xref: /openbmc/linux/block/blk-mq.c (revision b755c25f)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Block multiqueue core code
4  *
5  * Copyright (C) 2013-2014 Jens Axboe
6  * Copyright (C) 2013-2014 Christoph Hellwig
7  */
8 #include <linux/kernel.h>
9 #include <linux/module.h>
10 #include <linux/backing-dev.h>
11 #include <linux/bio.h>
12 #include <linux/blkdev.h>
13 #include <linux/blk-integrity.h>
14 #include <linux/kmemleak.h>
15 #include <linux/mm.h>
16 #include <linux/init.h>
17 #include <linux/slab.h>
18 #include <linux/workqueue.h>
19 #include <linux/smp.h>
20 #include <linux/interrupt.h>
21 #include <linux/llist.h>
22 #include <linux/cpu.h>
23 #include <linux/cache.h>
24 #include <linux/sched/sysctl.h>
25 #include <linux/sched/topology.h>
26 #include <linux/sched/signal.h>
27 #include <linux/delay.h>
28 #include <linux/crash_dump.h>
29 #include <linux/prefetch.h>
30 #include <linux/blk-crypto.h>
31 #include <linux/part_stat.h>
32 
33 #include <trace/events/block.h>
34 
35 #include <linux/t10-pi.h>
36 #include "blk.h"
37 #include "blk-mq.h"
38 #include "blk-mq-debugfs.h"
39 #include "blk-pm.h"
40 #include "blk-stat.h"
41 #include "blk-mq-sched.h"
42 #include "blk-rq-qos.h"
43 #include "blk-ioprio.h"
44 
45 static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
46 
47 static void blk_mq_insert_request(struct request *rq, blk_insert_t flags);
48 static void blk_mq_request_bypass_insert(struct request *rq,
49 		blk_insert_t flags);
50 static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
51 		struct list_head *list);
52 static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
53 			 struct io_comp_batch *iob, unsigned int flags);
54 
55 /*
56  * Check if any of the ctx, dispatch list or elevator
57  * have pending work in this hardware queue.
58  */
59 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
60 {
61 	return !list_empty_careful(&hctx->dispatch) ||
62 		sbitmap_any_bit_set(&hctx->ctx_map) ||
63 			blk_mq_sched_has_work(hctx);
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 	const int bit = ctx->index_hw[hctx->type];
73 
74 	if (!sbitmap_test_bit(&hctx->ctx_map, bit))
75 		sbitmap_set_bit(&hctx->ctx_map, bit);
76 }
77 
78 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
79 				      struct blk_mq_ctx *ctx)
80 {
81 	const int bit = ctx->index_hw[hctx->type];
82 
83 	sbitmap_clear_bit(&hctx->ctx_map, bit);
84 }
85 
86 struct mq_inflight {
87 	struct block_device *part;
88 	unsigned int inflight[2];
89 };
90 
91 static bool blk_mq_check_inflight(struct request *rq, void *priv)
92 {
93 	struct mq_inflight *mi = priv;
94 
95 	if (rq->part && blk_do_io_stat(rq) &&
96 	    (!mi->part->bd_partno || rq->part == mi->part) &&
97 	    blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
98 		mi->inflight[rq_data_dir(rq)]++;
99 
100 	return true;
101 }
102 
103 unsigned int blk_mq_in_flight(struct request_queue *q,
104 		struct block_device *part)
105 {
106 	struct mq_inflight mi = { .part = part };
107 
108 	blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
109 
110 	return mi.inflight[0] + mi.inflight[1];
111 }
112 
113 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
114 		unsigned int inflight[2])
115 {
116 	struct mq_inflight mi = { .part = part };
117 
118 	blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
119 	inflight[0] = mi.inflight[0];
120 	inflight[1] = mi.inflight[1];
121 }
122 
123 void blk_freeze_queue_start(struct request_queue *q)
124 {
125 	mutex_lock(&q->mq_freeze_lock);
126 	if (++q->mq_freeze_depth == 1) {
127 		percpu_ref_kill(&q->q_usage_counter);
128 		mutex_unlock(&q->mq_freeze_lock);
129 		if (queue_is_mq(q))
130 			blk_mq_run_hw_queues(q, false);
131 	} else {
132 		mutex_unlock(&q->mq_freeze_lock);
133 	}
134 }
135 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
136 
137 void blk_mq_freeze_queue_wait(struct request_queue *q)
138 {
139 	wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
140 }
141 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
142 
143 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
144 				     unsigned long timeout)
145 {
146 	return wait_event_timeout(q->mq_freeze_wq,
147 					percpu_ref_is_zero(&q->q_usage_counter),
148 					timeout);
149 }
150 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
151 
152 /*
153  * Guarantee no request is in use, so we can change any data structure of
154  * the queue afterward.
155  */
156 void blk_freeze_queue(struct request_queue *q)
157 {
158 	/*
159 	 * In the !blk_mq case we are only calling this to kill the
160 	 * q_usage_counter, otherwise this increases the freeze depth
161 	 * and waits for it to return to zero.  For this reason there is
162 	 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
163 	 * exported to drivers as the only user for unfreeze is blk_mq.
164 	 */
165 	blk_freeze_queue_start(q);
166 	blk_mq_freeze_queue_wait(q);
167 }
168 
169 void blk_mq_freeze_queue(struct request_queue *q)
170 {
171 	/*
172 	 * ...just an alias to keep freeze and unfreeze actions balanced
173 	 * in the blk_mq_* namespace
174 	 */
175 	blk_freeze_queue(q);
176 }
177 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
178 
179 void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
180 {
181 	mutex_lock(&q->mq_freeze_lock);
182 	if (force_atomic)
183 		q->q_usage_counter.data->force_atomic = true;
184 	q->mq_freeze_depth--;
185 	WARN_ON_ONCE(q->mq_freeze_depth < 0);
186 	if (!q->mq_freeze_depth) {
187 		percpu_ref_resurrect(&q->q_usage_counter);
188 		wake_up_all(&q->mq_freeze_wq);
189 	}
190 	mutex_unlock(&q->mq_freeze_lock);
191 }
192 
193 void blk_mq_unfreeze_queue(struct request_queue *q)
194 {
195 	__blk_mq_unfreeze_queue(q, false);
196 }
197 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
198 
199 /*
200  * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
201  * mpt3sas driver such that this function can be removed.
202  */
203 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
204 {
205 	unsigned long flags;
206 
207 	spin_lock_irqsave(&q->queue_lock, flags);
208 	if (!q->quiesce_depth++)
209 		blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
210 	spin_unlock_irqrestore(&q->queue_lock, flags);
211 }
212 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
213 
214 /**
215  * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
216  * @set: tag_set to wait on
217  *
218  * Note: it is driver's responsibility for making sure that quiesce has
219  * been started on or more of the request_queues of the tag_set.  This
220  * function only waits for the quiesce on those request_queues that had
221  * the quiesce flag set using blk_mq_quiesce_queue_nowait.
222  */
223 void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set)
224 {
225 	if (set->flags & BLK_MQ_F_BLOCKING)
226 		synchronize_srcu(set->srcu);
227 	else
228 		synchronize_rcu();
229 }
230 EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
231 
232 /**
233  * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
234  * @q: request queue.
235  *
236  * Note: this function does not prevent that the struct request end_io()
237  * callback function is invoked. Once this function is returned, we make
238  * sure no dispatch can happen until the queue is unquiesced via
239  * blk_mq_unquiesce_queue().
240  */
241 void blk_mq_quiesce_queue(struct request_queue *q)
242 {
243 	blk_mq_quiesce_queue_nowait(q);
244 	/* nothing to wait for non-mq queues */
245 	if (queue_is_mq(q))
246 		blk_mq_wait_quiesce_done(q->tag_set);
247 }
248 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
249 
250 /*
251  * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
252  * @q: request queue.
253  *
254  * This function recovers queue into the state before quiescing
255  * which is done by blk_mq_quiesce_queue.
256  */
257 void blk_mq_unquiesce_queue(struct request_queue *q)
258 {
259 	unsigned long flags;
260 	bool run_queue = false;
261 
262 	spin_lock_irqsave(&q->queue_lock, flags);
263 	if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
264 		;
265 	} else if (!--q->quiesce_depth) {
266 		blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
267 		run_queue = true;
268 	}
269 	spin_unlock_irqrestore(&q->queue_lock, flags);
270 
271 	/* dispatch requests which are inserted during quiescing */
272 	if (run_queue)
273 		blk_mq_run_hw_queues(q, true);
274 }
275 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
276 
277 void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set)
278 {
279 	struct request_queue *q;
280 
281 	mutex_lock(&set->tag_list_lock);
282 	list_for_each_entry(q, &set->tag_list, tag_set_list) {
283 		if (!blk_queue_skip_tagset_quiesce(q))
284 			blk_mq_quiesce_queue_nowait(q);
285 	}
286 	blk_mq_wait_quiesce_done(set);
287 	mutex_unlock(&set->tag_list_lock);
288 }
289 EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset);
290 
291 void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set)
292 {
293 	struct request_queue *q;
294 
295 	mutex_lock(&set->tag_list_lock);
296 	list_for_each_entry(q, &set->tag_list, tag_set_list) {
297 		if (!blk_queue_skip_tagset_quiesce(q))
298 			blk_mq_unquiesce_queue(q);
299 	}
300 	mutex_unlock(&set->tag_list_lock);
301 }
302 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset);
303 
304 void blk_mq_wake_waiters(struct request_queue *q)
305 {
306 	struct blk_mq_hw_ctx *hctx;
307 	unsigned long i;
308 
309 	queue_for_each_hw_ctx(q, hctx, i)
310 		if (blk_mq_hw_queue_mapped(hctx))
311 			blk_mq_tag_wakeup_all(hctx->tags, true);
312 }
313 
314 void blk_rq_init(struct request_queue *q, struct request *rq)
315 {
316 	memset(rq, 0, sizeof(*rq));
317 
318 	INIT_LIST_HEAD(&rq->queuelist);
319 	rq->q = q;
320 	rq->__sector = (sector_t) -1;
321 	INIT_HLIST_NODE(&rq->hash);
322 	RB_CLEAR_NODE(&rq->rb_node);
323 	rq->tag = BLK_MQ_NO_TAG;
324 	rq->internal_tag = BLK_MQ_NO_TAG;
325 	rq->start_time_ns = ktime_get_ns();
326 	rq->part = NULL;
327 	blk_crypto_rq_set_defaults(rq);
328 }
329 EXPORT_SYMBOL(blk_rq_init);
330 
331 /* Set start and alloc time when the allocated request is actually used */
332 static inline void blk_mq_rq_time_init(struct request *rq, u64 alloc_time_ns)
333 {
334 	if (blk_mq_need_time_stamp(rq))
335 		rq->start_time_ns = ktime_get_ns();
336 	else
337 		rq->start_time_ns = 0;
338 
339 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
340 	if (blk_queue_rq_alloc_time(rq->q))
341 		rq->alloc_time_ns = alloc_time_ns ?: rq->start_time_ns;
342 	else
343 		rq->alloc_time_ns = 0;
344 #endif
345 }
346 
347 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
348 		struct blk_mq_tags *tags, unsigned int tag)
349 {
350 	struct blk_mq_ctx *ctx = data->ctx;
351 	struct blk_mq_hw_ctx *hctx = data->hctx;
352 	struct request_queue *q = data->q;
353 	struct request *rq = tags->static_rqs[tag];
354 
355 	rq->q = q;
356 	rq->mq_ctx = ctx;
357 	rq->mq_hctx = hctx;
358 	rq->cmd_flags = data->cmd_flags;
359 
360 	if (data->flags & BLK_MQ_REQ_PM)
361 		data->rq_flags |= RQF_PM;
362 	if (blk_queue_io_stat(q))
363 		data->rq_flags |= RQF_IO_STAT;
364 	rq->rq_flags = data->rq_flags;
365 
366 	if (data->rq_flags & RQF_SCHED_TAGS) {
367 		rq->tag = BLK_MQ_NO_TAG;
368 		rq->internal_tag = tag;
369 	} else {
370 		rq->tag = tag;
371 		rq->internal_tag = BLK_MQ_NO_TAG;
372 	}
373 	rq->timeout = 0;
374 
375 	rq->part = NULL;
376 	rq->io_start_time_ns = 0;
377 	rq->stats_sectors = 0;
378 	rq->nr_phys_segments = 0;
379 #if defined(CONFIG_BLK_DEV_INTEGRITY)
380 	rq->nr_integrity_segments = 0;
381 #endif
382 	rq->end_io = NULL;
383 	rq->end_io_data = NULL;
384 
385 	blk_crypto_rq_set_defaults(rq);
386 	INIT_LIST_HEAD(&rq->queuelist);
387 	/* tag was already set */
388 	WRITE_ONCE(rq->deadline, 0);
389 	req_ref_set(rq, 1);
390 
391 	if (rq->rq_flags & RQF_USE_SCHED) {
392 		struct elevator_queue *e = data->q->elevator;
393 
394 		INIT_HLIST_NODE(&rq->hash);
395 		RB_CLEAR_NODE(&rq->rb_node);
396 
397 		if (e->type->ops.prepare_request)
398 			e->type->ops.prepare_request(rq);
399 	}
400 
401 	return rq;
402 }
403 
404 static inline struct request *
405 __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data)
406 {
407 	unsigned int tag, tag_offset;
408 	struct blk_mq_tags *tags;
409 	struct request *rq;
410 	unsigned long tag_mask;
411 	int i, nr = 0;
412 
413 	tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset);
414 	if (unlikely(!tag_mask))
415 		return NULL;
416 
417 	tags = blk_mq_tags_from_data(data);
418 	for (i = 0; tag_mask; i++) {
419 		if (!(tag_mask & (1UL << i)))
420 			continue;
421 		tag = tag_offset + i;
422 		prefetch(tags->static_rqs[tag]);
423 		tag_mask &= ~(1UL << i);
424 		rq = blk_mq_rq_ctx_init(data, tags, tag);
425 		rq_list_add(data->cached_rq, rq);
426 		nr++;
427 	}
428 	/* caller already holds a reference, add for remainder */
429 	percpu_ref_get_many(&data->q->q_usage_counter, nr - 1);
430 	data->nr_tags -= nr;
431 
432 	return rq_list_pop(data->cached_rq);
433 }
434 
435 static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
436 {
437 	struct request_queue *q = data->q;
438 	u64 alloc_time_ns = 0;
439 	struct request *rq;
440 	unsigned int tag;
441 
442 	/* alloc_time includes depth and tag waits */
443 	if (blk_queue_rq_alloc_time(q))
444 		alloc_time_ns = ktime_get_ns();
445 
446 	if (data->cmd_flags & REQ_NOWAIT)
447 		data->flags |= BLK_MQ_REQ_NOWAIT;
448 
449 	if (q->elevator) {
450 		/*
451 		 * All requests use scheduler tags when an I/O scheduler is
452 		 * enabled for the queue.
453 		 */
454 		data->rq_flags |= RQF_SCHED_TAGS;
455 
456 		/*
457 		 * Flush/passthrough requests are special and go directly to the
458 		 * dispatch list.
459 		 */
460 		if ((data->cmd_flags & REQ_OP_MASK) != REQ_OP_FLUSH &&
461 		    !blk_op_is_passthrough(data->cmd_flags)) {
462 			struct elevator_mq_ops *ops = &q->elevator->type->ops;
463 
464 			WARN_ON_ONCE(data->flags & BLK_MQ_REQ_RESERVED);
465 
466 			data->rq_flags |= RQF_USE_SCHED;
467 			if (ops->limit_depth)
468 				ops->limit_depth(data->cmd_flags, data);
469 		}
470 	}
471 
472 retry:
473 	data->ctx = blk_mq_get_ctx(q);
474 	data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
475 	if (!(data->rq_flags & RQF_SCHED_TAGS))
476 		blk_mq_tag_busy(data->hctx);
477 
478 	if (data->flags & BLK_MQ_REQ_RESERVED)
479 		data->rq_flags |= RQF_RESV;
480 
481 	/*
482 	 * Try batched alloc if we want more than 1 tag.
483 	 */
484 	if (data->nr_tags > 1) {
485 		rq = __blk_mq_alloc_requests_batch(data);
486 		if (rq) {
487 			blk_mq_rq_time_init(rq, alloc_time_ns);
488 			return rq;
489 		}
490 		data->nr_tags = 1;
491 	}
492 
493 	/*
494 	 * Waiting allocations only fail because of an inactive hctx.  In that
495 	 * case just retry the hctx assignment and tag allocation as CPU hotplug
496 	 * should have migrated us to an online CPU by now.
497 	 */
498 	tag = blk_mq_get_tag(data);
499 	if (tag == BLK_MQ_NO_TAG) {
500 		if (data->flags & BLK_MQ_REQ_NOWAIT)
501 			return NULL;
502 		/*
503 		 * Give up the CPU and sleep for a random short time to
504 		 * ensure that thread using a realtime scheduling class
505 		 * are migrated off the CPU, and thus off the hctx that
506 		 * is going away.
507 		 */
508 		msleep(3);
509 		goto retry;
510 	}
511 
512 	rq = blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag);
513 	blk_mq_rq_time_init(rq, alloc_time_ns);
514 	return rq;
515 }
516 
517 static struct request *blk_mq_rq_cache_fill(struct request_queue *q,
518 					    struct blk_plug *plug,
519 					    blk_opf_t opf,
520 					    blk_mq_req_flags_t flags)
521 {
522 	struct blk_mq_alloc_data data = {
523 		.q		= q,
524 		.flags		= flags,
525 		.cmd_flags	= opf,
526 		.nr_tags	= plug->nr_ios,
527 		.cached_rq	= &plug->cached_rq,
528 	};
529 	struct request *rq;
530 
531 	if (blk_queue_enter(q, flags))
532 		return NULL;
533 
534 	plug->nr_ios = 1;
535 
536 	rq = __blk_mq_alloc_requests(&data);
537 	if (unlikely(!rq))
538 		blk_queue_exit(q);
539 	return rq;
540 }
541 
542 static struct request *blk_mq_alloc_cached_request(struct request_queue *q,
543 						   blk_opf_t opf,
544 						   blk_mq_req_flags_t flags)
545 {
546 	struct blk_plug *plug = current->plug;
547 	struct request *rq;
548 
549 	if (!plug)
550 		return NULL;
551 
552 	if (rq_list_empty(plug->cached_rq)) {
553 		if (plug->nr_ios == 1)
554 			return NULL;
555 		rq = blk_mq_rq_cache_fill(q, plug, opf, flags);
556 		if (!rq)
557 			return NULL;
558 	} else {
559 		rq = rq_list_peek(&plug->cached_rq);
560 		if (!rq || rq->q != q)
561 			return NULL;
562 
563 		if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type)
564 			return NULL;
565 		if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
566 			return NULL;
567 
568 		plug->cached_rq = rq_list_next(rq);
569 		blk_mq_rq_time_init(rq, 0);
570 	}
571 
572 	rq->cmd_flags = opf;
573 	INIT_LIST_HEAD(&rq->queuelist);
574 	return rq;
575 }
576 
577 struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf,
578 		blk_mq_req_flags_t flags)
579 {
580 	struct request *rq;
581 
582 	rq = blk_mq_alloc_cached_request(q, opf, flags);
583 	if (!rq) {
584 		struct blk_mq_alloc_data data = {
585 			.q		= q,
586 			.flags		= flags,
587 			.cmd_flags	= opf,
588 			.nr_tags	= 1,
589 		};
590 		int ret;
591 
592 		ret = blk_queue_enter(q, flags);
593 		if (ret)
594 			return ERR_PTR(ret);
595 
596 		rq = __blk_mq_alloc_requests(&data);
597 		if (!rq)
598 			goto out_queue_exit;
599 	}
600 	rq->__data_len = 0;
601 	rq->__sector = (sector_t) -1;
602 	rq->bio = rq->biotail = NULL;
603 	return rq;
604 out_queue_exit:
605 	blk_queue_exit(q);
606 	return ERR_PTR(-EWOULDBLOCK);
607 }
608 EXPORT_SYMBOL(blk_mq_alloc_request);
609 
610 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
611 	blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx)
612 {
613 	struct blk_mq_alloc_data data = {
614 		.q		= q,
615 		.flags		= flags,
616 		.cmd_flags	= opf,
617 		.nr_tags	= 1,
618 	};
619 	u64 alloc_time_ns = 0;
620 	struct request *rq;
621 	unsigned int cpu;
622 	unsigned int tag;
623 	int ret;
624 
625 	/* alloc_time includes depth and tag waits */
626 	if (blk_queue_rq_alloc_time(q))
627 		alloc_time_ns = ktime_get_ns();
628 
629 	/*
630 	 * If the tag allocator sleeps we could get an allocation for a
631 	 * different hardware context.  No need to complicate the low level
632 	 * allocator for this for the rare use case of a command tied to
633 	 * a specific queue.
634 	 */
635 	if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) ||
636 	    WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED)))
637 		return ERR_PTR(-EINVAL);
638 
639 	if (hctx_idx >= q->nr_hw_queues)
640 		return ERR_PTR(-EIO);
641 
642 	ret = blk_queue_enter(q, flags);
643 	if (ret)
644 		return ERR_PTR(ret);
645 
646 	/*
647 	 * Check if the hardware context is actually mapped to anything.
648 	 * If not tell the caller that it should skip this queue.
649 	 */
650 	ret = -EXDEV;
651 	data.hctx = xa_load(&q->hctx_table, hctx_idx);
652 	if (!blk_mq_hw_queue_mapped(data.hctx))
653 		goto out_queue_exit;
654 	cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
655 	if (cpu >= nr_cpu_ids)
656 		goto out_queue_exit;
657 	data.ctx = __blk_mq_get_ctx(q, cpu);
658 
659 	if (q->elevator)
660 		data.rq_flags |= RQF_SCHED_TAGS;
661 	else
662 		blk_mq_tag_busy(data.hctx);
663 
664 	if (flags & BLK_MQ_REQ_RESERVED)
665 		data.rq_flags |= RQF_RESV;
666 
667 	ret = -EWOULDBLOCK;
668 	tag = blk_mq_get_tag(&data);
669 	if (tag == BLK_MQ_NO_TAG)
670 		goto out_queue_exit;
671 	rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag);
672 	blk_mq_rq_time_init(rq, alloc_time_ns);
673 	rq->__data_len = 0;
674 	rq->__sector = (sector_t) -1;
675 	rq->bio = rq->biotail = NULL;
676 	return rq;
677 
678 out_queue_exit:
679 	blk_queue_exit(q);
680 	return ERR_PTR(ret);
681 }
682 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
683 
684 static void __blk_mq_free_request(struct request *rq)
685 {
686 	struct request_queue *q = rq->q;
687 	struct blk_mq_ctx *ctx = rq->mq_ctx;
688 	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
689 	const int sched_tag = rq->internal_tag;
690 
691 	blk_crypto_free_request(rq);
692 	blk_pm_mark_last_busy(rq);
693 	rq->mq_hctx = NULL;
694 
695 	if (rq->rq_flags & RQF_MQ_INFLIGHT)
696 		__blk_mq_dec_active_requests(hctx);
697 
698 	if (rq->tag != BLK_MQ_NO_TAG)
699 		blk_mq_put_tag(hctx->tags, ctx, rq->tag);
700 	if (sched_tag != BLK_MQ_NO_TAG)
701 		blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
702 	blk_mq_sched_restart(hctx);
703 	blk_queue_exit(q);
704 }
705 
706 void blk_mq_free_request(struct request *rq)
707 {
708 	struct request_queue *q = rq->q;
709 
710 	if ((rq->rq_flags & RQF_USE_SCHED) &&
711 	    q->elevator->type->ops.finish_request)
712 		q->elevator->type->ops.finish_request(rq);
713 
714 	if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
715 		laptop_io_completion(q->disk->bdi);
716 
717 	rq_qos_done(q, rq);
718 
719 	WRITE_ONCE(rq->state, MQ_RQ_IDLE);
720 	if (req_ref_put_and_test(rq))
721 		__blk_mq_free_request(rq);
722 }
723 EXPORT_SYMBOL_GPL(blk_mq_free_request);
724 
725 void blk_mq_free_plug_rqs(struct blk_plug *plug)
726 {
727 	struct request *rq;
728 
729 	while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
730 		blk_mq_free_request(rq);
731 }
732 
733 void blk_dump_rq_flags(struct request *rq, char *msg)
734 {
735 	printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
736 		rq->q->disk ? rq->q->disk->disk_name : "?",
737 		(__force unsigned long long) rq->cmd_flags);
738 
739 	printk(KERN_INFO "  sector %llu, nr/cnr %u/%u\n",
740 	       (unsigned long long)blk_rq_pos(rq),
741 	       blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
742 	printk(KERN_INFO "  bio %p, biotail %p, len %u\n",
743 	       rq->bio, rq->biotail, blk_rq_bytes(rq));
744 }
745 EXPORT_SYMBOL(blk_dump_rq_flags);
746 
747 static void req_bio_endio(struct request *rq, struct bio *bio,
748 			  unsigned int nbytes, blk_status_t error)
749 {
750 	if (unlikely(error)) {
751 		bio->bi_status = error;
752 	} else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
753 		/*
754 		 * Partial zone append completions cannot be supported as the
755 		 * BIO fragments may end up not being written sequentially.
756 		 */
757 		if (bio->bi_iter.bi_size != nbytes)
758 			bio->bi_status = BLK_STS_IOERR;
759 		else
760 			bio->bi_iter.bi_sector = rq->__sector;
761 	}
762 
763 	bio_advance(bio, nbytes);
764 
765 	if (unlikely(rq->rq_flags & RQF_QUIET))
766 		bio_set_flag(bio, BIO_QUIET);
767 	/* don't actually finish bio if it's part of flush sequence */
768 	if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
769 		bio_endio(bio);
770 }
771 
772 static void blk_account_io_completion(struct request *req, unsigned int bytes)
773 {
774 	if (req->part && blk_do_io_stat(req)) {
775 		const int sgrp = op_stat_group(req_op(req));
776 
777 		part_stat_lock();
778 		part_stat_add(req->part, sectors[sgrp], bytes >> 9);
779 		part_stat_unlock();
780 	}
781 }
782 
783 static void blk_print_req_error(struct request *req, blk_status_t status)
784 {
785 	printk_ratelimited(KERN_ERR
786 		"%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
787 		"phys_seg %u prio class %u\n",
788 		blk_status_to_str(status),
789 		req->q->disk ? req->q->disk->disk_name : "?",
790 		blk_rq_pos(req), (__force u32)req_op(req),
791 		blk_op_str(req_op(req)),
792 		(__force u32)(req->cmd_flags & ~REQ_OP_MASK),
793 		req->nr_phys_segments,
794 		IOPRIO_PRIO_CLASS(req->ioprio));
795 }
796 
797 /*
798  * Fully end IO on a request. Does not support partial completions, or
799  * errors.
800  */
801 static void blk_complete_request(struct request *req)
802 {
803 	const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
804 	int total_bytes = blk_rq_bytes(req);
805 	struct bio *bio = req->bio;
806 
807 	trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
808 
809 	if (!bio)
810 		return;
811 
812 #ifdef CONFIG_BLK_DEV_INTEGRITY
813 	if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
814 		req->q->integrity.profile->complete_fn(req, total_bytes);
815 #endif
816 
817 	/*
818 	 * Upper layers may call blk_crypto_evict_key() anytime after the last
819 	 * bio_endio().  Therefore, the keyslot must be released before that.
820 	 */
821 	blk_crypto_rq_put_keyslot(req);
822 
823 	blk_account_io_completion(req, total_bytes);
824 
825 	do {
826 		struct bio *next = bio->bi_next;
827 
828 		/* Completion has already been traced */
829 		bio_clear_flag(bio, BIO_TRACE_COMPLETION);
830 
831 		if (req_op(req) == REQ_OP_ZONE_APPEND)
832 			bio->bi_iter.bi_sector = req->__sector;
833 
834 		if (!is_flush)
835 			bio_endio(bio);
836 		bio = next;
837 	} while (bio);
838 
839 	/*
840 	 * Reset counters so that the request stacking driver
841 	 * can find how many bytes remain in the request
842 	 * later.
843 	 */
844 	if (!req->end_io) {
845 		req->bio = NULL;
846 		req->__data_len = 0;
847 	}
848 }
849 
850 /**
851  * blk_update_request - Complete multiple bytes without completing the request
852  * @req:      the request being processed
853  * @error:    block status code
854  * @nr_bytes: number of bytes to complete for @req
855  *
856  * Description:
857  *     Ends I/O on a number of bytes attached to @req, but doesn't complete
858  *     the request structure even if @req doesn't have leftover.
859  *     If @req has leftover, sets it up for the next range of segments.
860  *
861  *     Passing the result of blk_rq_bytes() as @nr_bytes guarantees
862  *     %false return from this function.
863  *
864  * Note:
865  *	The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
866  *      except in the consistency check at the end of this function.
867  *
868  * Return:
869  *     %false - this request doesn't have any more data
870  *     %true  - this request has more data
871  **/
872 bool blk_update_request(struct request *req, blk_status_t error,
873 		unsigned int nr_bytes)
874 {
875 	int total_bytes;
876 
877 	trace_block_rq_complete(req, error, nr_bytes);
878 
879 	if (!req->bio)
880 		return false;
881 
882 #ifdef CONFIG_BLK_DEV_INTEGRITY
883 	if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
884 	    error == BLK_STS_OK)
885 		req->q->integrity.profile->complete_fn(req, nr_bytes);
886 #endif
887 
888 	/*
889 	 * Upper layers may call blk_crypto_evict_key() anytime after the last
890 	 * bio_endio().  Therefore, the keyslot must be released before that.
891 	 */
892 	if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req))
893 		__blk_crypto_rq_put_keyslot(req);
894 
895 	if (unlikely(error && !blk_rq_is_passthrough(req) &&
896 		     !(req->rq_flags & RQF_QUIET)) &&
897 		     !test_bit(GD_DEAD, &req->q->disk->state)) {
898 		blk_print_req_error(req, error);
899 		trace_block_rq_error(req, error, nr_bytes);
900 	}
901 
902 	blk_account_io_completion(req, nr_bytes);
903 
904 	total_bytes = 0;
905 	while (req->bio) {
906 		struct bio *bio = req->bio;
907 		unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
908 
909 		if (bio_bytes == bio->bi_iter.bi_size)
910 			req->bio = bio->bi_next;
911 
912 		/* Completion has already been traced */
913 		bio_clear_flag(bio, BIO_TRACE_COMPLETION);
914 		req_bio_endio(req, bio, bio_bytes, error);
915 
916 		total_bytes += bio_bytes;
917 		nr_bytes -= bio_bytes;
918 
919 		if (!nr_bytes)
920 			break;
921 	}
922 
923 	/*
924 	 * completely done
925 	 */
926 	if (!req->bio) {
927 		/*
928 		 * Reset counters so that the request stacking driver
929 		 * can find how many bytes remain in the request
930 		 * later.
931 		 */
932 		req->__data_len = 0;
933 		return false;
934 	}
935 
936 	req->__data_len -= total_bytes;
937 
938 	/* update sector only for requests with clear definition of sector */
939 	if (!blk_rq_is_passthrough(req))
940 		req->__sector += total_bytes >> 9;
941 
942 	/* mixed attributes always follow the first bio */
943 	if (req->rq_flags & RQF_MIXED_MERGE) {
944 		req->cmd_flags &= ~REQ_FAILFAST_MASK;
945 		req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
946 	}
947 
948 	if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
949 		/*
950 		 * If total number of sectors is less than the first segment
951 		 * size, something has gone terribly wrong.
952 		 */
953 		if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
954 			blk_dump_rq_flags(req, "request botched");
955 			req->__data_len = blk_rq_cur_bytes(req);
956 		}
957 
958 		/* recalculate the number of segments */
959 		req->nr_phys_segments = blk_recalc_rq_segments(req);
960 	}
961 
962 	return true;
963 }
964 EXPORT_SYMBOL_GPL(blk_update_request);
965 
966 static inline void blk_account_io_done(struct request *req, u64 now)
967 {
968 	trace_block_io_done(req);
969 
970 	/*
971 	 * Account IO completion.  flush_rq isn't accounted as a
972 	 * normal IO on queueing nor completion.  Accounting the
973 	 * containing request is enough.
974 	 */
975 	if (blk_do_io_stat(req) && req->part &&
976 	    !(req->rq_flags & RQF_FLUSH_SEQ)) {
977 		const int sgrp = op_stat_group(req_op(req));
978 
979 		part_stat_lock();
980 		update_io_ticks(req->part, jiffies, true);
981 		part_stat_inc(req->part, ios[sgrp]);
982 		part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
983 		part_stat_unlock();
984 	}
985 }
986 
987 static inline void blk_account_io_start(struct request *req)
988 {
989 	trace_block_io_start(req);
990 
991 	if (blk_do_io_stat(req)) {
992 		/*
993 		 * All non-passthrough requests are created from a bio with one
994 		 * exception: when a flush command that is part of a flush sequence
995 		 * generated by the state machine in blk-flush.c is cloned onto the
996 		 * lower device by dm-multipath we can get here without a bio.
997 		 */
998 		if (req->bio)
999 			req->part = req->bio->bi_bdev;
1000 		else
1001 			req->part = req->q->disk->part0;
1002 
1003 		part_stat_lock();
1004 		update_io_ticks(req->part, jiffies, false);
1005 		part_stat_unlock();
1006 	}
1007 }
1008 
1009 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
1010 {
1011 	if (rq->rq_flags & RQF_STATS)
1012 		blk_stat_add(rq, now);
1013 
1014 	blk_mq_sched_completed_request(rq, now);
1015 	blk_account_io_done(rq, now);
1016 }
1017 
1018 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
1019 {
1020 	if (blk_mq_need_time_stamp(rq))
1021 		__blk_mq_end_request_acct(rq, ktime_get_ns());
1022 
1023 	if (rq->end_io) {
1024 		rq_qos_done(rq->q, rq);
1025 		if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1026 			blk_mq_free_request(rq);
1027 	} else {
1028 		blk_mq_free_request(rq);
1029 	}
1030 }
1031 EXPORT_SYMBOL(__blk_mq_end_request);
1032 
1033 void blk_mq_end_request(struct request *rq, blk_status_t error)
1034 {
1035 	if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1036 		BUG();
1037 	__blk_mq_end_request(rq, error);
1038 }
1039 EXPORT_SYMBOL(blk_mq_end_request);
1040 
1041 #define TAG_COMP_BATCH		32
1042 
1043 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1044 					  int *tag_array, int nr_tags)
1045 {
1046 	struct request_queue *q = hctx->queue;
1047 
1048 	/*
1049 	 * All requests should have been marked as RQF_MQ_INFLIGHT, so
1050 	 * update hctx->nr_active in batch
1051 	 */
1052 	if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
1053 		__blk_mq_sub_active_requests(hctx, nr_tags);
1054 
1055 	blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
1056 	percpu_ref_put_many(&q->q_usage_counter, nr_tags);
1057 }
1058 
1059 void blk_mq_end_request_batch(struct io_comp_batch *iob)
1060 {
1061 	int tags[TAG_COMP_BATCH], nr_tags = 0;
1062 	struct blk_mq_hw_ctx *cur_hctx = NULL;
1063 	struct request *rq;
1064 	u64 now = 0;
1065 
1066 	if (iob->need_ts)
1067 		now = ktime_get_ns();
1068 
1069 	while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1070 		prefetch(rq->bio);
1071 		prefetch(rq->rq_next);
1072 
1073 		blk_complete_request(rq);
1074 		if (iob->need_ts)
1075 			__blk_mq_end_request_acct(rq, now);
1076 
1077 		rq_qos_done(rq->q, rq);
1078 
1079 		/*
1080 		 * If end_io handler returns NONE, then it still has
1081 		 * ownership of the request.
1082 		 */
1083 		if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
1084 			continue;
1085 
1086 		WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1087 		if (!req_ref_put_and_test(rq))
1088 			continue;
1089 
1090 		blk_crypto_free_request(rq);
1091 		blk_pm_mark_last_busy(rq);
1092 
1093 		if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1094 			if (cur_hctx)
1095 				blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1096 			nr_tags = 0;
1097 			cur_hctx = rq->mq_hctx;
1098 		}
1099 		tags[nr_tags++] = rq->tag;
1100 	}
1101 
1102 	if (nr_tags)
1103 		blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1104 }
1105 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1106 
1107 static void blk_complete_reqs(struct llist_head *list)
1108 {
1109 	struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1110 	struct request *rq, *next;
1111 
1112 	llist_for_each_entry_safe(rq, next, entry, ipi_list)
1113 		rq->q->mq_ops->complete(rq);
1114 }
1115 
1116 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1117 {
1118 	blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1119 }
1120 
1121 static int blk_softirq_cpu_dead(unsigned int cpu)
1122 {
1123 	blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1124 	return 0;
1125 }
1126 
1127 static void __blk_mq_complete_request_remote(void *data)
1128 {
1129 	__raise_softirq_irqoff(BLOCK_SOFTIRQ);
1130 }
1131 
1132 static inline bool blk_mq_complete_need_ipi(struct request *rq)
1133 {
1134 	int cpu = raw_smp_processor_id();
1135 
1136 	if (!IS_ENABLED(CONFIG_SMP) ||
1137 	    !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1138 		return false;
1139 	/*
1140 	 * With force threaded interrupts enabled, raising softirq from an SMP
1141 	 * function call will always result in waking the ksoftirqd thread.
1142 	 * This is probably worse than completing the request on a different
1143 	 * cache domain.
1144 	 */
1145 	if (force_irqthreads())
1146 		return false;
1147 
1148 	/* same CPU or cache domain?  Complete locally */
1149 	if (cpu == rq->mq_ctx->cpu ||
1150 	    (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1151 	     cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1152 		return false;
1153 
1154 	/* don't try to IPI to an offline CPU */
1155 	return cpu_online(rq->mq_ctx->cpu);
1156 }
1157 
1158 static void blk_mq_complete_send_ipi(struct request *rq)
1159 {
1160 	struct llist_head *list;
1161 	unsigned int cpu;
1162 
1163 	cpu = rq->mq_ctx->cpu;
1164 	list = &per_cpu(blk_cpu_done, cpu);
1165 	if (llist_add(&rq->ipi_list, list)) {
1166 		INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
1167 		smp_call_function_single_async(cpu, &rq->csd);
1168 	}
1169 }
1170 
1171 static void blk_mq_raise_softirq(struct request *rq)
1172 {
1173 	struct llist_head *list;
1174 
1175 	preempt_disable();
1176 	list = this_cpu_ptr(&blk_cpu_done);
1177 	if (llist_add(&rq->ipi_list, list))
1178 		raise_softirq(BLOCK_SOFTIRQ);
1179 	preempt_enable();
1180 }
1181 
1182 bool blk_mq_complete_request_remote(struct request *rq)
1183 {
1184 	WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1185 
1186 	/*
1187 	 * For request which hctx has only one ctx mapping,
1188 	 * or a polled request, always complete locally,
1189 	 * it's pointless to redirect the completion.
1190 	 */
1191 	if ((rq->mq_hctx->nr_ctx == 1 &&
1192 	     rq->mq_ctx->cpu == raw_smp_processor_id()) ||
1193 	     rq->cmd_flags & REQ_POLLED)
1194 		return false;
1195 
1196 	if (blk_mq_complete_need_ipi(rq)) {
1197 		blk_mq_complete_send_ipi(rq);
1198 		return true;
1199 	}
1200 
1201 	if (rq->q->nr_hw_queues == 1) {
1202 		blk_mq_raise_softirq(rq);
1203 		return true;
1204 	}
1205 	return false;
1206 }
1207 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1208 
1209 /**
1210  * blk_mq_complete_request - end I/O on a request
1211  * @rq:		the request being processed
1212  *
1213  * Description:
1214  *	Complete a request by scheduling the ->complete_rq operation.
1215  **/
1216 void blk_mq_complete_request(struct request *rq)
1217 {
1218 	if (!blk_mq_complete_request_remote(rq))
1219 		rq->q->mq_ops->complete(rq);
1220 }
1221 EXPORT_SYMBOL(blk_mq_complete_request);
1222 
1223 /**
1224  * blk_mq_start_request - Start processing a request
1225  * @rq: Pointer to request to be started
1226  *
1227  * Function used by device drivers to notify the block layer that a request
1228  * is going to be processed now, so blk layer can do proper initializations
1229  * such as starting the timeout timer.
1230  */
1231 void blk_mq_start_request(struct request *rq)
1232 {
1233 	struct request_queue *q = rq->q;
1234 
1235 	trace_block_rq_issue(rq);
1236 
1237 	if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1238 		rq->io_start_time_ns = ktime_get_ns();
1239 		rq->stats_sectors = blk_rq_sectors(rq);
1240 		rq->rq_flags |= RQF_STATS;
1241 		rq_qos_issue(q, rq);
1242 	}
1243 
1244 	WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1245 
1246 	blk_add_timer(rq);
1247 	WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1248 
1249 #ifdef CONFIG_BLK_DEV_INTEGRITY
1250 	if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1251 		q->integrity.profile->prepare_fn(rq);
1252 #endif
1253 	if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1254 	        WRITE_ONCE(rq->bio->bi_cookie, rq->mq_hctx->queue_num);
1255 }
1256 EXPORT_SYMBOL(blk_mq_start_request);
1257 
1258 /*
1259  * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1260  * queues. This is important for md arrays to benefit from merging
1261  * requests.
1262  */
1263 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1264 {
1265 	if (plug->multiple_queues)
1266 		return BLK_MAX_REQUEST_COUNT * 2;
1267 	return BLK_MAX_REQUEST_COUNT;
1268 }
1269 
1270 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1271 {
1272 	struct request *last = rq_list_peek(&plug->mq_list);
1273 
1274 	if (!plug->rq_count) {
1275 		trace_block_plug(rq->q);
1276 	} else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1277 		   (!blk_queue_nomerges(rq->q) &&
1278 		    blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1279 		blk_mq_flush_plug_list(plug, false);
1280 		last = NULL;
1281 		trace_block_plug(rq->q);
1282 	}
1283 
1284 	if (!plug->multiple_queues && last && last->q != rq->q)
1285 		plug->multiple_queues = true;
1286 	/*
1287 	 * Any request allocated from sched tags can't be issued to
1288 	 * ->queue_rqs() directly
1289 	 */
1290 	if (!plug->has_elevator && (rq->rq_flags & RQF_SCHED_TAGS))
1291 		plug->has_elevator = true;
1292 	rq->rq_next = NULL;
1293 	rq_list_add(&plug->mq_list, rq);
1294 	plug->rq_count++;
1295 }
1296 
1297 /**
1298  * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1299  * @rq:		request to insert
1300  * @at_head:    insert request at head or tail of queue
1301  *
1302  * Description:
1303  *    Insert a fully prepared request at the back of the I/O scheduler queue
1304  *    for execution.  Don't wait for completion.
1305  *
1306  * Note:
1307  *    This function will invoke @done directly if the queue is dead.
1308  */
1309 void blk_execute_rq_nowait(struct request *rq, bool at_head)
1310 {
1311 	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1312 
1313 	WARN_ON(irqs_disabled());
1314 	WARN_ON(!blk_rq_is_passthrough(rq));
1315 
1316 	blk_account_io_start(rq);
1317 
1318 	/*
1319 	 * As plugging can be enabled for passthrough requests on a zoned
1320 	 * device, directly accessing the plug instead of using blk_mq_plug()
1321 	 * should not have any consequences.
1322 	 */
1323 	if (current->plug && !at_head) {
1324 		blk_add_rq_to_plug(current->plug, rq);
1325 		return;
1326 	}
1327 
1328 	blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1329 	blk_mq_run_hw_queue(hctx, false);
1330 }
1331 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1332 
1333 struct blk_rq_wait {
1334 	struct completion done;
1335 	blk_status_t ret;
1336 };
1337 
1338 static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
1339 {
1340 	struct blk_rq_wait *wait = rq->end_io_data;
1341 
1342 	wait->ret = ret;
1343 	complete(&wait->done);
1344 	return RQ_END_IO_NONE;
1345 }
1346 
1347 bool blk_rq_is_poll(struct request *rq)
1348 {
1349 	if (!rq->mq_hctx)
1350 		return false;
1351 	if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1352 		return false;
1353 	return true;
1354 }
1355 EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1356 
1357 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1358 {
1359 	do {
1360 		blk_hctx_poll(rq->q, rq->mq_hctx, NULL, 0);
1361 		cond_resched();
1362 	} while (!completion_done(wait));
1363 }
1364 
1365 /**
1366  * blk_execute_rq - insert a request into queue for execution
1367  * @rq:		request to insert
1368  * @at_head:    insert request at head or tail of queue
1369  *
1370  * Description:
1371  *    Insert a fully prepared request at the back of the I/O scheduler queue
1372  *    for execution and wait for completion.
1373  * Return: The blk_status_t result provided to blk_mq_end_request().
1374  */
1375 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1376 {
1377 	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1378 	struct blk_rq_wait wait = {
1379 		.done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1380 	};
1381 
1382 	WARN_ON(irqs_disabled());
1383 	WARN_ON(!blk_rq_is_passthrough(rq));
1384 
1385 	rq->end_io_data = &wait;
1386 	rq->end_io = blk_end_sync_rq;
1387 
1388 	blk_account_io_start(rq);
1389 	blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1390 	blk_mq_run_hw_queue(hctx, false);
1391 
1392 	if (blk_rq_is_poll(rq)) {
1393 		blk_rq_poll_completion(rq, &wait.done);
1394 	} else {
1395 		/*
1396 		 * Prevent hang_check timer from firing at us during very long
1397 		 * I/O
1398 		 */
1399 		unsigned long hang_check = sysctl_hung_task_timeout_secs;
1400 
1401 		if (hang_check)
1402 			while (!wait_for_completion_io_timeout(&wait.done,
1403 					hang_check * (HZ/2)))
1404 				;
1405 		else
1406 			wait_for_completion_io(&wait.done);
1407 	}
1408 
1409 	return wait.ret;
1410 }
1411 EXPORT_SYMBOL(blk_execute_rq);
1412 
1413 static void __blk_mq_requeue_request(struct request *rq)
1414 {
1415 	struct request_queue *q = rq->q;
1416 
1417 	blk_mq_put_driver_tag(rq);
1418 
1419 	trace_block_rq_requeue(rq);
1420 	rq_qos_requeue(q, rq);
1421 
1422 	if (blk_mq_request_started(rq)) {
1423 		WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1424 		rq->rq_flags &= ~RQF_TIMED_OUT;
1425 	}
1426 }
1427 
1428 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1429 {
1430 	struct request_queue *q = rq->q;
1431 	unsigned long flags;
1432 
1433 	__blk_mq_requeue_request(rq);
1434 
1435 	/* this request will be re-inserted to io scheduler queue */
1436 	blk_mq_sched_requeue_request(rq);
1437 
1438 	spin_lock_irqsave(&q->requeue_lock, flags);
1439 	list_add_tail(&rq->queuelist, &q->requeue_list);
1440 	spin_unlock_irqrestore(&q->requeue_lock, flags);
1441 
1442 	if (kick_requeue_list)
1443 		blk_mq_kick_requeue_list(q);
1444 }
1445 EXPORT_SYMBOL(blk_mq_requeue_request);
1446 
1447 static void blk_mq_requeue_work(struct work_struct *work)
1448 {
1449 	struct request_queue *q =
1450 		container_of(work, struct request_queue, requeue_work.work);
1451 	LIST_HEAD(rq_list);
1452 	LIST_HEAD(flush_list);
1453 	struct request *rq;
1454 
1455 	spin_lock_irq(&q->requeue_lock);
1456 	list_splice_init(&q->requeue_list, &rq_list);
1457 	list_splice_init(&q->flush_list, &flush_list);
1458 	spin_unlock_irq(&q->requeue_lock);
1459 
1460 	while (!list_empty(&rq_list)) {
1461 		rq = list_entry(rq_list.next, struct request, queuelist);
1462 		/*
1463 		 * If RQF_DONTPREP ist set, the request has been started by the
1464 		 * driver already and might have driver-specific data allocated
1465 		 * already.  Insert it into the hctx dispatch list to avoid
1466 		 * block layer merges for the request.
1467 		 */
1468 		if (rq->rq_flags & RQF_DONTPREP) {
1469 			list_del_init(&rq->queuelist);
1470 			blk_mq_request_bypass_insert(rq, 0);
1471 		} else {
1472 			list_del_init(&rq->queuelist);
1473 			blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD);
1474 		}
1475 	}
1476 
1477 	while (!list_empty(&flush_list)) {
1478 		rq = list_entry(flush_list.next, struct request, queuelist);
1479 		list_del_init(&rq->queuelist);
1480 		blk_mq_insert_request(rq, 0);
1481 	}
1482 
1483 	blk_mq_run_hw_queues(q, false);
1484 }
1485 
1486 void blk_mq_kick_requeue_list(struct request_queue *q)
1487 {
1488 	kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1489 }
1490 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1491 
1492 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1493 				    unsigned long msecs)
1494 {
1495 	kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1496 				    msecs_to_jiffies(msecs));
1497 }
1498 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1499 
1500 static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1501 {
1502 	/*
1503 	 * If we find a request that isn't idle we know the queue is busy
1504 	 * as it's checked in the iter.
1505 	 * Return false to stop the iteration.
1506 	 */
1507 	if (blk_mq_request_started(rq)) {
1508 		bool *busy = priv;
1509 
1510 		*busy = true;
1511 		return false;
1512 	}
1513 
1514 	return true;
1515 }
1516 
1517 bool blk_mq_queue_inflight(struct request_queue *q)
1518 {
1519 	bool busy = false;
1520 
1521 	blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1522 	return busy;
1523 }
1524 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1525 
1526 static void blk_mq_rq_timed_out(struct request *req)
1527 {
1528 	req->rq_flags |= RQF_TIMED_OUT;
1529 	if (req->q->mq_ops->timeout) {
1530 		enum blk_eh_timer_return ret;
1531 
1532 		ret = req->q->mq_ops->timeout(req);
1533 		if (ret == BLK_EH_DONE)
1534 			return;
1535 		WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1536 	}
1537 
1538 	blk_add_timer(req);
1539 }
1540 
1541 struct blk_expired_data {
1542 	bool has_timedout_rq;
1543 	unsigned long next;
1544 	unsigned long timeout_start;
1545 };
1546 
1547 static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
1548 {
1549 	unsigned long deadline;
1550 
1551 	if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1552 		return false;
1553 	if (rq->rq_flags & RQF_TIMED_OUT)
1554 		return false;
1555 
1556 	deadline = READ_ONCE(rq->deadline);
1557 	if (time_after_eq(expired->timeout_start, deadline))
1558 		return true;
1559 
1560 	if (expired->next == 0)
1561 		expired->next = deadline;
1562 	else if (time_after(expired->next, deadline))
1563 		expired->next = deadline;
1564 	return false;
1565 }
1566 
1567 void blk_mq_put_rq_ref(struct request *rq)
1568 {
1569 	if (is_flush_rq(rq)) {
1570 		if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1571 			blk_mq_free_request(rq);
1572 	} else if (req_ref_put_and_test(rq)) {
1573 		__blk_mq_free_request(rq);
1574 	}
1575 }
1576 
1577 static bool blk_mq_check_expired(struct request *rq, void *priv)
1578 {
1579 	struct blk_expired_data *expired = priv;
1580 
1581 	/*
1582 	 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1583 	 * be reallocated underneath the timeout handler's processing, then
1584 	 * the expire check is reliable. If the request is not expired, then
1585 	 * it was completed and reallocated as a new request after returning
1586 	 * from blk_mq_check_expired().
1587 	 */
1588 	if (blk_mq_req_expired(rq, expired)) {
1589 		expired->has_timedout_rq = true;
1590 		return false;
1591 	}
1592 	return true;
1593 }
1594 
1595 static bool blk_mq_handle_expired(struct request *rq, void *priv)
1596 {
1597 	struct blk_expired_data *expired = priv;
1598 
1599 	if (blk_mq_req_expired(rq, expired))
1600 		blk_mq_rq_timed_out(rq);
1601 	return true;
1602 }
1603 
1604 static void blk_mq_timeout_work(struct work_struct *work)
1605 {
1606 	struct request_queue *q =
1607 		container_of(work, struct request_queue, timeout_work);
1608 	struct blk_expired_data expired = {
1609 		.timeout_start = jiffies,
1610 	};
1611 	struct blk_mq_hw_ctx *hctx;
1612 	unsigned long i;
1613 
1614 	/* A deadlock might occur if a request is stuck requiring a
1615 	 * timeout at the same time a queue freeze is waiting
1616 	 * completion, since the timeout code would not be able to
1617 	 * acquire the queue reference here.
1618 	 *
1619 	 * That's why we don't use blk_queue_enter here; instead, we use
1620 	 * percpu_ref_tryget directly, because we need to be able to
1621 	 * obtain a reference even in the short window between the queue
1622 	 * starting to freeze, by dropping the first reference in
1623 	 * blk_freeze_queue_start, and the moment the last request is
1624 	 * consumed, marked by the instant q_usage_counter reaches
1625 	 * zero.
1626 	 */
1627 	if (!percpu_ref_tryget(&q->q_usage_counter))
1628 		return;
1629 
1630 	/* check if there is any timed-out request */
1631 	blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
1632 	if (expired.has_timedout_rq) {
1633 		/*
1634 		 * Before walking tags, we must ensure any submit started
1635 		 * before the current time has finished. Since the submit
1636 		 * uses srcu or rcu, wait for a synchronization point to
1637 		 * ensure all running submits have finished
1638 		 */
1639 		blk_mq_wait_quiesce_done(q->tag_set);
1640 
1641 		expired.next = 0;
1642 		blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
1643 	}
1644 
1645 	if (expired.next != 0) {
1646 		mod_timer(&q->timeout, expired.next);
1647 	} else {
1648 		/*
1649 		 * Request timeouts are handled as a forward rolling timer. If
1650 		 * we end up here it means that no requests are pending and
1651 		 * also that no request has been pending for a while. Mark
1652 		 * each hctx as idle.
1653 		 */
1654 		queue_for_each_hw_ctx(q, hctx, i) {
1655 			/* the hctx may be unmapped, so check it here */
1656 			if (blk_mq_hw_queue_mapped(hctx))
1657 				blk_mq_tag_idle(hctx);
1658 		}
1659 	}
1660 	blk_queue_exit(q);
1661 }
1662 
1663 struct flush_busy_ctx_data {
1664 	struct blk_mq_hw_ctx *hctx;
1665 	struct list_head *list;
1666 };
1667 
1668 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1669 {
1670 	struct flush_busy_ctx_data *flush_data = data;
1671 	struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1672 	struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1673 	enum hctx_type type = hctx->type;
1674 
1675 	spin_lock(&ctx->lock);
1676 	list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1677 	sbitmap_clear_bit(sb, bitnr);
1678 	spin_unlock(&ctx->lock);
1679 	return true;
1680 }
1681 
1682 /*
1683  * Process software queues that have been marked busy, splicing them
1684  * to the for-dispatch
1685  */
1686 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1687 {
1688 	struct flush_busy_ctx_data data = {
1689 		.hctx = hctx,
1690 		.list = list,
1691 	};
1692 
1693 	sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1694 }
1695 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1696 
1697 struct dispatch_rq_data {
1698 	struct blk_mq_hw_ctx *hctx;
1699 	struct request *rq;
1700 };
1701 
1702 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1703 		void *data)
1704 {
1705 	struct dispatch_rq_data *dispatch_data = data;
1706 	struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1707 	struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1708 	enum hctx_type type = hctx->type;
1709 
1710 	spin_lock(&ctx->lock);
1711 	if (!list_empty(&ctx->rq_lists[type])) {
1712 		dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1713 		list_del_init(&dispatch_data->rq->queuelist);
1714 		if (list_empty(&ctx->rq_lists[type]))
1715 			sbitmap_clear_bit(sb, bitnr);
1716 	}
1717 	spin_unlock(&ctx->lock);
1718 
1719 	return !dispatch_data->rq;
1720 }
1721 
1722 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1723 					struct blk_mq_ctx *start)
1724 {
1725 	unsigned off = start ? start->index_hw[hctx->type] : 0;
1726 	struct dispatch_rq_data data = {
1727 		.hctx = hctx,
1728 		.rq   = NULL,
1729 	};
1730 
1731 	__sbitmap_for_each_set(&hctx->ctx_map, off,
1732 			       dispatch_rq_from_ctx, &data);
1733 
1734 	return data.rq;
1735 }
1736 
1737 static bool __blk_mq_alloc_driver_tag(struct request *rq)
1738 {
1739 	struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1740 	unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1741 	int tag;
1742 
1743 	blk_mq_tag_busy(rq->mq_hctx);
1744 
1745 	if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1746 		bt = &rq->mq_hctx->tags->breserved_tags;
1747 		tag_offset = 0;
1748 	} else {
1749 		if (!hctx_may_queue(rq->mq_hctx, bt))
1750 			return false;
1751 	}
1752 
1753 	tag = __sbitmap_queue_get(bt);
1754 	if (tag == BLK_MQ_NO_TAG)
1755 		return false;
1756 
1757 	rq->tag = tag + tag_offset;
1758 	return true;
1759 }
1760 
1761 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1762 {
1763 	if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1764 		return false;
1765 
1766 	if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1767 			!(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1768 		rq->rq_flags |= RQF_MQ_INFLIGHT;
1769 		__blk_mq_inc_active_requests(hctx);
1770 	}
1771 	hctx->tags->rqs[rq->tag] = rq;
1772 	return true;
1773 }
1774 
1775 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1776 				int flags, void *key)
1777 {
1778 	struct blk_mq_hw_ctx *hctx;
1779 
1780 	hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1781 
1782 	spin_lock(&hctx->dispatch_wait_lock);
1783 	if (!list_empty(&wait->entry)) {
1784 		struct sbitmap_queue *sbq;
1785 
1786 		list_del_init(&wait->entry);
1787 		sbq = &hctx->tags->bitmap_tags;
1788 		atomic_dec(&sbq->ws_active);
1789 	}
1790 	spin_unlock(&hctx->dispatch_wait_lock);
1791 
1792 	blk_mq_run_hw_queue(hctx, true);
1793 	return 1;
1794 }
1795 
1796 /*
1797  * Mark us waiting for a tag. For shared tags, this involves hooking us into
1798  * the tag wakeups. For non-shared tags, we can simply mark us needing a
1799  * restart. For both cases, take care to check the condition again after
1800  * marking us as waiting.
1801  */
1802 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1803 				 struct request *rq)
1804 {
1805 	struct sbitmap_queue *sbq;
1806 	struct wait_queue_head *wq;
1807 	wait_queue_entry_t *wait;
1808 	bool ret;
1809 
1810 	if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1811 	    !(blk_mq_is_shared_tags(hctx->flags))) {
1812 		blk_mq_sched_mark_restart_hctx(hctx);
1813 
1814 		/*
1815 		 * It's possible that a tag was freed in the window between the
1816 		 * allocation failure and adding the hardware queue to the wait
1817 		 * queue.
1818 		 *
1819 		 * Don't clear RESTART here, someone else could have set it.
1820 		 * At most this will cost an extra queue run.
1821 		 */
1822 		return blk_mq_get_driver_tag(rq);
1823 	}
1824 
1825 	wait = &hctx->dispatch_wait;
1826 	if (!list_empty_careful(&wait->entry))
1827 		return false;
1828 
1829 	if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag))
1830 		sbq = &hctx->tags->breserved_tags;
1831 	else
1832 		sbq = &hctx->tags->bitmap_tags;
1833 	wq = &bt_wait_ptr(sbq, hctx)->wait;
1834 
1835 	spin_lock_irq(&wq->lock);
1836 	spin_lock(&hctx->dispatch_wait_lock);
1837 	if (!list_empty(&wait->entry)) {
1838 		spin_unlock(&hctx->dispatch_wait_lock);
1839 		spin_unlock_irq(&wq->lock);
1840 		return false;
1841 	}
1842 
1843 	atomic_inc(&sbq->ws_active);
1844 	wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1845 	__add_wait_queue(wq, wait);
1846 
1847 	/*
1848 	 * It's possible that a tag was freed in the window between the
1849 	 * allocation failure and adding the hardware queue to the wait
1850 	 * queue.
1851 	 */
1852 	ret = blk_mq_get_driver_tag(rq);
1853 	if (!ret) {
1854 		spin_unlock(&hctx->dispatch_wait_lock);
1855 		spin_unlock_irq(&wq->lock);
1856 		return false;
1857 	}
1858 
1859 	/*
1860 	 * We got a tag, remove ourselves from the wait queue to ensure
1861 	 * someone else gets the wakeup.
1862 	 */
1863 	list_del_init(&wait->entry);
1864 	atomic_dec(&sbq->ws_active);
1865 	spin_unlock(&hctx->dispatch_wait_lock);
1866 	spin_unlock_irq(&wq->lock);
1867 
1868 	return true;
1869 }
1870 
1871 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT  8
1872 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR  4
1873 /*
1874  * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1875  * - EWMA is one simple way to compute running average value
1876  * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1877  * - take 4 as factor for avoiding to get too small(0) result, and this
1878  *   factor doesn't matter because EWMA decreases exponentially
1879  */
1880 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1881 {
1882 	unsigned int ewma;
1883 
1884 	ewma = hctx->dispatch_busy;
1885 
1886 	if (!ewma && !busy)
1887 		return;
1888 
1889 	ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1890 	if (busy)
1891 		ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1892 	ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1893 
1894 	hctx->dispatch_busy = ewma;
1895 }
1896 
1897 #define BLK_MQ_RESOURCE_DELAY	3		/* ms units */
1898 
1899 static void blk_mq_handle_dev_resource(struct request *rq,
1900 				       struct list_head *list)
1901 {
1902 	list_add(&rq->queuelist, list);
1903 	__blk_mq_requeue_request(rq);
1904 }
1905 
1906 static void blk_mq_handle_zone_resource(struct request *rq,
1907 					struct list_head *zone_list)
1908 {
1909 	/*
1910 	 * If we end up here it is because we cannot dispatch a request to a
1911 	 * specific zone due to LLD level zone-write locking or other zone
1912 	 * related resource not being available. In this case, set the request
1913 	 * aside in zone_list for retrying it later.
1914 	 */
1915 	list_add(&rq->queuelist, zone_list);
1916 	__blk_mq_requeue_request(rq);
1917 }
1918 
1919 enum prep_dispatch {
1920 	PREP_DISPATCH_OK,
1921 	PREP_DISPATCH_NO_TAG,
1922 	PREP_DISPATCH_NO_BUDGET,
1923 };
1924 
1925 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1926 						  bool need_budget)
1927 {
1928 	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1929 	int budget_token = -1;
1930 
1931 	if (need_budget) {
1932 		budget_token = blk_mq_get_dispatch_budget(rq->q);
1933 		if (budget_token < 0) {
1934 			blk_mq_put_driver_tag(rq);
1935 			return PREP_DISPATCH_NO_BUDGET;
1936 		}
1937 		blk_mq_set_rq_budget_token(rq, budget_token);
1938 	}
1939 
1940 	if (!blk_mq_get_driver_tag(rq)) {
1941 		/*
1942 		 * The initial allocation attempt failed, so we need to
1943 		 * rerun the hardware queue when a tag is freed. The
1944 		 * waitqueue takes care of that. If the queue is run
1945 		 * before we add this entry back on the dispatch list,
1946 		 * we'll re-run it below.
1947 		 */
1948 		if (!blk_mq_mark_tag_wait(hctx, rq)) {
1949 			/*
1950 			 * All budgets not got from this function will be put
1951 			 * together during handling partial dispatch
1952 			 */
1953 			if (need_budget)
1954 				blk_mq_put_dispatch_budget(rq->q, budget_token);
1955 			return PREP_DISPATCH_NO_TAG;
1956 		}
1957 	}
1958 
1959 	return PREP_DISPATCH_OK;
1960 }
1961 
1962 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1963 static void blk_mq_release_budgets(struct request_queue *q,
1964 		struct list_head *list)
1965 {
1966 	struct request *rq;
1967 
1968 	list_for_each_entry(rq, list, queuelist) {
1969 		int budget_token = blk_mq_get_rq_budget_token(rq);
1970 
1971 		if (budget_token >= 0)
1972 			blk_mq_put_dispatch_budget(q, budget_token);
1973 	}
1974 }
1975 
1976 /*
1977  * blk_mq_commit_rqs will notify driver using bd->last that there is no
1978  * more requests. (See comment in struct blk_mq_ops for commit_rqs for
1979  * details)
1980  * Attention, we should explicitly call this in unusual cases:
1981  *  1) did not queue everything initially scheduled to queue
1982  *  2) the last attempt to queue a request failed
1983  */
1984 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued,
1985 			      bool from_schedule)
1986 {
1987 	if (hctx->queue->mq_ops->commit_rqs && queued) {
1988 		trace_block_unplug(hctx->queue, queued, !from_schedule);
1989 		hctx->queue->mq_ops->commit_rqs(hctx);
1990 	}
1991 }
1992 
1993 /*
1994  * Returns true if we did some work AND can potentially do more.
1995  */
1996 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1997 			     unsigned int nr_budgets)
1998 {
1999 	enum prep_dispatch prep;
2000 	struct request_queue *q = hctx->queue;
2001 	struct request *rq;
2002 	int queued;
2003 	blk_status_t ret = BLK_STS_OK;
2004 	LIST_HEAD(zone_list);
2005 	bool needs_resource = false;
2006 
2007 	if (list_empty(list))
2008 		return false;
2009 
2010 	/*
2011 	 * Now process all the entries, sending them to the driver.
2012 	 */
2013 	queued = 0;
2014 	do {
2015 		struct blk_mq_queue_data bd;
2016 
2017 		rq = list_first_entry(list, struct request, queuelist);
2018 
2019 		WARN_ON_ONCE(hctx != rq->mq_hctx);
2020 		prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
2021 		if (prep != PREP_DISPATCH_OK)
2022 			break;
2023 
2024 		list_del_init(&rq->queuelist);
2025 
2026 		bd.rq = rq;
2027 		bd.last = list_empty(list);
2028 
2029 		/*
2030 		 * once the request is queued to lld, no need to cover the
2031 		 * budget any more
2032 		 */
2033 		if (nr_budgets)
2034 			nr_budgets--;
2035 		ret = q->mq_ops->queue_rq(hctx, &bd);
2036 		switch (ret) {
2037 		case BLK_STS_OK:
2038 			queued++;
2039 			break;
2040 		case BLK_STS_RESOURCE:
2041 			needs_resource = true;
2042 			fallthrough;
2043 		case BLK_STS_DEV_RESOURCE:
2044 			blk_mq_handle_dev_resource(rq, list);
2045 			goto out;
2046 		case BLK_STS_ZONE_RESOURCE:
2047 			/*
2048 			 * Move the request to zone_list and keep going through
2049 			 * the dispatch list to find more requests the drive can
2050 			 * accept.
2051 			 */
2052 			blk_mq_handle_zone_resource(rq, &zone_list);
2053 			needs_resource = true;
2054 			break;
2055 		default:
2056 			blk_mq_end_request(rq, ret);
2057 		}
2058 	} while (!list_empty(list));
2059 out:
2060 	if (!list_empty(&zone_list))
2061 		list_splice_tail_init(&zone_list, list);
2062 
2063 	/* If we didn't flush the entire list, we could have told the driver
2064 	 * there was more coming, but that turned out to be a lie.
2065 	 */
2066 	if (!list_empty(list) || ret != BLK_STS_OK)
2067 		blk_mq_commit_rqs(hctx, queued, false);
2068 
2069 	/*
2070 	 * Any items that need requeuing? Stuff them into hctx->dispatch,
2071 	 * that is where we will continue on next queue run.
2072 	 */
2073 	if (!list_empty(list)) {
2074 		bool needs_restart;
2075 		/* For non-shared tags, the RESTART check will suffice */
2076 		bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2077 			((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
2078 			blk_mq_is_shared_tags(hctx->flags));
2079 
2080 		if (nr_budgets)
2081 			blk_mq_release_budgets(q, list);
2082 
2083 		spin_lock(&hctx->lock);
2084 		list_splice_tail_init(list, &hctx->dispatch);
2085 		spin_unlock(&hctx->lock);
2086 
2087 		/*
2088 		 * Order adding requests to hctx->dispatch and checking
2089 		 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2090 		 * in blk_mq_sched_restart(). Avoid restart code path to
2091 		 * miss the new added requests to hctx->dispatch, meantime
2092 		 * SCHED_RESTART is observed here.
2093 		 */
2094 		smp_mb();
2095 
2096 		/*
2097 		 * If SCHED_RESTART was set by the caller of this function and
2098 		 * it is no longer set that means that it was cleared by another
2099 		 * thread and hence that a queue rerun is needed.
2100 		 *
2101 		 * If 'no_tag' is set, that means that we failed getting
2102 		 * a driver tag with an I/O scheduler attached. If our dispatch
2103 		 * waitqueue is no longer active, ensure that we run the queue
2104 		 * AFTER adding our entries back to the list.
2105 		 *
2106 		 * If no I/O scheduler has been configured it is possible that
2107 		 * the hardware queue got stopped and restarted before requests
2108 		 * were pushed back onto the dispatch list. Rerun the queue to
2109 		 * avoid starvation. Notes:
2110 		 * - blk_mq_run_hw_queue() checks whether or not a queue has
2111 		 *   been stopped before rerunning a queue.
2112 		 * - Some but not all block drivers stop a queue before
2113 		 *   returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2114 		 *   and dm-rq.
2115 		 *
2116 		 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2117 		 * bit is set, run queue after a delay to avoid IO stalls
2118 		 * that could otherwise occur if the queue is idle.  We'll do
2119 		 * similar if we couldn't get budget or couldn't lock a zone
2120 		 * and SCHED_RESTART is set.
2121 		 */
2122 		needs_restart = blk_mq_sched_needs_restart(hctx);
2123 		if (prep == PREP_DISPATCH_NO_BUDGET)
2124 			needs_resource = true;
2125 		if (!needs_restart ||
2126 		    (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2127 			blk_mq_run_hw_queue(hctx, true);
2128 		else if (needs_resource)
2129 			blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2130 
2131 		blk_mq_update_dispatch_busy(hctx, true);
2132 		return false;
2133 	}
2134 
2135 	blk_mq_update_dispatch_busy(hctx, false);
2136 	return true;
2137 }
2138 
2139 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2140 {
2141 	int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2142 
2143 	if (cpu >= nr_cpu_ids)
2144 		cpu = cpumask_first(hctx->cpumask);
2145 	return cpu;
2146 }
2147 
2148 /*
2149  * It'd be great if the workqueue API had a way to pass
2150  * in a mask and had some smarts for more clever placement.
2151  * For now we just round-robin here, switching for every
2152  * BLK_MQ_CPU_WORK_BATCH queued items.
2153  */
2154 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2155 {
2156 	bool tried = false;
2157 	int next_cpu = hctx->next_cpu;
2158 
2159 	if (hctx->queue->nr_hw_queues == 1)
2160 		return WORK_CPU_UNBOUND;
2161 
2162 	if (--hctx->next_cpu_batch <= 0) {
2163 select_cpu:
2164 		next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2165 				cpu_online_mask);
2166 		if (next_cpu >= nr_cpu_ids)
2167 			next_cpu = blk_mq_first_mapped_cpu(hctx);
2168 		hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2169 	}
2170 
2171 	/*
2172 	 * Do unbound schedule if we can't find a online CPU for this hctx,
2173 	 * and it should only happen in the path of handling CPU DEAD.
2174 	 */
2175 	if (!cpu_online(next_cpu)) {
2176 		if (!tried) {
2177 			tried = true;
2178 			goto select_cpu;
2179 		}
2180 
2181 		/*
2182 		 * Make sure to re-select CPU next time once after CPUs
2183 		 * in hctx->cpumask become online again.
2184 		 */
2185 		hctx->next_cpu = next_cpu;
2186 		hctx->next_cpu_batch = 1;
2187 		return WORK_CPU_UNBOUND;
2188 	}
2189 
2190 	hctx->next_cpu = next_cpu;
2191 	return next_cpu;
2192 }
2193 
2194 /**
2195  * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2196  * @hctx: Pointer to the hardware queue to run.
2197  * @msecs: Milliseconds of delay to wait before running the queue.
2198  *
2199  * Run a hardware queue asynchronously with a delay of @msecs.
2200  */
2201 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2202 {
2203 	if (unlikely(blk_mq_hctx_stopped(hctx)))
2204 		return;
2205 	kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2206 				    msecs_to_jiffies(msecs));
2207 }
2208 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2209 
2210 /**
2211  * blk_mq_run_hw_queue - Start to run a hardware queue.
2212  * @hctx: Pointer to the hardware queue to run.
2213  * @async: If we want to run the queue asynchronously.
2214  *
2215  * Check if the request queue is not in a quiesced state and if there are
2216  * pending requests to be sent. If this is true, run the queue to send requests
2217  * to hardware.
2218  */
2219 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2220 {
2221 	bool need_run;
2222 
2223 	/*
2224 	 * We can't run the queue inline with interrupts disabled.
2225 	 */
2226 	WARN_ON_ONCE(!async && in_interrupt());
2227 
2228 	/*
2229 	 * When queue is quiesced, we may be switching io scheduler, or
2230 	 * updating nr_hw_queues, or other things, and we can't run queue
2231 	 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2232 	 *
2233 	 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2234 	 * quiesced.
2235 	 */
2236 	__blk_mq_run_dispatch_ops(hctx->queue, false,
2237 		need_run = !blk_queue_quiesced(hctx->queue) &&
2238 		blk_mq_hctx_has_pending(hctx));
2239 
2240 	if (!need_run)
2241 		return;
2242 
2243 	if (async || (hctx->flags & BLK_MQ_F_BLOCKING) ||
2244 	    !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2245 		blk_mq_delay_run_hw_queue(hctx, 0);
2246 		return;
2247 	}
2248 
2249 	blk_mq_run_dispatch_ops(hctx->queue,
2250 				blk_mq_sched_dispatch_requests(hctx));
2251 }
2252 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2253 
2254 /*
2255  * Return prefered queue to dispatch from (if any) for non-mq aware IO
2256  * scheduler.
2257  */
2258 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2259 {
2260 	struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2261 	/*
2262 	 * If the IO scheduler does not respect hardware queues when
2263 	 * dispatching, we just don't bother with multiple HW queues and
2264 	 * dispatch from hctx for the current CPU since running multiple queues
2265 	 * just causes lock contention inside the scheduler and pointless cache
2266 	 * bouncing.
2267 	 */
2268 	struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2269 
2270 	if (!blk_mq_hctx_stopped(hctx))
2271 		return hctx;
2272 	return NULL;
2273 }
2274 
2275 /**
2276  * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2277  * @q: Pointer to the request queue to run.
2278  * @async: If we want to run the queue asynchronously.
2279  */
2280 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2281 {
2282 	struct blk_mq_hw_ctx *hctx, *sq_hctx;
2283 	unsigned long i;
2284 
2285 	sq_hctx = NULL;
2286 	if (blk_queue_sq_sched(q))
2287 		sq_hctx = blk_mq_get_sq_hctx(q);
2288 	queue_for_each_hw_ctx(q, hctx, i) {
2289 		if (blk_mq_hctx_stopped(hctx))
2290 			continue;
2291 		/*
2292 		 * Dispatch from this hctx either if there's no hctx preferred
2293 		 * by IO scheduler or if it has requests that bypass the
2294 		 * scheduler.
2295 		 */
2296 		if (!sq_hctx || sq_hctx == hctx ||
2297 		    !list_empty_careful(&hctx->dispatch))
2298 			blk_mq_run_hw_queue(hctx, async);
2299 	}
2300 }
2301 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2302 
2303 /**
2304  * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2305  * @q: Pointer to the request queue to run.
2306  * @msecs: Milliseconds of delay to wait before running the queues.
2307  */
2308 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2309 {
2310 	struct blk_mq_hw_ctx *hctx, *sq_hctx;
2311 	unsigned long i;
2312 
2313 	sq_hctx = NULL;
2314 	if (blk_queue_sq_sched(q))
2315 		sq_hctx = blk_mq_get_sq_hctx(q);
2316 	queue_for_each_hw_ctx(q, hctx, i) {
2317 		if (blk_mq_hctx_stopped(hctx))
2318 			continue;
2319 		/*
2320 		 * If there is already a run_work pending, leave the
2321 		 * pending delay untouched. Otherwise, a hctx can stall
2322 		 * if another hctx is re-delaying the other's work
2323 		 * before the work executes.
2324 		 */
2325 		if (delayed_work_pending(&hctx->run_work))
2326 			continue;
2327 		/*
2328 		 * Dispatch from this hctx either if there's no hctx preferred
2329 		 * by IO scheduler or if it has requests that bypass the
2330 		 * scheduler.
2331 		 */
2332 		if (!sq_hctx || sq_hctx == hctx ||
2333 		    !list_empty_careful(&hctx->dispatch))
2334 			blk_mq_delay_run_hw_queue(hctx, msecs);
2335 	}
2336 }
2337 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2338 
2339 /*
2340  * This function is often used for pausing .queue_rq() by driver when
2341  * there isn't enough resource or some conditions aren't satisfied, and
2342  * BLK_STS_RESOURCE is usually returned.
2343  *
2344  * We do not guarantee that dispatch can be drained or blocked
2345  * after blk_mq_stop_hw_queue() returns. Please use
2346  * blk_mq_quiesce_queue() for that requirement.
2347  */
2348 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2349 {
2350 	cancel_delayed_work(&hctx->run_work);
2351 
2352 	set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2353 }
2354 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2355 
2356 /*
2357  * This function is often used for pausing .queue_rq() by driver when
2358  * there isn't enough resource or some conditions aren't satisfied, and
2359  * BLK_STS_RESOURCE is usually returned.
2360  *
2361  * We do not guarantee that dispatch can be drained or blocked
2362  * after blk_mq_stop_hw_queues() returns. Please use
2363  * blk_mq_quiesce_queue() for that requirement.
2364  */
2365 void blk_mq_stop_hw_queues(struct request_queue *q)
2366 {
2367 	struct blk_mq_hw_ctx *hctx;
2368 	unsigned long i;
2369 
2370 	queue_for_each_hw_ctx(q, hctx, i)
2371 		blk_mq_stop_hw_queue(hctx);
2372 }
2373 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2374 
2375 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2376 {
2377 	clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2378 
2379 	blk_mq_run_hw_queue(hctx, false);
2380 }
2381 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2382 
2383 void blk_mq_start_hw_queues(struct request_queue *q)
2384 {
2385 	struct blk_mq_hw_ctx *hctx;
2386 	unsigned long i;
2387 
2388 	queue_for_each_hw_ctx(q, hctx, i)
2389 		blk_mq_start_hw_queue(hctx);
2390 }
2391 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2392 
2393 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2394 {
2395 	if (!blk_mq_hctx_stopped(hctx))
2396 		return;
2397 
2398 	clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2399 	blk_mq_run_hw_queue(hctx, async);
2400 }
2401 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2402 
2403 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2404 {
2405 	struct blk_mq_hw_ctx *hctx;
2406 	unsigned long i;
2407 
2408 	queue_for_each_hw_ctx(q, hctx, i)
2409 		blk_mq_start_stopped_hw_queue(hctx, async);
2410 }
2411 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2412 
2413 static void blk_mq_run_work_fn(struct work_struct *work)
2414 {
2415 	struct blk_mq_hw_ctx *hctx =
2416 		container_of(work, struct blk_mq_hw_ctx, run_work.work);
2417 
2418 	blk_mq_run_dispatch_ops(hctx->queue,
2419 				blk_mq_sched_dispatch_requests(hctx));
2420 }
2421 
2422 /**
2423  * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2424  * @rq: Pointer to request to be inserted.
2425  * @flags: BLK_MQ_INSERT_*
2426  *
2427  * Should only be used carefully, when the caller knows we want to
2428  * bypass a potential IO scheduler on the target device.
2429  */
2430 static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags)
2431 {
2432 	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2433 
2434 	spin_lock(&hctx->lock);
2435 	if (flags & BLK_MQ_INSERT_AT_HEAD)
2436 		list_add(&rq->queuelist, &hctx->dispatch);
2437 	else
2438 		list_add_tail(&rq->queuelist, &hctx->dispatch);
2439 	spin_unlock(&hctx->lock);
2440 }
2441 
2442 static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx,
2443 		struct blk_mq_ctx *ctx, struct list_head *list,
2444 		bool run_queue_async)
2445 {
2446 	struct request *rq;
2447 	enum hctx_type type = hctx->type;
2448 
2449 	/*
2450 	 * Try to issue requests directly if the hw queue isn't busy to save an
2451 	 * extra enqueue & dequeue to the sw queue.
2452 	 */
2453 	if (!hctx->dispatch_busy && !run_queue_async) {
2454 		blk_mq_run_dispatch_ops(hctx->queue,
2455 			blk_mq_try_issue_list_directly(hctx, list));
2456 		if (list_empty(list))
2457 			goto out;
2458 	}
2459 
2460 	/*
2461 	 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2462 	 * offline now
2463 	 */
2464 	list_for_each_entry(rq, list, queuelist) {
2465 		BUG_ON(rq->mq_ctx != ctx);
2466 		trace_block_rq_insert(rq);
2467 	}
2468 
2469 	spin_lock(&ctx->lock);
2470 	list_splice_tail_init(list, &ctx->rq_lists[type]);
2471 	blk_mq_hctx_mark_pending(hctx, ctx);
2472 	spin_unlock(&ctx->lock);
2473 out:
2474 	blk_mq_run_hw_queue(hctx, run_queue_async);
2475 }
2476 
2477 static void blk_mq_insert_request(struct request *rq, blk_insert_t flags)
2478 {
2479 	struct request_queue *q = rq->q;
2480 	struct blk_mq_ctx *ctx = rq->mq_ctx;
2481 	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2482 
2483 	if (blk_rq_is_passthrough(rq)) {
2484 		/*
2485 		 * Passthrough request have to be added to hctx->dispatch
2486 		 * directly.  The device may be in a situation where it can't
2487 		 * handle FS request, and always returns BLK_STS_RESOURCE for
2488 		 * them, which gets them added to hctx->dispatch.
2489 		 *
2490 		 * If a passthrough request is required to unblock the queues,
2491 		 * and it is added to the scheduler queue, there is no chance to
2492 		 * dispatch it given we prioritize requests in hctx->dispatch.
2493 		 */
2494 		blk_mq_request_bypass_insert(rq, flags);
2495 	} else if (req_op(rq) == REQ_OP_FLUSH) {
2496 		/*
2497 		 * Firstly normal IO request is inserted to scheduler queue or
2498 		 * sw queue, meantime we add flush request to dispatch queue(
2499 		 * hctx->dispatch) directly and there is at most one in-flight
2500 		 * flush request for each hw queue, so it doesn't matter to add
2501 		 * flush request to tail or front of the dispatch queue.
2502 		 *
2503 		 * Secondly in case of NCQ, flush request belongs to non-NCQ
2504 		 * command, and queueing it will fail when there is any
2505 		 * in-flight normal IO request(NCQ command). When adding flush
2506 		 * rq to the front of hctx->dispatch, it is easier to introduce
2507 		 * extra time to flush rq's latency because of S_SCHED_RESTART
2508 		 * compared with adding to the tail of dispatch queue, then
2509 		 * chance of flush merge is increased, and less flush requests
2510 		 * will be issued to controller. It is observed that ~10% time
2511 		 * is saved in blktests block/004 on disk attached to AHCI/NCQ
2512 		 * drive when adding flush rq to the front of hctx->dispatch.
2513 		 *
2514 		 * Simply queue flush rq to the front of hctx->dispatch so that
2515 		 * intensive flush workloads can benefit in case of NCQ HW.
2516 		 */
2517 		blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD);
2518 	} else if (q->elevator) {
2519 		LIST_HEAD(list);
2520 
2521 		WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG);
2522 
2523 		list_add(&rq->queuelist, &list);
2524 		q->elevator->type->ops.insert_requests(hctx, &list, flags);
2525 	} else {
2526 		trace_block_rq_insert(rq);
2527 
2528 		spin_lock(&ctx->lock);
2529 		if (flags & BLK_MQ_INSERT_AT_HEAD)
2530 			list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]);
2531 		else
2532 			list_add_tail(&rq->queuelist,
2533 				      &ctx->rq_lists[hctx->type]);
2534 		blk_mq_hctx_mark_pending(hctx, ctx);
2535 		spin_unlock(&ctx->lock);
2536 	}
2537 }
2538 
2539 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2540 		unsigned int nr_segs)
2541 {
2542 	int err;
2543 
2544 	if (bio->bi_opf & REQ_RAHEAD)
2545 		rq->cmd_flags |= REQ_FAILFAST_MASK;
2546 
2547 	rq->__sector = bio->bi_iter.bi_sector;
2548 	blk_rq_bio_prep(rq, bio, nr_segs);
2549 
2550 	/* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2551 	err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2552 	WARN_ON_ONCE(err);
2553 
2554 	blk_account_io_start(rq);
2555 }
2556 
2557 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2558 					    struct request *rq, bool last)
2559 {
2560 	struct request_queue *q = rq->q;
2561 	struct blk_mq_queue_data bd = {
2562 		.rq = rq,
2563 		.last = last,
2564 	};
2565 	blk_status_t ret;
2566 
2567 	/*
2568 	 * For OK queue, we are done. For error, caller may kill it.
2569 	 * Any other error (busy), just add it to our list as we
2570 	 * previously would have done.
2571 	 */
2572 	ret = q->mq_ops->queue_rq(hctx, &bd);
2573 	switch (ret) {
2574 	case BLK_STS_OK:
2575 		blk_mq_update_dispatch_busy(hctx, false);
2576 		break;
2577 	case BLK_STS_RESOURCE:
2578 	case BLK_STS_DEV_RESOURCE:
2579 		blk_mq_update_dispatch_busy(hctx, true);
2580 		__blk_mq_requeue_request(rq);
2581 		break;
2582 	default:
2583 		blk_mq_update_dispatch_busy(hctx, false);
2584 		break;
2585 	}
2586 
2587 	return ret;
2588 }
2589 
2590 static bool blk_mq_get_budget_and_tag(struct request *rq)
2591 {
2592 	int budget_token;
2593 
2594 	budget_token = blk_mq_get_dispatch_budget(rq->q);
2595 	if (budget_token < 0)
2596 		return false;
2597 	blk_mq_set_rq_budget_token(rq, budget_token);
2598 	if (!blk_mq_get_driver_tag(rq)) {
2599 		blk_mq_put_dispatch_budget(rq->q, budget_token);
2600 		return false;
2601 	}
2602 	return true;
2603 }
2604 
2605 /**
2606  * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2607  * @hctx: Pointer of the associated hardware queue.
2608  * @rq: Pointer to request to be sent.
2609  *
2610  * If the device has enough resources to accept a new request now, send the
2611  * request directly to device driver. Else, insert at hctx->dispatch queue, so
2612  * we can try send it another time in the future. Requests inserted at this
2613  * queue have higher priority.
2614  */
2615 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2616 		struct request *rq)
2617 {
2618 	blk_status_t ret;
2619 
2620 	if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2621 		blk_mq_insert_request(rq, 0);
2622 		return;
2623 	}
2624 
2625 	if ((rq->rq_flags & RQF_USE_SCHED) || !blk_mq_get_budget_and_tag(rq)) {
2626 		blk_mq_insert_request(rq, 0);
2627 		blk_mq_run_hw_queue(hctx, false);
2628 		return;
2629 	}
2630 
2631 	ret = __blk_mq_issue_directly(hctx, rq, true);
2632 	switch (ret) {
2633 	case BLK_STS_OK:
2634 		break;
2635 	case BLK_STS_RESOURCE:
2636 	case BLK_STS_DEV_RESOURCE:
2637 		blk_mq_request_bypass_insert(rq, 0);
2638 		blk_mq_run_hw_queue(hctx, false);
2639 		break;
2640 	default:
2641 		blk_mq_end_request(rq, ret);
2642 		break;
2643 	}
2644 }
2645 
2646 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2647 {
2648 	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2649 
2650 	if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2651 		blk_mq_insert_request(rq, 0);
2652 		return BLK_STS_OK;
2653 	}
2654 
2655 	if (!blk_mq_get_budget_and_tag(rq))
2656 		return BLK_STS_RESOURCE;
2657 	return __blk_mq_issue_directly(hctx, rq, last);
2658 }
2659 
2660 static void blk_mq_plug_issue_direct(struct blk_plug *plug)
2661 {
2662 	struct blk_mq_hw_ctx *hctx = NULL;
2663 	struct request *rq;
2664 	int queued = 0;
2665 	blk_status_t ret = BLK_STS_OK;
2666 
2667 	while ((rq = rq_list_pop(&plug->mq_list))) {
2668 		bool last = rq_list_empty(plug->mq_list);
2669 
2670 		if (hctx != rq->mq_hctx) {
2671 			if (hctx) {
2672 				blk_mq_commit_rqs(hctx, queued, false);
2673 				queued = 0;
2674 			}
2675 			hctx = rq->mq_hctx;
2676 		}
2677 
2678 		ret = blk_mq_request_issue_directly(rq, last);
2679 		switch (ret) {
2680 		case BLK_STS_OK:
2681 			queued++;
2682 			break;
2683 		case BLK_STS_RESOURCE:
2684 		case BLK_STS_DEV_RESOURCE:
2685 			blk_mq_request_bypass_insert(rq, 0);
2686 			blk_mq_run_hw_queue(hctx, false);
2687 			goto out;
2688 		default:
2689 			blk_mq_end_request(rq, ret);
2690 			break;
2691 		}
2692 	}
2693 
2694 out:
2695 	if (ret != BLK_STS_OK)
2696 		blk_mq_commit_rqs(hctx, queued, false);
2697 }
2698 
2699 static void __blk_mq_flush_plug_list(struct request_queue *q,
2700 				     struct blk_plug *plug)
2701 {
2702 	if (blk_queue_quiesced(q))
2703 		return;
2704 	q->mq_ops->queue_rqs(&plug->mq_list);
2705 }
2706 
2707 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2708 {
2709 	struct blk_mq_hw_ctx *this_hctx = NULL;
2710 	struct blk_mq_ctx *this_ctx = NULL;
2711 	struct request *requeue_list = NULL;
2712 	struct request **requeue_lastp = &requeue_list;
2713 	unsigned int depth = 0;
2714 	bool is_passthrough = false;
2715 	LIST_HEAD(list);
2716 
2717 	do {
2718 		struct request *rq = rq_list_pop(&plug->mq_list);
2719 
2720 		if (!this_hctx) {
2721 			this_hctx = rq->mq_hctx;
2722 			this_ctx = rq->mq_ctx;
2723 			is_passthrough = blk_rq_is_passthrough(rq);
2724 		} else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx ||
2725 			   is_passthrough != blk_rq_is_passthrough(rq)) {
2726 			rq_list_add_tail(&requeue_lastp, rq);
2727 			continue;
2728 		}
2729 		list_add(&rq->queuelist, &list);
2730 		depth++;
2731 	} while (!rq_list_empty(plug->mq_list));
2732 
2733 	plug->mq_list = requeue_list;
2734 	trace_block_unplug(this_hctx->queue, depth, !from_sched);
2735 
2736 	percpu_ref_get(&this_hctx->queue->q_usage_counter);
2737 	/* passthrough requests should never be issued to the I/O scheduler */
2738 	if (is_passthrough) {
2739 		spin_lock(&this_hctx->lock);
2740 		list_splice_tail_init(&list, &this_hctx->dispatch);
2741 		spin_unlock(&this_hctx->lock);
2742 		blk_mq_run_hw_queue(this_hctx, from_sched);
2743 	} else if (this_hctx->queue->elevator) {
2744 		this_hctx->queue->elevator->type->ops.insert_requests(this_hctx,
2745 				&list, 0);
2746 		blk_mq_run_hw_queue(this_hctx, from_sched);
2747 	} else {
2748 		blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched);
2749 	}
2750 	percpu_ref_put(&this_hctx->queue->q_usage_counter);
2751 }
2752 
2753 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2754 {
2755 	struct request *rq;
2756 
2757 	/*
2758 	 * We may have been called recursively midway through handling
2759 	 * plug->mq_list via a schedule() in the driver's queue_rq() callback.
2760 	 * To avoid mq_list changing under our feet, clear rq_count early and
2761 	 * bail out specifically if rq_count is 0 rather than checking
2762 	 * whether the mq_list is empty.
2763 	 */
2764 	if (plug->rq_count == 0)
2765 		return;
2766 	plug->rq_count = 0;
2767 
2768 	if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2769 		struct request_queue *q;
2770 
2771 		rq = rq_list_peek(&plug->mq_list);
2772 		q = rq->q;
2773 
2774 		/*
2775 		 * Peek first request and see if we have a ->queue_rqs() hook.
2776 		 * If we do, we can dispatch the whole plug list in one go. We
2777 		 * already know at this point that all requests belong to the
2778 		 * same queue, caller must ensure that's the case.
2779 		 *
2780 		 * Since we pass off the full list to the driver at this point,
2781 		 * we do not increment the active request count for the queue.
2782 		 * Bypass shared tags for now because of that.
2783 		 */
2784 		if (q->mq_ops->queue_rqs &&
2785 		    !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2786 			blk_mq_run_dispatch_ops(q,
2787 				__blk_mq_flush_plug_list(q, plug));
2788 			if (rq_list_empty(plug->mq_list))
2789 				return;
2790 		}
2791 
2792 		blk_mq_run_dispatch_ops(q,
2793 				blk_mq_plug_issue_direct(plug));
2794 		if (rq_list_empty(plug->mq_list))
2795 			return;
2796 	}
2797 
2798 	do {
2799 		blk_mq_dispatch_plug_list(plug, from_schedule);
2800 	} while (!rq_list_empty(plug->mq_list));
2801 }
2802 
2803 static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2804 		struct list_head *list)
2805 {
2806 	int queued = 0;
2807 	blk_status_t ret = BLK_STS_OK;
2808 
2809 	while (!list_empty(list)) {
2810 		struct request *rq = list_first_entry(list, struct request,
2811 				queuelist);
2812 
2813 		list_del_init(&rq->queuelist);
2814 		ret = blk_mq_request_issue_directly(rq, list_empty(list));
2815 		switch (ret) {
2816 		case BLK_STS_OK:
2817 			queued++;
2818 			break;
2819 		case BLK_STS_RESOURCE:
2820 		case BLK_STS_DEV_RESOURCE:
2821 			blk_mq_request_bypass_insert(rq, 0);
2822 			if (list_empty(list))
2823 				blk_mq_run_hw_queue(hctx, false);
2824 			goto out;
2825 		default:
2826 			blk_mq_end_request(rq, ret);
2827 			break;
2828 		}
2829 	}
2830 
2831 out:
2832 	if (ret != BLK_STS_OK)
2833 		blk_mq_commit_rqs(hctx, queued, false);
2834 }
2835 
2836 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2837 				     struct bio *bio, unsigned int nr_segs)
2838 {
2839 	if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2840 		if (blk_attempt_plug_merge(q, bio, nr_segs))
2841 			return true;
2842 		if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2843 			return true;
2844 	}
2845 	return false;
2846 }
2847 
2848 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2849 					       struct blk_plug *plug,
2850 					       struct bio *bio,
2851 					       unsigned int nsegs)
2852 {
2853 	struct blk_mq_alloc_data data = {
2854 		.q		= q,
2855 		.nr_tags	= 1,
2856 		.cmd_flags	= bio->bi_opf,
2857 	};
2858 	struct request *rq;
2859 
2860 	if (unlikely(bio_queue_enter(bio)))
2861 		return NULL;
2862 
2863 	if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2864 		goto queue_exit;
2865 
2866 	rq_qos_throttle(q, bio);
2867 
2868 	if (plug) {
2869 		data.nr_tags = plug->nr_ios;
2870 		plug->nr_ios = 1;
2871 		data.cached_rq = &plug->cached_rq;
2872 	}
2873 
2874 	rq = __blk_mq_alloc_requests(&data);
2875 	if (rq)
2876 		return rq;
2877 	rq_qos_cleanup(q, bio);
2878 	if (bio->bi_opf & REQ_NOWAIT)
2879 		bio_wouldblock_error(bio);
2880 queue_exit:
2881 	blk_queue_exit(q);
2882 	return NULL;
2883 }
2884 
2885 static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
2886 		struct blk_plug *plug, struct bio **bio, unsigned int nsegs)
2887 {
2888 	struct request *rq;
2889 	enum hctx_type type, hctx_type;
2890 
2891 	if (!plug)
2892 		return NULL;
2893 	rq = rq_list_peek(&plug->cached_rq);
2894 	if (!rq || rq->q != q)
2895 		return NULL;
2896 
2897 	if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) {
2898 		*bio = NULL;
2899 		return NULL;
2900 	}
2901 
2902 	type = blk_mq_get_hctx_type((*bio)->bi_opf);
2903 	hctx_type = rq->mq_hctx->type;
2904 	if (type != hctx_type &&
2905 	    !(type == HCTX_TYPE_READ && hctx_type == HCTX_TYPE_DEFAULT))
2906 		return NULL;
2907 	if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf))
2908 		return NULL;
2909 
2910 	/*
2911 	 * If any qos ->throttle() end up blocking, we will have flushed the
2912 	 * plug and hence killed the cached_rq list as well. Pop this entry
2913 	 * before we throttle.
2914 	 */
2915 	plug->cached_rq = rq_list_next(rq);
2916 	rq_qos_throttle(q, *bio);
2917 
2918 	blk_mq_rq_time_init(rq, 0);
2919 	rq->cmd_flags = (*bio)->bi_opf;
2920 	INIT_LIST_HEAD(&rq->queuelist);
2921 	return rq;
2922 }
2923 
2924 static void bio_set_ioprio(struct bio *bio)
2925 {
2926 	/* Nobody set ioprio so far? Initialize it based on task's nice value */
2927 	if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE)
2928 		bio->bi_ioprio = get_current_ioprio();
2929 	blkcg_set_ioprio(bio);
2930 }
2931 
2932 /**
2933  * blk_mq_submit_bio - Create and send a request to block device.
2934  * @bio: Bio pointer.
2935  *
2936  * Builds up a request structure from @q and @bio and send to the device. The
2937  * request may not be queued directly to hardware if:
2938  * * This request can be merged with another one
2939  * * We want to place request at plug queue for possible future merging
2940  * * There is an IO scheduler active at this queue
2941  *
2942  * It will not queue the request if there is an error with the bio, or at the
2943  * request creation.
2944  */
2945 void blk_mq_submit_bio(struct bio *bio)
2946 {
2947 	struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2948 	struct blk_plug *plug = blk_mq_plug(bio);
2949 	const int is_sync = op_is_sync(bio->bi_opf);
2950 	struct blk_mq_hw_ctx *hctx;
2951 	struct request *rq;
2952 	unsigned int nr_segs = 1;
2953 	blk_status_t ret;
2954 
2955 	bio = blk_queue_bounce(bio, q);
2956 	if (bio_may_exceed_limits(bio, &q->limits)) {
2957 		bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2958 		if (!bio)
2959 			return;
2960 	}
2961 
2962 	if (!bio_integrity_prep(bio))
2963 		return;
2964 
2965 	bio_set_ioprio(bio);
2966 
2967 	rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs);
2968 	if (!rq) {
2969 		if (!bio)
2970 			return;
2971 		rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2972 		if (unlikely(!rq))
2973 			return;
2974 	}
2975 
2976 	trace_block_getrq(bio);
2977 
2978 	rq_qos_track(q, rq, bio);
2979 
2980 	blk_mq_bio_to_request(rq, bio, nr_segs);
2981 
2982 	ret = blk_crypto_rq_get_keyslot(rq);
2983 	if (ret != BLK_STS_OK) {
2984 		bio->bi_status = ret;
2985 		bio_endio(bio);
2986 		blk_mq_free_request(rq);
2987 		return;
2988 	}
2989 
2990 	if (op_is_flush(bio->bi_opf) && blk_insert_flush(rq))
2991 		return;
2992 
2993 	if (plug) {
2994 		blk_add_rq_to_plug(plug, rq);
2995 		return;
2996 	}
2997 
2998 	hctx = rq->mq_hctx;
2999 	if ((rq->rq_flags & RQF_USE_SCHED) ||
3000 	    (hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) {
3001 		blk_mq_insert_request(rq, 0);
3002 		blk_mq_run_hw_queue(hctx, true);
3003 	} else {
3004 		blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq));
3005 	}
3006 }
3007 
3008 #ifdef CONFIG_BLK_MQ_STACKING
3009 /**
3010  * blk_insert_cloned_request - Helper for stacking drivers to submit a request
3011  * @rq: the request being queued
3012  */
3013 blk_status_t blk_insert_cloned_request(struct request *rq)
3014 {
3015 	struct request_queue *q = rq->q;
3016 	unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
3017 	unsigned int max_segments = blk_rq_get_max_segments(rq);
3018 	blk_status_t ret;
3019 
3020 	if (blk_rq_sectors(rq) > max_sectors) {
3021 		/*
3022 		 * SCSI device does not have a good way to return if
3023 		 * Write Same/Zero is actually supported. If a device rejects
3024 		 * a non-read/write command (discard, write same,etc.) the
3025 		 * low-level device driver will set the relevant queue limit to
3026 		 * 0 to prevent blk-lib from issuing more of the offending
3027 		 * operations. Commands queued prior to the queue limit being
3028 		 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3029 		 * errors being propagated to upper layers.
3030 		 */
3031 		if (max_sectors == 0)
3032 			return BLK_STS_NOTSUPP;
3033 
3034 		printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3035 			__func__, blk_rq_sectors(rq), max_sectors);
3036 		return BLK_STS_IOERR;
3037 	}
3038 
3039 	/*
3040 	 * The queue settings related to segment counting may differ from the
3041 	 * original queue.
3042 	 */
3043 	rq->nr_phys_segments = blk_recalc_rq_segments(rq);
3044 	if (rq->nr_phys_segments > max_segments) {
3045 		printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n",
3046 			__func__, rq->nr_phys_segments, max_segments);
3047 		return BLK_STS_IOERR;
3048 	}
3049 
3050 	if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
3051 		return BLK_STS_IOERR;
3052 
3053 	ret = blk_crypto_rq_get_keyslot(rq);
3054 	if (ret != BLK_STS_OK)
3055 		return ret;
3056 
3057 	blk_account_io_start(rq);
3058 
3059 	/*
3060 	 * Since we have a scheduler attached on the top device,
3061 	 * bypass a potential scheduler on the bottom device for
3062 	 * insert.
3063 	 */
3064 	blk_mq_run_dispatch_ops(q,
3065 			ret = blk_mq_request_issue_directly(rq, true));
3066 	if (ret)
3067 		blk_account_io_done(rq, ktime_get_ns());
3068 	return ret;
3069 }
3070 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3071 
3072 /**
3073  * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3074  * @rq: the clone request to be cleaned up
3075  *
3076  * Description:
3077  *     Free all bios in @rq for a cloned request.
3078  */
3079 void blk_rq_unprep_clone(struct request *rq)
3080 {
3081 	struct bio *bio;
3082 
3083 	while ((bio = rq->bio) != NULL) {
3084 		rq->bio = bio->bi_next;
3085 
3086 		bio_put(bio);
3087 	}
3088 }
3089 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3090 
3091 /**
3092  * blk_rq_prep_clone - Helper function to setup clone request
3093  * @rq: the request to be setup
3094  * @rq_src: original request to be cloned
3095  * @bs: bio_set that bios for clone are allocated from
3096  * @gfp_mask: memory allocation mask for bio
3097  * @bio_ctr: setup function to be called for each clone bio.
3098  *           Returns %0 for success, non %0 for failure.
3099  * @data: private data to be passed to @bio_ctr
3100  *
3101  * Description:
3102  *     Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3103  *     Also, pages which the original bios are pointing to are not copied
3104  *     and the cloned bios just point same pages.
3105  *     So cloned bios must be completed before original bios, which means
3106  *     the caller must complete @rq before @rq_src.
3107  */
3108 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3109 		      struct bio_set *bs, gfp_t gfp_mask,
3110 		      int (*bio_ctr)(struct bio *, struct bio *, void *),
3111 		      void *data)
3112 {
3113 	struct bio *bio, *bio_src;
3114 
3115 	if (!bs)
3116 		bs = &fs_bio_set;
3117 
3118 	__rq_for_each_bio(bio_src, rq_src) {
3119 		bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3120 				      bs);
3121 		if (!bio)
3122 			goto free_and_out;
3123 
3124 		if (bio_ctr && bio_ctr(bio, bio_src, data))
3125 			goto free_and_out;
3126 
3127 		if (rq->bio) {
3128 			rq->biotail->bi_next = bio;
3129 			rq->biotail = bio;
3130 		} else {
3131 			rq->bio = rq->biotail = bio;
3132 		}
3133 		bio = NULL;
3134 	}
3135 
3136 	/* Copy attributes of the original request to the clone request. */
3137 	rq->__sector = blk_rq_pos(rq_src);
3138 	rq->__data_len = blk_rq_bytes(rq_src);
3139 	if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3140 		rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3141 		rq->special_vec = rq_src->special_vec;
3142 	}
3143 	rq->nr_phys_segments = rq_src->nr_phys_segments;
3144 	rq->ioprio = rq_src->ioprio;
3145 
3146 	if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3147 		goto free_and_out;
3148 
3149 	return 0;
3150 
3151 free_and_out:
3152 	if (bio)
3153 		bio_put(bio);
3154 	blk_rq_unprep_clone(rq);
3155 
3156 	return -ENOMEM;
3157 }
3158 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3159 #endif /* CONFIG_BLK_MQ_STACKING */
3160 
3161 /*
3162  * Steal bios from a request and add them to a bio list.
3163  * The request must not have been partially completed before.
3164  */
3165 void blk_steal_bios(struct bio_list *list, struct request *rq)
3166 {
3167 	if (rq->bio) {
3168 		if (list->tail)
3169 			list->tail->bi_next = rq->bio;
3170 		else
3171 			list->head = rq->bio;
3172 		list->tail = rq->biotail;
3173 
3174 		rq->bio = NULL;
3175 		rq->biotail = NULL;
3176 	}
3177 
3178 	rq->__data_len = 0;
3179 }
3180 EXPORT_SYMBOL_GPL(blk_steal_bios);
3181 
3182 static size_t order_to_size(unsigned int order)
3183 {
3184 	return (size_t)PAGE_SIZE << order;
3185 }
3186 
3187 /* called before freeing request pool in @tags */
3188 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3189 				    struct blk_mq_tags *tags)
3190 {
3191 	struct page *page;
3192 	unsigned long flags;
3193 
3194 	/*
3195 	 * There is no need to clear mapping if driver tags is not initialized
3196 	 * or the mapping belongs to the driver tags.
3197 	 */
3198 	if (!drv_tags || drv_tags == tags)
3199 		return;
3200 
3201 	list_for_each_entry(page, &tags->page_list, lru) {
3202 		unsigned long start = (unsigned long)page_address(page);
3203 		unsigned long end = start + order_to_size(page->private);
3204 		int i;
3205 
3206 		for (i = 0; i < drv_tags->nr_tags; i++) {
3207 			struct request *rq = drv_tags->rqs[i];
3208 			unsigned long rq_addr = (unsigned long)rq;
3209 
3210 			if (rq_addr >= start && rq_addr < end) {
3211 				WARN_ON_ONCE(req_ref_read(rq) != 0);
3212 				cmpxchg(&drv_tags->rqs[i], rq, NULL);
3213 			}
3214 		}
3215 	}
3216 
3217 	/*
3218 	 * Wait until all pending iteration is done.
3219 	 *
3220 	 * Request reference is cleared and it is guaranteed to be observed
3221 	 * after the ->lock is released.
3222 	 */
3223 	spin_lock_irqsave(&drv_tags->lock, flags);
3224 	spin_unlock_irqrestore(&drv_tags->lock, flags);
3225 }
3226 
3227 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3228 		     unsigned int hctx_idx)
3229 {
3230 	struct blk_mq_tags *drv_tags;
3231 	struct page *page;
3232 
3233 	if (list_empty(&tags->page_list))
3234 		return;
3235 
3236 	if (blk_mq_is_shared_tags(set->flags))
3237 		drv_tags = set->shared_tags;
3238 	else
3239 		drv_tags = set->tags[hctx_idx];
3240 
3241 	if (tags->static_rqs && set->ops->exit_request) {
3242 		int i;
3243 
3244 		for (i = 0; i < tags->nr_tags; i++) {
3245 			struct request *rq = tags->static_rqs[i];
3246 
3247 			if (!rq)
3248 				continue;
3249 			set->ops->exit_request(set, rq, hctx_idx);
3250 			tags->static_rqs[i] = NULL;
3251 		}
3252 	}
3253 
3254 	blk_mq_clear_rq_mapping(drv_tags, tags);
3255 
3256 	while (!list_empty(&tags->page_list)) {
3257 		page = list_first_entry(&tags->page_list, struct page, lru);
3258 		list_del_init(&page->lru);
3259 		/*
3260 		 * Remove kmemleak object previously allocated in
3261 		 * blk_mq_alloc_rqs().
3262 		 */
3263 		kmemleak_free(page_address(page));
3264 		__free_pages(page, page->private);
3265 	}
3266 }
3267 
3268 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3269 {
3270 	kfree(tags->rqs);
3271 	tags->rqs = NULL;
3272 	kfree(tags->static_rqs);
3273 	tags->static_rqs = NULL;
3274 
3275 	blk_mq_free_tags(tags);
3276 }
3277 
3278 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3279 		unsigned int hctx_idx)
3280 {
3281 	int i;
3282 
3283 	for (i = 0; i < set->nr_maps; i++) {
3284 		unsigned int start = set->map[i].queue_offset;
3285 		unsigned int end = start + set->map[i].nr_queues;
3286 
3287 		if (hctx_idx >= start && hctx_idx < end)
3288 			break;
3289 	}
3290 
3291 	if (i >= set->nr_maps)
3292 		i = HCTX_TYPE_DEFAULT;
3293 
3294 	return i;
3295 }
3296 
3297 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3298 		unsigned int hctx_idx)
3299 {
3300 	enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3301 
3302 	return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3303 }
3304 
3305 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3306 					       unsigned int hctx_idx,
3307 					       unsigned int nr_tags,
3308 					       unsigned int reserved_tags)
3309 {
3310 	int node = blk_mq_get_hctx_node(set, hctx_idx);
3311 	struct blk_mq_tags *tags;
3312 
3313 	if (node == NUMA_NO_NODE)
3314 		node = set->numa_node;
3315 
3316 	tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3317 				BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3318 	if (!tags)
3319 		return NULL;
3320 
3321 	tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3322 				 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3323 				 node);
3324 	if (!tags->rqs)
3325 		goto err_free_tags;
3326 
3327 	tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3328 					GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3329 					node);
3330 	if (!tags->static_rqs)
3331 		goto err_free_rqs;
3332 
3333 	return tags;
3334 
3335 err_free_rqs:
3336 	kfree(tags->rqs);
3337 err_free_tags:
3338 	blk_mq_free_tags(tags);
3339 	return NULL;
3340 }
3341 
3342 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3343 			       unsigned int hctx_idx, int node)
3344 {
3345 	int ret;
3346 
3347 	if (set->ops->init_request) {
3348 		ret = set->ops->init_request(set, rq, hctx_idx, node);
3349 		if (ret)
3350 			return ret;
3351 	}
3352 
3353 	WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3354 	return 0;
3355 }
3356 
3357 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3358 			    struct blk_mq_tags *tags,
3359 			    unsigned int hctx_idx, unsigned int depth)
3360 {
3361 	unsigned int i, j, entries_per_page, max_order = 4;
3362 	int node = blk_mq_get_hctx_node(set, hctx_idx);
3363 	size_t rq_size, left;
3364 
3365 	if (node == NUMA_NO_NODE)
3366 		node = set->numa_node;
3367 
3368 	INIT_LIST_HEAD(&tags->page_list);
3369 
3370 	/*
3371 	 * rq_size is the size of the request plus driver payload, rounded
3372 	 * to the cacheline size
3373 	 */
3374 	rq_size = round_up(sizeof(struct request) + set->cmd_size,
3375 				cache_line_size());
3376 	left = rq_size * depth;
3377 
3378 	for (i = 0; i < depth; ) {
3379 		int this_order = max_order;
3380 		struct page *page;
3381 		int to_do;
3382 		void *p;
3383 
3384 		while (this_order && left < order_to_size(this_order - 1))
3385 			this_order--;
3386 
3387 		do {
3388 			page = alloc_pages_node(node,
3389 				GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3390 				this_order);
3391 			if (page)
3392 				break;
3393 			if (!this_order--)
3394 				break;
3395 			if (order_to_size(this_order) < rq_size)
3396 				break;
3397 		} while (1);
3398 
3399 		if (!page)
3400 			goto fail;
3401 
3402 		page->private = this_order;
3403 		list_add_tail(&page->lru, &tags->page_list);
3404 
3405 		p = page_address(page);
3406 		/*
3407 		 * Allow kmemleak to scan these pages as they contain pointers
3408 		 * to additional allocations like via ops->init_request().
3409 		 */
3410 		kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3411 		entries_per_page = order_to_size(this_order) / rq_size;
3412 		to_do = min(entries_per_page, depth - i);
3413 		left -= to_do * rq_size;
3414 		for (j = 0; j < to_do; j++) {
3415 			struct request *rq = p;
3416 
3417 			tags->static_rqs[i] = rq;
3418 			if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3419 				tags->static_rqs[i] = NULL;
3420 				goto fail;
3421 			}
3422 
3423 			p += rq_size;
3424 			i++;
3425 		}
3426 	}
3427 	return 0;
3428 
3429 fail:
3430 	blk_mq_free_rqs(set, tags, hctx_idx);
3431 	return -ENOMEM;
3432 }
3433 
3434 struct rq_iter_data {
3435 	struct blk_mq_hw_ctx *hctx;
3436 	bool has_rq;
3437 };
3438 
3439 static bool blk_mq_has_request(struct request *rq, void *data)
3440 {
3441 	struct rq_iter_data *iter_data = data;
3442 
3443 	if (rq->mq_hctx != iter_data->hctx)
3444 		return true;
3445 	iter_data->has_rq = true;
3446 	return false;
3447 }
3448 
3449 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3450 {
3451 	struct blk_mq_tags *tags = hctx->sched_tags ?
3452 			hctx->sched_tags : hctx->tags;
3453 	struct rq_iter_data data = {
3454 		.hctx	= hctx,
3455 	};
3456 
3457 	blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3458 	return data.has_rq;
3459 }
3460 
3461 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3462 		struct blk_mq_hw_ctx *hctx)
3463 {
3464 	if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3465 		return false;
3466 	if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3467 		return false;
3468 	return true;
3469 }
3470 
3471 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3472 {
3473 	struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3474 			struct blk_mq_hw_ctx, cpuhp_online);
3475 
3476 	if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3477 	    !blk_mq_last_cpu_in_hctx(cpu, hctx))
3478 		return 0;
3479 
3480 	/*
3481 	 * Prevent new request from being allocated on the current hctx.
3482 	 *
3483 	 * The smp_mb__after_atomic() Pairs with the implied barrier in
3484 	 * test_and_set_bit_lock in sbitmap_get().  Ensures the inactive flag is
3485 	 * seen once we return from the tag allocator.
3486 	 */
3487 	set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3488 	smp_mb__after_atomic();
3489 
3490 	/*
3491 	 * Try to grab a reference to the queue and wait for any outstanding
3492 	 * requests.  If we could not grab a reference the queue has been
3493 	 * frozen and there are no requests.
3494 	 */
3495 	if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3496 		while (blk_mq_hctx_has_requests(hctx))
3497 			msleep(5);
3498 		percpu_ref_put(&hctx->queue->q_usage_counter);
3499 	}
3500 
3501 	return 0;
3502 }
3503 
3504 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3505 {
3506 	struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3507 			struct blk_mq_hw_ctx, cpuhp_online);
3508 
3509 	if (cpumask_test_cpu(cpu, hctx->cpumask))
3510 		clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3511 	return 0;
3512 }
3513 
3514 /*
3515  * 'cpu' is going away. splice any existing rq_list entries from this
3516  * software queue to the hw queue dispatch list, and ensure that it
3517  * gets run.
3518  */
3519 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3520 {
3521 	struct blk_mq_hw_ctx *hctx;
3522 	struct blk_mq_ctx *ctx;
3523 	LIST_HEAD(tmp);
3524 	enum hctx_type type;
3525 
3526 	hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3527 	if (!cpumask_test_cpu(cpu, hctx->cpumask))
3528 		return 0;
3529 
3530 	ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3531 	type = hctx->type;
3532 
3533 	spin_lock(&ctx->lock);
3534 	if (!list_empty(&ctx->rq_lists[type])) {
3535 		list_splice_init(&ctx->rq_lists[type], &tmp);
3536 		blk_mq_hctx_clear_pending(hctx, ctx);
3537 	}
3538 	spin_unlock(&ctx->lock);
3539 
3540 	if (list_empty(&tmp))
3541 		return 0;
3542 
3543 	spin_lock(&hctx->lock);
3544 	list_splice_tail_init(&tmp, &hctx->dispatch);
3545 	spin_unlock(&hctx->lock);
3546 
3547 	blk_mq_run_hw_queue(hctx, true);
3548 	return 0;
3549 }
3550 
3551 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3552 {
3553 	if (!(hctx->flags & BLK_MQ_F_STACKING))
3554 		cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3555 						    &hctx->cpuhp_online);
3556 	cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3557 					    &hctx->cpuhp_dead);
3558 }
3559 
3560 /*
3561  * Before freeing hw queue, clearing the flush request reference in
3562  * tags->rqs[] for avoiding potential UAF.
3563  */
3564 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3565 		unsigned int queue_depth, struct request *flush_rq)
3566 {
3567 	int i;
3568 	unsigned long flags;
3569 
3570 	/* The hw queue may not be mapped yet */
3571 	if (!tags)
3572 		return;
3573 
3574 	WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3575 
3576 	for (i = 0; i < queue_depth; i++)
3577 		cmpxchg(&tags->rqs[i], flush_rq, NULL);
3578 
3579 	/*
3580 	 * Wait until all pending iteration is done.
3581 	 *
3582 	 * Request reference is cleared and it is guaranteed to be observed
3583 	 * after the ->lock is released.
3584 	 */
3585 	spin_lock_irqsave(&tags->lock, flags);
3586 	spin_unlock_irqrestore(&tags->lock, flags);
3587 }
3588 
3589 /* hctx->ctxs will be freed in queue's release handler */
3590 static void blk_mq_exit_hctx(struct request_queue *q,
3591 		struct blk_mq_tag_set *set,
3592 		struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3593 {
3594 	struct request *flush_rq = hctx->fq->flush_rq;
3595 
3596 	if (blk_mq_hw_queue_mapped(hctx))
3597 		blk_mq_tag_idle(hctx);
3598 
3599 	if (blk_queue_init_done(q))
3600 		blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3601 				set->queue_depth, flush_rq);
3602 	if (set->ops->exit_request)
3603 		set->ops->exit_request(set, flush_rq, hctx_idx);
3604 
3605 	if (set->ops->exit_hctx)
3606 		set->ops->exit_hctx(hctx, hctx_idx);
3607 
3608 	blk_mq_remove_cpuhp(hctx);
3609 
3610 	xa_erase(&q->hctx_table, hctx_idx);
3611 
3612 	spin_lock(&q->unused_hctx_lock);
3613 	list_add(&hctx->hctx_list, &q->unused_hctx_list);
3614 	spin_unlock(&q->unused_hctx_lock);
3615 }
3616 
3617 static void blk_mq_exit_hw_queues(struct request_queue *q,
3618 		struct blk_mq_tag_set *set, int nr_queue)
3619 {
3620 	struct blk_mq_hw_ctx *hctx;
3621 	unsigned long i;
3622 
3623 	queue_for_each_hw_ctx(q, hctx, i) {
3624 		if (i == nr_queue)
3625 			break;
3626 		blk_mq_exit_hctx(q, set, hctx, i);
3627 	}
3628 }
3629 
3630 static int blk_mq_init_hctx(struct request_queue *q,
3631 		struct blk_mq_tag_set *set,
3632 		struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3633 {
3634 	hctx->queue_num = hctx_idx;
3635 
3636 	if (!(hctx->flags & BLK_MQ_F_STACKING))
3637 		cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3638 				&hctx->cpuhp_online);
3639 	cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3640 
3641 	hctx->tags = set->tags[hctx_idx];
3642 
3643 	if (set->ops->init_hctx &&
3644 	    set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3645 		goto unregister_cpu_notifier;
3646 
3647 	if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3648 				hctx->numa_node))
3649 		goto exit_hctx;
3650 
3651 	if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3652 		goto exit_flush_rq;
3653 
3654 	return 0;
3655 
3656  exit_flush_rq:
3657 	if (set->ops->exit_request)
3658 		set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3659  exit_hctx:
3660 	if (set->ops->exit_hctx)
3661 		set->ops->exit_hctx(hctx, hctx_idx);
3662  unregister_cpu_notifier:
3663 	blk_mq_remove_cpuhp(hctx);
3664 	return -1;
3665 }
3666 
3667 static struct blk_mq_hw_ctx *
3668 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3669 		int node)
3670 {
3671 	struct blk_mq_hw_ctx *hctx;
3672 	gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3673 
3674 	hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3675 	if (!hctx)
3676 		goto fail_alloc_hctx;
3677 
3678 	if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3679 		goto free_hctx;
3680 
3681 	atomic_set(&hctx->nr_active, 0);
3682 	if (node == NUMA_NO_NODE)
3683 		node = set->numa_node;
3684 	hctx->numa_node = node;
3685 
3686 	INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3687 	spin_lock_init(&hctx->lock);
3688 	INIT_LIST_HEAD(&hctx->dispatch);
3689 	hctx->queue = q;
3690 	hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3691 
3692 	INIT_LIST_HEAD(&hctx->hctx_list);
3693 
3694 	/*
3695 	 * Allocate space for all possible cpus to avoid allocation at
3696 	 * runtime
3697 	 */
3698 	hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3699 			gfp, node);
3700 	if (!hctx->ctxs)
3701 		goto free_cpumask;
3702 
3703 	if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3704 				gfp, node, false, false))
3705 		goto free_ctxs;
3706 	hctx->nr_ctx = 0;
3707 
3708 	spin_lock_init(&hctx->dispatch_wait_lock);
3709 	init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3710 	INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3711 
3712 	hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3713 	if (!hctx->fq)
3714 		goto free_bitmap;
3715 
3716 	blk_mq_hctx_kobj_init(hctx);
3717 
3718 	return hctx;
3719 
3720  free_bitmap:
3721 	sbitmap_free(&hctx->ctx_map);
3722  free_ctxs:
3723 	kfree(hctx->ctxs);
3724  free_cpumask:
3725 	free_cpumask_var(hctx->cpumask);
3726  free_hctx:
3727 	kfree(hctx);
3728  fail_alloc_hctx:
3729 	return NULL;
3730 }
3731 
3732 static void blk_mq_init_cpu_queues(struct request_queue *q,
3733 				   unsigned int nr_hw_queues)
3734 {
3735 	struct blk_mq_tag_set *set = q->tag_set;
3736 	unsigned int i, j;
3737 
3738 	for_each_possible_cpu(i) {
3739 		struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3740 		struct blk_mq_hw_ctx *hctx;
3741 		int k;
3742 
3743 		__ctx->cpu = i;
3744 		spin_lock_init(&__ctx->lock);
3745 		for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3746 			INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3747 
3748 		__ctx->queue = q;
3749 
3750 		/*
3751 		 * Set local node, IFF we have more than one hw queue. If
3752 		 * not, we remain on the home node of the device
3753 		 */
3754 		for (j = 0; j < set->nr_maps; j++) {
3755 			hctx = blk_mq_map_queue_type(q, j, i);
3756 			if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3757 				hctx->numa_node = cpu_to_node(i);
3758 		}
3759 	}
3760 }
3761 
3762 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3763 					     unsigned int hctx_idx,
3764 					     unsigned int depth)
3765 {
3766 	struct blk_mq_tags *tags;
3767 	int ret;
3768 
3769 	tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3770 	if (!tags)
3771 		return NULL;
3772 
3773 	ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3774 	if (ret) {
3775 		blk_mq_free_rq_map(tags);
3776 		return NULL;
3777 	}
3778 
3779 	return tags;
3780 }
3781 
3782 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3783 				       int hctx_idx)
3784 {
3785 	if (blk_mq_is_shared_tags(set->flags)) {
3786 		set->tags[hctx_idx] = set->shared_tags;
3787 
3788 		return true;
3789 	}
3790 
3791 	set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3792 						       set->queue_depth);
3793 
3794 	return set->tags[hctx_idx];
3795 }
3796 
3797 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3798 			     struct blk_mq_tags *tags,
3799 			     unsigned int hctx_idx)
3800 {
3801 	if (tags) {
3802 		blk_mq_free_rqs(set, tags, hctx_idx);
3803 		blk_mq_free_rq_map(tags);
3804 	}
3805 }
3806 
3807 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3808 				      unsigned int hctx_idx)
3809 {
3810 	if (!blk_mq_is_shared_tags(set->flags))
3811 		blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3812 
3813 	set->tags[hctx_idx] = NULL;
3814 }
3815 
3816 static void blk_mq_map_swqueue(struct request_queue *q)
3817 {
3818 	unsigned int j, hctx_idx;
3819 	unsigned long i;
3820 	struct blk_mq_hw_ctx *hctx;
3821 	struct blk_mq_ctx *ctx;
3822 	struct blk_mq_tag_set *set = q->tag_set;
3823 
3824 	queue_for_each_hw_ctx(q, hctx, i) {
3825 		cpumask_clear(hctx->cpumask);
3826 		hctx->nr_ctx = 0;
3827 		hctx->dispatch_from = NULL;
3828 	}
3829 
3830 	/*
3831 	 * Map software to hardware queues.
3832 	 *
3833 	 * If the cpu isn't present, the cpu is mapped to first hctx.
3834 	 */
3835 	for_each_possible_cpu(i) {
3836 
3837 		ctx = per_cpu_ptr(q->queue_ctx, i);
3838 		for (j = 0; j < set->nr_maps; j++) {
3839 			if (!set->map[j].nr_queues) {
3840 				ctx->hctxs[j] = blk_mq_map_queue_type(q,
3841 						HCTX_TYPE_DEFAULT, i);
3842 				continue;
3843 			}
3844 			hctx_idx = set->map[j].mq_map[i];
3845 			/* unmapped hw queue can be remapped after CPU topo changed */
3846 			if (!set->tags[hctx_idx] &&
3847 			    !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3848 				/*
3849 				 * If tags initialization fail for some hctx,
3850 				 * that hctx won't be brought online.  In this
3851 				 * case, remap the current ctx to hctx[0] which
3852 				 * is guaranteed to always have tags allocated
3853 				 */
3854 				set->map[j].mq_map[i] = 0;
3855 			}
3856 
3857 			hctx = blk_mq_map_queue_type(q, j, i);
3858 			ctx->hctxs[j] = hctx;
3859 			/*
3860 			 * If the CPU is already set in the mask, then we've
3861 			 * mapped this one already. This can happen if
3862 			 * devices share queues across queue maps.
3863 			 */
3864 			if (cpumask_test_cpu(i, hctx->cpumask))
3865 				continue;
3866 
3867 			cpumask_set_cpu(i, hctx->cpumask);
3868 			hctx->type = j;
3869 			ctx->index_hw[hctx->type] = hctx->nr_ctx;
3870 			hctx->ctxs[hctx->nr_ctx++] = ctx;
3871 
3872 			/*
3873 			 * If the nr_ctx type overflows, we have exceeded the
3874 			 * amount of sw queues we can support.
3875 			 */
3876 			BUG_ON(!hctx->nr_ctx);
3877 		}
3878 
3879 		for (; j < HCTX_MAX_TYPES; j++)
3880 			ctx->hctxs[j] = blk_mq_map_queue_type(q,
3881 					HCTX_TYPE_DEFAULT, i);
3882 	}
3883 
3884 	queue_for_each_hw_ctx(q, hctx, i) {
3885 		/*
3886 		 * If no software queues are mapped to this hardware queue,
3887 		 * disable it and free the request entries.
3888 		 */
3889 		if (!hctx->nr_ctx) {
3890 			/* Never unmap queue 0.  We need it as a
3891 			 * fallback in case of a new remap fails
3892 			 * allocation
3893 			 */
3894 			if (i)
3895 				__blk_mq_free_map_and_rqs(set, i);
3896 
3897 			hctx->tags = NULL;
3898 			continue;
3899 		}
3900 
3901 		hctx->tags = set->tags[i];
3902 		WARN_ON(!hctx->tags);
3903 
3904 		/*
3905 		 * Set the map size to the number of mapped software queues.
3906 		 * This is more accurate and more efficient than looping
3907 		 * over all possibly mapped software queues.
3908 		 */
3909 		sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3910 
3911 		/*
3912 		 * Initialize batch roundrobin counts
3913 		 */
3914 		hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3915 		hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3916 	}
3917 }
3918 
3919 /*
3920  * Caller needs to ensure that we're either frozen/quiesced, or that
3921  * the queue isn't live yet.
3922  */
3923 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3924 {
3925 	struct blk_mq_hw_ctx *hctx;
3926 	unsigned long i;
3927 
3928 	queue_for_each_hw_ctx(q, hctx, i) {
3929 		if (shared) {
3930 			hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3931 		} else {
3932 			blk_mq_tag_idle(hctx);
3933 			hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3934 		}
3935 	}
3936 }
3937 
3938 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3939 					 bool shared)
3940 {
3941 	struct request_queue *q;
3942 
3943 	lockdep_assert_held(&set->tag_list_lock);
3944 
3945 	list_for_each_entry(q, &set->tag_list, tag_set_list) {
3946 		blk_mq_freeze_queue(q);
3947 		queue_set_hctx_shared(q, shared);
3948 		blk_mq_unfreeze_queue(q);
3949 	}
3950 }
3951 
3952 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3953 {
3954 	struct blk_mq_tag_set *set = q->tag_set;
3955 
3956 	mutex_lock(&set->tag_list_lock);
3957 	list_del(&q->tag_set_list);
3958 	if (list_is_singular(&set->tag_list)) {
3959 		/* just transitioned to unshared */
3960 		set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3961 		/* update existing queue */
3962 		blk_mq_update_tag_set_shared(set, false);
3963 	}
3964 	mutex_unlock(&set->tag_list_lock);
3965 	INIT_LIST_HEAD(&q->tag_set_list);
3966 }
3967 
3968 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3969 				     struct request_queue *q)
3970 {
3971 	mutex_lock(&set->tag_list_lock);
3972 
3973 	/*
3974 	 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3975 	 */
3976 	if (!list_empty(&set->tag_list) &&
3977 	    !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3978 		set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3979 		/* update existing queue */
3980 		blk_mq_update_tag_set_shared(set, true);
3981 	}
3982 	if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3983 		queue_set_hctx_shared(q, true);
3984 	list_add_tail(&q->tag_set_list, &set->tag_list);
3985 
3986 	mutex_unlock(&set->tag_list_lock);
3987 }
3988 
3989 /* All allocations will be freed in release handler of q->mq_kobj */
3990 static int blk_mq_alloc_ctxs(struct request_queue *q)
3991 {
3992 	struct blk_mq_ctxs *ctxs;
3993 	int cpu;
3994 
3995 	ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3996 	if (!ctxs)
3997 		return -ENOMEM;
3998 
3999 	ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
4000 	if (!ctxs->queue_ctx)
4001 		goto fail;
4002 
4003 	for_each_possible_cpu(cpu) {
4004 		struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
4005 		ctx->ctxs = ctxs;
4006 	}
4007 
4008 	q->mq_kobj = &ctxs->kobj;
4009 	q->queue_ctx = ctxs->queue_ctx;
4010 
4011 	return 0;
4012  fail:
4013 	kfree(ctxs);
4014 	return -ENOMEM;
4015 }
4016 
4017 /*
4018  * It is the actual release handler for mq, but we do it from
4019  * request queue's release handler for avoiding use-after-free
4020  * and headache because q->mq_kobj shouldn't have been introduced,
4021  * but we can't group ctx/kctx kobj without it.
4022  */
4023 void blk_mq_release(struct request_queue *q)
4024 {
4025 	struct blk_mq_hw_ctx *hctx, *next;
4026 	unsigned long i;
4027 
4028 	queue_for_each_hw_ctx(q, hctx, i)
4029 		WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4030 
4031 	/* all hctx are in .unused_hctx_list now */
4032 	list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4033 		list_del_init(&hctx->hctx_list);
4034 		kobject_put(&hctx->kobj);
4035 	}
4036 
4037 	xa_destroy(&q->hctx_table);
4038 
4039 	/*
4040 	 * release .mq_kobj and sw queue's kobject now because
4041 	 * both share lifetime with request queue.
4042 	 */
4043 	blk_mq_sysfs_deinit(q);
4044 }
4045 
4046 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
4047 		void *queuedata)
4048 {
4049 	struct request_queue *q;
4050 	int ret;
4051 
4052 	q = blk_alloc_queue(set->numa_node);
4053 	if (!q)
4054 		return ERR_PTR(-ENOMEM);
4055 	q->queuedata = queuedata;
4056 	ret = blk_mq_init_allocated_queue(set, q);
4057 	if (ret) {
4058 		blk_put_queue(q);
4059 		return ERR_PTR(ret);
4060 	}
4061 	return q;
4062 }
4063 
4064 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
4065 {
4066 	return blk_mq_init_queue_data(set, NULL);
4067 }
4068 EXPORT_SYMBOL(blk_mq_init_queue);
4069 
4070 /**
4071  * blk_mq_destroy_queue - shutdown a request queue
4072  * @q: request queue to shutdown
4073  *
4074  * This shuts down a request queue allocated by blk_mq_init_queue(). All future
4075  * requests will be failed with -ENODEV. The caller is responsible for dropping
4076  * the reference from blk_mq_init_queue() by calling blk_put_queue().
4077  *
4078  * Context: can sleep
4079  */
4080 void blk_mq_destroy_queue(struct request_queue *q)
4081 {
4082 	WARN_ON_ONCE(!queue_is_mq(q));
4083 	WARN_ON_ONCE(blk_queue_registered(q));
4084 
4085 	might_sleep();
4086 
4087 	blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4088 	blk_queue_start_drain(q);
4089 	blk_mq_freeze_queue_wait(q);
4090 
4091 	blk_sync_queue(q);
4092 	blk_mq_cancel_work_sync(q);
4093 	blk_mq_exit_queue(q);
4094 }
4095 EXPORT_SYMBOL(blk_mq_destroy_queue);
4096 
4097 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
4098 		struct lock_class_key *lkclass)
4099 {
4100 	struct request_queue *q;
4101 	struct gendisk *disk;
4102 
4103 	q = blk_mq_init_queue_data(set, queuedata);
4104 	if (IS_ERR(q))
4105 		return ERR_CAST(q);
4106 
4107 	disk = __alloc_disk_node(q, set->numa_node, lkclass);
4108 	if (!disk) {
4109 		blk_mq_destroy_queue(q);
4110 		blk_put_queue(q);
4111 		return ERR_PTR(-ENOMEM);
4112 	}
4113 	set_bit(GD_OWNS_QUEUE, &disk->state);
4114 	return disk;
4115 }
4116 EXPORT_SYMBOL(__blk_mq_alloc_disk);
4117 
4118 struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4119 		struct lock_class_key *lkclass)
4120 {
4121 	struct gendisk *disk;
4122 
4123 	if (!blk_get_queue(q))
4124 		return NULL;
4125 	disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4126 	if (!disk)
4127 		blk_put_queue(q);
4128 	return disk;
4129 }
4130 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4131 
4132 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4133 		struct blk_mq_tag_set *set, struct request_queue *q,
4134 		int hctx_idx, int node)
4135 {
4136 	struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4137 
4138 	/* reuse dead hctx first */
4139 	spin_lock(&q->unused_hctx_lock);
4140 	list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4141 		if (tmp->numa_node == node) {
4142 			hctx = tmp;
4143 			break;
4144 		}
4145 	}
4146 	if (hctx)
4147 		list_del_init(&hctx->hctx_list);
4148 	spin_unlock(&q->unused_hctx_lock);
4149 
4150 	if (!hctx)
4151 		hctx = blk_mq_alloc_hctx(q, set, node);
4152 	if (!hctx)
4153 		goto fail;
4154 
4155 	if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4156 		goto free_hctx;
4157 
4158 	return hctx;
4159 
4160  free_hctx:
4161 	kobject_put(&hctx->kobj);
4162  fail:
4163 	return NULL;
4164 }
4165 
4166 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4167 						struct request_queue *q)
4168 {
4169 	struct blk_mq_hw_ctx *hctx;
4170 	unsigned long i, j;
4171 
4172 	/* protect against switching io scheduler  */
4173 	mutex_lock(&q->sysfs_lock);
4174 	for (i = 0; i < set->nr_hw_queues; i++) {
4175 		int old_node;
4176 		int node = blk_mq_get_hctx_node(set, i);
4177 		struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4178 
4179 		if (old_hctx) {
4180 			old_node = old_hctx->numa_node;
4181 			blk_mq_exit_hctx(q, set, old_hctx, i);
4182 		}
4183 
4184 		if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4185 			if (!old_hctx)
4186 				break;
4187 			pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4188 					node, old_node);
4189 			hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4190 			WARN_ON_ONCE(!hctx);
4191 		}
4192 	}
4193 	/*
4194 	 * Increasing nr_hw_queues fails. Free the newly allocated
4195 	 * hctxs and keep the previous q->nr_hw_queues.
4196 	 */
4197 	if (i != set->nr_hw_queues) {
4198 		j = q->nr_hw_queues;
4199 	} else {
4200 		j = i;
4201 		q->nr_hw_queues = set->nr_hw_queues;
4202 	}
4203 
4204 	xa_for_each_start(&q->hctx_table, j, hctx, j)
4205 		blk_mq_exit_hctx(q, set, hctx, j);
4206 	mutex_unlock(&q->sysfs_lock);
4207 }
4208 
4209 static void blk_mq_update_poll_flag(struct request_queue *q)
4210 {
4211 	struct blk_mq_tag_set *set = q->tag_set;
4212 
4213 	if (set->nr_maps > HCTX_TYPE_POLL &&
4214 	    set->map[HCTX_TYPE_POLL].nr_queues)
4215 		blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4216 	else
4217 		blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4218 }
4219 
4220 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4221 		struct request_queue *q)
4222 {
4223 	/* mark the queue as mq asap */
4224 	q->mq_ops = set->ops;
4225 
4226 	if (blk_mq_alloc_ctxs(q))
4227 		goto err_exit;
4228 
4229 	/* init q->mq_kobj and sw queues' kobjects */
4230 	blk_mq_sysfs_init(q);
4231 
4232 	INIT_LIST_HEAD(&q->unused_hctx_list);
4233 	spin_lock_init(&q->unused_hctx_lock);
4234 
4235 	xa_init(&q->hctx_table);
4236 
4237 	blk_mq_realloc_hw_ctxs(set, q);
4238 	if (!q->nr_hw_queues)
4239 		goto err_hctxs;
4240 
4241 	INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4242 	blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4243 
4244 	q->tag_set = set;
4245 
4246 	q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4247 	blk_mq_update_poll_flag(q);
4248 
4249 	INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4250 	INIT_LIST_HEAD(&q->flush_list);
4251 	INIT_LIST_HEAD(&q->requeue_list);
4252 	spin_lock_init(&q->requeue_lock);
4253 
4254 	q->nr_requests = set->queue_depth;
4255 
4256 	blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4257 	blk_mq_add_queue_tag_set(set, q);
4258 	blk_mq_map_swqueue(q);
4259 	return 0;
4260 
4261 err_hctxs:
4262 	blk_mq_release(q);
4263 err_exit:
4264 	q->mq_ops = NULL;
4265 	return -ENOMEM;
4266 }
4267 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4268 
4269 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4270 void blk_mq_exit_queue(struct request_queue *q)
4271 {
4272 	struct blk_mq_tag_set *set = q->tag_set;
4273 
4274 	/* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4275 	blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4276 	/* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4277 	blk_mq_del_queue_tag_set(q);
4278 }
4279 
4280 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4281 {
4282 	int i;
4283 
4284 	if (blk_mq_is_shared_tags(set->flags)) {
4285 		set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4286 						BLK_MQ_NO_HCTX_IDX,
4287 						set->queue_depth);
4288 		if (!set->shared_tags)
4289 			return -ENOMEM;
4290 	}
4291 
4292 	for (i = 0; i < set->nr_hw_queues; i++) {
4293 		if (!__blk_mq_alloc_map_and_rqs(set, i))
4294 			goto out_unwind;
4295 		cond_resched();
4296 	}
4297 
4298 	return 0;
4299 
4300 out_unwind:
4301 	while (--i >= 0)
4302 		__blk_mq_free_map_and_rqs(set, i);
4303 
4304 	if (blk_mq_is_shared_tags(set->flags)) {
4305 		blk_mq_free_map_and_rqs(set, set->shared_tags,
4306 					BLK_MQ_NO_HCTX_IDX);
4307 	}
4308 
4309 	return -ENOMEM;
4310 }
4311 
4312 /*
4313  * Allocate the request maps associated with this tag_set. Note that this
4314  * may reduce the depth asked for, if memory is tight. set->queue_depth
4315  * will be updated to reflect the allocated depth.
4316  */
4317 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4318 {
4319 	unsigned int depth;
4320 	int err;
4321 
4322 	depth = set->queue_depth;
4323 	do {
4324 		err = __blk_mq_alloc_rq_maps(set);
4325 		if (!err)
4326 			break;
4327 
4328 		set->queue_depth >>= 1;
4329 		if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4330 			err = -ENOMEM;
4331 			break;
4332 		}
4333 	} while (set->queue_depth);
4334 
4335 	if (!set->queue_depth || err) {
4336 		pr_err("blk-mq: failed to allocate request map\n");
4337 		return -ENOMEM;
4338 	}
4339 
4340 	if (depth != set->queue_depth)
4341 		pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4342 						depth, set->queue_depth);
4343 
4344 	return 0;
4345 }
4346 
4347 static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4348 {
4349 	/*
4350 	 * blk_mq_map_queues() and multiple .map_queues() implementations
4351 	 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4352 	 * number of hardware queues.
4353 	 */
4354 	if (set->nr_maps == 1)
4355 		set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4356 
4357 	if (set->ops->map_queues && !is_kdump_kernel()) {
4358 		int i;
4359 
4360 		/*
4361 		 * transport .map_queues is usually done in the following
4362 		 * way:
4363 		 *
4364 		 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4365 		 * 	mask = get_cpu_mask(queue)
4366 		 * 	for_each_cpu(cpu, mask)
4367 		 * 		set->map[x].mq_map[cpu] = queue;
4368 		 * }
4369 		 *
4370 		 * When we need to remap, the table has to be cleared for
4371 		 * killing stale mapping since one CPU may not be mapped
4372 		 * to any hw queue.
4373 		 */
4374 		for (i = 0; i < set->nr_maps; i++)
4375 			blk_mq_clear_mq_map(&set->map[i]);
4376 
4377 		set->ops->map_queues(set);
4378 	} else {
4379 		BUG_ON(set->nr_maps > 1);
4380 		blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4381 	}
4382 }
4383 
4384 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4385 				       int new_nr_hw_queues)
4386 {
4387 	struct blk_mq_tags **new_tags;
4388 
4389 	if (set->nr_hw_queues >= new_nr_hw_queues)
4390 		goto done;
4391 
4392 	new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4393 				GFP_KERNEL, set->numa_node);
4394 	if (!new_tags)
4395 		return -ENOMEM;
4396 
4397 	if (set->tags)
4398 		memcpy(new_tags, set->tags, set->nr_hw_queues *
4399 		       sizeof(*set->tags));
4400 	kfree(set->tags);
4401 	set->tags = new_tags;
4402 done:
4403 	set->nr_hw_queues = new_nr_hw_queues;
4404 	return 0;
4405 }
4406 
4407 /*
4408  * Alloc a tag set to be associated with one or more request queues.
4409  * May fail with EINVAL for various error conditions. May adjust the
4410  * requested depth down, if it's too large. In that case, the set
4411  * value will be stored in set->queue_depth.
4412  */
4413 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4414 {
4415 	int i, ret;
4416 
4417 	BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4418 
4419 	if (!set->nr_hw_queues)
4420 		return -EINVAL;
4421 	if (!set->queue_depth)
4422 		return -EINVAL;
4423 	if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4424 		return -EINVAL;
4425 
4426 	if (!set->ops->queue_rq)
4427 		return -EINVAL;
4428 
4429 	if (!set->ops->get_budget ^ !set->ops->put_budget)
4430 		return -EINVAL;
4431 
4432 	if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4433 		pr_info("blk-mq: reduced tag depth to %u\n",
4434 			BLK_MQ_MAX_DEPTH);
4435 		set->queue_depth = BLK_MQ_MAX_DEPTH;
4436 	}
4437 
4438 	if (!set->nr_maps)
4439 		set->nr_maps = 1;
4440 	else if (set->nr_maps > HCTX_MAX_TYPES)
4441 		return -EINVAL;
4442 
4443 	/*
4444 	 * If a crashdump is active, then we are potentially in a very
4445 	 * memory constrained environment. Limit us to 1 queue and
4446 	 * 64 tags to prevent using too much memory.
4447 	 */
4448 	if (is_kdump_kernel()) {
4449 		set->nr_hw_queues = 1;
4450 		set->nr_maps = 1;
4451 		set->queue_depth = min(64U, set->queue_depth);
4452 	}
4453 	/*
4454 	 * There is no use for more h/w queues than cpus if we just have
4455 	 * a single map
4456 	 */
4457 	if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4458 		set->nr_hw_queues = nr_cpu_ids;
4459 
4460 	if (set->flags & BLK_MQ_F_BLOCKING) {
4461 		set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL);
4462 		if (!set->srcu)
4463 			return -ENOMEM;
4464 		ret = init_srcu_struct(set->srcu);
4465 		if (ret)
4466 			goto out_free_srcu;
4467 	}
4468 
4469 	ret = -ENOMEM;
4470 	set->tags = kcalloc_node(set->nr_hw_queues,
4471 				 sizeof(struct blk_mq_tags *), GFP_KERNEL,
4472 				 set->numa_node);
4473 	if (!set->tags)
4474 		goto out_cleanup_srcu;
4475 
4476 	for (i = 0; i < set->nr_maps; i++) {
4477 		set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4478 						  sizeof(set->map[i].mq_map[0]),
4479 						  GFP_KERNEL, set->numa_node);
4480 		if (!set->map[i].mq_map)
4481 			goto out_free_mq_map;
4482 		set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4483 	}
4484 
4485 	blk_mq_update_queue_map(set);
4486 
4487 	ret = blk_mq_alloc_set_map_and_rqs(set);
4488 	if (ret)
4489 		goto out_free_mq_map;
4490 
4491 	mutex_init(&set->tag_list_lock);
4492 	INIT_LIST_HEAD(&set->tag_list);
4493 
4494 	return 0;
4495 
4496 out_free_mq_map:
4497 	for (i = 0; i < set->nr_maps; i++) {
4498 		kfree(set->map[i].mq_map);
4499 		set->map[i].mq_map = NULL;
4500 	}
4501 	kfree(set->tags);
4502 	set->tags = NULL;
4503 out_cleanup_srcu:
4504 	if (set->flags & BLK_MQ_F_BLOCKING)
4505 		cleanup_srcu_struct(set->srcu);
4506 out_free_srcu:
4507 	if (set->flags & BLK_MQ_F_BLOCKING)
4508 		kfree(set->srcu);
4509 	return ret;
4510 }
4511 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4512 
4513 /* allocate and initialize a tagset for a simple single-queue device */
4514 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4515 		const struct blk_mq_ops *ops, unsigned int queue_depth,
4516 		unsigned int set_flags)
4517 {
4518 	memset(set, 0, sizeof(*set));
4519 	set->ops = ops;
4520 	set->nr_hw_queues = 1;
4521 	set->nr_maps = 1;
4522 	set->queue_depth = queue_depth;
4523 	set->numa_node = NUMA_NO_NODE;
4524 	set->flags = set_flags;
4525 	return blk_mq_alloc_tag_set(set);
4526 }
4527 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4528 
4529 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4530 {
4531 	int i, j;
4532 
4533 	for (i = 0; i < set->nr_hw_queues; i++)
4534 		__blk_mq_free_map_and_rqs(set, i);
4535 
4536 	if (blk_mq_is_shared_tags(set->flags)) {
4537 		blk_mq_free_map_and_rqs(set, set->shared_tags,
4538 					BLK_MQ_NO_HCTX_IDX);
4539 	}
4540 
4541 	for (j = 0; j < set->nr_maps; j++) {
4542 		kfree(set->map[j].mq_map);
4543 		set->map[j].mq_map = NULL;
4544 	}
4545 
4546 	kfree(set->tags);
4547 	set->tags = NULL;
4548 	if (set->flags & BLK_MQ_F_BLOCKING) {
4549 		cleanup_srcu_struct(set->srcu);
4550 		kfree(set->srcu);
4551 	}
4552 }
4553 EXPORT_SYMBOL(blk_mq_free_tag_set);
4554 
4555 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4556 {
4557 	struct blk_mq_tag_set *set = q->tag_set;
4558 	struct blk_mq_hw_ctx *hctx;
4559 	int ret;
4560 	unsigned long i;
4561 
4562 	if (!set)
4563 		return -EINVAL;
4564 
4565 	if (q->nr_requests == nr)
4566 		return 0;
4567 
4568 	blk_mq_freeze_queue(q);
4569 	blk_mq_quiesce_queue(q);
4570 
4571 	ret = 0;
4572 	queue_for_each_hw_ctx(q, hctx, i) {
4573 		if (!hctx->tags)
4574 			continue;
4575 		/*
4576 		 * If we're using an MQ scheduler, just update the scheduler
4577 		 * queue depth. This is similar to what the old code would do.
4578 		 */
4579 		if (hctx->sched_tags) {
4580 			ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4581 						      nr, true);
4582 		} else {
4583 			ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4584 						      false);
4585 		}
4586 		if (ret)
4587 			break;
4588 		if (q->elevator && q->elevator->type->ops.depth_updated)
4589 			q->elevator->type->ops.depth_updated(hctx);
4590 	}
4591 	if (!ret) {
4592 		q->nr_requests = nr;
4593 		if (blk_mq_is_shared_tags(set->flags)) {
4594 			if (q->elevator)
4595 				blk_mq_tag_update_sched_shared_tags(q);
4596 			else
4597 				blk_mq_tag_resize_shared_tags(set, nr);
4598 		}
4599 	}
4600 
4601 	blk_mq_unquiesce_queue(q);
4602 	blk_mq_unfreeze_queue(q);
4603 
4604 	return ret;
4605 }
4606 
4607 /*
4608  * request_queue and elevator_type pair.
4609  * It is just used by __blk_mq_update_nr_hw_queues to cache
4610  * the elevator_type associated with a request_queue.
4611  */
4612 struct blk_mq_qe_pair {
4613 	struct list_head node;
4614 	struct request_queue *q;
4615 	struct elevator_type *type;
4616 };
4617 
4618 /*
4619  * Cache the elevator_type in qe pair list and switch the
4620  * io scheduler to 'none'
4621  */
4622 static bool blk_mq_elv_switch_none(struct list_head *head,
4623 		struct request_queue *q)
4624 {
4625 	struct blk_mq_qe_pair *qe;
4626 
4627 	qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4628 	if (!qe)
4629 		return false;
4630 
4631 	/* q->elevator needs protection from ->sysfs_lock */
4632 	mutex_lock(&q->sysfs_lock);
4633 
4634 	/* the check has to be done with holding sysfs_lock */
4635 	if (!q->elevator) {
4636 		kfree(qe);
4637 		goto unlock;
4638 	}
4639 
4640 	INIT_LIST_HEAD(&qe->node);
4641 	qe->q = q;
4642 	qe->type = q->elevator->type;
4643 	/* keep a reference to the elevator module as we'll switch back */
4644 	__elevator_get(qe->type);
4645 	list_add(&qe->node, head);
4646 	elevator_disable(q);
4647 unlock:
4648 	mutex_unlock(&q->sysfs_lock);
4649 
4650 	return true;
4651 }
4652 
4653 static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4654 						struct request_queue *q)
4655 {
4656 	struct blk_mq_qe_pair *qe;
4657 
4658 	list_for_each_entry(qe, head, node)
4659 		if (qe->q == q)
4660 			return qe;
4661 
4662 	return NULL;
4663 }
4664 
4665 static void blk_mq_elv_switch_back(struct list_head *head,
4666 				  struct request_queue *q)
4667 {
4668 	struct blk_mq_qe_pair *qe;
4669 	struct elevator_type *t;
4670 
4671 	qe = blk_lookup_qe_pair(head, q);
4672 	if (!qe)
4673 		return;
4674 	t = qe->type;
4675 	list_del(&qe->node);
4676 	kfree(qe);
4677 
4678 	mutex_lock(&q->sysfs_lock);
4679 	elevator_switch(q, t);
4680 	/* drop the reference acquired in blk_mq_elv_switch_none */
4681 	elevator_put(t);
4682 	mutex_unlock(&q->sysfs_lock);
4683 }
4684 
4685 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4686 							int nr_hw_queues)
4687 {
4688 	struct request_queue *q;
4689 	LIST_HEAD(head);
4690 	int prev_nr_hw_queues;
4691 
4692 	lockdep_assert_held(&set->tag_list_lock);
4693 
4694 	if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4695 		nr_hw_queues = nr_cpu_ids;
4696 	if (nr_hw_queues < 1)
4697 		return;
4698 	if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4699 		return;
4700 
4701 	list_for_each_entry(q, &set->tag_list, tag_set_list)
4702 		blk_mq_freeze_queue(q);
4703 	/*
4704 	 * Switch IO scheduler to 'none', cleaning up the data associated
4705 	 * with the previous scheduler. We will switch back once we are done
4706 	 * updating the new sw to hw queue mappings.
4707 	 */
4708 	list_for_each_entry(q, &set->tag_list, tag_set_list)
4709 		if (!blk_mq_elv_switch_none(&head, q))
4710 			goto switch_back;
4711 
4712 	list_for_each_entry(q, &set->tag_list, tag_set_list) {
4713 		blk_mq_debugfs_unregister_hctxs(q);
4714 		blk_mq_sysfs_unregister_hctxs(q);
4715 	}
4716 
4717 	prev_nr_hw_queues = set->nr_hw_queues;
4718 	if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0)
4719 		goto reregister;
4720 
4721 fallback:
4722 	blk_mq_update_queue_map(set);
4723 	list_for_each_entry(q, &set->tag_list, tag_set_list) {
4724 		blk_mq_realloc_hw_ctxs(set, q);
4725 		blk_mq_update_poll_flag(q);
4726 		if (q->nr_hw_queues != set->nr_hw_queues) {
4727 			int i = prev_nr_hw_queues;
4728 
4729 			pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4730 					nr_hw_queues, prev_nr_hw_queues);
4731 			for (; i < set->nr_hw_queues; i++)
4732 				__blk_mq_free_map_and_rqs(set, i);
4733 
4734 			set->nr_hw_queues = prev_nr_hw_queues;
4735 			blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4736 			goto fallback;
4737 		}
4738 		blk_mq_map_swqueue(q);
4739 	}
4740 
4741 reregister:
4742 	list_for_each_entry(q, &set->tag_list, tag_set_list) {
4743 		blk_mq_sysfs_register_hctxs(q);
4744 		blk_mq_debugfs_register_hctxs(q);
4745 	}
4746 
4747 switch_back:
4748 	list_for_each_entry(q, &set->tag_list, tag_set_list)
4749 		blk_mq_elv_switch_back(&head, q);
4750 
4751 	list_for_each_entry(q, &set->tag_list, tag_set_list)
4752 		blk_mq_unfreeze_queue(q);
4753 }
4754 
4755 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4756 {
4757 	mutex_lock(&set->tag_list_lock);
4758 	__blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4759 	mutex_unlock(&set->tag_list_lock);
4760 }
4761 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4762 
4763 static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
4764 			 struct io_comp_batch *iob, unsigned int flags)
4765 {
4766 	long state = get_current_state();
4767 	int ret;
4768 
4769 	do {
4770 		ret = q->mq_ops->poll(hctx, iob);
4771 		if (ret > 0) {
4772 			__set_current_state(TASK_RUNNING);
4773 			return ret;
4774 		}
4775 
4776 		if (signal_pending_state(state, current))
4777 			__set_current_state(TASK_RUNNING);
4778 		if (task_is_running(current))
4779 			return 1;
4780 
4781 		if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4782 			break;
4783 		cpu_relax();
4784 	} while (!need_resched());
4785 
4786 	__set_current_state(TASK_RUNNING);
4787 	return 0;
4788 }
4789 
4790 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie,
4791 		struct io_comp_batch *iob, unsigned int flags)
4792 {
4793 	struct blk_mq_hw_ctx *hctx = xa_load(&q->hctx_table, cookie);
4794 
4795 	return blk_hctx_poll(q, hctx, iob, flags);
4796 }
4797 
4798 int blk_rq_poll(struct request *rq, struct io_comp_batch *iob,
4799 		unsigned int poll_flags)
4800 {
4801 	struct request_queue *q = rq->q;
4802 	int ret;
4803 
4804 	if (!blk_rq_is_poll(rq))
4805 		return 0;
4806 	if (!percpu_ref_tryget(&q->q_usage_counter))
4807 		return 0;
4808 
4809 	ret = blk_hctx_poll(q, rq->mq_hctx, iob, poll_flags);
4810 	blk_queue_exit(q);
4811 
4812 	return ret;
4813 }
4814 EXPORT_SYMBOL_GPL(blk_rq_poll);
4815 
4816 unsigned int blk_mq_rq_cpu(struct request *rq)
4817 {
4818 	return rq->mq_ctx->cpu;
4819 }
4820 EXPORT_SYMBOL(blk_mq_rq_cpu);
4821 
4822 void blk_mq_cancel_work_sync(struct request_queue *q)
4823 {
4824 	struct blk_mq_hw_ctx *hctx;
4825 	unsigned long i;
4826 
4827 	cancel_delayed_work_sync(&q->requeue_work);
4828 
4829 	queue_for_each_hw_ctx(q, hctx, i)
4830 		cancel_delayed_work_sync(&hctx->run_work);
4831 }
4832 
4833 static int __init blk_mq_init(void)
4834 {
4835 	int i;
4836 
4837 	for_each_possible_cpu(i)
4838 		init_llist_head(&per_cpu(blk_cpu_done, i));
4839 	open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4840 
4841 	cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4842 				  "block/softirq:dead", NULL,
4843 				  blk_softirq_cpu_dead);
4844 	cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4845 				blk_mq_hctx_notify_dead);
4846 	cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4847 				blk_mq_hctx_notify_online,
4848 				blk_mq_hctx_notify_offline);
4849 	return 0;
4850 }
4851 subsys_initcall(blk_mq_init);
4852