58 
59 	for (i = 0; i < hctx->nr_ctx_map; i++)
60 		if (hctx->ctx_map[i])
61 			return true;
62 
63 	return false;
64 }
65 
66 /*
67  * Mark this ctx as having pending work in this hardware queue
68  */
69 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
70 				     struct blk_mq_ctx *ctx)
71 {
72 	if (!test_bit(ctx->index_hw, hctx->ctx_map))
73 		set_bit(ctx->index_hw, hctx->ctx_map);
74 }
75 
76 static struct request *blk_mq_alloc_rq(struct blk_mq_hw_ctx *hctx, gfp_t gfp,
77 				       bool reserved)
78 {
79 	struct request *rq;
80 	unsigned int tag;
81 
82 	tag = blk_mq_get_tag(hctx->tags, gfp, reserved);
83 	if (tag != BLK_MQ_TAG_FAIL) {
84 		rq = hctx->rqs[tag];
85 		rq->tag = tag;
86 
87 		return rq;
88 	}
89 
90 	return NULL;
91 }
92 
93 static int blk_mq_queue_enter(struct request_queue *q)
94 {
95 	int ret;
96 
97 	__percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
98 	smp_wmb();
99 	/* we have problems to freeze the queue if it's initializing */
100 	if (!blk_queue_bypass(q) || !blk_queue_init_done(q))
101 		return 0;
102 
103 	__percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
104 
105 	spin_lock_irq(q->queue_lock);
106 	ret = wait_event_interruptible_lock_irq(q->mq_freeze_wq,
107 		!blk_queue_bypass(q) || blk_queue_dying(q),
108 		*q->queue_lock);
109 	/* inc usage with lock hold to avoid freeze_queue runs here */
110 	if (!ret && !blk_queue_dying(q))
111 		__percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
112 	else if (blk_queue_dying(q))
113 		ret = -ENODEV;
114 	spin_unlock_irq(q->queue_lock);
115 
116 	return ret;
117 }
118 
119 static void blk_mq_queue_exit(struct request_queue *q)
120 {
121 	__percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
122 }
123 
124 static void __blk_mq_drain_queue(struct request_queue *q)
125 {
126 	while (true) {
127 		s64 count;
128 
129 		spin_lock_irq(q->queue_lock);
130 		count = percpu_counter_sum(&q->mq_usage_counter);
131 		spin_unlock_irq(q->queue_lock);
132 
133 		if (count == 0)
134 			break;
135 		blk_mq_run_queues(q, false);
136 		msleep(10);
137 	}
138 }
139 
140 /*
141  * Guarantee no request is in use, so we can change any data structure of
142  * the queue afterward.
143  */
144 static void blk_mq_freeze_queue(struct request_queue *q)
145 {
146 	bool drain;
147 
148 	spin_lock_irq(q->queue_lock);
149 	drain = !q->bypass_depth++;
150 	queue_flag_set(QUEUE_FLAG_BYPASS, q);
151 	spin_unlock_irq(q->queue_lock);
152 
153 	if (drain)
154 		__blk_mq_drain_queue(q);
155 }
156 
157 void blk_mq_drain_queue(struct request_queue *q)
158 {
159 	__blk_mq_drain_queue(q);
160 }
161 
162 static void blk_mq_unfreeze_queue(struct request_queue *q)
163 {
164 	bool wake = false;
165 
166 	spin_lock_irq(q->queue_lock);
167 	if (!--q->bypass_depth) {
168 		queue_flag_clear(QUEUE_FLAG_BYPASS, q);
169 		wake = true;
170 	}
171 	WARN_ON_ONCE(q->bypass_depth < 0);
172 	spin_unlock_irq(q->queue_lock);
173 	if (wake)
174 		wake_up_all(&q->mq_freeze_wq);
175 }
176 
177 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
178 {
179 	return blk_mq_has_free_tags(hctx->tags);
180 }
181 EXPORT_SYMBOL(blk_mq_can_queue);
182 
183 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
184 			       struct request *rq, unsigned int rw_flags)
185 {
186 	if (blk_queue_io_stat(q))
187 		rw_flags |= REQ_IO_STAT;
188 
189 	rq->mq_ctx = ctx;
190 	rq->cmd_flags = rw_flags;
191 	rq->start_time = jiffies;
192 	set_start_time_ns(rq);
193 	ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
194 }
195 
196 static struct request *__blk_mq_alloc_request(struct blk_mq_hw_ctx *hctx,
197 					      gfp_t gfp, bool reserved)
198 {
199 	return blk_mq_alloc_rq(hctx, gfp, reserved);
200 }
201 
202 static struct request *blk_mq_alloc_request_pinned(struct request_queue *q,
203 						   int rw, gfp_t gfp,
204 						   bool reserved)
205 {
206 	struct request *rq;
207 
208 	do {
209 		struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
210 		struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q, ctx->cpu);
211 
212 		rq = __blk_mq_alloc_request(hctx, gfp & ~__GFP_WAIT, reserved);
213 		if (rq) {
214 			blk_mq_rq_ctx_init(q, ctx, rq, rw);
215 			break;
216 		}
217 
218 		blk_mq_put_ctx(ctx);
219 		if (!(gfp & __GFP_WAIT))
220 			break;
221 
222 		__blk_mq_run_hw_queue(hctx);
223 		blk_mq_wait_for_tags(hctx->tags);
224 	} while (1);
225 
226 	return rq;
227 }
228 
229 struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
230 		gfp_t gfp, bool reserved)
231 {
232 	struct request *rq;
233 
234 	if (blk_mq_queue_enter(q))
235 		return NULL;
236 
237 	rq = blk_mq_alloc_request_pinned(q, rw, gfp, reserved);
238 	if (rq)
239 		blk_mq_put_ctx(rq->mq_ctx);
240 	return rq;
241 }
242 
243 struct request *blk_mq_alloc_reserved_request(struct request_queue *q, int rw,
244 					      gfp_t gfp)
245 {
246 	struct request *rq;
247 
248 	if (blk_mq_queue_enter(q))
249 		return NULL;
250 
251 	rq = blk_mq_alloc_request_pinned(q, rw, gfp, true);
252 	if (rq)
253 		blk_mq_put_ctx(rq->mq_ctx);
254 	return rq;
255 }
256 EXPORT_SYMBOL(blk_mq_alloc_reserved_request);
257 
258 /*
259  * Re-init and set pdu, if we have it
260  */
261 static void blk_mq_rq_init(struct blk_mq_hw_ctx *hctx, struct request *rq)
262 {
263 	blk_rq_init(hctx->queue, rq);
264 
265 	if (hctx->cmd_size)
266 		rq->special = blk_mq_rq_to_pdu(rq);
267 }
268 
269 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
270 				  struct blk_mq_ctx *ctx, struct request *rq)
271 {
272 	const int tag = rq->tag;
273 	struct request_queue *q = rq->q;
274 
275 	blk_mq_rq_init(hctx, rq);
276 	blk_mq_put_tag(hctx->tags, tag);
277 
278 	blk_mq_queue_exit(q);
279 }
280 
281 void blk_mq_free_request(struct request *rq)
282 {
283 	struct blk_mq_ctx *ctx = rq->mq_ctx;
284 	struct blk_mq_hw_ctx *hctx;
285 	struct request_queue *q = rq->q;
286 
287 	ctx->rq_completed[rq_is_sync(rq)]++;
288 
289 	hctx = q->mq_ops->map_queue(q, ctx->cpu);
290 	__blk_mq_free_request(hctx, ctx, rq);
291 }
292 
293 static void blk_mq_bio_endio(struct request *rq, struct bio *bio, int error)
294 {
295 	if (error)
296 		clear_bit(BIO_UPTODATE, &bio->bi_flags);
297 	else if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
298 		error = -EIO;
299 
300 	if (unlikely(rq->cmd_flags & REQ_QUIET))
301 		set_bit(BIO_QUIET, &bio->bi_flags);
302 
303 	/* don't actually finish bio if it's part of flush sequence */
304 	if (!(rq->cmd_flags & REQ_FLUSH_SEQ))
305 		bio_endio(bio, error);
306 }
307 
308 void blk_mq_complete_request(struct request *rq, int error)
309 {
310 	struct bio *bio = rq->bio;
311 	unsigned int bytes = 0;
312 
313 	trace_block_rq_complete(rq->q, rq);
314 
315 	while (bio) {
316 		struct bio *next = bio->bi_next;
317 
318 		bio->bi_next = NULL;
319 		bytes += bio->bi_iter.bi_size;
320 		blk_mq_bio_endio(rq, bio, error);
321 		bio = next;
322 	}
323 
324 	blk_account_io_completion(rq, bytes);
325 
326 	blk_account_io_done(rq);
327 
328 	if (rq->end_io)
329 		rq->end_io(rq, error);
330 	else
331 		blk_mq_free_request(rq);
332 }
333 
334 void __blk_mq_end_io(struct request *rq, int error)
335 {
336 	if (!blk_mark_rq_complete(rq))
337 		blk_mq_complete_request(rq, error);
338 }
339 
340 static void blk_mq_end_io_remote(void *data)
341 {
342 	struct request *rq = data;
343 
344 	__blk_mq_end_io(rq, rq->errors);
345 }
346 
347 /*
348  * End IO on this request on a multiqueue enabled driver. We'll either do
349  * it directly inline, or punt to a local IPI handler on the matching
350  * remote CPU.
351  */
352 void blk_mq_end_io(struct request *rq, int error)
353 {
354 	struct blk_mq_ctx *ctx = rq->mq_ctx;
355 	int cpu;
356 
357 	if (!ctx->ipi_redirect)
358 		return __blk_mq_end_io(rq, error);
359 
360 	cpu = get_cpu();
361 	if (cpu != ctx->cpu && cpu_online(ctx->cpu)) {
362 		rq->errors = error;
363 		rq->csd.func = blk_mq_end_io_remote;
364 		rq->csd.info = rq;
365 		rq->csd.flags = 0;
366 		__smp_call_function_single(ctx->cpu, &rq->csd, 0);
367 	} else {
368 		__blk_mq_end_io(rq, error);
369 	}
370 	put_cpu();
371 }
372 EXPORT_SYMBOL(blk_mq_end_io);
373 
374 static void blk_mq_start_request(struct request *rq)
375 {
376 	struct request_queue *q = rq->q;
377 
378 	trace_block_rq_issue(q, rq);
379 
380 	/*
381 	 * Just mark start time and set the started bit. Due to memory
382 	 * ordering, we know we'll see the correct deadline as long as
383 	 * REQ_ATOMIC_STARTED is seen.
384 	 */
385 	rq->deadline = jiffies + q->rq_timeout;
386 	set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
387 }
388 
389 static void blk_mq_requeue_request(struct request *rq)
390 {
391 	struct request_queue *q = rq->q;
392 
393 	trace_block_rq_requeue(q, rq);
394 	clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
395 }
396 
397 struct blk_mq_timeout_data {
398 	struct blk_mq_hw_ctx *hctx;
399 	unsigned long *next;
400 	unsigned int *next_set;
401 };
402 
403 static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
404 {
405 	struct blk_mq_timeout_data *data = __data;
406 	struct blk_mq_hw_ctx *hctx = data->hctx;
407 	unsigned int tag;
408 
409 	 /* It may not be in flight yet (this is where
410 	 * the REQ_ATOMIC_STARTED flag comes in). The requests are
411 	 * statically allocated, so we know it's always safe to access the
412 	 * memory associated with a bit offset into ->rqs[].
413 	 */
414 	tag = 0;
415 	do {
416 		struct request *rq;
417 
418 		tag = find_next_zero_bit(free_tags, hctx->queue_depth, tag);
419 		if (tag >= hctx->queue_depth)
420 			break;
421 
422 		rq = hctx->rqs[tag++];
423 
424 		if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
425 			continue;
426 
427 		blk_rq_check_expired(rq, data->next, data->next_set);
428 	} while (1);
429 }
430 
431 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
432 					unsigned long *next,
433 					unsigned int *next_set)
434 {
435 	struct blk_mq_timeout_data data = {
436 		.hctx		= hctx,
437 		.next		= next,
438 		.next_set	= next_set,
439 	};
440 
441 	/*
442 	 * Ask the tagging code to iterate busy requests, so we can
443 	 * check them for timeout.
444 	 */
445 	blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
446 }
447 
448 static void blk_mq_rq_timer(unsigned long data)
449 {
450 	struct request_queue *q = (struct request_queue *) data;
451 	struct blk_mq_hw_ctx *hctx;
452 	unsigned long next = 0;
453 	int i, next_set = 0;
454 
455 	queue_for_each_hw_ctx(q, hctx, i)
456 		blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
457 
458 	if (next_set)
459 		mod_timer(&q->timeout, round_jiffies_up(next));
460 }
461 
462 /*
463  * Reverse check our software queue for entries that we could potentially
464  * merge with. Currently includes a hand-wavy stop count of 8, to not spend
465  * too much time checking for merges.
466  */
467 static bool blk_mq_attempt_merge(struct request_queue *q,
468 				 struct blk_mq_ctx *ctx, struct bio *bio)
469 {
470 	struct request *rq;
471 	int checked = 8;
472 
473 	list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
474 		int el_ret;
475 
476 		if (!checked--)
477 			break;
478 
479 		if (!blk_rq_merge_ok(rq, bio))
480 			continue;
481 
482 		el_ret = blk_try_merge(rq, bio);
483 		if (el_ret == ELEVATOR_BACK_MERGE) {
484 			if (bio_attempt_back_merge(q, rq, bio)) {
485 				ctx->rq_merged++;
486 				return true;
487 			}
488 			break;
489 		} else if (el_ret == ELEVATOR_FRONT_MERGE) {
490 			if (bio_attempt_front_merge(q, rq, bio)) {
491 				ctx->rq_merged++;
492 				return true;
493 			}
494 			break;
495 		}
496 	}
497 
498 	return false;
499 }
500 
501 void blk_mq_add_timer(struct request *rq)
502 {
503 	__blk_add_timer(rq, NULL);
504 }
505 
506 /*
507  * Run this hardware queue, pulling any software queues mapped to it in.
508  * Note that this function currently has various problems around ordering
509  * of IO. In particular, we'd like FIFO behaviour on handling existing
510  * items on the hctx->dispatch list. Ignore that for now.
511  */
512 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
513 {
514 	struct request_queue *q = hctx->queue;
515 	struct blk_mq_ctx *ctx;
516 	struct request *rq;
517 	LIST_HEAD(rq_list);
518 	int bit, queued;
519 
520 	if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->flags)))
521 		return;
522 
523 	hctx->run++;
524 
525 	/*
526 	 * Touch any software queue that has pending entries.
527 	 */
528 	for_each_set_bit(bit, hctx->ctx_map, hctx->nr_ctx) {
529 		clear_bit(bit, hctx->ctx_map);
530 		ctx = hctx->ctxs[bit];
531 		BUG_ON(bit != ctx->index_hw);
532 
533 		spin_lock(&ctx->lock);
534 		list_splice_tail_init(&ctx->rq_list, &rq_list);
535 		spin_unlock(&ctx->lock);
536 	}
537 
538 	/*
539 	 * If we have previous entries on our dispatch list, grab them
540 	 * and stuff them at the front for more fair dispatch.
541 	 */
542 	if (!list_empty_careful(&hctx->dispatch)) {
543 		spin_lock(&hctx->lock);
544 		if (!list_empty(&hctx->dispatch))
545 			list_splice_init(&hctx->dispatch, &rq_list);
546 		spin_unlock(&hctx->lock);
547 	}
548 
549 	/*
550 	 * Delete and return all entries from our dispatch list
551 	 */
552 	queued = 0;
553 
554 	/*
555 	 * Now process all the entries, sending them to the driver.
556 	 */
557 	while (!list_empty(&rq_list)) {
558 		int ret;
559 
560 		rq = list_first_entry(&rq_list, struct request, queuelist);
561 		list_del_init(&rq->queuelist);
562 		blk_mq_start_request(rq);
563 
564 		/*
565 		 * Last request in the series. Flag it as such, this
566 		 * enables drivers to know when IO should be kicked off,
567 		 * if they don't do it on a per-request basis.
568 		 *
569 		 * Note: the flag isn't the only condition drivers
570 		 * should do kick off. If drive is busy, the last
571 		 * request might not have the bit set.
572 		 */
573 		if (list_empty(&rq_list))
574 			rq->cmd_flags |= REQ_END;
575 
576 		ret = q->mq_ops->queue_rq(hctx, rq);
577 		switch (ret) {
578 		case BLK_MQ_RQ_QUEUE_OK:
579 			queued++;
580 			continue;
581 		case BLK_MQ_RQ_QUEUE_BUSY:
582 			/*
583 			 * FIXME: we should have a mechanism to stop the queue
584 			 * like blk_stop_queue, otherwise we will waste cpu
585 			 * time
586 			 */
587 			list_add(&rq->queuelist, &rq_list);
588 			blk_mq_requeue_request(rq);
589 			break;
590 		default:
591 			pr_err("blk-mq: bad return on queue: %d\n", ret);
592 			rq->errors = -EIO;
593 		case BLK_MQ_RQ_QUEUE_ERROR:
594 			blk_mq_end_io(rq, rq->errors);
595 			break;
596 		}
597 
598 		if (ret == BLK_MQ_RQ_QUEUE_BUSY)
599 			break;
600 	}
601 
602 	if (!queued)
603 		hctx->dispatched[0]++;
604 	else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
605 		hctx->dispatched[ilog2(queued) + 1]++;
606 
607 	/*
608 	 * Any items that need requeuing? Stuff them into hctx->dispatch,
609 	 * that is where we will continue on next queue run.
610 	 */
611 	if (!list_empty(&rq_list)) {
612 		spin_lock(&hctx->lock);
613 		list_splice(&rq_list, &hctx->dispatch);
614 		spin_unlock(&hctx->lock);
615 	}
616 }
617 
618 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
619 {
620 	if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->flags)))
621 		return;
622 
623 	if (!async)
624 		__blk_mq_run_hw_queue(hctx);
625 	else {
626 		struct request_queue *q = hctx->queue;
627 
628 		kblockd_schedule_delayed_work(q, &hctx->delayed_work, 0);
629 	}
630 }
631 
632 void blk_mq_run_queues(struct request_queue *q, bool async)
633 {
634 	struct blk_mq_hw_ctx *hctx;
635 	int i;
636 
637 	queue_for_each_hw_ctx(q, hctx, i) {
638 		if ((!blk_mq_hctx_has_pending(hctx) &&
639 		    list_empty_careful(&hctx->dispatch)) ||
640 		    test_bit(BLK_MQ_S_STOPPED, &hctx->flags))
641 			continue;
642 
643 		blk_mq_run_hw_queue(hctx, async);
644 	}
645 }
646 EXPORT_SYMBOL(blk_mq_run_queues);
647 
648 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
649 {
650 	cancel_delayed_work(&hctx->delayed_work);
651 	set_bit(BLK_MQ_S_STOPPED, &hctx->state);
652 }
653 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
654 
655 void blk_mq_stop_hw_queues(struct request_queue *q)
656 {
657 	struct blk_mq_hw_ctx *hctx;
658 	int i;
659 
660 	queue_for_each_hw_ctx(q, hctx, i)
661 		blk_mq_stop_hw_queue(hctx);
662 }
663 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
664 
665 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
666 {
667 	clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
668 	__blk_mq_run_hw_queue(hctx);
669 }
670 EXPORT_SYMBOL(blk_mq_start_hw_queue);
671 
672 void blk_mq_start_stopped_hw_queues(struct request_queue *q)
673 {
674 	struct blk_mq_hw_ctx *hctx;
675 	int i;
676 
677 	queue_for_each_hw_ctx(q, hctx, i) {
678 		if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
679 			continue;
680 
681 		clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
682 		blk_mq_run_hw_queue(hctx, true);
683 	}
684 }
685 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
686 
687 static void blk_mq_work_fn(struct work_struct *work)
688 {
689 	struct blk_mq_hw_ctx *hctx;
690 
691 	hctx = container_of(work, struct blk_mq_hw_ctx, delayed_work.work);
692 	__blk_mq_run_hw_queue(hctx);
693 }
694 
695 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
696 				    struct request *rq)
697 {
698 	struct blk_mq_ctx *ctx = rq->mq_ctx;
699 
700 	trace_block_rq_insert(hctx->queue, rq);
701 
702 	list_add_tail(&rq->queuelist, &ctx->rq_list);
703 	blk_mq_hctx_mark_pending(hctx, ctx);
704 
705 	/*
706 	 * We do this early, to ensure we are on the right CPU.
707 	 */
708 	blk_mq_add_timer(rq);
709 }
710 
711 void blk_mq_insert_request(struct request_queue *q, struct request *rq,
712 			   bool run_queue)
713 {
714 	struct blk_mq_hw_ctx *hctx;
715 	struct blk_mq_ctx *ctx, *current_ctx;
716 
717 	ctx = rq->mq_ctx;
718 	hctx = q->mq_ops->map_queue(q, ctx->cpu);
719 
720 	if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA)) {
721 		blk_insert_flush(rq);
722 	} else {
723 		current_ctx = blk_mq_get_ctx(q);
724 
725 		if (!cpu_online(ctx->cpu)) {
726 			ctx = current_ctx;
727 			hctx = q->mq_ops->map_queue(q, ctx->cpu);
728 			rq->mq_ctx = ctx;
729 		}
730 		spin_lock(&ctx->lock);
731 		__blk_mq_insert_request(hctx, rq);
732 		spin_unlock(&ctx->lock);
733 
734 		blk_mq_put_ctx(current_ctx);
735 	}
736 
737 	if (run_queue)
738 		__blk_mq_run_hw_queue(hctx);
739 }
740 EXPORT_SYMBOL(blk_mq_insert_request);
741 
742 /*
743  * This is a special version of blk_mq_insert_request to bypass FLUSH request
744  * check. Should only be used internally.
745  */
746 void blk_mq_run_request(struct request *rq, bool run_queue, bool async)
747 {
748 	struct request_queue *q = rq->q;
749 	struct blk_mq_hw_ctx *hctx;
750 	struct blk_mq_ctx *ctx, *current_ctx;
751 
752 	current_ctx = blk_mq_get_ctx(q);
753 
754 	ctx = rq->mq_ctx;
755 	if (!cpu_online(ctx->cpu)) {
756 		ctx = current_ctx;
757 		rq->mq_ctx = ctx;
758 	}
759 	hctx = q->mq_ops->map_queue(q, ctx->cpu);
760 
761 	/* ctx->cpu might be offline */
762 	spin_lock(&ctx->lock);
763 	__blk_mq_insert_request(hctx, rq);
764 	spin_unlock(&ctx->lock);
765 
766 	blk_mq_put_ctx(current_ctx);
767 
768 	if (run_queue)
769 		blk_mq_run_hw_queue(hctx, async);
770 }
771 
772 static void blk_mq_insert_requests(struct request_queue *q,
773 				     struct blk_mq_ctx *ctx,
774 				     struct list_head *list,
775 				     int depth,
776 				     bool from_schedule)
777 
778 {
779 	struct blk_mq_hw_ctx *hctx;
780 	struct blk_mq_ctx *current_ctx;
781 
782 	trace_block_unplug(q, depth, !from_schedule);
783 
784 	current_ctx = blk_mq_get_ctx(q);
785 
786 	if (!cpu_online(ctx->cpu))
787 		ctx = current_ctx;
788 	hctx = q->mq_ops->map_queue(q, ctx->cpu);
789 
790 	/*
791 	 * preemption doesn't flush plug list, so it's possible ctx->cpu is
792 	 * offline now
793 	 */
794 	spin_lock(&ctx->lock);
795 	while (!list_empty(list)) {
796 		struct request *rq;
797 
798 		rq = list_first_entry(list, struct request, queuelist);
799 		list_del_init(&rq->queuelist);
800 		rq->mq_ctx = ctx;
801 		__blk_mq_insert_request(hctx, rq);
802 	}
803 	spin_unlock(&ctx->lock);
804 
805 	blk_mq_put_ctx(current_ctx);
806 
807 	blk_mq_run_hw_queue(hctx, from_schedule);
808 }
809 
810 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
811 {
812 	struct request *rqa = container_of(a, struct request, queuelist);
813 	struct request *rqb = container_of(b, struct request, queuelist);
814 
815 	return !(rqa->mq_ctx < rqb->mq_ctx ||
816 		 (rqa->mq_ctx == rqb->mq_ctx &&
817 		  blk_rq_pos(rqa) < blk_rq_pos(rqb)));
818 }
819 
820 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
821 {
822 	struct blk_mq_ctx *this_ctx;
823 	struct request_queue *this_q;
824 	struct request *rq;
825 	LIST_HEAD(list);
826 	LIST_HEAD(ctx_list);
827 	unsigned int depth;
828 
829 	list_splice_init(&plug->mq_list, &list);
830 
831 	list_sort(NULL, &list, plug_ctx_cmp);
832 
833 	this_q = NULL;
834 	this_ctx = NULL;
835 	depth = 0;
836 
837 	while (!list_empty(&list)) {
838 		rq = list_entry_rq(list.next);
839 		list_del_init(&rq->queuelist);
840 		BUG_ON(!rq->q);
841 		if (rq->mq_ctx != this_ctx) {
842 			if (this_ctx) {
843 				blk_mq_insert_requests(this_q, this_ctx,
844 							&ctx_list, depth,
845 							from_schedule);
846 			}
847 
848 			this_ctx = rq->mq_ctx;
849 			this_q = rq->q;
850 			depth = 0;
851 		}
852 
853 		depth++;
854 		list_add_tail(&rq->queuelist, &ctx_list);
855 	}
856 
857 	/*
858 	 * If 'this_ctx' is set, we know we have entries to complete
859 	 * on 'ctx_list'. Do those.
860 	 */
861 	if (this_ctx) {
862 		blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
863 				       from_schedule);
864 	}
865 }
866 
867 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
868 {
869 	init_request_from_bio(rq, bio);
870 	blk_account_io_start(rq, 1);
871 }
872 
873 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
874 {
875 	struct blk_mq_hw_ctx *hctx;
876 	struct blk_mq_ctx *ctx;
877 	const int is_sync = rw_is_sync(bio->bi_rw);
878 	const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
879 	int rw = bio_data_dir(bio);
880 	struct request *rq;
881 	unsigned int use_plug, request_count = 0;
882 
883 	/*
884 	 * If we have multiple hardware queues, just go directly to
885 	 * one of those for sync IO.
886 	 */
887 	use_plug = !is_flush_fua && ((q->nr_hw_queues == 1) || !is_sync);
888 
889 	blk_queue_bounce(q, &bio);
890 
891 	if (use_plug && blk_attempt_plug_merge(q, bio, &request_count))
892 		return;
893 
894 	if (blk_mq_queue_enter(q)) {
895 		bio_endio(bio, -EIO);
896 		return;
897 	}
898 
899 	ctx = blk_mq_get_ctx(q);
900 	hctx = q->mq_ops->map_queue(q, ctx->cpu);
901 
902 	trace_block_getrq(q, bio, rw);
903 	rq = __blk_mq_alloc_request(hctx, GFP_ATOMIC, false);
904 	if (likely(rq))
905 		blk_mq_rq_ctx_init(q, ctx, rq, rw);
906 	else {
907 		blk_mq_put_ctx(ctx);
908 		trace_block_sleeprq(q, bio, rw);
909 		rq = blk_mq_alloc_request_pinned(q, rw, __GFP_WAIT|GFP_ATOMIC,
910 							false);
911 		ctx = rq->mq_ctx;
912 		hctx = q->mq_ops->map_queue(q, ctx->cpu);
913 	}
914 
915 	hctx->queued++;
916 
917 	if (unlikely(is_flush_fua)) {
918 		blk_mq_bio_to_request(rq, bio);
919 		blk_mq_put_ctx(ctx);
920 		blk_insert_flush(rq);
921 		goto run_queue;
922 	}
923 
924 	/*
925 	 * A task plug currently exists. Since this is completely lockless,
926 	 * utilize that to temporarily store requests until the task is
927 	 * either done or scheduled away.
928 	 */
929 	if (use_plug) {
930 		struct blk_plug *plug = current->plug;
931 
932 		if (plug) {
933 			blk_mq_bio_to_request(rq, bio);
934 			if (list_empty(&plug->mq_list))
935 				trace_block_plug(q);
936 			else if (request_count >= BLK_MAX_REQUEST_COUNT) {
937 				blk_flush_plug_list(plug, false);
938 				trace_block_plug(q);
939 			}
940 			list_add_tail(&rq->queuelist, &plug->mq_list);
941 			blk_mq_put_ctx(ctx);
942 			return;
943 		}
944 	}
945 
946 	spin_lock(&ctx->lock);
947 
948 	if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
949 	    blk_mq_attempt_merge(q, ctx, bio))
950 		__blk_mq_free_request(hctx, ctx, rq);
951 	else {
952 		blk_mq_bio_to_request(rq, bio);
953 		__blk_mq_insert_request(hctx, rq);
954 	}
955 
956 	spin_unlock(&ctx->lock);
957 	blk_mq_put_ctx(ctx);
958 
959 	/*
960 	 * For a SYNC request, send it to the hardware immediately. For an
961 	 * ASYNC request, just ensure that we run it later on. The latter
962 	 * allows for merging opportunities and more efficient dispatching.
963 	 */
964 run_queue:
965 	blk_mq_run_hw_queue(hctx, !is_sync || is_flush_fua);
966 }
967 
968 /*
969  * Default mapping to a software queue, since we use one per CPU.
970  */
971 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
972 {
973 	return q->queue_hw_ctx[q->mq_map[cpu]];
974 }
975 EXPORT_SYMBOL(blk_mq_map_queue);
976 
977 struct blk_mq_hw_ctx *blk_mq_alloc_single_hw_queue(struct blk_mq_reg *reg,
978 						   unsigned int hctx_index)
979 {
980 	return kmalloc_node(sizeof(struct blk_mq_hw_ctx),
981 				GFP_KERNEL | __GFP_ZERO, reg->numa_node);
982 }
983 EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue);
984 
985 void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx *hctx,
986 				 unsigned int hctx_index)
987 {
988 	kfree(hctx);
989 }
990 EXPORT_SYMBOL(blk_mq_free_single_hw_queue);
991 
992 static void blk_mq_hctx_notify(void *data, unsigned long action,
993 			       unsigned int cpu)
994 {
995 	struct blk_mq_hw_ctx *hctx = data;
996 	struct blk_mq_ctx *ctx;
997 	LIST_HEAD(tmp);
998 
999 	if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
1000 		return;
1001 
1002 	/*
1003 	 * Move ctx entries to new CPU, if this one is going away.
1004 	 */
1005 	ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1006 
1007 	spin_lock(&ctx->lock);
1008 	if (!list_empty(&ctx->rq_list)) {
1009 		list_splice_init(&ctx->rq_list, &tmp);
1010 		clear_bit(ctx->index_hw, hctx->ctx_map);
1011 	}
1012 	spin_unlock(&ctx->lock);
1013 
1014 	if (list_empty(&tmp))
1015 		return;
1016 
1017 	ctx = blk_mq_get_ctx(hctx->queue);
1018 	spin_lock(&ctx->lock);
1019 
1020 	while (!list_empty(&tmp)) {
1021 		struct request *rq;
1022 
1023 		rq = list_first_entry(&tmp, struct request, queuelist);
1024 		rq->mq_ctx = ctx;
1025 		list_move_tail(&rq->queuelist, &ctx->rq_list);
1026 	}
1027 
1028 	blk_mq_hctx_mark_pending(hctx, ctx);
1029 
1030 	spin_unlock(&ctx->lock);
1031 	blk_mq_put_ctx(ctx);
1032 }
1033 
1034 static void blk_mq_init_hw_commands(struct blk_mq_hw_ctx *hctx,
1035 				    void (*init)(void *, struct blk_mq_hw_ctx *,
1036 					struct request *, unsigned int),
1037 				    void *data)
1038 {
1039 	unsigned int i;
1040 
1041 	for (i = 0; i < hctx->queue_depth; i++) {
1042 		struct request *rq = hctx->rqs[i];
1043 
1044 		init(data, hctx, rq, i);
1045 	}
1046 }
1047 
1048 void blk_mq_init_commands(struct request_queue *q,
1049 			  void (*init)(void *, struct blk_mq_hw_ctx *,
1050 					struct request *, unsigned int),
1051 			  void *data)
1052 {
1053 	struct blk_mq_hw_ctx *hctx;
1054 	unsigned int i;
1055 
1056 	queue_for_each_hw_ctx(q, hctx, i)
1057 		blk_mq_init_hw_commands(hctx, init, data);
1058 }
1059 EXPORT_SYMBOL(blk_mq_init_commands);
1060 
1061 static void blk_mq_free_rq_map(struct blk_mq_hw_ctx *hctx)
1062 {
1063 	struct page *page;
1064 
1065 	while (!list_empty(&hctx->page_list)) {
1066 		page = list_first_entry(&hctx->page_list, struct page, lru);
1067 		list_del_init(&page->lru);
1068 		__free_pages(page, page->private);
1069 	}
1070 
1071 	kfree(hctx->rqs);
1072 
1073 	if (hctx->tags)
1074 		blk_mq_free_tags(hctx->tags);
1075 }
1076 
1077 static size_t order_to_size(unsigned int order)
1078 {
1079 	size_t ret = PAGE_SIZE;
1080 
1081 	while (order--)
1082 		ret *= 2;
1083 
1084 	return ret;
1085 }
1086 
1087 static int blk_mq_init_rq_map(struct blk_mq_hw_ctx *hctx,
1088 			      unsigned int reserved_tags, int node)
1089 {
1090 	unsigned int i, j, entries_per_page, max_order = 4;
1091 	size_t rq_size, left;
1092 
1093 	INIT_LIST_HEAD(&hctx->page_list);
1094 
1095 	hctx->rqs = kmalloc_node(hctx->queue_depth * sizeof(struct request *),
1096 					GFP_KERNEL, node);
1097 	if (!hctx->rqs)
1098 		return -ENOMEM;
1099 
1100 	/*
1101 	 * rq_size is the size of the request plus driver payload, rounded
1102 	 * to the cacheline size
1103 	 */
1104 	rq_size = round_up(sizeof(struct request) + hctx->cmd_size,
1105 				cache_line_size());
1106 	left = rq_size * hctx->queue_depth;
1107 
1108 	for (i = 0; i < hctx->queue_depth;) {
1109 		int this_order = max_order;
1110 		struct page *page;
1111 		int to_do;
1112 		void *p;
1113 
1114 		while (left < order_to_size(this_order - 1) && this_order)
1115 			this_order--;
1116 
1117 		do {
1118 			page = alloc_pages_node(node, GFP_KERNEL, this_order);
1119 			if (page)
1120 				break;
1121 			if (!this_order--)
1122 				break;
1123 			if (order_to_size(this_order) < rq_size)
1124 				break;
1125 		} while (1);
1126 
1127 		if (!page)
1128 			break;
1129 
1130 		page->private = this_order;
1131 		list_add_tail(&page->lru, &hctx->page_list);
1132 
1133 		p = page_address(page);
1134 		entries_per_page = order_to_size(this_order) / rq_size;
1135 		to_do = min(entries_per_page, hctx->queue_depth - i);
1136 		left -= to_do * rq_size;
1137 		for (j = 0; j < to_do; j++) {
1138 			hctx->rqs[i] = p;
1139 			blk_mq_rq_init(hctx, hctx->rqs[i]);
1140 			p += rq_size;
1141 			i++;
1142 		}
1143 	}
1144 
1145 	if (i < (reserved_tags + BLK_MQ_TAG_MIN))
1146 		goto err_rq_map;
1147 	else if (i != hctx->queue_depth) {
1148 		hctx->queue_depth = i;
1149 		pr_warn("%s: queue depth set to %u because of low memory\n",
1150 					__func__, i);
1151 	}
1152 
1153 	hctx->tags = blk_mq_init_tags(hctx->queue_depth, reserved_tags, node);
1154 	if (!hctx->tags) {
1155 err_rq_map:
1156 		blk_mq_free_rq_map(hctx);
1157 		return -ENOMEM;
1158 	}
1159 
1160 	return 0;
1161 }
1162 
1163 static int blk_mq_init_hw_queues(struct request_queue *q,
1164 				 struct blk_mq_reg *reg, void *driver_data)
1165 {
1166 	struct blk_mq_hw_ctx *hctx;
1167 	unsigned int i, j;
1168 
1169 	/*
1170 	 * Initialize hardware queues
1171 	 */
1172 	queue_for_each_hw_ctx(q, hctx, i) {
1173 		unsigned int num_maps;
1174 		int node;
1175 
1176 		node = hctx->numa_node;
1177 		if (node == NUMA_NO_NODE)
1178 			node = hctx->numa_node = reg->numa_node;
1179 
1180 		INIT_DELAYED_WORK(&hctx->delayed_work, blk_mq_work_fn);
1181 		spin_lock_init(&hctx->lock);
1182 		INIT_LIST_HEAD(&hctx->dispatch);
1183 		hctx->queue = q;
1184 		hctx->queue_num = i;
1185 		hctx->flags = reg->flags;
1186 		hctx->queue_depth = reg->queue_depth;
1187 		hctx->cmd_size = reg->cmd_size;
1188 
1189 		blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1190 						blk_mq_hctx_notify, hctx);
1191 		blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1192 
1193 		if (blk_mq_init_rq_map(hctx, reg->reserved_tags, node))
1194 			break;
1195 
1196 		/*
1197 		 * Allocate space for all possible cpus to avoid allocation in
1198 		 * runtime
1199 		 */
1200 		hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1201 						GFP_KERNEL, node);
1202 		if (!hctx->ctxs)
1203 			break;
1204 
1205 		num_maps = ALIGN(nr_cpu_ids, BITS_PER_LONG) / BITS_PER_LONG;
1206 		hctx->ctx_map = kzalloc_node(num_maps * sizeof(unsigned long),
1207 						GFP_KERNEL, node);
1208 		if (!hctx->ctx_map)
1209 			break;
1210 
1211 		hctx->nr_ctx_map = num_maps;
1212 		hctx->nr_ctx = 0;
1213 
1214 		if (reg->ops->init_hctx &&
1215 		    reg->ops->init_hctx(hctx, driver_data, i))
1216 			break;
1217 	}
1218 
1219 	if (i == q->nr_hw_queues)
1220 		return 0;
1221 
1222 	/*
1223 	 * Init failed
1224 	 */
1225 	queue_for_each_hw_ctx(q, hctx, j) {
1226 		if (i == j)
1227 			break;
1228 
1229 		if (reg->ops->exit_hctx)
1230 			reg->ops->exit_hctx(hctx, j);
1231 
1232 		blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1233 		blk_mq_free_rq_map(hctx);
1234 		kfree(hctx->ctxs);
1235 	}
1236 
1237 	return 1;
1238 }
1239 
1240 static void blk_mq_init_cpu_queues(struct request_queue *q,
1241 				   unsigned int nr_hw_queues)
1242 {
1243 	unsigned int i;
1244 
1245 	for_each_possible_cpu(i) {
1246 		struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1247 		struct blk_mq_hw_ctx *hctx;
1248 
1249 		memset(__ctx, 0, sizeof(*__ctx));
1250 		__ctx->cpu = i;
1251 		spin_lock_init(&__ctx->lock);
1252 		INIT_LIST_HEAD(&__ctx->rq_list);
1253 		__ctx->queue = q;
1254 
1255 		/* If the cpu isn't online, the cpu is mapped to first hctx */
1256 		hctx = q->mq_ops->map_queue(q, i);
1257 		hctx->nr_ctx++;
1258 
1259 		if (!cpu_online(i))
1260 			continue;
1261 
1262 		/*
1263 		 * Set local node, IFF we have more than one hw queue. If
1264 		 * not, we remain on the home node of the device
1265 		 */
1266 		if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1267 			hctx->numa_node = cpu_to_node(i);
1268 	}
1269 }
1270 
1271 static void blk_mq_map_swqueue(struct request_queue *q)
1272 {
1273 	unsigned int i;
1274 	struct blk_mq_hw_ctx *hctx;
1275 	struct blk_mq_ctx *ctx;
1276 
1277 	queue_for_each_hw_ctx(q, hctx, i) {
1278 		hctx->nr_ctx = 0;
1279 	}
1280 
1281 	/*
1282 	 * Map software to hardware queues
1283 	 */
1284 	queue_for_each_ctx(q, ctx, i) {
1285 		/* If the cpu isn't online, the cpu is mapped to first hctx */
1286 		hctx = q->mq_ops->map_queue(q, i);
1287 		ctx->index_hw = hctx->nr_ctx;
1288 		hctx->ctxs[hctx->nr_ctx++] = ctx;
1289 	}
1290 }
1291 
1292 struct request_queue *blk_mq_init_queue(struct blk_mq_reg *reg,
1293 					void *driver_data)
1294 {
1295 	struct blk_mq_hw_ctx **hctxs;
1296 	struct blk_mq_ctx *ctx;
1297 	struct request_queue *q;
1298 	int i;
1299 
1300 	if (!reg->nr_hw_queues ||
1301 	    !reg->ops->queue_rq || !reg->ops->map_queue ||
1302 	    !reg->ops->alloc_hctx || !reg->ops->free_hctx)
1303 		return ERR_PTR(-EINVAL);
1304 
1305 	if (!reg->queue_depth)
1306 		reg->queue_depth = BLK_MQ_MAX_DEPTH;
1307 	else if (reg->queue_depth > BLK_MQ_MAX_DEPTH) {
1308 		pr_err("blk-mq: queuedepth too large (%u)\n", reg->queue_depth);
1309 		reg->queue_depth = BLK_MQ_MAX_DEPTH;
1310 	}
1311 
1312 	/*
1313 	 * Set aside a tag for flush requests.  It will only be used while
1314 	 * another flush request is in progress but outside the driver.
1315 	 *
1316 	 * TODO: only allocate if flushes are supported
1317 	 */
1318 	reg->queue_depth++;
1319 	reg->reserved_tags++;
1320 
1321 	if (reg->queue_depth < (reg->reserved_tags + BLK_MQ_TAG_MIN))
1322 		return ERR_PTR(-EINVAL);
1323 
1324 	ctx = alloc_percpu(struct blk_mq_ctx);
1325 	if (!ctx)
1326 		return ERR_PTR(-ENOMEM);
1327 
1328 	hctxs = kmalloc_node(reg->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1329 			reg->numa_node);
1330 
1331 	if (!hctxs)
1332 		goto err_percpu;
1333 
1334 	for (i = 0; i < reg->nr_hw_queues; i++) {
1335 		hctxs[i] = reg->ops->alloc_hctx(reg, i);
1336 		if (!hctxs[i])
1337 			goto err_hctxs;
1338 
1339 		hctxs[i]->numa_node = NUMA_NO_NODE;
1340 		hctxs[i]->queue_num = i;
1341 	}
1342 
1343 	q = blk_alloc_queue_node(GFP_KERNEL, reg->numa_node);
1344 	if (!q)
1345 		goto err_hctxs;
1346 
1347 	q->mq_map = blk_mq_make_queue_map(reg);
1348 	if (!q->mq_map)
1349 		goto err_map;
1350 
1351 	setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1352 	blk_queue_rq_timeout(q, 30000);
1353 
1354 	q->nr_queues = nr_cpu_ids;
1355 	q->nr_hw_queues = reg->nr_hw_queues;
1356 
1357 	q->queue_ctx = ctx;
1358 	q->queue_hw_ctx = hctxs;
1359 
1360 	q->mq_ops = reg->ops;
1361 	q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1362 
1363 	blk_queue_make_request(q, blk_mq_make_request);
1364 	blk_queue_rq_timed_out(q, reg->ops->timeout);
1365 	if (reg->timeout)
1366 		blk_queue_rq_timeout(q, reg->timeout);
1367 
1368 	blk_mq_init_flush(q);
1369 	blk_mq_init_cpu_queues(q, reg->nr_hw_queues);
1370 
1371 	if (blk_mq_init_hw_queues(q, reg, driver_data))
1372 		goto err_hw;
1373 
1374 	blk_mq_map_swqueue(q);
1375 
1376 	mutex_lock(&all_q_mutex);
1377 	list_add_tail(&q->all_q_node, &all_q_list);
1378 	mutex_unlock(&all_q_mutex);
1379 
1380 	return q;
1381 err_hw:
1382 	kfree(q->mq_map);
1383 err_map:
1384 	blk_cleanup_queue(q);
1385 err_hctxs:
1386 	for (i = 0; i < reg->nr_hw_queues; i++) {
1387 		if (!hctxs[i])
1388 			break;
1389 		reg->ops->free_hctx(hctxs[i], i);
1390 	}
1391 	kfree(hctxs);
1392 err_percpu:
1393 	free_percpu(ctx);
1394 	return ERR_PTR(-ENOMEM);
1395 }
1396 EXPORT_SYMBOL(blk_mq_init_queue);
1397 
1398 void blk_mq_free_queue(struct request_queue *q)
1399 {
1400 	struct blk_mq_hw_ctx *hctx;
1401 	int i;
1402 
1403 	queue_for_each_hw_ctx(q, hctx, i) {
1404 		kfree(hctx->ctx_map);
1405 		kfree(hctx->ctxs);
1406 		blk_mq_free_rq_map(hctx);
1407 		blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1408 		if (q->mq_ops->exit_hctx)
1409 			q->mq_ops->exit_hctx(hctx, i);
1410 		q->mq_ops->free_hctx(hctx, i);
1411 	}
1412 
1413 	free_percpu(q->queue_ctx);
1414 	kfree(q->queue_hw_ctx);
1415 	kfree(q->mq_map);
1416 
1417 	q->queue_ctx = NULL;
1418 	q->queue_hw_ctx = NULL;
1419 	q->mq_map = NULL;
1420 
1421 	mutex_lock(&all_q_mutex);
1422 	list_del_init(&q->all_q_node);
1423 	mutex_unlock(&all_q_mutex);
1424 }
1425 
1426 /* Basically redo blk_mq_init_queue with queue frozen */
1427 static void blk_mq_queue_reinit(struct request_queue *q)
1428 {
1429 	blk_mq_freeze_queue(q);
1430 
1431 	blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1432 
1433 	/*
1434 	 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1435 	 * we should change hctx numa_node according to new topology (this
1436 	 * involves free and re-allocate memory, worthy doing?)
1437 	 */
1438 
1439 	blk_mq_map_swqueue(q);
1440 
1441 	blk_mq_unfreeze_queue(q);
1442 }
1443 
1444 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1445 				      unsigned long action, void *hcpu)
1446 {
1447 	struct request_queue *q;
1448 
1449 	/*
1450 	 * Before new mapping is established, hotadded cpu might already start
1451 	 * handling requests. This doesn't break anything as we map offline
1452 	 * CPUs to first hardware queue. We will re-init queue below to get
1453 	 * optimal settings.
1454 	 */
1455 	if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1456 	    action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1457 		return NOTIFY_OK;
1458 
1459 	mutex_lock(&all_q_mutex);
1460 	list_for_each_entry(q, &all_q_list, all_q_node)
1461 		blk_mq_queue_reinit(q);
1462 	mutex_unlock(&all_q_mutex);
1463 	return NOTIFY_OK;
1464 }
1465 
1466 static int __init blk_mq_init(void)
1467 {
1468 	blk_mq_cpu_init();
1469 
1470 	/* Must be called after percpu_counter_hotcpu_callback() */
1471 	hotcpu_notifier(blk_mq_queue_reinit_notify, -10);
1472 
1473 	return 0;
1474 }
1475 subsys_initcall(blk_mq_init);
1476