xref: /openbmc/linux/block/blk-mq.c (revision b34e08d5)
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_request(struct blk_mq_hw_ctx *hctx,
77 					      gfp_t gfp, 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_pinned(struct request_queue *q,
197 						   int rw, gfp_t gfp,
198 						   bool reserved)
199 {
200 	struct request *rq;
201 
202 	do {
203 		struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
204 		struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q, ctx->cpu);
205 
206 		rq = __blk_mq_alloc_request(hctx, gfp & ~__GFP_WAIT, reserved);
207 		if (rq) {
208 			blk_mq_rq_ctx_init(q, ctx, rq, rw);
209 			break;
210 		}
211 
212 		blk_mq_put_ctx(ctx);
213 		if (!(gfp & __GFP_WAIT))
214 			break;
215 
216 		__blk_mq_run_hw_queue(hctx);
217 		blk_mq_wait_for_tags(hctx->tags);
218 	} while (1);
219 
220 	return rq;
221 }
222 
223 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp)
224 {
225 	struct request *rq;
226 
227 	if (blk_mq_queue_enter(q))
228 		return NULL;
229 
230 	rq = blk_mq_alloc_request_pinned(q, rw, gfp, false);
231 	if (rq)
232 		blk_mq_put_ctx(rq->mq_ctx);
233 	return rq;
234 }
235 
236 struct request *blk_mq_alloc_reserved_request(struct request_queue *q, int rw,
237 					      gfp_t gfp)
238 {
239 	struct request *rq;
240 
241 	if (blk_mq_queue_enter(q))
242 		return NULL;
243 
244 	rq = blk_mq_alloc_request_pinned(q, rw, gfp, true);
245 	if (rq)
246 		blk_mq_put_ctx(rq->mq_ctx);
247 	return rq;
248 }
249 EXPORT_SYMBOL(blk_mq_alloc_reserved_request);
250 
251 /*
252  * Re-init and set pdu, if we have it
253  */
254 void blk_mq_rq_init(struct blk_mq_hw_ctx *hctx, struct request *rq)
255 {
256 	blk_rq_init(hctx->queue, rq);
257 
258 	if (hctx->cmd_size)
259 		rq->special = blk_mq_rq_to_pdu(rq);
260 }
261 
262 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
263 				  struct blk_mq_ctx *ctx, struct request *rq)
264 {
265 	const int tag = rq->tag;
266 	struct request_queue *q = rq->q;
267 
268 	blk_mq_rq_init(hctx, rq);
269 	blk_mq_put_tag(hctx->tags, tag);
270 
271 	blk_mq_queue_exit(q);
272 }
273 
274 void blk_mq_free_request(struct request *rq)
275 {
276 	struct blk_mq_ctx *ctx = rq->mq_ctx;
277 	struct blk_mq_hw_ctx *hctx;
278 	struct request_queue *q = rq->q;
279 
280 	ctx->rq_completed[rq_is_sync(rq)]++;
281 
282 	hctx = q->mq_ops->map_queue(q, ctx->cpu);
283 	__blk_mq_free_request(hctx, ctx, rq);
284 }
285 
286 bool blk_mq_end_io_partial(struct request *rq, int error, unsigned int nr_bytes)
287 {
288 	if (blk_update_request(rq, error, blk_rq_bytes(rq)))
289 		return true;
290 
291 	blk_account_io_done(rq);
292 
293 	if (rq->end_io)
294 		rq->end_io(rq, error);
295 	else
296 		blk_mq_free_request(rq);
297 	return false;
298 }
299 EXPORT_SYMBOL(blk_mq_end_io_partial);
300 
301 static void __blk_mq_complete_request_remote(void *data)
302 {
303 	struct request *rq = data;
304 
305 	rq->q->softirq_done_fn(rq);
306 }
307 
308 void __blk_mq_complete_request(struct request *rq)
309 {
310 	struct blk_mq_ctx *ctx = rq->mq_ctx;
311 	int cpu;
312 
313 	if (!ctx->ipi_redirect) {
314 		rq->q->softirq_done_fn(rq);
315 		return;
316 	}
317 
318 	cpu = get_cpu();
319 	if (cpu != ctx->cpu && cpu_online(ctx->cpu)) {
320 		rq->csd.func = __blk_mq_complete_request_remote;
321 		rq->csd.info = rq;
322 		rq->csd.flags = 0;
323 		smp_call_function_single_async(ctx->cpu, &rq->csd);
324 	} else {
325 		rq->q->softirq_done_fn(rq);
326 	}
327 	put_cpu();
328 }
329 
330 /**
331  * blk_mq_complete_request - end I/O on a request
332  * @rq:		the request being processed
333  *
334  * Description:
335  *	Ends all I/O on a request. It does not handle partial completions.
336  *	The actual completion happens out-of-order, through a IPI handler.
337  **/
338 void blk_mq_complete_request(struct request *rq)
339 {
340 	if (unlikely(blk_should_fake_timeout(rq->q)))
341 		return;
342 	if (!blk_mark_rq_complete(rq))
343 		__blk_mq_complete_request(rq);
344 }
345 EXPORT_SYMBOL(blk_mq_complete_request);
346 
347 static void blk_mq_start_request(struct request *rq, bool last)
348 {
349 	struct request_queue *q = rq->q;
350 
351 	trace_block_rq_issue(q, rq);
352 
353 	/*
354 	 * Just mark start time and set the started bit. Due to memory
355 	 * ordering, we know we'll see the correct deadline as long as
356 	 * REQ_ATOMIC_STARTED is seen.
357 	 */
358 	rq->deadline = jiffies + q->rq_timeout;
359 	set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
360 
361 	if (q->dma_drain_size && blk_rq_bytes(rq)) {
362 		/*
363 		 * Make sure space for the drain appears.  We know we can do
364 		 * this because max_hw_segments has been adjusted to be one
365 		 * fewer than the device can handle.
366 		 */
367 		rq->nr_phys_segments++;
368 	}
369 
370 	/*
371 	 * Flag the last request in the series so that drivers know when IO
372 	 * should be kicked off, if they don't do it on a per-request basis.
373 	 *
374 	 * Note: the flag isn't the only condition drivers should do kick off.
375 	 * If drive is busy, the last request might not have the bit set.
376 	 */
377 	if (last)
378 		rq->cmd_flags |= REQ_END;
379 }
380 
381 static void blk_mq_requeue_request(struct request *rq)
382 {
383 	struct request_queue *q = rq->q;
384 
385 	trace_block_rq_requeue(q, rq);
386 	clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
387 
388 	rq->cmd_flags &= ~REQ_END;
389 
390 	if (q->dma_drain_size && blk_rq_bytes(rq))
391 		rq->nr_phys_segments--;
392 }
393 
394 struct blk_mq_timeout_data {
395 	struct blk_mq_hw_ctx *hctx;
396 	unsigned long *next;
397 	unsigned int *next_set;
398 };
399 
400 static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
401 {
402 	struct blk_mq_timeout_data *data = __data;
403 	struct blk_mq_hw_ctx *hctx = data->hctx;
404 	unsigned int tag;
405 
406 	 /* It may not be in flight yet (this is where
407 	 * the REQ_ATOMIC_STARTED flag comes in). The requests are
408 	 * statically allocated, so we know it's always safe to access the
409 	 * memory associated with a bit offset into ->rqs[].
410 	 */
411 	tag = 0;
412 	do {
413 		struct request *rq;
414 
415 		tag = find_next_zero_bit(free_tags, hctx->queue_depth, tag);
416 		if (tag >= hctx->queue_depth)
417 			break;
418 
419 		rq = hctx->rqs[tag++];
420 
421 		if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
422 			continue;
423 
424 		blk_rq_check_expired(rq, data->next, data->next_set);
425 	} while (1);
426 }
427 
428 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
429 					unsigned long *next,
430 					unsigned int *next_set)
431 {
432 	struct blk_mq_timeout_data data = {
433 		.hctx		= hctx,
434 		.next		= next,
435 		.next_set	= next_set,
436 	};
437 
438 	/*
439 	 * Ask the tagging code to iterate busy requests, so we can
440 	 * check them for timeout.
441 	 */
442 	blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
443 }
444 
445 static void blk_mq_rq_timer(unsigned long data)
446 {
447 	struct request_queue *q = (struct request_queue *) data;
448 	struct blk_mq_hw_ctx *hctx;
449 	unsigned long next = 0;
450 	int i, next_set = 0;
451 
452 	queue_for_each_hw_ctx(q, hctx, i)
453 		blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
454 
455 	if (next_set)
456 		mod_timer(&q->timeout, round_jiffies_up(next));
457 }
458 
459 /*
460  * Reverse check our software queue for entries that we could potentially
461  * merge with. Currently includes a hand-wavy stop count of 8, to not spend
462  * too much time checking for merges.
463  */
464 static bool blk_mq_attempt_merge(struct request_queue *q,
465 				 struct blk_mq_ctx *ctx, struct bio *bio)
466 {
467 	struct request *rq;
468 	int checked = 8;
469 
470 	list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
471 		int el_ret;
472 
473 		if (!checked--)
474 			break;
475 
476 		if (!blk_rq_merge_ok(rq, bio))
477 			continue;
478 
479 		el_ret = blk_try_merge(rq, bio);
480 		if (el_ret == ELEVATOR_BACK_MERGE) {
481 			if (bio_attempt_back_merge(q, rq, bio)) {
482 				ctx->rq_merged++;
483 				return true;
484 			}
485 			break;
486 		} else if (el_ret == ELEVATOR_FRONT_MERGE) {
487 			if (bio_attempt_front_merge(q, rq, bio)) {
488 				ctx->rq_merged++;
489 				return true;
490 			}
491 			break;
492 		}
493 	}
494 
495 	return false;
496 }
497 
498 void blk_mq_add_timer(struct request *rq)
499 {
500 	__blk_add_timer(rq, NULL);
501 }
502 
503 /*
504  * Run this hardware queue, pulling any software queues mapped to it in.
505  * Note that this function currently has various problems around ordering
506  * of IO. In particular, we'd like FIFO behaviour on handling existing
507  * items on the hctx->dispatch list. Ignore that for now.
508  */
509 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
510 {
511 	struct request_queue *q = hctx->queue;
512 	struct blk_mq_ctx *ctx;
513 	struct request *rq;
514 	LIST_HEAD(rq_list);
515 	int bit, queued;
516 
517 	if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
518 		return;
519 
520 	hctx->run++;
521 
522 	/*
523 	 * Touch any software queue that has pending entries.
524 	 */
525 	for_each_set_bit(bit, hctx->ctx_map, hctx->nr_ctx) {
526 		clear_bit(bit, hctx->ctx_map);
527 		ctx = hctx->ctxs[bit];
528 		BUG_ON(bit != ctx->index_hw);
529 
530 		spin_lock(&ctx->lock);
531 		list_splice_tail_init(&ctx->rq_list, &rq_list);
532 		spin_unlock(&ctx->lock);
533 	}
534 
535 	/*
536 	 * If we have previous entries on our dispatch list, grab them
537 	 * and stuff them at the front for more fair dispatch.
538 	 */
539 	if (!list_empty_careful(&hctx->dispatch)) {
540 		spin_lock(&hctx->lock);
541 		if (!list_empty(&hctx->dispatch))
542 			list_splice_init(&hctx->dispatch, &rq_list);
543 		spin_unlock(&hctx->lock);
544 	}
545 
546 	/*
547 	 * Delete and return all entries from our dispatch list
548 	 */
549 	queued = 0;
550 
551 	/*
552 	 * Now process all the entries, sending them to the driver.
553 	 */
554 	while (!list_empty(&rq_list)) {
555 		int ret;
556 
557 		rq = list_first_entry(&rq_list, struct request, queuelist);
558 		list_del_init(&rq->queuelist);
559 
560 		blk_mq_start_request(rq, list_empty(&rq_list));
561 
562 		ret = q->mq_ops->queue_rq(hctx, rq);
563 		switch (ret) {
564 		case BLK_MQ_RQ_QUEUE_OK:
565 			queued++;
566 			continue;
567 		case BLK_MQ_RQ_QUEUE_BUSY:
568 			/*
569 			 * FIXME: we should have a mechanism to stop the queue
570 			 * like blk_stop_queue, otherwise we will waste cpu
571 			 * time
572 			 */
573 			list_add(&rq->queuelist, &rq_list);
574 			blk_mq_requeue_request(rq);
575 			break;
576 		default:
577 			pr_err("blk-mq: bad return on queue: %d\n", ret);
578 		case BLK_MQ_RQ_QUEUE_ERROR:
579 			rq->errors = -EIO;
580 			blk_mq_end_io(rq, rq->errors);
581 			break;
582 		}
583 
584 		if (ret == BLK_MQ_RQ_QUEUE_BUSY)
585 			break;
586 	}
587 
588 	if (!queued)
589 		hctx->dispatched[0]++;
590 	else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
591 		hctx->dispatched[ilog2(queued) + 1]++;
592 
593 	/*
594 	 * Any items that need requeuing? Stuff them into hctx->dispatch,
595 	 * that is where we will continue on next queue run.
596 	 */
597 	if (!list_empty(&rq_list)) {
598 		spin_lock(&hctx->lock);
599 		list_splice(&rq_list, &hctx->dispatch);
600 		spin_unlock(&hctx->lock);
601 	}
602 }
603 
604 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
605 {
606 	if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
607 		return;
608 
609 	if (!async)
610 		__blk_mq_run_hw_queue(hctx);
611 	else {
612 		struct request_queue *q = hctx->queue;
613 
614 		kblockd_schedule_delayed_work(q, &hctx->delayed_work, 0);
615 	}
616 }
617 
618 void blk_mq_run_queues(struct request_queue *q, bool async)
619 {
620 	struct blk_mq_hw_ctx *hctx;
621 	int i;
622 
623 	queue_for_each_hw_ctx(q, hctx, i) {
624 		if ((!blk_mq_hctx_has_pending(hctx) &&
625 		    list_empty_careful(&hctx->dispatch)) ||
626 		    test_bit(BLK_MQ_S_STOPPED, &hctx->state))
627 			continue;
628 
629 		blk_mq_run_hw_queue(hctx, async);
630 	}
631 }
632 EXPORT_SYMBOL(blk_mq_run_queues);
633 
634 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
635 {
636 	cancel_delayed_work(&hctx->delayed_work);
637 	set_bit(BLK_MQ_S_STOPPED, &hctx->state);
638 }
639 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
640 
641 void blk_mq_stop_hw_queues(struct request_queue *q)
642 {
643 	struct blk_mq_hw_ctx *hctx;
644 	int i;
645 
646 	queue_for_each_hw_ctx(q, hctx, i)
647 		blk_mq_stop_hw_queue(hctx);
648 }
649 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
650 
651 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
652 {
653 	clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
654 	__blk_mq_run_hw_queue(hctx);
655 }
656 EXPORT_SYMBOL(blk_mq_start_hw_queue);
657 
658 void blk_mq_start_stopped_hw_queues(struct request_queue *q)
659 {
660 	struct blk_mq_hw_ctx *hctx;
661 	int i;
662 
663 	queue_for_each_hw_ctx(q, hctx, i) {
664 		if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
665 			continue;
666 
667 		clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
668 		blk_mq_run_hw_queue(hctx, true);
669 	}
670 }
671 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
672 
673 static void blk_mq_work_fn(struct work_struct *work)
674 {
675 	struct blk_mq_hw_ctx *hctx;
676 
677 	hctx = container_of(work, struct blk_mq_hw_ctx, delayed_work.work);
678 	__blk_mq_run_hw_queue(hctx);
679 }
680 
681 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
682 				    struct request *rq, bool at_head)
683 {
684 	struct blk_mq_ctx *ctx = rq->mq_ctx;
685 
686 	trace_block_rq_insert(hctx->queue, rq);
687 
688 	if (at_head)
689 		list_add(&rq->queuelist, &ctx->rq_list);
690 	else
691 		list_add_tail(&rq->queuelist, &ctx->rq_list);
692 	blk_mq_hctx_mark_pending(hctx, ctx);
693 
694 	/*
695 	 * We do this early, to ensure we are on the right CPU.
696 	 */
697 	blk_mq_add_timer(rq);
698 }
699 
700 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
701 		bool async)
702 {
703 	struct request_queue *q = rq->q;
704 	struct blk_mq_hw_ctx *hctx;
705 	struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
706 
707 	current_ctx = blk_mq_get_ctx(q);
708 	if (!cpu_online(ctx->cpu))
709 		rq->mq_ctx = ctx = current_ctx;
710 
711 	hctx = q->mq_ops->map_queue(q, ctx->cpu);
712 
713 	if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA) &&
714 	    !(rq->cmd_flags & (REQ_FLUSH_SEQ))) {
715 		blk_insert_flush(rq);
716 	} else {
717 		spin_lock(&ctx->lock);
718 		__blk_mq_insert_request(hctx, rq, at_head);
719 		spin_unlock(&ctx->lock);
720 	}
721 
722 	blk_mq_put_ctx(current_ctx);
723 
724 	if (run_queue)
725 		blk_mq_run_hw_queue(hctx, async);
726 }
727 
728 static void blk_mq_insert_requests(struct request_queue *q,
729 				     struct blk_mq_ctx *ctx,
730 				     struct list_head *list,
731 				     int depth,
732 				     bool from_schedule)
733 
734 {
735 	struct blk_mq_hw_ctx *hctx;
736 	struct blk_mq_ctx *current_ctx;
737 
738 	trace_block_unplug(q, depth, !from_schedule);
739 
740 	current_ctx = blk_mq_get_ctx(q);
741 
742 	if (!cpu_online(ctx->cpu))
743 		ctx = current_ctx;
744 	hctx = q->mq_ops->map_queue(q, ctx->cpu);
745 
746 	/*
747 	 * preemption doesn't flush plug list, so it's possible ctx->cpu is
748 	 * offline now
749 	 */
750 	spin_lock(&ctx->lock);
751 	while (!list_empty(list)) {
752 		struct request *rq;
753 
754 		rq = list_first_entry(list, struct request, queuelist);
755 		list_del_init(&rq->queuelist);
756 		rq->mq_ctx = ctx;
757 		__blk_mq_insert_request(hctx, rq, false);
758 	}
759 	spin_unlock(&ctx->lock);
760 
761 	blk_mq_put_ctx(current_ctx);
762 
763 	blk_mq_run_hw_queue(hctx, from_schedule);
764 }
765 
766 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
767 {
768 	struct request *rqa = container_of(a, struct request, queuelist);
769 	struct request *rqb = container_of(b, struct request, queuelist);
770 
771 	return !(rqa->mq_ctx < rqb->mq_ctx ||
772 		 (rqa->mq_ctx == rqb->mq_ctx &&
773 		  blk_rq_pos(rqa) < blk_rq_pos(rqb)));
774 }
775 
776 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
777 {
778 	struct blk_mq_ctx *this_ctx;
779 	struct request_queue *this_q;
780 	struct request *rq;
781 	LIST_HEAD(list);
782 	LIST_HEAD(ctx_list);
783 	unsigned int depth;
784 
785 	list_splice_init(&plug->mq_list, &list);
786 
787 	list_sort(NULL, &list, plug_ctx_cmp);
788 
789 	this_q = NULL;
790 	this_ctx = NULL;
791 	depth = 0;
792 
793 	while (!list_empty(&list)) {
794 		rq = list_entry_rq(list.next);
795 		list_del_init(&rq->queuelist);
796 		BUG_ON(!rq->q);
797 		if (rq->mq_ctx != this_ctx) {
798 			if (this_ctx) {
799 				blk_mq_insert_requests(this_q, this_ctx,
800 							&ctx_list, depth,
801 							from_schedule);
802 			}
803 
804 			this_ctx = rq->mq_ctx;
805 			this_q = rq->q;
806 			depth = 0;
807 		}
808 
809 		depth++;
810 		list_add_tail(&rq->queuelist, &ctx_list);
811 	}
812 
813 	/*
814 	 * If 'this_ctx' is set, we know we have entries to complete
815 	 * on 'ctx_list'. Do those.
816 	 */
817 	if (this_ctx) {
818 		blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
819 				       from_schedule);
820 	}
821 }
822 
823 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
824 {
825 	init_request_from_bio(rq, bio);
826 	blk_account_io_start(rq, 1);
827 }
828 
829 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
830 {
831 	struct blk_mq_hw_ctx *hctx;
832 	struct blk_mq_ctx *ctx;
833 	const int is_sync = rw_is_sync(bio->bi_rw);
834 	const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
835 	int rw = bio_data_dir(bio);
836 	struct request *rq;
837 	unsigned int use_plug, request_count = 0;
838 
839 	/*
840 	 * If we have multiple hardware queues, just go directly to
841 	 * one of those for sync IO.
842 	 */
843 	use_plug = !is_flush_fua && ((q->nr_hw_queues == 1) || !is_sync);
844 
845 	blk_queue_bounce(q, &bio);
846 
847 	if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
848 		bio_endio(bio, -EIO);
849 		return;
850 	}
851 
852 	if (use_plug && blk_attempt_plug_merge(q, bio, &request_count))
853 		return;
854 
855 	if (blk_mq_queue_enter(q)) {
856 		bio_endio(bio, -EIO);
857 		return;
858 	}
859 
860 	ctx = blk_mq_get_ctx(q);
861 	hctx = q->mq_ops->map_queue(q, ctx->cpu);
862 
863 	if (is_sync)
864 		rw |= REQ_SYNC;
865 	trace_block_getrq(q, bio, rw);
866 	rq = __blk_mq_alloc_request(hctx, GFP_ATOMIC, false);
867 	if (likely(rq))
868 		blk_mq_rq_ctx_init(q, ctx, rq, rw);
869 	else {
870 		blk_mq_put_ctx(ctx);
871 		trace_block_sleeprq(q, bio, rw);
872 		rq = blk_mq_alloc_request_pinned(q, rw, __GFP_WAIT|GFP_ATOMIC,
873 							false);
874 		ctx = rq->mq_ctx;
875 		hctx = q->mq_ops->map_queue(q, ctx->cpu);
876 	}
877 
878 	hctx->queued++;
879 
880 	if (unlikely(is_flush_fua)) {
881 		blk_mq_bio_to_request(rq, bio);
882 		blk_mq_put_ctx(ctx);
883 		blk_insert_flush(rq);
884 		goto run_queue;
885 	}
886 
887 	/*
888 	 * A task plug currently exists. Since this is completely lockless,
889 	 * utilize that to temporarily store requests until the task is
890 	 * either done or scheduled away.
891 	 */
892 	if (use_plug) {
893 		struct blk_plug *plug = current->plug;
894 
895 		if (plug) {
896 			blk_mq_bio_to_request(rq, bio);
897 			if (list_empty(&plug->mq_list))
898 				trace_block_plug(q);
899 			else if (request_count >= BLK_MAX_REQUEST_COUNT) {
900 				blk_flush_plug_list(plug, false);
901 				trace_block_plug(q);
902 			}
903 			list_add_tail(&rq->queuelist, &plug->mq_list);
904 			blk_mq_put_ctx(ctx);
905 			return;
906 		}
907 	}
908 
909 	spin_lock(&ctx->lock);
910 
911 	if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
912 	    blk_mq_attempt_merge(q, ctx, bio))
913 		__blk_mq_free_request(hctx, ctx, rq);
914 	else {
915 		blk_mq_bio_to_request(rq, bio);
916 		__blk_mq_insert_request(hctx, rq, false);
917 	}
918 
919 	spin_unlock(&ctx->lock);
920 	blk_mq_put_ctx(ctx);
921 
922 	/*
923 	 * For a SYNC request, send it to the hardware immediately. For an
924 	 * ASYNC request, just ensure that we run it later on. The latter
925 	 * allows for merging opportunities and more efficient dispatching.
926 	 */
927 run_queue:
928 	blk_mq_run_hw_queue(hctx, !is_sync || is_flush_fua);
929 }
930 
931 /*
932  * Default mapping to a software queue, since we use one per CPU.
933  */
934 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
935 {
936 	return q->queue_hw_ctx[q->mq_map[cpu]];
937 }
938 EXPORT_SYMBOL(blk_mq_map_queue);
939 
940 struct blk_mq_hw_ctx *blk_mq_alloc_single_hw_queue(struct blk_mq_reg *reg,
941 						   unsigned int hctx_index)
942 {
943 	return kmalloc_node(sizeof(struct blk_mq_hw_ctx),
944 				GFP_KERNEL | __GFP_ZERO, reg->numa_node);
945 }
946 EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue);
947 
948 void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx *hctx,
949 				 unsigned int hctx_index)
950 {
951 	kfree(hctx);
952 }
953 EXPORT_SYMBOL(blk_mq_free_single_hw_queue);
954 
955 static void blk_mq_hctx_notify(void *data, unsigned long action,
956 			       unsigned int cpu)
957 {
958 	struct blk_mq_hw_ctx *hctx = data;
959 	struct request_queue *q = hctx->queue;
960 	struct blk_mq_ctx *ctx;
961 	LIST_HEAD(tmp);
962 
963 	if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
964 		return;
965 
966 	/*
967 	 * Move ctx entries to new CPU, if this one is going away.
968 	 */
969 	ctx = __blk_mq_get_ctx(q, cpu);
970 
971 	spin_lock(&ctx->lock);
972 	if (!list_empty(&ctx->rq_list)) {
973 		list_splice_init(&ctx->rq_list, &tmp);
974 		clear_bit(ctx->index_hw, hctx->ctx_map);
975 	}
976 	spin_unlock(&ctx->lock);
977 
978 	if (list_empty(&tmp))
979 		return;
980 
981 	ctx = blk_mq_get_ctx(q);
982 	spin_lock(&ctx->lock);
983 
984 	while (!list_empty(&tmp)) {
985 		struct request *rq;
986 
987 		rq = list_first_entry(&tmp, struct request, queuelist);
988 		rq->mq_ctx = ctx;
989 		list_move_tail(&rq->queuelist, &ctx->rq_list);
990 	}
991 
992 	hctx = q->mq_ops->map_queue(q, ctx->cpu);
993 	blk_mq_hctx_mark_pending(hctx, ctx);
994 
995 	spin_unlock(&ctx->lock);
996 	blk_mq_put_ctx(ctx);
997 
998 	blk_mq_run_hw_queue(hctx, true);
999 }
1000 
1001 static int blk_mq_init_hw_commands(struct blk_mq_hw_ctx *hctx,
1002 				   int (*init)(void *, struct blk_mq_hw_ctx *,
1003 					struct request *, unsigned int),
1004 				   void *data)
1005 {
1006 	unsigned int i;
1007 	int ret = 0;
1008 
1009 	for (i = 0; i < hctx->queue_depth; i++) {
1010 		struct request *rq = hctx->rqs[i];
1011 
1012 		ret = init(data, hctx, rq, i);
1013 		if (ret)
1014 			break;
1015 	}
1016 
1017 	return ret;
1018 }
1019 
1020 int blk_mq_init_commands(struct request_queue *q,
1021 			 int (*init)(void *, struct blk_mq_hw_ctx *,
1022 					struct request *, unsigned int),
1023 			 void *data)
1024 {
1025 	struct blk_mq_hw_ctx *hctx;
1026 	unsigned int i;
1027 	int ret = 0;
1028 
1029 	queue_for_each_hw_ctx(q, hctx, i) {
1030 		ret = blk_mq_init_hw_commands(hctx, init, data);
1031 		if (ret)
1032 			break;
1033 	}
1034 
1035 	return ret;
1036 }
1037 EXPORT_SYMBOL(blk_mq_init_commands);
1038 
1039 static void blk_mq_free_hw_commands(struct blk_mq_hw_ctx *hctx,
1040 				    void (*free)(void *, struct blk_mq_hw_ctx *,
1041 					struct request *, unsigned int),
1042 				    void *data)
1043 {
1044 	unsigned int i;
1045 
1046 	for (i = 0; i < hctx->queue_depth; i++) {
1047 		struct request *rq = hctx->rqs[i];
1048 
1049 		free(data, hctx, rq, i);
1050 	}
1051 }
1052 
1053 void blk_mq_free_commands(struct request_queue *q,
1054 			  void (*free)(void *, struct blk_mq_hw_ctx *,
1055 					struct request *, unsigned int),
1056 			  void *data)
1057 {
1058 	struct blk_mq_hw_ctx *hctx;
1059 	unsigned int i;
1060 
1061 	queue_for_each_hw_ctx(q, hctx, i)
1062 		blk_mq_free_hw_commands(hctx, free, data);
1063 }
1064 EXPORT_SYMBOL(blk_mq_free_commands);
1065 
1066 static void blk_mq_free_rq_map(struct blk_mq_hw_ctx *hctx)
1067 {
1068 	struct page *page;
1069 
1070 	while (!list_empty(&hctx->page_list)) {
1071 		page = list_first_entry(&hctx->page_list, struct page, lru);
1072 		list_del_init(&page->lru);
1073 		__free_pages(page, page->private);
1074 	}
1075 
1076 	kfree(hctx->rqs);
1077 
1078 	if (hctx->tags)
1079 		blk_mq_free_tags(hctx->tags);
1080 }
1081 
1082 static size_t order_to_size(unsigned int order)
1083 {
1084 	size_t ret = PAGE_SIZE;
1085 
1086 	while (order--)
1087 		ret *= 2;
1088 
1089 	return ret;
1090 }
1091 
1092 static int blk_mq_init_rq_map(struct blk_mq_hw_ctx *hctx,
1093 			      unsigned int reserved_tags, int node)
1094 {
1095 	unsigned int i, j, entries_per_page, max_order = 4;
1096 	size_t rq_size, left;
1097 
1098 	INIT_LIST_HEAD(&hctx->page_list);
1099 
1100 	hctx->rqs = kmalloc_node(hctx->queue_depth * sizeof(struct request *),
1101 					GFP_KERNEL, node);
1102 	if (!hctx->rqs)
1103 		return -ENOMEM;
1104 
1105 	/*
1106 	 * rq_size is the size of the request plus driver payload, rounded
1107 	 * to the cacheline size
1108 	 */
1109 	rq_size = round_up(sizeof(struct request) + hctx->cmd_size,
1110 				cache_line_size());
1111 	left = rq_size * hctx->queue_depth;
1112 
1113 	for (i = 0; i < hctx->queue_depth;) {
1114 		int this_order = max_order;
1115 		struct page *page;
1116 		int to_do;
1117 		void *p;
1118 
1119 		while (left < order_to_size(this_order - 1) && this_order)
1120 			this_order--;
1121 
1122 		do {
1123 			page = alloc_pages_node(node, GFP_KERNEL, this_order);
1124 			if (page)
1125 				break;
1126 			if (!this_order--)
1127 				break;
1128 			if (order_to_size(this_order) < rq_size)
1129 				break;
1130 		} while (1);
1131 
1132 		if (!page)
1133 			break;
1134 
1135 		page->private = this_order;
1136 		list_add_tail(&page->lru, &hctx->page_list);
1137 
1138 		p = page_address(page);
1139 		entries_per_page = order_to_size(this_order) / rq_size;
1140 		to_do = min(entries_per_page, hctx->queue_depth - i);
1141 		left -= to_do * rq_size;
1142 		for (j = 0; j < to_do; j++) {
1143 			hctx->rqs[i] = p;
1144 			blk_mq_rq_init(hctx, hctx->rqs[i]);
1145 			p += rq_size;
1146 			i++;
1147 		}
1148 	}
1149 
1150 	if (i < (reserved_tags + BLK_MQ_TAG_MIN))
1151 		goto err_rq_map;
1152 	else if (i != hctx->queue_depth) {
1153 		hctx->queue_depth = i;
1154 		pr_warn("%s: queue depth set to %u because of low memory\n",
1155 					__func__, i);
1156 	}
1157 
1158 	hctx->tags = blk_mq_init_tags(hctx->queue_depth, reserved_tags, node);
1159 	if (!hctx->tags) {
1160 err_rq_map:
1161 		blk_mq_free_rq_map(hctx);
1162 		return -ENOMEM;
1163 	}
1164 
1165 	return 0;
1166 }
1167 
1168 static int blk_mq_init_hw_queues(struct request_queue *q,
1169 				 struct blk_mq_reg *reg, void *driver_data)
1170 {
1171 	struct blk_mq_hw_ctx *hctx;
1172 	unsigned int i, j;
1173 
1174 	/*
1175 	 * Initialize hardware queues
1176 	 */
1177 	queue_for_each_hw_ctx(q, hctx, i) {
1178 		unsigned int num_maps;
1179 		int node;
1180 
1181 		node = hctx->numa_node;
1182 		if (node == NUMA_NO_NODE)
1183 			node = hctx->numa_node = reg->numa_node;
1184 
1185 		INIT_DELAYED_WORK(&hctx->delayed_work, blk_mq_work_fn);
1186 		spin_lock_init(&hctx->lock);
1187 		INIT_LIST_HEAD(&hctx->dispatch);
1188 		hctx->queue = q;
1189 		hctx->queue_num = i;
1190 		hctx->flags = reg->flags;
1191 		hctx->queue_depth = reg->queue_depth;
1192 		hctx->cmd_size = reg->cmd_size;
1193 
1194 		blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1195 						blk_mq_hctx_notify, hctx);
1196 		blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1197 
1198 		if (blk_mq_init_rq_map(hctx, reg->reserved_tags, node))
1199 			break;
1200 
1201 		/*
1202 		 * Allocate space for all possible cpus to avoid allocation in
1203 		 * runtime
1204 		 */
1205 		hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1206 						GFP_KERNEL, node);
1207 		if (!hctx->ctxs)
1208 			break;
1209 
1210 		num_maps = ALIGN(nr_cpu_ids, BITS_PER_LONG) / BITS_PER_LONG;
1211 		hctx->ctx_map = kzalloc_node(num_maps * sizeof(unsigned long),
1212 						GFP_KERNEL, node);
1213 		if (!hctx->ctx_map)
1214 			break;
1215 
1216 		hctx->nr_ctx_map = num_maps;
1217 		hctx->nr_ctx = 0;
1218 
1219 		if (reg->ops->init_hctx &&
1220 		    reg->ops->init_hctx(hctx, driver_data, i))
1221 			break;
1222 	}
1223 
1224 	if (i == q->nr_hw_queues)
1225 		return 0;
1226 
1227 	/*
1228 	 * Init failed
1229 	 */
1230 	queue_for_each_hw_ctx(q, hctx, j) {
1231 		if (i == j)
1232 			break;
1233 
1234 		if (reg->ops->exit_hctx)
1235 			reg->ops->exit_hctx(hctx, j);
1236 
1237 		blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1238 		blk_mq_free_rq_map(hctx);
1239 		kfree(hctx->ctxs);
1240 	}
1241 
1242 	return 1;
1243 }
1244 
1245 static void blk_mq_init_cpu_queues(struct request_queue *q,
1246 				   unsigned int nr_hw_queues)
1247 {
1248 	unsigned int i;
1249 
1250 	for_each_possible_cpu(i) {
1251 		struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1252 		struct blk_mq_hw_ctx *hctx;
1253 
1254 		memset(__ctx, 0, sizeof(*__ctx));
1255 		__ctx->cpu = i;
1256 		spin_lock_init(&__ctx->lock);
1257 		INIT_LIST_HEAD(&__ctx->rq_list);
1258 		__ctx->queue = q;
1259 
1260 		/* If the cpu isn't online, the cpu is mapped to first hctx */
1261 		hctx = q->mq_ops->map_queue(q, i);
1262 		hctx->nr_ctx++;
1263 
1264 		if (!cpu_online(i))
1265 			continue;
1266 
1267 		/*
1268 		 * Set local node, IFF we have more than one hw queue. If
1269 		 * not, we remain on the home node of the device
1270 		 */
1271 		if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1272 			hctx->numa_node = cpu_to_node(i);
1273 	}
1274 }
1275 
1276 static void blk_mq_map_swqueue(struct request_queue *q)
1277 {
1278 	unsigned int i;
1279 	struct blk_mq_hw_ctx *hctx;
1280 	struct blk_mq_ctx *ctx;
1281 
1282 	queue_for_each_hw_ctx(q, hctx, i) {
1283 		hctx->nr_ctx = 0;
1284 	}
1285 
1286 	/*
1287 	 * Map software to hardware queues
1288 	 */
1289 	queue_for_each_ctx(q, ctx, i) {
1290 		/* If the cpu isn't online, the cpu is mapped to first hctx */
1291 		hctx = q->mq_ops->map_queue(q, i);
1292 		ctx->index_hw = hctx->nr_ctx;
1293 		hctx->ctxs[hctx->nr_ctx++] = ctx;
1294 	}
1295 }
1296 
1297 struct request_queue *blk_mq_init_queue(struct blk_mq_reg *reg,
1298 					void *driver_data)
1299 {
1300 	struct blk_mq_hw_ctx **hctxs;
1301 	struct blk_mq_ctx *ctx;
1302 	struct request_queue *q;
1303 	int i;
1304 
1305 	if (!reg->nr_hw_queues ||
1306 	    !reg->ops->queue_rq || !reg->ops->map_queue ||
1307 	    !reg->ops->alloc_hctx || !reg->ops->free_hctx)
1308 		return ERR_PTR(-EINVAL);
1309 
1310 	if (!reg->queue_depth)
1311 		reg->queue_depth = BLK_MQ_MAX_DEPTH;
1312 	else if (reg->queue_depth > BLK_MQ_MAX_DEPTH) {
1313 		pr_err("blk-mq: queuedepth too large (%u)\n", reg->queue_depth);
1314 		reg->queue_depth = BLK_MQ_MAX_DEPTH;
1315 	}
1316 
1317 	if (reg->queue_depth < (reg->reserved_tags + BLK_MQ_TAG_MIN))
1318 		return ERR_PTR(-EINVAL);
1319 
1320 	ctx = alloc_percpu(struct blk_mq_ctx);
1321 	if (!ctx)
1322 		return ERR_PTR(-ENOMEM);
1323 
1324 	hctxs = kmalloc_node(reg->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1325 			reg->numa_node);
1326 
1327 	if (!hctxs)
1328 		goto err_percpu;
1329 
1330 	for (i = 0; i < reg->nr_hw_queues; i++) {
1331 		hctxs[i] = reg->ops->alloc_hctx(reg, i);
1332 		if (!hctxs[i])
1333 			goto err_hctxs;
1334 
1335 		hctxs[i]->numa_node = NUMA_NO_NODE;
1336 		hctxs[i]->queue_num = i;
1337 	}
1338 
1339 	q = blk_alloc_queue_node(GFP_KERNEL, reg->numa_node);
1340 	if (!q)
1341 		goto err_hctxs;
1342 
1343 	q->mq_map = blk_mq_make_queue_map(reg);
1344 	if (!q->mq_map)
1345 		goto err_map;
1346 
1347 	setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1348 	blk_queue_rq_timeout(q, 30000);
1349 
1350 	q->nr_queues = nr_cpu_ids;
1351 	q->nr_hw_queues = reg->nr_hw_queues;
1352 
1353 	q->queue_ctx = ctx;
1354 	q->queue_hw_ctx = hctxs;
1355 
1356 	q->mq_ops = reg->ops;
1357 	q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1358 
1359 	q->sg_reserved_size = INT_MAX;
1360 
1361 	blk_queue_make_request(q, blk_mq_make_request);
1362 	blk_queue_rq_timed_out(q, reg->ops->timeout);
1363 	if (reg->timeout)
1364 		blk_queue_rq_timeout(q, reg->timeout);
1365 
1366 	if (reg->ops->complete)
1367 		blk_queue_softirq_done(q, reg->ops->complete);
1368 
1369 	blk_mq_init_flush(q);
1370 	blk_mq_init_cpu_queues(q, reg->nr_hw_queues);
1371 
1372 	q->flush_rq = kzalloc(round_up(sizeof(struct request) + reg->cmd_size,
1373 				cache_line_size()), GFP_KERNEL);
1374 	if (!q->flush_rq)
1375 		goto err_hw;
1376 
1377 	if (blk_mq_init_hw_queues(q, reg, driver_data))
1378 		goto err_flush_rq;
1379 
1380 	blk_mq_map_swqueue(q);
1381 
1382 	mutex_lock(&all_q_mutex);
1383 	list_add_tail(&q->all_q_node, &all_q_list);
1384 	mutex_unlock(&all_q_mutex);
1385 
1386 	return q;
1387 
1388 err_flush_rq:
1389 	kfree(q->flush_rq);
1390 err_hw:
1391 	kfree(q->mq_map);
1392 err_map:
1393 	blk_cleanup_queue(q);
1394 err_hctxs:
1395 	for (i = 0; i < reg->nr_hw_queues; i++) {
1396 		if (!hctxs[i])
1397 			break;
1398 		reg->ops->free_hctx(hctxs[i], i);
1399 	}
1400 	kfree(hctxs);
1401 err_percpu:
1402 	free_percpu(ctx);
1403 	return ERR_PTR(-ENOMEM);
1404 }
1405 EXPORT_SYMBOL(blk_mq_init_queue);
1406 
1407 void blk_mq_free_queue(struct request_queue *q)
1408 {
1409 	struct blk_mq_hw_ctx *hctx;
1410 	int i;
1411 
1412 	queue_for_each_hw_ctx(q, hctx, i) {
1413 		kfree(hctx->ctx_map);
1414 		kfree(hctx->ctxs);
1415 		blk_mq_free_rq_map(hctx);
1416 		blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1417 		if (q->mq_ops->exit_hctx)
1418 			q->mq_ops->exit_hctx(hctx, i);
1419 		q->mq_ops->free_hctx(hctx, i);
1420 	}
1421 
1422 	free_percpu(q->queue_ctx);
1423 	kfree(q->queue_hw_ctx);
1424 	kfree(q->mq_map);
1425 
1426 	q->queue_ctx = NULL;
1427 	q->queue_hw_ctx = NULL;
1428 	q->mq_map = NULL;
1429 
1430 	mutex_lock(&all_q_mutex);
1431 	list_del_init(&q->all_q_node);
1432 	mutex_unlock(&all_q_mutex);
1433 }
1434 
1435 /* Basically redo blk_mq_init_queue with queue frozen */
1436 static void blk_mq_queue_reinit(struct request_queue *q)
1437 {
1438 	blk_mq_freeze_queue(q);
1439 
1440 	blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1441 
1442 	/*
1443 	 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1444 	 * we should change hctx numa_node according to new topology (this
1445 	 * involves free and re-allocate memory, worthy doing?)
1446 	 */
1447 
1448 	blk_mq_map_swqueue(q);
1449 
1450 	blk_mq_unfreeze_queue(q);
1451 }
1452 
1453 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1454 				      unsigned long action, void *hcpu)
1455 {
1456 	struct request_queue *q;
1457 
1458 	/*
1459 	 * Before new mapping is established, hotadded cpu might already start
1460 	 * handling requests. This doesn't break anything as we map offline
1461 	 * CPUs to first hardware queue. We will re-init queue below to get
1462 	 * optimal settings.
1463 	 */
1464 	if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1465 	    action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1466 		return NOTIFY_OK;
1467 
1468 	mutex_lock(&all_q_mutex);
1469 	list_for_each_entry(q, &all_q_list, all_q_node)
1470 		blk_mq_queue_reinit(q);
1471 	mutex_unlock(&all_q_mutex);
1472 	return NOTIFY_OK;
1473 }
1474 
1475 void blk_mq_disable_hotplug(void)
1476 {
1477 	mutex_lock(&all_q_mutex);
1478 }
1479 
1480 void blk_mq_enable_hotplug(void)
1481 {
1482 	mutex_unlock(&all_q_mutex);
1483 }
1484 
1485 static int __init blk_mq_init(void)
1486 {
1487 	blk_mq_cpu_init();
1488 
1489 	/* Must be called after percpu_counter_hotcpu_callback() */
1490 	hotcpu_notifier(blk_mq_queue_reinit_notify, -10);
1491 
1492 	return 0;
1493 }
1494 subsys_initcall(blk_mq_init);
1495