xref: /openbmc/linux/block/blk-mq.c (revision f17f06a0)
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/kmemleak.h>
14 #include <linux/mm.h>
15 #include <linux/init.h>
16 #include <linux/slab.h>
17 #include <linux/workqueue.h>
18 #include <linux/smp.h>
19 #include <linux/llist.h>
20 #include <linux/list_sort.h>
21 #include <linux/cpu.h>
22 #include <linux/cache.h>
23 #include <linux/sched/sysctl.h>
24 #include <linux/sched/topology.h>
25 #include <linux/sched/signal.h>
26 #include <linux/delay.h>
27 #include <linux/crash_dump.h>
28 #include <linux/prefetch.h>
29 
30 #include <trace/events/block.h>
31 
32 #include <linux/blk-mq.h>
33 #include <linux/t10-pi.h>
34 #include "blk.h"
35 #include "blk-mq.h"
36 #include "blk-mq-debugfs.h"
37 #include "blk-mq-tag.h"
38 #include "blk-pm.h"
39 #include "blk-stat.h"
40 #include "blk-mq-sched.h"
41 #include "blk-rq-qos.h"
42 
43 static void blk_mq_poll_stats_start(struct request_queue *q);
44 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
45 
46 static int blk_mq_poll_stats_bkt(const struct request *rq)
47 {
48 	int ddir, sectors, bucket;
49 
50 	ddir = rq_data_dir(rq);
51 	sectors = blk_rq_stats_sectors(rq);
52 
53 	bucket = ddir + 2 * ilog2(sectors);
54 
55 	if (bucket < 0)
56 		return -1;
57 	else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
58 		return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
59 
60 	return bucket;
61 }
62 
63 /*
64  * Check if any of the ctx, dispatch list or elevator
65  * have pending work in this hardware queue.
66  */
67 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
68 {
69 	return !list_empty_careful(&hctx->dispatch) ||
70 		sbitmap_any_bit_set(&hctx->ctx_map) ||
71 			blk_mq_sched_has_work(hctx);
72 }
73 
74 /*
75  * Mark this ctx as having pending work in this hardware queue
76  */
77 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
78 				     struct blk_mq_ctx *ctx)
79 {
80 	const int bit = ctx->index_hw[hctx->type];
81 
82 	if (!sbitmap_test_bit(&hctx->ctx_map, bit))
83 		sbitmap_set_bit(&hctx->ctx_map, bit);
84 }
85 
86 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
87 				      struct blk_mq_ctx *ctx)
88 {
89 	const int bit = ctx->index_hw[hctx->type];
90 
91 	sbitmap_clear_bit(&hctx->ctx_map, bit);
92 }
93 
94 struct mq_inflight {
95 	struct hd_struct *part;
96 	unsigned int inflight[2];
97 };
98 
99 static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
100 				  struct request *rq, void *priv,
101 				  bool reserved)
102 {
103 	struct mq_inflight *mi = priv;
104 
105 	if (rq->part == mi->part)
106 		mi->inflight[rq_data_dir(rq)]++;
107 
108 	return true;
109 }
110 
111 unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part)
112 {
113 	struct mq_inflight mi = { .part = part };
114 
115 	blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
116 
117 	return mi.inflight[0] + mi.inflight[1];
118 }
119 
120 void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
121 			 unsigned int inflight[2])
122 {
123 	struct mq_inflight mi = { .part = part };
124 
125 	blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
126 	inflight[0] = mi.inflight[0];
127 	inflight[1] = mi.inflight[1];
128 }
129 
130 void blk_freeze_queue_start(struct request_queue *q)
131 {
132 	mutex_lock(&q->mq_freeze_lock);
133 	if (++q->mq_freeze_depth == 1) {
134 		percpu_ref_kill(&q->q_usage_counter);
135 		mutex_unlock(&q->mq_freeze_lock);
136 		if (queue_is_mq(q))
137 			blk_mq_run_hw_queues(q, false);
138 	} else {
139 		mutex_unlock(&q->mq_freeze_lock);
140 	}
141 }
142 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
143 
144 void blk_mq_freeze_queue_wait(struct request_queue *q)
145 {
146 	wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
147 }
148 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
149 
150 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
151 				     unsigned long timeout)
152 {
153 	return wait_event_timeout(q->mq_freeze_wq,
154 					percpu_ref_is_zero(&q->q_usage_counter),
155 					timeout);
156 }
157 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
158 
159 /*
160  * Guarantee no request is in use, so we can change any data structure of
161  * the queue afterward.
162  */
163 void blk_freeze_queue(struct request_queue *q)
164 {
165 	/*
166 	 * In the !blk_mq case we are only calling this to kill the
167 	 * q_usage_counter, otherwise this increases the freeze depth
168 	 * and waits for it to return to zero.  For this reason there is
169 	 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
170 	 * exported to drivers as the only user for unfreeze is blk_mq.
171 	 */
172 	blk_freeze_queue_start(q);
173 	blk_mq_freeze_queue_wait(q);
174 }
175 
176 void blk_mq_freeze_queue(struct request_queue *q)
177 {
178 	/*
179 	 * ...just an alias to keep freeze and unfreeze actions balanced
180 	 * in the blk_mq_* namespace
181 	 */
182 	blk_freeze_queue(q);
183 }
184 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
185 
186 void blk_mq_unfreeze_queue(struct request_queue *q)
187 {
188 	mutex_lock(&q->mq_freeze_lock);
189 	q->mq_freeze_depth--;
190 	WARN_ON_ONCE(q->mq_freeze_depth < 0);
191 	if (!q->mq_freeze_depth) {
192 		percpu_ref_resurrect(&q->q_usage_counter);
193 		wake_up_all(&q->mq_freeze_wq);
194 	}
195 	mutex_unlock(&q->mq_freeze_lock);
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 	blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
206 }
207 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
208 
209 /**
210  * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
211  * @q: request queue.
212  *
213  * Note: this function does not prevent that the struct request end_io()
214  * callback function is invoked. Once this function is returned, we make
215  * sure no dispatch can happen until the queue is unquiesced via
216  * blk_mq_unquiesce_queue().
217  */
218 void blk_mq_quiesce_queue(struct request_queue *q)
219 {
220 	struct blk_mq_hw_ctx *hctx;
221 	unsigned int i;
222 	bool rcu = false;
223 
224 	blk_mq_quiesce_queue_nowait(q);
225 
226 	queue_for_each_hw_ctx(q, hctx, i) {
227 		if (hctx->flags & BLK_MQ_F_BLOCKING)
228 			synchronize_srcu(hctx->srcu);
229 		else
230 			rcu = true;
231 	}
232 	if (rcu)
233 		synchronize_rcu();
234 }
235 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
236 
237 /*
238  * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
239  * @q: request queue.
240  *
241  * This function recovers queue into the state before quiescing
242  * which is done by blk_mq_quiesce_queue.
243  */
244 void blk_mq_unquiesce_queue(struct request_queue *q)
245 {
246 	blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
247 
248 	/* dispatch requests which are inserted during quiescing */
249 	blk_mq_run_hw_queues(q, true);
250 }
251 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
252 
253 void blk_mq_wake_waiters(struct request_queue *q)
254 {
255 	struct blk_mq_hw_ctx *hctx;
256 	unsigned int i;
257 
258 	queue_for_each_hw_ctx(q, hctx, i)
259 		if (blk_mq_hw_queue_mapped(hctx))
260 			blk_mq_tag_wakeup_all(hctx->tags, true);
261 }
262 
263 /*
264  * Only need start/end time stamping if we have iostat or
265  * blk stats enabled, or using an IO scheduler.
266  */
267 static inline bool blk_mq_need_time_stamp(struct request *rq)
268 {
269 	return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator;
270 }
271 
272 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
273 		unsigned int tag, unsigned int op, u64 alloc_time_ns)
274 {
275 	struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
276 	struct request *rq = tags->static_rqs[tag];
277 	req_flags_t rq_flags = 0;
278 
279 	if (data->flags & BLK_MQ_REQ_INTERNAL) {
280 		rq->tag = -1;
281 		rq->internal_tag = tag;
282 	} else {
283 		if (data->hctx->flags & BLK_MQ_F_TAG_SHARED) {
284 			rq_flags = RQF_MQ_INFLIGHT;
285 			atomic_inc(&data->hctx->nr_active);
286 		}
287 		rq->tag = tag;
288 		rq->internal_tag = -1;
289 		data->hctx->tags->rqs[rq->tag] = rq;
290 	}
291 
292 	/* csd/requeue_work/fifo_time is initialized before use */
293 	rq->q = data->q;
294 	rq->mq_ctx = data->ctx;
295 	rq->mq_hctx = data->hctx;
296 	rq->rq_flags = rq_flags;
297 	rq->cmd_flags = op;
298 	if (data->flags & BLK_MQ_REQ_PREEMPT)
299 		rq->rq_flags |= RQF_PREEMPT;
300 	if (blk_queue_io_stat(data->q))
301 		rq->rq_flags |= RQF_IO_STAT;
302 	INIT_LIST_HEAD(&rq->queuelist);
303 	INIT_HLIST_NODE(&rq->hash);
304 	RB_CLEAR_NODE(&rq->rb_node);
305 	rq->rq_disk = NULL;
306 	rq->part = NULL;
307 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
308 	rq->alloc_time_ns = alloc_time_ns;
309 #endif
310 	if (blk_mq_need_time_stamp(rq))
311 		rq->start_time_ns = ktime_get_ns();
312 	else
313 		rq->start_time_ns = 0;
314 	rq->io_start_time_ns = 0;
315 	rq->stats_sectors = 0;
316 	rq->nr_phys_segments = 0;
317 #if defined(CONFIG_BLK_DEV_INTEGRITY)
318 	rq->nr_integrity_segments = 0;
319 #endif
320 	/* tag was already set */
321 	rq->extra_len = 0;
322 	WRITE_ONCE(rq->deadline, 0);
323 
324 	rq->timeout = 0;
325 
326 	rq->end_io = NULL;
327 	rq->end_io_data = NULL;
328 
329 	data->ctx->rq_dispatched[op_is_sync(op)]++;
330 	refcount_set(&rq->ref, 1);
331 	return rq;
332 }
333 
334 static struct request *blk_mq_get_request(struct request_queue *q,
335 					  struct bio *bio,
336 					  struct blk_mq_alloc_data *data)
337 {
338 	struct elevator_queue *e = q->elevator;
339 	struct request *rq;
340 	unsigned int tag;
341 	bool clear_ctx_on_error = false;
342 	u64 alloc_time_ns = 0;
343 
344 	blk_queue_enter_live(q);
345 
346 	/* alloc_time includes depth and tag waits */
347 	if (blk_queue_rq_alloc_time(q))
348 		alloc_time_ns = ktime_get_ns();
349 
350 	data->q = q;
351 	if (likely(!data->ctx)) {
352 		data->ctx = blk_mq_get_ctx(q);
353 		clear_ctx_on_error = true;
354 	}
355 	if (likely(!data->hctx))
356 		data->hctx = blk_mq_map_queue(q, data->cmd_flags,
357 						data->ctx);
358 	if (data->cmd_flags & REQ_NOWAIT)
359 		data->flags |= BLK_MQ_REQ_NOWAIT;
360 
361 	if (e) {
362 		data->flags |= BLK_MQ_REQ_INTERNAL;
363 
364 		/*
365 		 * Flush requests are special and go directly to the
366 		 * dispatch list. Don't include reserved tags in the
367 		 * limiting, as it isn't useful.
368 		 */
369 		if (!op_is_flush(data->cmd_flags) &&
370 		    e->type->ops.limit_depth &&
371 		    !(data->flags & BLK_MQ_REQ_RESERVED))
372 			e->type->ops.limit_depth(data->cmd_flags, data);
373 	} else {
374 		blk_mq_tag_busy(data->hctx);
375 	}
376 
377 	tag = blk_mq_get_tag(data);
378 	if (tag == BLK_MQ_TAG_FAIL) {
379 		if (clear_ctx_on_error)
380 			data->ctx = NULL;
381 		blk_queue_exit(q);
382 		return NULL;
383 	}
384 
385 	rq = blk_mq_rq_ctx_init(data, tag, data->cmd_flags, alloc_time_ns);
386 	if (!op_is_flush(data->cmd_flags)) {
387 		rq->elv.icq = NULL;
388 		if (e && e->type->ops.prepare_request) {
389 			if (e->type->icq_cache)
390 				blk_mq_sched_assign_ioc(rq);
391 
392 			e->type->ops.prepare_request(rq, bio);
393 			rq->rq_flags |= RQF_ELVPRIV;
394 		}
395 	}
396 	data->hctx->queued++;
397 	return rq;
398 }
399 
400 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
401 		blk_mq_req_flags_t flags)
402 {
403 	struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
404 	struct request *rq;
405 	int ret;
406 
407 	ret = blk_queue_enter(q, flags);
408 	if (ret)
409 		return ERR_PTR(ret);
410 
411 	rq = blk_mq_get_request(q, NULL, &alloc_data);
412 	blk_queue_exit(q);
413 
414 	if (!rq)
415 		return ERR_PTR(-EWOULDBLOCK);
416 
417 	rq->__data_len = 0;
418 	rq->__sector = (sector_t) -1;
419 	rq->bio = rq->biotail = NULL;
420 	return rq;
421 }
422 EXPORT_SYMBOL(blk_mq_alloc_request);
423 
424 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
425 	unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
426 {
427 	struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
428 	struct request *rq;
429 	unsigned int cpu;
430 	int ret;
431 
432 	/*
433 	 * If the tag allocator sleeps we could get an allocation for a
434 	 * different hardware context.  No need to complicate the low level
435 	 * allocator for this for the rare use case of a command tied to
436 	 * a specific queue.
437 	 */
438 	if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
439 		return ERR_PTR(-EINVAL);
440 
441 	if (hctx_idx >= q->nr_hw_queues)
442 		return ERR_PTR(-EIO);
443 
444 	ret = blk_queue_enter(q, flags);
445 	if (ret)
446 		return ERR_PTR(ret);
447 
448 	/*
449 	 * Check if the hardware context is actually mapped to anything.
450 	 * If not tell the caller that it should skip this queue.
451 	 */
452 	alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
453 	if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
454 		blk_queue_exit(q);
455 		return ERR_PTR(-EXDEV);
456 	}
457 	cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
458 	alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
459 
460 	rq = blk_mq_get_request(q, NULL, &alloc_data);
461 	blk_queue_exit(q);
462 
463 	if (!rq)
464 		return ERR_PTR(-EWOULDBLOCK);
465 
466 	return rq;
467 }
468 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
469 
470 static void __blk_mq_free_request(struct request *rq)
471 {
472 	struct request_queue *q = rq->q;
473 	struct blk_mq_ctx *ctx = rq->mq_ctx;
474 	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
475 	const int sched_tag = rq->internal_tag;
476 
477 	blk_pm_mark_last_busy(rq);
478 	rq->mq_hctx = NULL;
479 	if (rq->tag != -1)
480 		blk_mq_put_tag(hctx->tags, ctx, rq->tag);
481 	if (sched_tag != -1)
482 		blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
483 	blk_mq_sched_restart(hctx);
484 	blk_queue_exit(q);
485 }
486 
487 void blk_mq_free_request(struct request *rq)
488 {
489 	struct request_queue *q = rq->q;
490 	struct elevator_queue *e = q->elevator;
491 	struct blk_mq_ctx *ctx = rq->mq_ctx;
492 	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
493 
494 	if (rq->rq_flags & RQF_ELVPRIV) {
495 		if (e && e->type->ops.finish_request)
496 			e->type->ops.finish_request(rq);
497 		if (rq->elv.icq) {
498 			put_io_context(rq->elv.icq->ioc);
499 			rq->elv.icq = NULL;
500 		}
501 	}
502 
503 	ctx->rq_completed[rq_is_sync(rq)]++;
504 	if (rq->rq_flags & RQF_MQ_INFLIGHT)
505 		atomic_dec(&hctx->nr_active);
506 
507 	if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
508 		laptop_io_completion(q->backing_dev_info);
509 
510 	rq_qos_done(q, rq);
511 
512 	WRITE_ONCE(rq->state, MQ_RQ_IDLE);
513 	if (refcount_dec_and_test(&rq->ref))
514 		__blk_mq_free_request(rq);
515 }
516 EXPORT_SYMBOL_GPL(blk_mq_free_request);
517 
518 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
519 {
520 	u64 now = 0;
521 
522 	if (blk_mq_need_time_stamp(rq))
523 		now = ktime_get_ns();
524 
525 	if (rq->rq_flags & RQF_STATS) {
526 		blk_mq_poll_stats_start(rq->q);
527 		blk_stat_add(rq, now);
528 	}
529 
530 	if (rq->internal_tag != -1)
531 		blk_mq_sched_completed_request(rq, now);
532 
533 	blk_account_io_done(rq, now);
534 
535 	if (rq->end_io) {
536 		rq_qos_done(rq->q, rq);
537 		rq->end_io(rq, error);
538 	} else {
539 		blk_mq_free_request(rq);
540 	}
541 }
542 EXPORT_SYMBOL(__blk_mq_end_request);
543 
544 void blk_mq_end_request(struct request *rq, blk_status_t error)
545 {
546 	if (blk_update_request(rq, error, blk_rq_bytes(rq)))
547 		BUG();
548 	__blk_mq_end_request(rq, error);
549 }
550 EXPORT_SYMBOL(blk_mq_end_request);
551 
552 static void __blk_mq_complete_request_remote(void *data)
553 {
554 	struct request *rq = data;
555 	struct request_queue *q = rq->q;
556 
557 	q->mq_ops->complete(rq);
558 }
559 
560 static void __blk_mq_complete_request(struct request *rq)
561 {
562 	struct blk_mq_ctx *ctx = rq->mq_ctx;
563 	struct request_queue *q = rq->q;
564 	bool shared = false;
565 	int cpu;
566 
567 	WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
568 	/*
569 	 * Most of single queue controllers, there is only one irq vector
570 	 * for handling IO completion, and the only irq's affinity is set
571 	 * as all possible CPUs. On most of ARCHs, this affinity means the
572 	 * irq is handled on one specific CPU.
573 	 *
574 	 * So complete IO reqeust in softirq context in case of single queue
575 	 * for not degrading IO performance by irqsoff latency.
576 	 */
577 	if (q->nr_hw_queues == 1) {
578 		__blk_complete_request(rq);
579 		return;
580 	}
581 
582 	/*
583 	 * For a polled request, always complete locallly, it's pointless
584 	 * to redirect the completion.
585 	 */
586 	if ((rq->cmd_flags & REQ_HIPRI) ||
587 	    !test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags)) {
588 		q->mq_ops->complete(rq);
589 		return;
590 	}
591 
592 	cpu = get_cpu();
593 	if (!test_bit(QUEUE_FLAG_SAME_FORCE, &q->queue_flags))
594 		shared = cpus_share_cache(cpu, ctx->cpu);
595 
596 	if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
597 		rq->csd.func = __blk_mq_complete_request_remote;
598 		rq->csd.info = rq;
599 		rq->csd.flags = 0;
600 		smp_call_function_single_async(ctx->cpu, &rq->csd);
601 	} else {
602 		q->mq_ops->complete(rq);
603 	}
604 	put_cpu();
605 }
606 
607 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
608 	__releases(hctx->srcu)
609 {
610 	if (!(hctx->flags & BLK_MQ_F_BLOCKING))
611 		rcu_read_unlock();
612 	else
613 		srcu_read_unlock(hctx->srcu, srcu_idx);
614 }
615 
616 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
617 	__acquires(hctx->srcu)
618 {
619 	if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
620 		/* shut up gcc false positive */
621 		*srcu_idx = 0;
622 		rcu_read_lock();
623 	} else
624 		*srcu_idx = srcu_read_lock(hctx->srcu);
625 }
626 
627 /**
628  * blk_mq_complete_request - end I/O on a request
629  * @rq:		the request being processed
630  *
631  * Description:
632  *	Ends all I/O on a request. It does not handle partial completions.
633  *	The actual completion happens out-of-order, through a IPI handler.
634  **/
635 bool blk_mq_complete_request(struct request *rq)
636 {
637 	if (unlikely(blk_should_fake_timeout(rq->q)))
638 		return false;
639 	__blk_mq_complete_request(rq);
640 	return true;
641 }
642 EXPORT_SYMBOL(blk_mq_complete_request);
643 
644 /**
645  * blk_mq_start_request - Start processing a request
646  * @rq: Pointer to request to be started
647  *
648  * Function used by device drivers to notify the block layer that a request
649  * is going to be processed now, so blk layer can do proper initializations
650  * such as starting the timeout timer.
651  */
652 void blk_mq_start_request(struct request *rq)
653 {
654 	struct request_queue *q = rq->q;
655 
656 	trace_block_rq_issue(q, rq);
657 
658 	if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
659 		rq->io_start_time_ns = ktime_get_ns();
660 		rq->stats_sectors = blk_rq_sectors(rq);
661 		rq->rq_flags |= RQF_STATS;
662 		rq_qos_issue(q, rq);
663 	}
664 
665 	WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
666 
667 	blk_add_timer(rq);
668 	WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
669 
670 	if (q->dma_drain_size && blk_rq_bytes(rq)) {
671 		/*
672 		 * Make sure space for the drain appears.  We know we can do
673 		 * this because max_hw_segments has been adjusted to be one
674 		 * fewer than the device can handle.
675 		 */
676 		rq->nr_phys_segments++;
677 	}
678 
679 #ifdef CONFIG_BLK_DEV_INTEGRITY
680 	if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
681 		q->integrity.profile->prepare_fn(rq);
682 #endif
683 }
684 EXPORT_SYMBOL(blk_mq_start_request);
685 
686 static void __blk_mq_requeue_request(struct request *rq)
687 {
688 	struct request_queue *q = rq->q;
689 
690 	blk_mq_put_driver_tag(rq);
691 
692 	trace_block_rq_requeue(q, rq);
693 	rq_qos_requeue(q, rq);
694 
695 	if (blk_mq_request_started(rq)) {
696 		WRITE_ONCE(rq->state, MQ_RQ_IDLE);
697 		rq->rq_flags &= ~RQF_TIMED_OUT;
698 		if (q->dma_drain_size && blk_rq_bytes(rq))
699 			rq->nr_phys_segments--;
700 	}
701 }
702 
703 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
704 {
705 	__blk_mq_requeue_request(rq);
706 
707 	/* this request will be re-inserted to io scheduler queue */
708 	blk_mq_sched_requeue_request(rq);
709 
710 	BUG_ON(!list_empty(&rq->queuelist));
711 	blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
712 }
713 EXPORT_SYMBOL(blk_mq_requeue_request);
714 
715 static void blk_mq_requeue_work(struct work_struct *work)
716 {
717 	struct request_queue *q =
718 		container_of(work, struct request_queue, requeue_work.work);
719 	LIST_HEAD(rq_list);
720 	struct request *rq, *next;
721 
722 	spin_lock_irq(&q->requeue_lock);
723 	list_splice_init(&q->requeue_list, &rq_list);
724 	spin_unlock_irq(&q->requeue_lock);
725 
726 	list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
727 		if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
728 			continue;
729 
730 		rq->rq_flags &= ~RQF_SOFTBARRIER;
731 		list_del_init(&rq->queuelist);
732 		/*
733 		 * If RQF_DONTPREP, rq has contained some driver specific
734 		 * data, so insert it to hctx dispatch list to avoid any
735 		 * merge.
736 		 */
737 		if (rq->rq_flags & RQF_DONTPREP)
738 			blk_mq_request_bypass_insert(rq, false, false);
739 		else
740 			blk_mq_sched_insert_request(rq, true, false, false);
741 	}
742 
743 	while (!list_empty(&rq_list)) {
744 		rq = list_entry(rq_list.next, struct request, queuelist);
745 		list_del_init(&rq->queuelist);
746 		blk_mq_sched_insert_request(rq, false, false, false);
747 	}
748 
749 	blk_mq_run_hw_queues(q, false);
750 }
751 
752 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
753 				bool kick_requeue_list)
754 {
755 	struct request_queue *q = rq->q;
756 	unsigned long flags;
757 
758 	/*
759 	 * We abuse this flag that is otherwise used by the I/O scheduler to
760 	 * request head insertion from the workqueue.
761 	 */
762 	BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
763 
764 	spin_lock_irqsave(&q->requeue_lock, flags);
765 	if (at_head) {
766 		rq->rq_flags |= RQF_SOFTBARRIER;
767 		list_add(&rq->queuelist, &q->requeue_list);
768 	} else {
769 		list_add_tail(&rq->queuelist, &q->requeue_list);
770 	}
771 	spin_unlock_irqrestore(&q->requeue_lock, flags);
772 
773 	if (kick_requeue_list)
774 		blk_mq_kick_requeue_list(q);
775 }
776 
777 void blk_mq_kick_requeue_list(struct request_queue *q)
778 {
779 	kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
780 }
781 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
782 
783 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
784 				    unsigned long msecs)
785 {
786 	kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
787 				    msecs_to_jiffies(msecs));
788 }
789 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
790 
791 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
792 {
793 	if (tag < tags->nr_tags) {
794 		prefetch(tags->rqs[tag]);
795 		return tags->rqs[tag];
796 	}
797 
798 	return NULL;
799 }
800 EXPORT_SYMBOL(blk_mq_tag_to_rq);
801 
802 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
803 			       void *priv, bool reserved)
804 {
805 	/*
806 	 * If we find a request that is inflight and the queue matches,
807 	 * we know the queue is busy. Return false to stop the iteration.
808 	 */
809 	if (rq->state == MQ_RQ_IN_FLIGHT && rq->q == hctx->queue) {
810 		bool *busy = priv;
811 
812 		*busy = true;
813 		return false;
814 	}
815 
816 	return true;
817 }
818 
819 bool blk_mq_queue_inflight(struct request_queue *q)
820 {
821 	bool busy = false;
822 
823 	blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
824 	return busy;
825 }
826 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
827 
828 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
829 {
830 	req->rq_flags |= RQF_TIMED_OUT;
831 	if (req->q->mq_ops->timeout) {
832 		enum blk_eh_timer_return ret;
833 
834 		ret = req->q->mq_ops->timeout(req, reserved);
835 		if (ret == BLK_EH_DONE)
836 			return;
837 		WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
838 	}
839 
840 	blk_add_timer(req);
841 }
842 
843 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
844 {
845 	unsigned long deadline;
846 
847 	if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
848 		return false;
849 	if (rq->rq_flags & RQF_TIMED_OUT)
850 		return false;
851 
852 	deadline = READ_ONCE(rq->deadline);
853 	if (time_after_eq(jiffies, deadline))
854 		return true;
855 
856 	if (*next == 0)
857 		*next = deadline;
858 	else if (time_after(*next, deadline))
859 		*next = deadline;
860 	return false;
861 }
862 
863 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
864 		struct request *rq, void *priv, bool reserved)
865 {
866 	unsigned long *next = priv;
867 
868 	/*
869 	 * Just do a quick check if it is expired before locking the request in
870 	 * so we're not unnecessarilly synchronizing across CPUs.
871 	 */
872 	if (!blk_mq_req_expired(rq, next))
873 		return true;
874 
875 	/*
876 	 * We have reason to believe the request may be expired. Take a
877 	 * reference on the request to lock this request lifetime into its
878 	 * currently allocated context to prevent it from being reallocated in
879 	 * the event the completion by-passes this timeout handler.
880 	 *
881 	 * If the reference was already released, then the driver beat the
882 	 * timeout handler to posting a natural completion.
883 	 */
884 	if (!refcount_inc_not_zero(&rq->ref))
885 		return true;
886 
887 	/*
888 	 * The request is now locked and cannot be reallocated underneath the
889 	 * timeout handler's processing. Re-verify this exact request is truly
890 	 * expired; if it is not expired, then the request was completed and
891 	 * reallocated as a new request.
892 	 */
893 	if (blk_mq_req_expired(rq, next))
894 		blk_mq_rq_timed_out(rq, reserved);
895 
896 	if (is_flush_rq(rq, hctx))
897 		rq->end_io(rq, 0);
898 	else if (refcount_dec_and_test(&rq->ref))
899 		__blk_mq_free_request(rq);
900 
901 	return true;
902 }
903 
904 static void blk_mq_timeout_work(struct work_struct *work)
905 {
906 	struct request_queue *q =
907 		container_of(work, struct request_queue, timeout_work);
908 	unsigned long next = 0;
909 	struct blk_mq_hw_ctx *hctx;
910 	int i;
911 
912 	/* A deadlock might occur if a request is stuck requiring a
913 	 * timeout at the same time a queue freeze is waiting
914 	 * completion, since the timeout code would not be able to
915 	 * acquire the queue reference here.
916 	 *
917 	 * That's why we don't use blk_queue_enter here; instead, we use
918 	 * percpu_ref_tryget directly, because we need to be able to
919 	 * obtain a reference even in the short window between the queue
920 	 * starting to freeze, by dropping the first reference in
921 	 * blk_freeze_queue_start, and the moment the last request is
922 	 * consumed, marked by the instant q_usage_counter reaches
923 	 * zero.
924 	 */
925 	if (!percpu_ref_tryget(&q->q_usage_counter))
926 		return;
927 
928 	blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
929 
930 	if (next != 0) {
931 		mod_timer(&q->timeout, next);
932 	} else {
933 		/*
934 		 * Request timeouts are handled as a forward rolling timer. If
935 		 * we end up here it means that no requests are pending and
936 		 * also that no request has been pending for a while. Mark
937 		 * each hctx as idle.
938 		 */
939 		queue_for_each_hw_ctx(q, hctx, i) {
940 			/* the hctx may be unmapped, so check it here */
941 			if (blk_mq_hw_queue_mapped(hctx))
942 				blk_mq_tag_idle(hctx);
943 		}
944 	}
945 	blk_queue_exit(q);
946 }
947 
948 struct flush_busy_ctx_data {
949 	struct blk_mq_hw_ctx *hctx;
950 	struct list_head *list;
951 };
952 
953 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
954 {
955 	struct flush_busy_ctx_data *flush_data = data;
956 	struct blk_mq_hw_ctx *hctx = flush_data->hctx;
957 	struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
958 	enum hctx_type type = hctx->type;
959 
960 	spin_lock(&ctx->lock);
961 	list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
962 	sbitmap_clear_bit(sb, bitnr);
963 	spin_unlock(&ctx->lock);
964 	return true;
965 }
966 
967 /*
968  * Process software queues that have been marked busy, splicing them
969  * to the for-dispatch
970  */
971 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
972 {
973 	struct flush_busy_ctx_data data = {
974 		.hctx = hctx,
975 		.list = list,
976 	};
977 
978 	sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
979 }
980 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
981 
982 struct dispatch_rq_data {
983 	struct blk_mq_hw_ctx *hctx;
984 	struct request *rq;
985 };
986 
987 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
988 		void *data)
989 {
990 	struct dispatch_rq_data *dispatch_data = data;
991 	struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
992 	struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
993 	enum hctx_type type = hctx->type;
994 
995 	spin_lock(&ctx->lock);
996 	if (!list_empty(&ctx->rq_lists[type])) {
997 		dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
998 		list_del_init(&dispatch_data->rq->queuelist);
999 		if (list_empty(&ctx->rq_lists[type]))
1000 			sbitmap_clear_bit(sb, bitnr);
1001 	}
1002 	spin_unlock(&ctx->lock);
1003 
1004 	return !dispatch_data->rq;
1005 }
1006 
1007 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1008 					struct blk_mq_ctx *start)
1009 {
1010 	unsigned off = start ? start->index_hw[hctx->type] : 0;
1011 	struct dispatch_rq_data data = {
1012 		.hctx = hctx,
1013 		.rq   = NULL,
1014 	};
1015 
1016 	__sbitmap_for_each_set(&hctx->ctx_map, off,
1017 			       dispatch_rq_from_ctx, &data);
1018 
1019 	return data.rq;
1020 }
1021 
1022 static inline unsigned int queued_to_index(unsigned int queued)
1023 {
1024 	if (!queued)
1025 		return 0;
1026 
1027 	return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1028 }
1029 
1030 bool blk_mq_get_driver_tag(struct request *rq)
1031 {
1032 	struct blk_mq_alloc_data data = {
1033 		.q = rq->q,
1034 		.hctx = rq->mq_hctx,
1035 		.flags = BLK_MQ_REQ_NOWAIT,
1036 		.cmd_flags = rq->cmd_flags,
1037 	};
1038 	bool shared;
1039 
1040 	if (rq->tag != -1)
1041 		return true;
1042 
1043 	if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1044 		data.flags |= BLK_MQ_REQ_RESERVED;
1045 
1046 	shared = blk_mq_tag_busy(data.hctx);
1047 	rq->tag = blk_mq_get_tag(&data);
1048 	if (rq->tag >= 0) {
1049 		if (shared) {
1050 			rq->rq_flags |= RQF_MQ_INFLIGHT;
1051 			atomic_inc(&data.hctx->nr_active);
1052 		}
1053 		data.hctx->tags->rqs[rq->tag] = rq;
1054 	}
1055 
1056 	return rq->tag != -1;
1057 }
1058 
1059 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1060 				int flags, void *key)
1061 {
1062 	struct blk_mq_hw_ctx *hctx;
1063 
1064 	hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1065 
1066 	spin_lock(&hctx->dispatch_wait_lock);
1067 	if (!list_empty(&wait->entry)) {
1068 		struct sbitmap_queue *sbq;
1069 
1070 		list_del_init(&wait->entry);
1071 		sbq = &hctx->tags->bitmap_tags;
1072 		atomic_dec(&sbq->ws_active);
1073 	}
1074 	spin_unlock(&hctx->dispatch_wait_lock);
1075 
1076 	blk_mq_run_hw_queue(hctx, true);
1077 	return 1;
1078 }
1079 
1080 /*
1081  * Mark us waiting for a tag. For shared tags, this involves hooking us into
1082  * the tag wakeups. For non-shared tags, we can simply mark us needing a
1083  * restart. For both cases, take care to check the condition again after
1084  * marking us as waiting.
1085  */
1086 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1087 				 struct request *rq)
1088 {
1089 	struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1090 	struct wait_queue_head *wq;
1091 	wait_queue_entry_t *wait;
1092 	bool ret;
1093 
1094 	if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1095 		blk_mq_sched_mark_restart_hctx(hctx);
1096 
1097 		/*
1098 		 * It's possible that a tag was freed in the window between the
1099 		 * allocation failure and adding the hardware queue to the wait
1100 		 * queue.
1101 		 *
1102 		 * Don't clear RESTART here, someone else could have set it.
1103 		 * At most this will cost an extra queue run.
1104 		 */
1105 		return blk_mq_get_driver_tag(rq);
1106 	}
1107 
1108 	wait = &hctx->dispatch_wait;
1109 	if (!list_empty_careful(&wait->entry))
1110 		return false;
1111 
1112 	wq = &bt_wait_ptr(sbq, hctx)->wait;
1113 
1114 	spin_lock_irq(&wq->lock);
1115 	spin_lock(&hctx->dispatch_wait_lock);
1116 	if (!list_empty(&wait->entry)) {
1117 		spin_unlock(&hctx->dispatch_wait_lock);
1118 		spin_unlock_irq(&wq->lock);
1119 		return false;
1120 	}
1121 
1122 	atomic_inc(&sbq->ws_active);
1123 	wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1124 	__add_wait_queue(wq, wait);
1125 
1126 	/*
1127 	 * It's possible that a tag was freed in the window between the
1128 	 * allocation failure and adding the hardware queue to the wait
1129 	 * queue.
1130 	 */
1131 	ret = blk_mq_get_driver_tag(rq);
1132 	if (!ret) {
1133 		spin_unlock(&hctx->dispatch_wait_lock);
1134 		spin_unlock_irq(&wq->lock);
1135 		return false;
1136 	}
1137 
1138 	/*
1139 	 * We got a tag, remove ourselves from the wait queue to ensure
1140 	 * someone else gets the wakeup.
1141 	 */
1142 	list_del_init(&wait->entry);
1143 	atomic_dec(&sbq->ws_active);
1144 	spin_unlock(&hctx->dispatch_wait_lock);
1145 	spin_unlock_irq(&wq->lock);
1146 
1147 	return true;
1148 }
1149 
1150 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT  8
1151 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR  4
1152 /*
1153  * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1154  * - EWMA is one simple way to compute running average value
1155  * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1156  * - take 4 as factor for avoiding to get too small(0) result, and this
1157  *   factor doesn't matter because EWMA decreases exponentially
1158  */
1159 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1160 {
1161 	unsigned int ewma;
1162 
1163 	if (hctx->queue->elevator)
1164 		return;
1165 
1166 	ewma = hctx->dispatch_busy;
1167 
1168 	if (!ewma && !busy)
1169 		return;
1170 
1171 	ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1172 	if (busy)
1173 		ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1174 	ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1175 
1176 	hctx->dispatch_busy = ewma;
1177 }
1178 
1179 #define BLK_MQ_RESOURCE_DELAY	3		/* ms units */
1180 
1181 /*
1182  * Returns true if we did some work AND can potentially do more.
1183  */
1184 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1185 			     bool got_budget)
1186 {
1187 	struct blk_mq_hw_ctx *hctx;
1188 	struct request *rq, *nxt;
1189 	bool no_tag = false;
1190 	int errors, queued;
1191 	blk_status_t ret = BLK_STS_OK;
1192 
1193 	if (list_empty(list))
1194 		return false;
1195 
1196 	WARN_ON(!list_is_singular(list) && got_budget);
1197 
1198 	/*
1199 	 * Now process all the entries, sending them to the driver.
1200 	 */
1201 	errors = queued = 0;
1202 	do {
1203 		struct blk_mq_queue_data bd;
1204 
1205 		rq = list_first_entry(list, struct request, queuelist);
1206 
1207 		hctx = rq->mq_hctx;
1208 		if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1209 			break;
1210 
1211 		if (!blk_mq_get_driver_tag(rq)) {
1212 			/*
1213 			 * The initial allocation attempt failed, so we need to
1214 			 * rerun the hardware queue when a tag is freed. The
1215 			 * waitqueue takes care of that. If the queue is run
1216 			 * before we add this entry back on the dispatch list,
1217 			 * we'll re-run it below.
1218 			 */
1219 			if (!blk_mq_mark_tag_wait(hctx, rq)) {
1220 				blk_mq_put_dispatch_budget(hctx);
1221 				/*
1222 				 * For non-shared tags, the RESTART check
1223 				 * will suffice.
1224 				 */
1225 				if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1226 					no_tag = true;
1227 				break;
1228 			}
1229 		}
1230 
1231 		list_del_init(&rq->queuelist);
1232 
1233 		bd.rq = rq;
1234 
1235 		/*
1236 		 * Flag last if we have no more requests, or if we have more
1237 		 * but can't assign a driver tag to it.
1238 		 */
1239 		if (list_empty(list))
1240 			bd.last = true;
1241 		else {
1242 			nxt = list_first_entry(list, struct request, queuelist);
1243 			bd.last = !blk_mq_get_driver_tag(nxt);
1244 		}
1245 
1246 		ret = q->mq_ops->queue_rq(hctx, &bd);
1247 		if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1248 			/*
1249 			 * If an I/O scheduler has been configured and we got a
1250 			 * driver tag for the next request already, free it
1251 			 * again.
1252 			 */
1253 			if (!list_empty(list)) {
1254 				nxt = list_first_entry(list, struct request, queuelist);
1255 				blk_mq_put_driver_tag(nxt);
1256 			}
1257 			list_add(&rq->queuelist, list);
1258 			__blk_mq_requeue_request(rq);
1259 			break;
1260 		}
1261 
1262 		if (unlikely(ret != BLK_STS_OK)) {
1263 			errors++;
1264 			blk_mq_end_request(rq, BLK_STS_IOERR);
1265 			continue;
1266 		}
1267 
1268 		queued++;
1269 	} while (!list_empty(list));
1270 
1271 	hctx->dispatched[queued_to_index(queued)]++;
1272 
1273 	/*
1274 	 * Any items that need requeuing? Stuff them into hctx->dispatch,
1275 	 * that is where we will continue on next queue run.
1276 	 */
1277 	if (!list_empty(list)) {
1278 		bool needs_restart;
1279 
1280 		/*
1281 		 * If we didn't flush the entire list, we could have told
1282 		 * the driver there was more coming, but that turned out to
1283 		 * be a lie.
1284 		 */
1285 		if (q->mq_ops->commit_rqs)
1286 			q->mq_ops->commit_rqs(hctx);
1287 
1288 		spin_lock(&hctx->lock);
1289 		list_splice_tail_init(list, &hctx->dispatch);
1290 		spin_unlock(&hctx->lock);
1291 
1292 		/*
1293 		 * If SCHED_RESTART was set by the caller of this function and
1294 		 * it is no longer set that means that it was cleared by another
1295 		 * thread and hence that a queue rerun is needed.
1296 		 *
1297 		 * If 'no_tag' is set, that means that we failed getting
1298 		 * a driver tag with an I/O scheduler attached. If our dispatch
1299 		 * waitqueue is no longer active, ensure that we run the queue
1300 		 * AFTER adding our entries back to the list.
1301 		 *
1302 		 * If no I/O scheduler has been configured it is possible that
1303 		 * the hardware queue got stopped and restarted before requests
1304 		 * were pushed back onto the dispatch list. Rerun the queue to
1305 		 * avoid starvation. Notes:
1306 		 * - blk_mq_run_hw_queue() checks whether or not a queue has
1307 		 *   been stopped before rerunning a queue.
1308 		 * - Some but not all block drivers stop a queue before
1309 		 *   returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1310 		 *   and dm-rq.
1311 		 *
1312 		 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1313 		 * bit is set, run queue after a delay to avoid IO stalls
1314 		 * that could otherwise occur if the queue is idle.
1315 		 */
1316 		needs_restart = blk_mq_sched_needs_restart(hctx);
1317 		if (!needs_restart ||
1318 		    (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1319 			blk_mq_run_hw_queue(hctx, true);
1320 		else if (needs_restart && (ret == BLK_STS_RESOURCE))
1321 			blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1322 
1323 		blk_mq_update_dispatch_busy(hctx, true);
1324 		return false;
1325 	} else
1326 		blk_mq_update_dispatch_busy(hctx, false);
1327 
1328 	/*
1329 	 * If the host/device is unable to accept more work, inform the
1330 	 * caller of that.
1331 	 */
1332 	if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1333 		return false;
1334 
1335 	return (queued + errors) != 0;
1336 }
1337 
1338 /**
1339  * __blk_mq_run_hw_queue - Run a hardware queue.
1340  * @hctx: Pointer to the hardware queue to run.
1341  *
1342  * Send pending requests to the hardware.
1343  */
1344 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1345 {
1346 	int srcu_idx;
1347 
1348 	/*
1349 	 * We should be running this queue from one of the CPUs that
1350 	 * are mapped to it.
1351 	 *
1352 	 * There are at least two related races now between setting
1353 	 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1354 	 * __blk_mq_run_hw_queue():
1355 	 *
1356 	 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1357 	 *   but later it becomes online, then this warning is harmless
1358 	 *   at all
1359 	 *
1360 	 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1361 	 *   but later it becomes offline, then the warning can't be
1362 	 *   triggered, and we depend on blk-mq timeout handler to
1363 	 *   handle dispatched requests to this hctx
1364 	 */
1365 	if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1366 		cpu_online(hctx->next_cpu)) {
1367 		printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1368 			raw_smp_processor_id(),
1369 			cpumask_empty(hctx->cpumask) ? "inactive": "active");
1370 		dump_stack();
1371 	}
1372 
1373 	/*
1374 	 * We can't run the queue inline with ints disabled. Ensure that
1375 	 * we catch bad users of this early.
1376 	 */
1377 	WARN_ON_ONCE(in_interrupt());
1378 
1379 	might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1380 
1381 	hctx_lock(hctx, &srcu_idx);
1382 	blk_mq_sched_dispatch_requests(hctx);
1383 	hctx_unlock(hctx, srcu_idx);
1384 }
1385 
1386 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1387 {
1388 	int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1389 
1390 	if (cpu >= nr_cpu_ids)
1391 		cpu = cpumask_first(hctx->cpumask);
1392 	return cpu;
1393 }
1394 
1395 /*
1396  * It'd be great if the workqueue API had a way to pass
1397  * in a mask and had some smarts for more clever placement.
1398  * For now we just round-robin here, switching for every
1399  * BLK_MQ_CPU_WORK_BATCH queued items.
1400  */
1401 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1402 {
1403 	bool tried = false;
1404 	int next_cpu = hctx->next_cpu;
1405 
1406 	if (hctx->queue->nr_hw_queues == 1)
1407 		return WORK_CPU_UNBOUND;
1408 
1409 	if (--hctx->next_cpu_batch <= 0) {
1410 select_cpu:
1411 		next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1412 				cpu_online_mask);
1413 		if (next_cpu >= nr_cpu_ids)
1414 			next_cpu = blk_mq_first_mapped_cpu(hctx);
1415 		hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1416 	}
1417 
1418 	/*
1419 	 * Do unbound schedule if we can't find a online CPU for this hctx,
1420 	 * and it should only happen in the path of handling CPU DEAD.
1421 	 */
1422 	if (!cpu_online(next_cpu)) {
1423 		if (!tried) {
1424 			tried = true;
1425 			goto select_cpu;
1426 		}
1427 
1428 		/*
1429 		 * Make sure to re-select CPU next time once after CPUs
1430 		 * in hctx->cpumask become online again.
1431 		 */
1432 		hctx->next_cpu = next_cpu;
1433 		hctx->next_cpu_batch = 1;
1434 		return WORK_CPU_UNBOUND;
1435 	}
1436 
1437 	hctx->next_cpu = next_cpu;
1438 	return next_cpu;
1439 }
1440 
1441 /**
1442  * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1443  * @hctx: Pointer to the hardware queue to run.
1444  * @async: If we want to run the queue asynchronously.
1445  * @msecs: Microseconds of delay to wait before running the queue.
1446  *
1447  * If !@async, try to run the queue now. Else, run the queue asynchronously and
1448  * with a delay of @msecs.
1449  */
1450 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1451 					unsigned long msecs)
1452 {
1453 	if (unlikely(blk_mq_hctx_stopped(hctx)))
1454 		return;
1455 
1456 	if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1457 		int cpu = get_cpu();
1458 		if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1459 			__blk_mq_run_hw_queue(hctx);
1460 			put_cpu();
1461 			return;
1462 		}
1463 
1464 		put_cpu();
1465 	}
1466 
1467 	kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1468 				    msecs_to_jiffies(msecs));
1469 }
1470 
1471 /**
1472  * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1473  * @hctx: Pointer to the hardware queue to run.
1474  * @msecs: Microseconds of delay to wait before running the queue.
1475  *
1476  * Run a hardware queue asynchronously with a delay of @msecs.
1477  */
1478 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1479 {
1480 	__blk_mq_delay_run_hw_queue(hctx, true, msecs);
1481 }
1482 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1483 
1484 /**
1485  * blk_mq_run_hw_queue - Start to run a hardware queue.
1486  * @hctx: Pointer to the hardware queue to run.
1487  * @async: If we want to run the queue asynchronously.
1488  *
1489  * Check if the request queue is not in a quiesced state and if there are
1490  * pending requests to be sent. If this is true, run the queue to send requests
1491  * to hardware.
1492  */
1493 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1494 {
1495 	int srcu_idx;
1496 	bool need_run;
1497 
1498 	/*
1499 	 * When queue is quiesced, we may be switching io scheduler, or
1500 	 * updating nr_hw_queues, or other things, and we can't run queue
1501 	 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1502 	 *
1503 	 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1504 	 * quiesced.
1505 	 */
1506 	hctx_lock(hctx, &srcu_idx);
1507 	need_run = !blk_queue_quiesced(hctx->queue) &&
1508 		blk_mq_hctx_has_pending(hctx);
1509 	hctx_unlock(hctx, srcu_idx);
1510 
1511 	if (need_run)
1512 		__blk_mq_delay_run_hw_queue(hctx, async, 0);
1513 }
1514 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1515 
1516 /**
1517  * blk_mq_run_hw_queue - Run all hardware queues in a request queue.
1518  * @q: Pointer to the request queue to run.
1519  * @async: If we want to run the queue asynchronously.
1520  */
1521 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1522 {
1523 	struct blk_mq_hw_ctx *hctx;
1524 	int i;
1525 
1526 	queue_for_each_hw_ctx(q, hctx, i) {
1527 		if (blk_mq_hctx_stopped(hctx))
1528 			continue;
1529 
1530 		blk_mq_run_hw_queue(hctx, async);
1531 	}
1532 }
1533 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1534 
1535 /**
1536  * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1537  * @q: request queue.
1538  *
1539  * The caller is responsible for serializing this function against
1540  * blk_mq_{start,stop}_hw_queue().
1541  */
1542 bool blk_mq_queue_stopped(struct request_queue *q)
1543 {
1544 	struct blk_mq_hw_ctx *hctx;
1545 	int i;
1546 
1547 	queue_for_each_hw_ctx(q, hctx, i)
1548 		if (blk_mq_hctx_stopped(hctx))
1549 			return true;
1550 
1551 	return false;
1552 }
1553 EXPORT_SYMBOL(blk_mq_queue_stopped);
1554 
1555 /*
1556  * This function is often used for pausing .queue_rq() by driver when
1557  * there isn't enough resource or some conditions aren't satisfied, and
1558  * BLK_STS_RESOURCE is usually returned.
1559  *
1560  * We do not guarantee that dispatch can be drained or blocked
1561  * after blk_mq_stop_hw_queue() returns. Please use
1562  * blk_mq_quiesce_queue() for that requirement.
1563  */
1564 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1565 {
1566 	cancel_delayed_work(&hctx->run_work);
1567 
1568 	set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1569 }
1570 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1571 
1572 /*
1573  * This function is often used for pausing .queue_rq() by driver when
1574  * there isn't enough resource or some conditions aren't satisfied, and
1575  * BLK_STS_RESOURCE is usually returned.
1576  *
1577  * We do not guarantee that dispatch can be drained or blocked
1578  * after blk_mq_stop_hw_queues() returns. Please use
1579  * blk_mq_quiesce_queue() for that requirement.
1580  */
1581 void blk_mq_stop_hw_queues(struct request_queue *q)
1582 {
1583 	struct blk_mq_hw_ctx *hctx;
1584 	int i;
1585 
1586 	queue_for_each_hw_ctx(q, hctx, i)
1587 		blk_mq_stop_hw_queue(hctx);
1588 }
1589 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1590 
1591 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1592 {
1593 	clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1594 
1595 	blk_mq_run_hw_queue(hctx, false);
1596 }
1597 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1598 
1599 void blk_mq_start_hw_queues(struct request_queue *q)
1600 {
1601 	struct blk_mq_hw_ctx *hctx;
1602 	int i;
1603 
1604 	queue_for_each_hw_ctx(q, hctx, i)
1605 		blk_mq_start_hw_queue(hctx);
1606 }
1607 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1608 
1609 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1610 {
1611 	if (!blk_mq_hctx_stopped(hctx))
1612 		return;
1613 
1614 	clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1615 	blk_mq_run_hw_queue(hctx, async);
1616 }
1617 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1618 
1619 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1620 {
1621 	struct blk_mq_hw_ctx *hctx;
1622 	int i;
1623 
1624 	queue_for_each_hw_ctx(q, hctx, i)
1625 		blk_mq_start_stopped_hw_queue(hctx, async);
1626 }
1627 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1628 
1629 static void blk_mq_run_work_fn(struct work_struct *work)
1630 {
1631 	struct blk_mq_hw_ctx *hctx;
1632 
1633 	hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1634 
1635 	/*
1636 	 * If we are stopped, don't run the queue.
1637 	 */
1638 	if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1639 		return;
1640 
1641 	__blk_mq_run_hw_queue(hctx);
1642 }
1643 
1644 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1645 					    struct request *rq,
1646 					    bool at_head)
1647 {
1648 	struct blk_mq_ctx *ctx = rq->mq_ctx;
1649 	enum hctx_type type = hctx->type;
1650 
1651 	lockdep_assert_held(&ctx->lock);
1652 
1653 	trace_block_rq_insert(hctx->queue, rq);
1654 
1655 	if (at_head)
1656 		list_add(&rq->queuelist, &ctx->rq_lists[type]);
1657 	else
1658 		list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1659 }
1660 
1661 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1662 			     bool at_head)
1663 {
1664 	struct blk_mq_ctx *ctx = rq->mq_ctx;
1665 
1666 	lockdep_assert_held(&ctx->lock);
1667 
1668 	__blk_mq_insert_req_list(hctx, rq, at_head);
1669 	blk_mq_hctx_mark_pending(hctx, ctx);
1670 }
1671 
1672 /**
1673  * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1674  * @rq: Pointer to request to be inserted.
1675  * @run_queue: If we should run the hardware queue after inserting the request.
1676  *
1677  * Should only be used carefully, when the caller knows we want to
1678  * bypass a potential IO scheduler on the target device.
1679  */
1680 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
1681 				  bool run_queue)
1682 {
1683 	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1684 
1685 	spin_lock(&hctx->lock);
1686 	if (at_head)
1687 		list_add(&rq->queuelist, &hctx->dispatch);
1688 	else
1689 		list_add_tail(&rq->queuelist, &hctx->dispatch);
1690 	spin_unlock(&hctx->lock);
1691 
1692 	if (run_queue)
1693 		blk_mq_run_hw_queue(hctx, false);
1694 }
1695 
1696 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1697 			    struct list_head *list)
1698 
1699 {
1700 	struct request *rq;
1701 	enum hctx_type type = hctx->type;
1702 
1703 	/*
1704 	 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1705 	 * offline now
1706 	 */
1707 	list_for_each_entry(rq, list, queuelist) {
1708 		BUG_ON(rq->mq_ctx != ctx);
1709 		trace_block_rq_insert(hctx->queue, rq);
1710 	}
1711 
1712 	spin_lock(&ctx->lock);
1713 	list_splice_tail_init(list, &ctx->rq_lists[type]);
1714 	blk_mq_hctx_mark_pending(hctx, ctx);
1715 	spin_unlock(&ctx->lock);
1716 }
1717 
1718 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1719 {
1720 	struct request *rqa = container_of(a, struct request, queuelist);
1721 	struct request *rqb = container_of(b, struct request, queuelist);
1722 
1723 	if (rqa->mq_ctx != rqb->mq_ctx)
1724 		return rqa->mq_ctx > rqb->mq_ctx;
1725 	if (rqa->mq_hctx != rqb->mq_hctx)
1726 		return rqa->mq_hctx > rqb->mq_hctx;
1727 
1728 	return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1729 }
1730 
1731 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1732 {
1733 	LIST_HEAD(list);
1734 
1735 	if (list_empty(&plug->mq_list))
1736 		return;
1737 	list_splice_init(&plug->mq_list, &list);
1738 
1739 	if (plug->rq_count > 2 && plug->multiple_queues)
1740 		list_sort(NULL, &list, plug_rq_cmp);
1741 
1742 	plug->rq_count = 0;
1743 
1744 	do {
1745 		struct list_head rq_list;
1746 		struct request *rq, *head_rq = list_entry_rq(list.next);
1747 		struct list_head *pos = &head_rq->queuelist; /* skip first */
1748 		struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx;
1749 		struct blk_mq_ctx *this_ctx = head_rq->mq_ctx;
1750 		unsigned int depth = 1;
1751 
1752 		list_for_each_continue(pos, &list) {
1753 			rq = list_entry_rq(pos);
1754 			BUG_ON(!rq->q);
1755 			if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx)
1756 				break;
1757 			depth++;
1758 		}
1759 
1760 		list_cut_before(&rq_list, &list, pos);
1761 		trace_block_unplug(head_rq->q, depth, !from_schedule);
1762 		blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1763 						from_schedule);
1764 	} while(!list_empty(&list));
1765 }
1766 
1767 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1768 		unsigned int nr_segs)
1769 {
1770 	if (bio->bi_opf & REQ_RAHEAD)
1771 		rq->cmd_flags |= REQ_FAILFAST_MASK;
1772 
1773 	rq->__sector = bio->bi_iter.bi_sector;
1774 	rq->write_hint = bio->bi_write_hint;
1775 	blk_rq_bio_prep(rq, bio, nr_segs);
1776 
1777 	blk_account_io_start(rq, true);
1778 }
1779 
1780 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1781 					    struct request *rq,
1782 					    blk_qc_t *cookie, bool last)
1783 {
1784 	struct request_queue *q = rq->q;
1785 	struct blk_mq_queue_data bd = {
1786 		.rq = rq,
1787 		.last = last,
1788 	};
1789 	blk_qc_t new_cookie;
1790 	blk_status_t ret;
1791 
1792 	new_cookie = request_to_qc_t(hctx, rq);
1793 
1794 	/*
1795 	 * For OK queue, we are done. For error, caller may kill it.
1796 	 * Any other error (busy), just add it to our list as we
1797 	 * previously would have done.
1798 	 */
1799 	ret = q->mq_ops->queue_rq(hctx, &bd);
1800 	switch (ret) {
1801 	case BLK_STS_OK:
1802 		blk_mq_update_dispatch_busy(hctx, false);
1803 		*cookie = new_cookie;
1804 		break;
1805 	case BLK_STS_RESOURCE:
1806 	case BLK_STS_DEV_RESOURCE:
1807 		blk_mq_update_dispatch_busy(hctx, true);
1808 		__blk_mq_requeue_request(rq);
1809 		break;
1810 	default:
1811 		blk_mq_update_dispatch_busy(hctx, false);
1812 		*cookie = BLK_QC_T_NONE;
1813 		break;
1814 	}
1815 
1816 	return ret;
1817 }
1818 
1819 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1820 						struct request *rq,
1821 						blk_qc_t *cookie,
1822 						bool bypass_insert, bool last)
1823 {
1824 	struct request_queue *q = rq->q;
1825 	bool run_queue = true;
1826 
1827 	/*
1828 	 * RCU or SRCU read lock is needed before checking quiesced flag.
1829 	 *
1830 	 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1831 	 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1832 	 * and avoid driver to try to dispatch again.
1833 	 */
1834 	if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1835 		run_queue = false;
1836 		bypass_insert = false;
1837 		goto insert;
1838 	}
1839 
1840 	if (q->elevator && !bypass_insert)
1841 		goto insert;
1842 
1843 	if (!blk_mq_get_dispatch_budget(hctx))
1844 		goto insert;
1845 
1846 	if (!blk_mq_get_driver_tag(rq)) {
1847 		blk_mq_put_dispatch_budget(hctx);
1848 		goto insert;
1849 	}
1850 
1851 	return __blk_mq_issue_directly(hctx, rq, cookie, last);
1852 insert:
1853 	if (bypass_insert)
1854 		return BLK_STS_RESOURCE;
1855 
1856 	blk_mq_request_bypass_insert(rq, false, run_queue);
1857 	return BLK_STS_OK;
1858 }
1859 
1860 /**
1861  * blk_mq_try_issue_directly - Try to send a request directly to device driver.
1862  * @hctx: Pointer of the associated hardware queue.
1863  * @rq: Pointer to request to be sent.
1864  * @cookie: Request queue cookie.
1865  *
1866  * If the device has enough resources to accept a new request now, send the
1867  * request directly to device driver. Else, insert at hctx->dispatch queue, so
1868  * we can try send it another time in the future. Requests inserted at this
1869  * queue have higher priority.
1870  */
1871 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1872 		struct request *rq, blk_qc_t *cookie)
1873 {
1874 	blk_status_t ret;
1875 	int srcu_idx;
1876 
1877 	might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1878 
1879 	hctx_lock(hctx, &srcu_idx);
1880 
1881 	ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
1882 	if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1883 		blk_mq_request_bypass_insert(rq, false, true);
1884 	else if (ret != BLK_STS_OK)
1885 		blk_mq_end_request(rq, ret);
1886 
1887 	hctx_unlock(hctx, srcu_idx);
1888 }
1889 
1890 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
1891 {
1892 	blk_status_t ret;
1893 	int srcu_idx;
1894 	blk_qc_t unused_cookie;
1895 	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1896 
1897 	hctx_lock(hctx, &srcu_idx);
1898 	ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
1899 	hctx_unlock(hctx, srcu_idx);
1900 
1901 	return ret;
1902 }
1903 
1904 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
1905 		struct list_head *list)
1906 {
1907 	while (!list_empty(list)) {
1908 		blk_status_t ret;
1909 		struct request *rq = list_first_entry(list, struct request,
1910 				queuelist);
1911 
1912 		list_del_init(&rq->queuelist);
1913 		ret = blk_mq_request_issue_directly(rq, list_empty(list));
1914 		if (ret != BLK_STS_OK) {
1915 			if (ret == BLK_STS_RESOURCE ||
1916 					ret == BLK_STS_DEV_RESOURCE) {
1917 				blk_mq_request_bypass_insert(rq, false,
1918 							list_empty(list));
1919 				break;
1920 			}
1921 			blk_mq_end_request(rq, ret);
1922 		}
1923 	}
1924 
1925 	/*
1926 	 * If we didn't flush the entire list, we could have told
1927 	 * the driver there was more coming, but that turned out to
1928 	 * be a lie.
1929 	 */
1930 	if (!list_empty(list) && hctx->queue->mq_ops->commit_rqs)
1931 		hctx->queue->mq_ops->commit_rqs(hctx);
1932 }
1933 
1934 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1935 {
1936 	list_add_tail(&rq->queuelist, &plug->mq_list);
1937 	plug->rq_count++;
1938 	if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
1939 		struct request *tmp;
1940 
1941 		tmp = list_first_entry(&plug->mq_list, struct request,
1942 						queuelist);
1943 		if (tmp->q != rq->q)
1944 			plug->multiple_queues = true;
1945 	}
1946 }
1947 
1948 /**
1949  * blk_mq_make_request - Create and send a request to block device.
1950  * @q: Request queue pointer.
1951  * @bio: Bio pointer.
1952  *
1953  * Builds up a request structure from @q and @bio and send to the device. The
1954  * request may not be queued directly to hardware if:
1955  * * This request can be merged with another one
1956  * * We want to place request at plug queue for possible future merging
1957  * * There is an IO scheduler active at this queue
1958  *
1959  * It will not queue the request if there is an error with the bio, or at the
1960  * request creation.
1961  *
1962  * Returns: Request queue cookie.
1963  */
1964 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1965 {
1966 	const int is_sync = op_is_sync(bio->bi_opf);
1967 	const int is_flush_fua = op_is_flush(bio->bi_opf);
1968 	struct blk_mq_alloc_data data = { .flags = 0};
1969 	struct request *rq;
1970 	struct blk_plug *plug;
1971 	struct request *same_queue_rq = NULL;
1972 	unsigned int nr_segs;
1973 	blk_qc_t cookie;
1974 
1975 	blk_queue_bounce(q, &bio);
1976 	__blk_queue_split(q, &bio, &nr_segs);
1977 
1978 	if (!bio_integrity_prep(bio))
1979 		return BLK_QC_T_NONE;
1980 
1981 	if (!is_flush_fua && !blk_queue_nomerges(q) &&
1982 	    blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
1983 		return BLK_QC_T_NONE;
1984 
1985 	if (blk_mq_sched_bio_merge(q, bio, nr_segs))
1986 		return BLK_QC_T_NONE;
1987 
1988 	rq_qos_throttle(q, bio);
1989 
1990 	data.cmd_flags = bio->bi_opf;
1991 	rq = blk_mq_get_request(q, bio, &data);
1992 	if (unlikely(!rq)) {
1993 		rq_qos_cleanup(q, bio);
1994 		if (bio->bi_opf & REQ_NOWAIT)
1995 			bio_wouldblock_error(bio);
1996 		return BLK_QC_T_NONE;
1997 	}
1998 
1999 	trace_block_getrq(q, bio, bio->bi_opf);
2000 
2001 	rq_qos_track(q, rq, bio);
2002 
2003 	cookie = request_to_qc_t(data.hctx, rq);
2004 
2005 	blk_mq_bio_to_request(rq, bio, nr_segs);
2006 
2007 	plug = blk_mq_plug(q, bio);
2008 	if (unlikely(is_flush_fua)) {
2009 		/* Bypass scheduler for flush requests */
2010 		blk_insert_flush(rq);
2011 		blk_mq_run_hw_queue(data.hctx, true);
2012 	} else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs ||
2013 				!blk_queue_nonrot(q))) {
2014 		/*
2015 		 * Use plugging if we have a ->commit_rqs() hook as well, as
2016 		 * we know the driver uses bd->last in a smart fashion.
2017 		 *
2018 		 * Use normal plugging if this disk is slow HDD, as sequential
2019 		 * IO may benefit a lot from plug merging.
2020 		 */
2021 		unsigned int request_count = plug->rq_count;
2022 		struct request *last = NULL;
2023 
2024 		if (!request_count)
2025 			trace_block_plug(q);
2026 		else
2027 			last = list_entry_rq(plug->mq_list.prev);
2028 
2029 		if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
2030 		    blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2031 			blk_flush_plug_list(plug, false);
2032 			trace_block_plug(q);
2033 		}
2034 
2035 		blk_add_rq_to_plug(plug, rq);
2036 	} else if (q->elevator) {
2037 		/* Insert the request at the IO scheduler queue */
2038 		blk_mq_sched_insert_request(rq, false, true, true);
2039 	} else if (plug && !blk_queue_nomerges(q)) {
2040 		/*
2041 		 * We do limited plugging. If the bio can be merged, do that.
2042 		 * Otherwise the existing request in the plug list will be
2043 		 * issued. So the plug list will have one request at most
2044 		 * The plug list might get flushed before this. If that happens,
2045 		 * the plug list is empty, and same_queue_rq is invalid.
2046 		 */
2047 		if (list_empty(&plug->mq_list))
2048 			same_queue_rq = NULL;
2049 		if (same_queue_rq) {
2050 			list_del_init(&same_queue_rq->queuelist);
2051 			plug->rq_count--;
2052 		}
2053 		blk_add_rq_to_plug(plug, rq);
2054 		trace_block_plug(q);
2055 
2056 		if (same_queue_rq) {
2057 			data.hctx = same_queue_rq->mq_hctx;
2058 			trace_block_unplug(q, 1, true);
2059 			blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2060 					&cookie);
2061 		}
2062 	} else if ((q->nr_hw_queues > 1 && is_sync) ||
2063 			!data.hctx->dispatch_busy) {
2064 		/*
2065 		 * There is no scheduler and we can try to send directly
2066 		 * to the hardware.
2067 		 */
2068 		blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2069 	} else {
2070 		/* Default case. */
2071 		blk_mq_sched_insert_request(rq, false, true, true);
2072 	}
2073 
2074 	return cookie;
2075 }
2076 
2077 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2078 		     unsigned int hctx_idx)
2079 {
2080 	struct page *page;
2081 
2082 	if (tags->rqs && set->ops->exit_request) {
2083 		int i;
2084 
2085 		for (i = 0; i < tags->nr_tags; i++) {
2086 			struct request *rq = tags->static_rqs[i];
2087 
2088 			if (!rq)
2089 				continue;
2090 			set->ops->exit_request(set, rq, hctx_idx);
2091 			tags->static_rqs[i] = NULL;
2092 		}
2093 	}
2094 
2095 	while (!list_empty(&tags->page_list)) {
2096 		page = list_first_entry(&tags->page_list, struct page, lru);
2097 		list_del_init(&page->lru);
2098 		/*
2099 		 * Remove kmemleak object previously allocated in
2100 		 * blk_mq_alloc_rqs().
2101 		 */
2102 		kmemleak_free(page_address(page));
2103 		__free_pages(page, page->private);
2104 	}
2105 }
2106 
2107 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2108 {
2109 	kfree(tags->rqs);
2110 	tags->rqs = NULL;
2111 	kfree(tags->static_rqs);
2112 	tags->static_rqs = NULL;
2113 
2114 	blk_mq_free_tags(tags);
2115 }
2116 
2117 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2118 					unsigned int hctx_idx,
2119 					unsigned int nr_tags,
2120 					unsigned int reserved_tags)
2121 {
2122 	struct blk_mq_tags *tags;
2123 	int node;
2124 
2125 	node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2126 	if (node == NUMA_NO_NODE)
2127 		node = set->numa_node;
2128 
2129 	tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2130 				BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2131 	if (!tags)
2132 		return NULL;
2133 
2134 	tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2135 				 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2136 				 node);
2137 	if (!tags->rqs) {
2138 		blk_mq_free_tags(tags);
2139 		return NULL;
2140 	}
2141 
2142 	tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2143 					GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2144 					node);
2145 	if (!tags->static_rqs) {
2146 		kfree(tags->rqs);
2147 		blk_mq_free_tags(tags);
2148 		return NULL;
2149 	}
2150 
2151 	return tags;
2152 }
2153 
2154 static size_t order_to_size(unsigned int order)
2155 {
2156 	return (size_t)PAGE_SIZE << order;
2157 }
2158 
2159 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2160 			       unsigned int hctx_idx, int node)
2161 {
2162 	int ret;
2163 
2164 	if (set->ops->init_request) {
2165 		ret = set->ops->init_request(set, rq, hctx_idx, node);
2166 		if (ret)
2167 			return ret;
2168 	}
2169 
2170 	WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2171 	return 0;
2172 }
2173 
2174 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2175 		     unsigned int hctx_idx, unsigned int depth)
2176 {
2177 	unsigned int i, j, entries_per_page, max_order = 4;
2178 	size_t rq_size, left;
2179 	int node;
2180 
2181 	node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2182 	if (node == NUMA_NO_NODE)
2183 		node = set->numa_node;
2184 
2185 	INIT_LIST_HEAD(&tags->page_list);
2186 
2187 	/*
2188 	 * rq_size is the size of the request plus driver payload, rounded
2189 	 * to the cacheline size
2190 	 */
2191 	rq_size = round_up(sizeof(struct request) + set->cmd_size,
2192 				cache_line_size());
2193 	left = rq_size * depth;
2194 
2195 	for (i = 0; i < depth; ) {
2196 		int this_order = max_order;
2197 		struct page *page;
2198 		int to_do;
2199 		void *p;
2200 
2201 		while (this_order && left < order_to_size(this_order - 1))
2202 			this_order--;
2203 
2204 		do {
2205 			page = alloc_pages_node(node,
2206 				GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2207 				this_order);
2208 			if (page)
2209 				break;
2210 			if (!this_order--)
2211 				break;
2212 			if (order_to_size(this_order) < rq_size)
2213 				break;
2214 		} while (1);
2215 
2216 		if (!page)
2217 			goto fail;
2218 
2219 		page->private = this_order;
2220 		list_add_tail(&page->lru, &tags->page_list);
2221 
2222 		p = page_address(page);
2223 		/*
2224 		 * Allow kmemleak to scan these pages as they contain pointers
2225 		 * to additional allocations like via ops->init_request().
2226 		 */
2227 		kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2228 		entries_per_page = order_to_size(this_order) / rq_size;
2229 		to_do = min(entries_per_page, depth - i);
2230 		left -= to_do * rq_size;
2231 		for (j = 0; j < to_do; j++) {
2232 			struct request *rq = p;
2233 
2234 			tags->static_rqs[i] = rq;
2235 			if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2236 				tags->static_rqs[i] = NULL;
2237 				goto fail;
2238 			}
2239 
2240 			p += rq_size;
2241 			i++;
2242 		}
2243 	}
2244 	return 0;
2245 
2246 fail:
2247 	blk_mq_free_rqs(set, tags, hctx_idx);
2248 	return -ENOMEM;
2249 }
2250 
2251 /*
2252  * 'cpu' is going away. splice any existing rq_list entries from this
2253  * software queue to the hw queue dispatch list, and ensure that it
2254  * gets run.
2255  */
2256 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2257 {
2258 	struct blk_mq_hw_ctx *hctx;
2259 	struct blk_mq_ctx *ctx;
2260 	LIST_HEAD(tmp);
2261 	enum hctx_type type;
2262 
2263 	hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2264 	ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2265 	type = hctx->type;
2266 
2267 	spin_lock(&ctx->lock);
2268 	if (!list_empty(&ctx->rq_lists[type])) {
2269 		list_splice_init(&ctx->rq_lists[type], &tmp);
2270 		blk_mq_hctx_clear_pending(hctx, ctx);
2271 	}
2272 	spin_unlock(&ctx->lock);
2273 
2274 	if (list_empty(&tmp))
2275 		return 0;
2276 
2277 	spin_lock(&hctx->lock);
2278 	list_splice_tail_init(&tmp, &hctx->dispatch);
2279 	spin_unlock(&hctx->lock);
2280 
2281 	blk_mq_run_hw_queue(hctx, true);
2282 	return 0;
2283 }
2284 
2285 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2286 {
2287 	cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2288 					    &hctx->cpuhp_dead);
2289 }
2290 
2291 /* hctx->ctxs will be freed in queue's release handler */
2292 static void blk_mq_exit_hctx(struct request_queue *q,
2293 		struct blk_mq_tag_set *set,
2294 		struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2295 {
2296 	if (blk_mq_hw_queue_mapped(hctx))
2297 		blk_mq_tag_idle(hctx);
2298 
2299 	if (set->ops->exit_request)
2300 		set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2301 
2302 	if (set->ops->exit_hctx)
2303 		set->ops->exit_hctx(hctx, hctx_idx);
2304 
2305 	blk_mq_remove_cpuhp(hctx);
2306 
2307 	spin_lock(&q->unused_hctx_lock);
2308 	list_add(&hctx->hctx_list, &q->unused_hctx_list);
2309 	spin_unlock(&q->unused_hctx_lock);
2310 }
2311 
2312 static void blk_mq_exit_hw_queues(struct request_queue *q,
2313 		struct blk_mq_tag_set *set, int nr_queue)
2314 {
2315 	struct blk_mq_hw_ctx *hctx;
2316 	unsigned int i;
2317 
2318 	queue_for_each_hw_ctx(q, hctx, i) {
2319 		if (i == nr_queue)
2320 			break;
2321 		blk_mq_debugfs_unregister_hctx(hctx);
2322 		blk_mq_exit_hctx(q, set, hctx, i);
2323 	}
2324 }
2325 
2326 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2327 {
2328 	int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2329 
2330 	BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2331 			   __alignof__(struct blk_mq_hw_ctx)) !=
2332 		     sizeof(struct blk_mq_hw_ctx));
2333 
2334 	if (tag_set->flags & BLK_MQ_F_BLOCKING)
2335 		hw_ctx_size += sizeof(struct srcu_struct);
2336 
2337 	return hw_ctx_size;
2338 }
2339 
2340 static int blk_mq_init_hctx(struct request_queue *q,
2341 		struct blk_mq_tag_set *set,
2342 		struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2343 {
2344 	hctx->queue_num = hctx_idx;
2345 
2346 	cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2347 
2348 	hctx->tags = set->tags[hctx_idx];
2349 
2350 	if (set->ops->init_hctx &&
2351 	    set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2352 		goto unregister_cpu_notifier;
2353 
2354 	if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2355 				hctx->numa_node))
2356 		goto exit_hctx;
2357 	return 0;
2358 
2359  exit_hctx:
2360 	if (set->ops->exit_hctx)
2361 		set->ops->exit_hctx(hctx, hctx_idx);
2362  unregister_cpu_notifier:
2363 	blk_mq_remove_cpuhp(hctx);
2364 	return -1;
2365 }
2366 
2367 static struct blk_mq_hw_ctx *
2368 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2369 		int node)
2370 {
2371 	struct blk_mq_hw_ctx *hctx;
2372 	gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2373 
2374 	hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2375 	if (!hctx)
2376 		goto fail_alloc_hctx;
2377 
2378 	if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2379 		goto free_hctx;
2380 
2381 	atomic_set(&hctx->nr_active, 0);
2382 	if (node == NUMA_NO_NODE)
2383 		node = set->numa_node;
2384 	hctx->numa_node = node;
2385 
2386 	INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2387 	spin_lock_init(&hctx->lock);
2388 	INIT_LIST_HEAD(&hctx->dispatch);
2389 	hctx->queue = q;
2390 	hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2391 
2392 	INIT_LIST_HEAD(&hctx->hctx_list);
2393 
2394 	/*
2395 	 * Allocate space for all possible cpus to avoid allocation at
2396 	 * runtime
2397 	 */
2398 	hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2399 			gfp, node);
2400 	if (!hctx->ctxs)
2401 		goto free_cpumask;
2402 
2403 	if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2404 				gfp, node))
2405 		goto free_ctxs;
2406 	hctx->nr_ctx = 0;
2407 
2408 	spin_lock_init(&hctx->dispatch_wait_lock);
2409 	init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2410 	INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2411 
2412 	hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size,
2413 			gfp);
2414 	if (!hctx->fq)
2415 		goto free_bitmap;
2416 
2417 	if (hctx->flags & BLK_MQ_F_BLOCKING)
2418 		init_srcu_struct(hctx->srcu);
2419 	blk_mq_hctx_kobj_init(hctx);
2420 
2421 	return hctx;
2422 
2423  free_bitmap:
2424 	sbitmap_free(&hctx->ctx_map);
2425  free_ctxs:
2426 	kfree(hctx->ctxs);
2427  free_cpumask:
2428 	free_cpumask_var(hctx->cpumask);
2429  free_hctx:
2430 	kfree(hctx);
2431  fail_alloc_hctx:
2432 	return NULL;
2433 }
2434 
2435 static void blk_mq_init_cpu_queues(struct request_queue *q,
2436 				   unsigned int nr_hw_queues)
2437 {
2438 	struct blk_mq_tag_set *set = q->tag_set;
2439 	unsigned int i, j;
2440 
2441 	for_each_possible_cpu(i) {
2442 		struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2443 		struct blk_mq_hw_ctx *hctx;
2444 		int k;
2445 
2446 		__ctx->cpu = i;
2447 		spin_lock_init(&__ctx->lock);
2448 		for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2449 			INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2450 
2451 		__ctx->queue = q;
2452 
2453 		/*
2454 		 * Set local node, IFF we have more than one hw queue. If
2455 		 * not, we remain on the home node of the device
2456 		 */
2457 		for (j = 0; j < set->nr_maps; j++) {
2458 			hctx = blk_mq_map_queue_type(q, j, i);
2459 			if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2460 				hctx->numa_node = local_memory_node(cpu_to_node(i));
2461 		}
2462 	}
2463 }
2464 
2465 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2466 {
2467 	int ret = 0;
2468 
2469 	set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2470 					set->queue_depth, set->reserved_tags);
2471 	if (!set->tags[hctx_idx])
2472 		return false;
2473 
2474 	ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2475 				set->queue_depth);
2476 	if (!ret)
2477 		return true;
2478 
2479 	blk_mq_free_rq_map(set->tags[hctx_idx]);
2480 	set->tags[hctx_idx] = NULL;
2481 	return false;
2482 }
2483 
2484 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2485 					 unsigned int hctx_idx)
2486 {
2487 	if (set->tags && set->tags[hctx_idx]) {
2488 		blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2489 		blk_mq_free_rq_map(set->tags[hctx_idx]);
2490 		set->tags[hctx_idx] = NULL;
2491 	}
2492 }
2493 
2494 static void blk_mq_map_swqueue(struct request_queue *q)
2495 {
2496 	unsigned int i, j, hctx_idx;
2497 	struct blk_mq_hw_ctx *hctx;
2498 	struct blk_mq_ctx *ctx;
2499 	struct blk_mq_tag_set *set = q->tag_set;
2500 
2501 	queue_for_each_hw_ctx(q, hctx, i) {
2502 		cpumask_clear(hctx->cpumask);
2503 		hctx->nr_ctx = 0;
2504 		hctx->dispatch_from = NULL;
2505 	}
2506 
2507 	/*
2508 	 * Map software to hardware queues.
2509 	 *
2510 	 * If the cpu isn't present, the cpu is mapped to first hctx.
2511 	 */
2512 	for_each_possible_cpu(i) {
2513 		hctx_idx = set->map[HCTX_TYPE_DEFAULT].mq_map[i];
2514 		/* unmapped hw queue can be remapped after CPU topo changed */
2515 		if (!set->tags[hctx_idx] &&
2516 		    !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2517 			/*
2518 			 * If tags initialization fail for some hctx,
2519 			 * that hctx won't be brought online.  In this
2520 			 * case, remap the current ctx to hctx[0] which
2521 			 * is guaranteed to always have tags allocated
2522 			 */
2523 			set->map[HCTX_TYPE_DEFAULT].mq_map[i] = 0;
2524 		}
2525 
2526 		ctx = per_cpu_ptr(q->queue_ctx, i);
2527 		for (j = 0; j < set->nr_maps; j++) {
2528 			if (!set->map[j].nr_queues) {
2529 				ctx->hctxs[j] = blk_mq_map_queue_type(q,
2530 						HCTX_TYPE_DEFAULT, i);
2531 				continue;
2532 			}
2533 
2534 			hctx = blk_mq_map_queue_type(q, j, i);
2535 			ctx->hctxs[j] = hctx;
2536 			/*
2537 			 * If the CPU is already set in the mask, then we've
2538 			 * mapped this one already. This can happen if
2539 			 * devices share queues across queue maps.
2540 			 */
2541 			if (cpumask_test_cpu(i, hctx->cpumask))
2542 				continue;
2543 
2544 			cpumask_set_cpu(i, hctx->cpumask);
2545 			hctx->type = j;
2546 			ctx->index_hw[hctx->type] = hctx->nr_ctx;
2547 			hctx->ctxs[hctx->nr_ctx++] = ctx;
2548 
2549 			/*
2550 			 * If the nr_ctx type overflows, we have exceeded the
2551 			 * amount of sw queues we can support.
2552 			 */
2553 			BUG_ON(!hctx->nr_ctx);
2554 		}
2555 
2556 		for (; j < HCTX_MAX_TYPES; j++)
2557 			ctx->hctxs[j] = blk_mq_map_queue_type(q,
2558 					HCTX_TYPE_DEFAULT, i);
2559 	}
2560 
2561 	queue_for_each_hw_ctx(q, hctx, i) {
2562 		/*
2563 		 * If no software queues are mapped to this hardware queue,
2564 		 * disable it and free the request entries.
2565 		 */
2566 		if (!hctx->nr_ctx) {
2567 			/* Never unmap queue 0.  We need it as a
2568 			 * fallback in case of a new remap fails
2569 			 * allocation
2570 			 */
2571 			if (i && set->tags[i])
2572 				blk_mq_free_map_and_requests(set, i);
2573 
2574 			hctx->tags = NULL;
2575 			continue;
2576 		}
2577 
2578 		hctx->tags = set->tags[i];
2579 		WARN_ON(!hctx->tags);
2580 
2581 		/*
2582 		 * Set the map size to the number of mapped software queues.
2583 		 * This is more accurate and more efficient than looping
2584 		 * over all possibly mapped software queues.
2585 		 */
2586 		sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2587 
2588 		/*
2589 		 * Initialize batch roundrobin counts
2590 		 */
2591 		hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2592 		hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2593 	}
2594 }
2595 
2596 /*
2597  * Caller needs to ensure that we're either frozen/quiesced, or that
2598  * the queue isn't live yet.
2599  */
2600 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2601 {
2602 	struct blk_mq_hw_ctx *hctx;
2603 	int i;
2604 
2605 	queue_for_each_hw_ctx(q, hctx, i) {
2606 		if (shared)
2607 			hctx->flags |= BLK_MQ_F_TAG_SHARED;
2608 		else
2609 			hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2610 	}
2611 }
2612 
2613 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2614 					bool shared)
2615 {
2616 	struct request_queue *q;
2617 
2618 	lockdep_assert_held(&set->tag_list_lock);
2619 
2620 	list_for_each_entry(q, &set->tag_list, tag_set_list) {
2621 		blk_mq_freeze_queue(q);
2622 		queue_set_hctx_shared(q, shared);
2623 		blk_mq_unfreeze_queue(q);
2624 	}
2625 }
2626 
2627 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2628 {
2629 	struct blk_mq_tag_set *set = q->tag_set;
2630 
2631 	mutex_lock(&set->tag_list_lock);
2632 	list_del_rcu(&q->tag_set_list);
2633 	if (list_is_singular(&set->tag_list)) {
2634 		/* just transitioned to unshared */
2635 		set->flags &= ~BLK_MQ_F_TAG_SHARED;
2636 		/* update existing queue */
2637 		blk_mq_update_tag_set_depth(set, false);
2638 	}
2639 	mutex_unlock(&set->tag_list_lock);
2640 	INIT_LIST_HEAD(&q->tag_set_list);
2641 }
2642 
2643 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2644 				     struct request_queue *q)
2645 {
2646 	mutex_lock(&set->tag_list_lock);
2647 
2648 	/*
2649 	 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2650 	 */
2651 	if (!list_empty(&set->tag_list) &&
2652 	    !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2653 		set->flags |= BLK_MQ_F_TAG_SHARED;
2654 		/* update existing queue */
2655 		blk_mq_update_tag_set_depth(set, true);
2656 	}
2657 	if (set->flags & BLK_MQ_F_TAG_SHARED)
2658 		queue_set_hctx_shared(q, true);
2659 	list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2660 
2661 	mutex_unlock(&set->tag_list_lock);
2662 }
2663 
2664 /* All allocations will be freed in release handler of q->mq_kobj */
2665 static int blk_mq_alloc_ctxs(struct request_queue *q)
2666 {
2667 	struct blk_mq_ctxs *ctxs;
2668 	int cpu;
2669 
2670 	ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2671 	if (!ctxs)
2672 		return -ENOMEM;
2673 
2674 	ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2675 	if (!ctxs->queue_ctx)
2676 		goto fail;
2677 
2678 	for_each_possible_cpu(cpu) {
2679 		struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2680 		ctx->ctxs = ctxs;
2681 	}
2682 
2683 	q->mq_kobj = &ctxs->kobj;
2684 	q->queue_ctx = ctxs->queue_ctx;
2685 
2686 	return 0;
2687  fail:
2688 	kfree(ctxs);
2689 	return -ENOMEM;
2690 }
2691 
2692 /*
2693  * It is the actual release handler for mq, but we do it from
2694  * request queue's release handler for avoiding use-after-free
2695  * and headache because q->mq_kobj shouldn't have been introduced,
2696  * but we can't group ctx/kctx kobj without it.
2697  */
2698 void blk_mq_release(struct request_queue *q)
2699 {
2700 	struct blk_mq_hw_ctx *hctx, *next;
2701 	int i;
2702 
2703 	queue_for_each_hw_ctx(q, hctx, i)
2704 		WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
2705 
2706 	/* all hctx are in .unused_hctx_list now */
2707 	list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
2708 		list_del_init(&hctx->hctx_list);
2709 		kobject_put(&hctx->kobj);
2710 	}
2711 
2712 	kfree(q->queue_hw_ctx);
2713 
2714 	/*
2715 	 * release .mq_kobj and sw queue's kobject now because
2716 	 * both share lifetime with request queue.
2717 	 */
2718 	blk_mq_sysfs_deinit(q);
2719 }
2720 
2721 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2722 {
2723 	struct request_queue *uninit_q, *q;
2724 
2725 	uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2726 	if (!uninit_q)
2727 		return ERR_PTR(-ENOMEM);
2728 
2729 	/*
2730 	 * Initialize the queue without an elevator. device_add_disk() will do
2731 	 * the initialization.
2732 	 */
2733 	q = blk_mq_init_allocated_queue(set, uninit_q, false);
2734 	if (IS_ERR(q))
2735 		blk_cleanup_queue(uninit_q);
2736 
2737 	return q;
2738 }
2739 EXPORT_SYMBOL(blk_mq_init_queue);
2740 
2741 /*
2742  * Helper for setting up a queue with mq ops, given queue depth, and
2743  * the passed in mq ops flags.
2744  */
2745 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
2746 					   const struct blk_mq_ops *ops,
2747 					   unsigned int queue_depth,
2748 					   unsigned int set_flags)
2749 {
2750 	struct request_queue *q;
2751 	int ret;
2752 
2753 	memset(set, 0, sizeof(*set));
2754 	set->ops = ops;
2755 	set->nr_hw_queues = 1;
2756 	set->nr_maps = 1;
2757 	set->queue_depth = queue_depth;
2758 	set->numa_node = NUMA_NO_NODE;
2759 	set->flags = set_flags;
2760 
2761 	ret = blk_mq_alloc_tag_set(set);
2762 	if (ret)
2763 		return ERR_PTR(ret);
2764 
2765 	q = blk_mq_init_queue(set);
2766 	if (IS_ERR(q)) {
2767 		blk_mq_free_tag_set(set);
2768 		return q;
2769 	}
2770 
2771 	return q;
2772 }
2773 EXPORT_SYMBOL(blk_mq_init_sq_queue);
2774 
2775 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
2776 		struct blk_mq_tag_set *set, struct request_queue *q,
2777 		int hctx_idx, int node)
2778 {
2779 	struct blk_mq_hw_ctx *hctx = NULL, *tmp;
2780 
2781 	/* reuse dead hctx first */
2782 	spin_lock(&q->unused_hctx_lock);
2783 	list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
2784 		if (tmp->numa_node == node) {
2785 			hctx = tmp;
2786 			break;
2787 		}
2788 	}
2789 	if (hctx)
2790 		list_del_init(&hctx->hctx_list);
2791 	spin_unlock(&q->unused_hctx_lock);
2792 
2793 	if (!hctx)
2794 		hctx = blk_mq_alloc_hctx(q, set, node);
2795 	if (!hctx)
2796 		goto fail;
2797 
2798 	if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
2799 		goto free_hctx;
2800 
2801 	return hctx;
2802 
2803  free_hctx:
2804 	kobject_put(&hctx->kobj);
2805  fail:
2806 	return NULL;
2807 }
2808 
2809 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2810 						struct request_queue *q)
2811 {
2812 	int i, j, end;
2813 	struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2814 
2815 	if (q->nr_hw_queues < set->nr_hw_queues) {
2816 		struct blk_mq_hw_ctx **new_hctxs;
2817 
2818 		new_hctxs = kcalloc_node(set->nr_hw_queues,
2819 				       sizeof(*new_hctxs), GFP_KERNEL,
2820 				       set->numa_node);
2821 		if (!new_hctxs)
2822 			return;
2823 		if (hctxs)
2824 			memcpy(new_hctxs, hctxs, q->nr_hw_queues *
2825 			       sizeof(*hctxs));
2826 		q->queue_hw_ctx = new_hctxs;
2827 		q->nr_hw_queues = set->nr_hw_queues;
2828 		kfree(hctxs);
2829 		hctxs = new_hctxs;
2830 	}
2831 
2832 	/* protect against switching io scheduler  */
2833 	mutex_lock(&q->sysfs_lock);
2834 	for (i = 0; i < set->nr_hw_queues; i++) {
2835 		int node;
2836 		struct blk_mq_hw_ctx *hctx;
2837 
2838 		node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
2839 		/*
2840 		 * If the hw queue has been mapped to another numa node,
2841 		 * we need to realloc the hctx. If allocation fails, fallback
2842 		 * to use the previous one.
2843 		 */
2844 		if (hctxs[i] && (hctxs[i]->numa_node == node))
2845 			continue;
2846 
2847 		hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
2848 		if (hctx) {
2849 			if (hctxs[i])
2850 				blk_mq_exit_hctx(q, set, hctxs[i], i);
2851 			hctxs[i] = hctx;
2852 		} else {
2853 			if (hctxs[i])
2854 				pr_warn("Allocate new hctx on node %d fails,\
2855 						fallback to previous one on node %d\n",
2856 						node, hctxs[i]->numa_node);
2857 			else
2858 				break;
2859 		}
2860 	}
2861 	/*
2862 	 * Increasing nr_hw_queues fails. Free the newly allocated
2863 	 * hctxs and keep the previous q->nr_hw_queues.
2864 	 */
2865 	if (i != set->nr_hw_queues) {
2866 		j = q->nr_hw_queues;
2867 		end = i;
2868 	} else {
2869 		j = i;
2870 		end = q->nr_hw_queues;
2871 		q->nr_hw_queues = set->nr_hw_queues;
2872 	}
2873 
2874 	for (; j < end; j++) {
2875 		struct blk_mq_hw_ctx *hctx = hctxs[j];
2876 
2877 		if (hctx) {
2878 			if (hctx->tags)
2879 				blk_mq_free_map_and_requests(set, j);
2880 			blk_mq_exit_hctx(q, set, hctx, j);
2881 			hctxs[j] = NULL;
2882 		}
2883 	}
2884 	mutex_unlock(&q->sysfs_lock);
2885 }
2886 
2887 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2888 						  struct request_queue *q,
2889 						  bool elevator_init)
2890 {
2891 	/* mark the queue as mq asap */
2892 	q->mq_ops = set->ops;
2893 
2894 	q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2895 					     blk_mq_poll_stats_bkt,
2896 					     BLK_MQ_POLL_STATS_BKTS, q);
2897 	if (!q->poll_cb)
2898 		goto err_exit;
2899 
2900 	if (blk_mq_alloc_ctxs(q))
2901 		goto err_poll;
2902 
2903 	/* init q->mq_kobj and sw queues' kobjects */
2904 	blk_mq_sysfs_init(q);
2905 
2906 	INIT_LIST_HEAD(&q->unused_hctx_list);
2907 	spin_lock_init(&q->unused_hctx_lock);
2908 
2909 	blk_mq_realloc_hw_ctxs(set, q);
2910 	if (!q->nr_hw_queues)
2911 		goto err_hctxs;
2912 
2913 	INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2914 	blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2915 
2916 	q->tag_set = set;
2917 
2918 	q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2919 	if (set->nr_maps > HCTX_TYPE_POLL &&
2920 	    set->map[HCTX_TYPE_POLL].nr_queues)
2921 		blk_queue_flag_set(QUEUE_FLAG_POLL, q);
2922 
2923 	q->sg_reserved_size = INT_MAX;
2924 
2925 	INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2926 	INIT_LIST_HEAD(&q->requeue_list);
2927 	spin_lock_init(&q->requeue_lock);
2928 
2929 	blk_queue_make_request(q, blk_mq_make_request);
2930 
2931 	/*
2932 	 * Do this after blk_queue_make_request() overrides it...
2933 	 */
2934 	q->nr_requests = set->queue_depth;
2935 
2936 	/*
2937 	 * Default to classic polling
2938 	 */
2939 	q->poll_nsec = BLK_MQ_POLL_CLASSIC;
2940 
2941 	blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2942 	blk_mq_add_queue_tag_set(set, q);
2943 	blk_mq_map_swqueue(q);
2944 
2945 	if (elevator_init)
2946 		elevator_init_mq(q);
2947 
2948 	return q;
2949 
2950 err_hctxs:
2951 	kfree(q->queue_hw_ctx);
2952 	q->nr_hw_queues = 0;
2953 	blk_mq_sysfs_deinit(q);
2954 err_poll:
2955 	blk_stat_free_callback(q->poll_cb);
2956 	q->poll_cb = NULL;
2957 err_exit:
2958 	q->mq_ops = NULL;
2959 	return ERR_PTR(-ENOMEM);
2960 }
2961 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2962 
2963 /* tags can _not_ be used after returning from blk_mq_exit_queue */
2964 void blk_mq_exit_queue(struct request_queue *q)
2965 {
2966 	struct blk_mq_tag_set	*set = q->tag_set;
2967 
2968 	blk_mq_del_queue_tag_set(q);
2969 	blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2970 }
2971 
2972 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2973 {
2974 	int i;
2975 
2976 	for (i = 0; i < set->nr_hw_queues; i++)
2977 		if (!__blk_mq_alloc_rq_map(set, i))
2978 			goto out_unwind;
2979 
2980 	return 0;
2981 
2982 out_unwind:
2983 	while (--i >= 0)
2984 		blk_mq_free_rq_map(set->tags[i]);
2985 
2986 	return -ENOMEM;
2987 }
2988 
2989 /*
2990  * Allocate the request maps associated with this tag_set. Note that this
2991  * may reduce the depth asked for, if memory is tight. set->queue_depth
2992  * will be updated to reflect the allocated depth.
2993  */
2994 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2995 {
2996 	unsigned int depth;
2997 	int err;
2998 
2999 	depth = set->queue_depth;
3000 	do {
3001 		err = __blk_mq_alloc_rq_maps(set);
3002 		if (!err)
3003 			break;
3004 
3005 		set->queue_depth >>= 1;
3006 		if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
3007 			err = -ENOMEM;
3008 			break;
3009 		}
3010 	} while (set->queue_depth);
3011 
3012 	if (!set->queue_depth || err) {
3013 		pr_err("blk-mq: failed to allocate request map\n");
3014 		return -ENOMEM;
3015 	}
3016 
3017 	if (depth != set->queue_depth)
3018 		pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3019 						depth, set->queue_depth);
3020 
3021 	return 0;
3022 }
3023 
3024 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3025 {
3026 	if (set->ops->map_queues && !is_kdump_kernel()) {
3027 		int i;
3028 
3029 		/*
3030 		 * transport .map_queues is usually done in the following
3031 		 * way:
3032 		 *
3033 		 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3034 		 * 	mask = get_cpu_mask(queue)
3035 		 * 	for_each_cpu(cpu, mask)
3036 		 * 		set->map[x].mq_map[cpu] = queue;
3037 		 * }
3038 		 *
3039 		 * When we need to remap, the table has to be cleared for
3040 		 * killing stale mapping since one CPU may not be mapped
3041 		 * to any hw queue.
3042 		 */
3043 		for (i = 0; i < set->nr_maps; i++)
3044 			blk_mq_clear_mq_map(&set->map[i]);
3045 
3046 		return set->ops->map_queues(set);
3047 	} else {
3048 		BUG_ON(set->nr_maps > 1);
3049 		return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3050 	}
3051 }
3052 
3053 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
3054 				  int cur_nr_hw_queues, int new_nr_hw_queues)
3055 {
3056 	struct blk_mq_tags **new_tags;
3057 
3058 	if (cur_nr_hw_queues >= new_nr_hw_queues)
3059 		return 0;
3060 
3061 	new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
3062 				GFP_KERNEL, set->numa_node);
3063 	if (!new_tags)
3064 		return -ENOMEM;
3065 
3066 	if (set->tags)
3067 		memcpy(new_tags, set->tags, cur_nr_hw_queues *
3068 		       sizeof(*set->tags));
3069 	kfree(set->tags);
3070 	set->tags = new_tags;
3071 	set->nr_hw_queues = new_nr_hw_queues;
3072 
3073 	return 0;
3074 }
3075 
3076 /*
3077  * Alloc a tag set to be associated with one or more request queues.
3078  * May fail with EINVAL for various error conditions. May adjust the
3079  * requested depth down, if it's too large. In that case, the set
3080  * value will be stored in set->queue_depth.
3081  */
3082 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3083 {
3084 	int i, ret;
3085 
3086 	BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3087 
3088 	if (!set->nr_hw_queues)
3089 		return -EINVAL;
3090 	if (!set->queue_depth)
3091 		return -EINVAL;
3092 	if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3093 		return -EINVAL;
3094 
3095 	if (!set->ops->queue_rq)
3096 		return -EINVAL;
3097 
3098 	if (!set->ops->get_budget ^ !set->ops->put_budget)
3099 		return -EINVAL;
3100 
3101 	if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3102 		pr_info("blk-mq: reduced tag depth to %u\n",
3103 			BLK_MQ_MAX_DEPTH);
3104 		set->queue_depth = BLK_MQ_MAX_DEPTH;
3105 	}
3106 
3107 	if (!set->nr_maps)
3108 		set->nr_maps = 1;
3109 	else if (set->nr_maps > HCTX_MAX_TYPES)
3110 		return -EINVAL;
3111 
3112 	/*
3113 	 * If a crashdump is active, then we are potentially in a very
3114 	 * memory constrained environment. Limit us to 1 queue and
3115 	 * 64 tags to prevent using too much memory.
3116 	 */
3117 	if (is_kdump_kernel()) {
3118 		set->nr_hw_queues = 1;
3119 		set->nr_maps = 1;
3120 		set->queue_depth = min(64U, set->queue_depth);
3121 	}
3122 	/*
3123 	 * There is no use for more h/w queues than cpus if we just have
3124 	 * a single map
3125 	 */
3126 	if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3127 		set->nr_hw_queues = nr_cpu_ids;
3128 
3129 	if (blk_mq_realloc_tag_set_tags(set, 0, set->nr_hw_queues) < 0)
3130 		return -ENOMEM;
3131 
3132 	ret = -ENOMEM;
3133 	for (i = 0; i < set->nr_maps; i++) {
3134 		set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3135 						  sizeof(set->map[i].mq_map[0]),
3136 						  GFP_KERNEL, set->numa_node);
3137 		if (!set->map[i].mq_map)
3138 			goto out_free_mq_map;
3139 		set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3140 	}
3141 
3142 	ret = blk_mq_update_queue_map(set);
3143 	if (ret)
3144 		goto out_free_mq_map;
3145 
3146 	ret = blk_mq_alloc_rq_maps(set);
3147 	if (ret)
3148 		goto out_free_mq_map;
3149 
3150 	mutex_init(&set->tag_list_lock);
3151 	INIT_LIST_HEAD(&set->tag_list);
3152 
3153 	return 0;
3154 
3155 out_free_mq_map:
3156 	for (i = 0; i < set->nr_maps; i++) {
3157 		kfree(set->map[i].mq_map);
3158 		set->map[i].mq_map = NULL;
3159 	}
3160 	kfree(set->tags);
3161 	set->tags = NULL;
3162 	return ret;
3163 }
3164 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3165 
3166 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3167 {
3168 	int i, j;
3169 
3170 	for (i = 0; i < set->nr_hw_queues; i++)
3171 		blk_mq_free_map_and_requests(set, i);
3172 
3173 	for (j = 0; j < set->nr_maps; j++) {
3174 		kfree(set->map[j].mq_map);
3175 		set->map[j].mq_map = NULL;
3176 	}
3177 
3178 	kfree(set->tags);
3179 	set->tags = NULL;
3180 }
3181 EXPORT_SYMBOL(blk_mq_free_tag_set);
3182 
3183 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3184 {
3185 	struct blk_mq_tag_set *set = q->tag_set;
3186 	struct blk_mq_hw_ctx *hctx;
3187 	int i, ret;
3188 
3189 	if (!set)
3190 		return -EINVAL;
3191 
3192 	if (q->nr_requests == nr)
3193 		return 0;
3194 
3195 	blk_mq_freeze_queue(q);
3196 	blk_mq_quiesce_queue(q);
3197 
3198 	ret = 0;
3199 	queue_for_each_hw_ctx(q, hctx, i) {
3200 		if (!hctx->tags)
3201 			continue;
3202 		/*
3203 		 * If we're using an MQ scheduler, just update the scheduler
3204 		 * queue depth. This is similar to what the old code would do.
3205 		 */
3206 		if (!hctx->sched_tags) {
3207 			ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3208 							false);
3209 		} else {
3210 			ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3211 							nr, true);
3212 		}
3213 		if (ret)
3214 			break;
3215 		if (q->elevator && q->elevator->type->ops.depth_updated)
3216 			q->elevator->type->ops.depth_updated(hctx);
3217 	}
3218 
3219 	if (!ret)
3220 		q->nr_requests = nr;
3221 
3222 	blk_mq_unquiesce_queue(q);
3223 	blk_mq_unfreeze_queue(q);
3224 
3225 	return ret;
3226 }
3227 
3228 /*
3229  * request_queue and elevator_type pair.
3230  * It is just used by __blk_mq_update_nr_hw_queues to cache
3231  * the elevator_type associated with a request_queue.
3232  */
3233 struct blk_mq_qe_pair {
3234 	struct list_head node;
3235 	struct request_queue *q;
3236 	struct elevator_type *type;
3237 };
3238 
3239 /*
3240  * Cache the elevator_type in qe pair list and switch the
3241  * io scheduler to 'none'
3242  */
3243 static bool blk_mq_elv_switch_none(struct list_head *head,
3244 		struct request_queue *q)
3245 {
3246 	struct blk_mq_qe_pair *qe;
3247 
3248 	if (!q->elevator)
3249 		return true;
3250 
3251 	qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3252 	if (!qe)
3253 		return false;
3254 
3255 	INIT_LIST_HEAD(&qe->node);
3256 	qe->q = q;
3257 	qe->type = q->elevator->type;
3258 	list_add(&qe->node, head);
3259 
3260 	mutex_lock(&q->sysfs_lock);
3261 	/*
3262 	 * After elevator_switch_mq, the previous elevator_queue will be
3263 	 * released by elevator_release. The reference of the io scheduler
3264 	 * module get by elevator_get will also be put. So we need to get
3265 	 * a reference of the io scheduler module here to prevent it to be
3266 	 * removed.
3267 	 */
3268 	__module_get(qe->type->elevator_owner);
3269 	elevator_switch_mq(q, NULL);
3270 	mutex_unlock(&q->sysfs_lock);
3271 
3272 	return true;
3273 }
3274 
3275 static void blk_mq_elv_switch_back(struct list_head *head,
3276 		struct request_queue *q)
3277 {
3278 	struct blk_mq_qe_pair *qe;
3279 	struct elevator_type *t = NULL;
3280 
3281 	list_for_each_entry(qe, head, node)
3282 		if (qe->q == q) {
3283 			t = qe->type;
3284 			break;
3285 		}
3286 
3287 	if (!t)
3288 		return;
3289 
3290 	list_del(&qe->node);
3291 	kfree(qe);
3292 
3293 	mutex_lock(&q->sysfs_lock);
3294 	elevator_switch_mq(q, t);
3295 	mutex_unlock(&q->sysfs_lock);
3296 }
3297 
3298 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3299 							int nr_hw_queues)
3300 {
3301 	struct request_queue *q;
3302 	LIST_HEAD(head);
3303 	int prev_nr_hw_queues;
3304 
3305 	lockdep_assert_held(&set->tag_list_lock);
3306 
3307 	if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3308 		nr_hw_queues = nr_cpu_ids;
3309 	if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
3310 		return;
3311 
3312 	list_for_each_entry(q, &set->tag_list, tag_set_list)
3313 		blk_mq_freeze_queue(q);
3314 	/*
3315 	 * Switch IO scheduler to 'none', cleaning up the data associated
3316 	 * with the previous scheduler. We will switch back once we are done
3317 	 * updating the new sw to hw queue mappings.
3318 	 */
3319 	list_for_each_entry(q, &set->tag_list, tag_set_list)
3320 		if (!blk_mq_elv_switch_none(&head, q))
3321 			goto switch_back;
3322 
3323 	list_for_each_entry(q, &set->tag_list, tag_set_list) {
3324 		blk_mq_debugfs_unregister_hctxs(q);
3325 		blk_mq_sysfs_unregister(q);
3326 	}
3327 
3328 	if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
3329 	    0)
3330 		goto reregister;
3331 
3332 	prev_nr_hw_queues = set->nr_hw_queues;
3333 	set->nr_hw_queues = nr_hw_queues;
3334 	blk_mq_update_queue_map(set);
3335 fallback:
3336 	list_for_each_entry(q, &set->tag_list, tag_set_list) {
3337 		blk_mq_realloc_hw_ctxs(set, q);
3338 		if (q->nr_hw_queues != set->nr_hw_queues) {
3339 			pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3340 					nr_hw_queues, prev_nr_hw_queues);
3341 			set->nr_hw_queues = prev_nr_hw_queues;
3342 			blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3343 			goto fallback;
3344 		}
3345 		blk_mq_map_swqueue(q);
3346 	}
3347 
3348 reregister:
3349 	list_for_each_entry(q, &set->tag_list, tag_set_list) {
3350 		blk_mq_sysfs_register(q);
3351 		blk_mq_debugfs_register_hctxs(q);
3352 	}
3353 
3354 switch_back:
3355 	list_for_each_entry(q, &set->tag_list, tag_set_list)
3356 		blk_mq_elv_switch_back(&head, q);
3357 
3358 	list_for_each_entry(q, &set->tag_list, tag_set_list)
3359 		blk_mq_unfreeze_queue(q);
3360 }
3361 
3362 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3363 {
3364 	mutex_lock(&set->tag_list_lock);
3365 	__blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3366 	mutex_unlock(&set->tag_list_lock);
3367 }
3368 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3369 
3370 /* Enable polling stats and return whether they were already enabled. */
3371 static bool blk_poll_stats_enable(struct request_queue *q)
3372 {
3373 	if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3374 	    blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3375 		return true;
3376 	blk_stat_add_callback(q, q->poll_cb);
3377 	return false;
3378 }
3379 
3380 static void blk_mq_poll_stats_start(struct request_queue *q)
3381 {
3382 	/*
3383 	 * We don't arm the callback if polling stats are not enabled or the
3384 	 * callback is already active.
3385 	 */
3386 	if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3387 	    blk_stat_is_active(q->poll_cb))
3388 		return;
3389 
3390 	blk_stat_activate_msecs(q->poll_cb, 100);
3391 }
3392 
3393 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3394 {
3395 	struct request_queue *q = cb->data;
3396 	int bucket;
3397 
3398 	for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3399 		if (cb->stat[bucket].nr_samples)
3400 			q->poll_stat[bucket] = cb->stat[bucket];
3401 	}
3402 }
3403 
3404 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3405 				       struct request *rq)
3406 {
3407 	unsigned long ret = 0;
3408 	int bucket;
3409 
3410 	/*
3411 	 * If stats collection isn't on, don't sleep but turn it on for
3412 	 * future users
3413 	 */
3414 	if (!blk_poll_stats_enable(q))
3415 		return 0;
3416 
3417 	/*
3418 	 * As an optimistic guess, use half of the mean service time
3419 	 * for this type of request. We can (and should) make this smarter.
3420 	 * For instance, if the completion latencies are tight, we can
3421 	 * get closer than just half the mean. This is especially
3422 	 * important on devices where the completion latencies are longer
3423 	 * than ~10 usec. We do use the stats for the relevant IO size
3424 	 * if available which does lead to better estimates.
3425 	 */
3426 	bucket = blk_mq_poll_stats_bkt(rq);
3427 	if (bucket < 0)
3428 		return ret;
3429 
3430 	if (q->poll_stat[bucket].nr_samples)
3431 		ret = (q->poll_stat[bucket].mean + 1) / 2;
3432 
3433 	return ret;
3434 }
3435 
3436 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3437 				     struct request *rq)
3438 {
3439 	struct hrtimer_sleeper hs;
3440 	enum hrtimer_mode mode;
3441 	unsigned int nsecs;
3442 	ktime_t kt;
3443 
3444 	if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3445 		return false;
3446 
3447 	/*
3448 	 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3449 	 *
3450 	 *  0:	use half of prev avg
3451 	 * >0:	use this specific value
3452 	 */
3453 	if (q->poll_nsec > 0)
3454 		nsecs = q->poll_nsec;
3455 	else
3456 		nsecs = blk_mq_poll_nsecs(q, rq);
3457 
3458 	if (!nsecs)
3459 		return false;
3460 
3461 	rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3462 
3463 	/*
3464 	 * This will be replaced with the stats tracking code, using
3465 	 * 'avg_completion_time / 2' as the pre-sleep target.
3466 	 */
3467 	kt = nsecs;
3468 
3469 	mode = HRTIMER_MODE_REL;
3470 	hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3471 	hrtimer_set_expires(&hs.timer, kt);
3472 
3473 	do {
3474 		if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3475 			break;
3476 		set_current_state(TASK_UNINTERRUPTIBLE);
3477 		hrtimer_sleeper_start_expires(&hs, mode);
3478 		if (hs.task)
3479 			io_schedule();
3480 		hrtimer_cancel(&hs.timer);
3481 		mode = HRTIMER_MODE_ABS;
3482 	} while (hs.task && !signal_pending(current));
3483 
3484 	__set_current_state(TASK_RUNNING);
3485 	destroy_hrtimer_on_stack(&hs.timer);
3486 	return true;
3487 }
3488 
3489 static bool blk_mq_poll_hybrid(struct request_queue *q,
3490 			       struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3491 {
3492 	struct request *rq;
3493 
3494 	if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3495 		return false;
3496 
3497 	if (!blk_qc_t_is_internal(cookie))
3498 		rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3499 	else {
3500 		rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3501 		/*
3502 		 * With scheduling, if the request has completed, we'll
3503 		 * get a NULL return here, as we clear the sched tag when
3504 		 * that happens. The request still remains valid, like always,
3505 		 * so we should be safe with just the NULL check.
3506 		 */
3507 		if (!rq)
3508 			return false;
3509 	}
3510 
3511 	return blk_mq_poll_hybrid_sleep(q, rq);
3512 }
3513 
3514 /**
3515  * blk_poll - poll for IO completions
3516  * @q:  the queue
3517  * @cookie: cookie passed back at IO submission time
3518  * @spin: whether to spin for completions
3519  *
3520  * Description:
3521  *    Poll for completions on the passed in queue. Returns number of
3522  *    completed entries found. If @spin is true, then blk_poll will continue
3523  *    looping until at least one completion is found, unless the task is
3524  *    otherwise marked running (or we need to reschedule).
3525  */
3526 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3527 {
3528 	struct blk_mq_hw_ctx *hctx;
3529 	long state;
3530 
3531 	if (!blk_qc_t_valid(cookie) ||
3532 	    !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3533 		return 0;
3534 
3535 	if (current->plug)
3536 		blk_flush_plug_list(current->plug, false);
3537 
3538 	hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3539 
3540 	/*
3541 	 * If we sleep, have the caller restart the poll loop to reset
3542 	 * the state. Like for the other success return cases, the
3543 	 * caller is responsible for checking if the IO completed. If
3544 	 * the IO isn't complete, we'll get called again and will go
3545 	 * straight to the busy poll loop.
3546 	 */
3547 	if (blk_mq_poll_hybrid(q, hctx, cookie))
3548 		return 1;
3549 
3550 	hctx->poll_considered++;
3551 
3552 	state = current->state;
3553 	do {
3554 		int ret;
3555 
3556 		hctx->poll_invoked++;
3557 
3558 		ret = q->mq_ops->poll(hctx);
3559 		if (ret > 0) {
3560 			hctx->poll_success++;
3561 			__set_current_state(TASK_RUNNING);
3562 			return ret;
3563 		}
3564 
3565 		if (signal_pending_state(state, current))
3566 			__set_current_state(TASK_RUNNING);
3567 
3568 		if (current->state == TASK_RUNNING)
3569 			return 1;
3570 		if (ret < 0 || !spin)
3571 			break;
3572 		cpu_relax();
3573 	} while (!need_resched());
3574 
3575 	__set_current_state(TASK_RUNNING);
3576 	return 0;
3577 }
3578 EXPORT_SYMBOL_GPL(blk_poll);
3579 
3580 unsigned int blk_mq_rq_cpu(struct request *rq)
3581 {
3582 	return rq->mq_ctx->cpu;
3583 }
3584 EXPORT_SYMBOL(blk_mq_rq_cpu);
3585 
3586 static int __init blk_mq_init(void)
3587 {
3588 	cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3589 				blk_mq_hctx_notify_dead);
3590 	return 0;
3591 }
3592 subsys_initcall(blk_mq_init);
3593