xref: /openbmc/linux/block/blk-mq.c (revision fa84cf09)
1 /*
2  * Block multiqueue core code
3  *
4  * Copyright (C) 2013-2014 Jens Axboe
5  * Copyright (C) 2013-2014 Christoph Hellwig
6  */
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/backing-dev.h>
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/kmemleak.h>
13 #include <linux/mm.h>
14 #include <linux/init.h>
15 #include <linux/slab.h>
16 #include <linux/workqueue.h>
17 #include <linux/smp.h>
18 #include <linux/llist.h>
19 #include <linux/list_sort.h>
20 #include <linux/cpu.h>
21 #include <linux/cache.h>
22 #include <linux/sched/sysctl.h>
23 #include <linux/sched/topology.h>
24 #include <linux/sched/signal.h>
25 #include <linux/delay.h>
26 #include <linux/crash_dump.h>
27 #include <linux/prefetch.h>
28 
29 #include <trace/events/block.h>
30 
31 #include <linux/blk-mq.h>
32 #include "blk.h"
33 #include "blk-mq.h"
34 #include "blk-mq-debugfs.h"
35 #include "blk-mq-tag.h"
36 #include "blk-stat.h"
37 #include "blk-wbt.h"
38 #include "blk-mq-sched.h"
39 
40 static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie);
41 static void blk_mq_poll_stats_start(struct request_queue *q);
42 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
43 
44 static int blk_mq_poll_stats_bkt(const struct request *rq)
45 {
46 	int ddir, bytes, bucket;
47 
48 	ddir = rq_data_dir(rq);
49 	bytes = blk_rq_bytes(rq);
50 
51 	bucket = ddir + 2*(ilog2(bytes) - 9);
52 
53 	if (bucket < 0)
54 		return -1;
55 	else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
56 		return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
57 
58 	return bucket;
59 }
60 
61 /*
62  * Check if any of the ctx's have pending work in this hardware queue
63  */
64 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
65 {
66 	return !list_empty_careful(&hctx->dispatch) ||
67 		sbitmap_any_bit_set(&hctx->ctx_map) ||
68 			blk_mq_sched_has_work(hctx);
69 }
70 
71 /*
72  * Mark this ctx as having pending work in this hardware queue
73  */
74 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
75 				     struct blk_mq_ctx *ctx)
76 {
77 	if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
78 		sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
79 }
80 
81 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
82 				      struct blk_mq_ctx *ctx)
83 {
84 	sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
85 }
86 
87 struct mq_inflight {
88 	struct hd_struct *part;
89 	unsigned int *inflight;
90 };
91 
92 static void blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
93 				  struct request *rq, void *priv,
94 				  bool reserved)
95 {
96 	struct mq_inflight *mi = priv;
97 
98 	/*
99 	 * index[0] counts the specific partition that was asked for. index[1]
100 	 * counts the ones that are active on the whole device, so increment
101 	 * that if mi->part is indeed a partition, and not a whole device.
102 	 */
103 	if (rq->part == mi->part)
104 		mi->inflight[0]++;
105 	if (mi->part->partno)
106 		mi->inflight[1]++;
107 }
108 
109 void blk_mq_in_flight(struct request_queue *q, struct hd_struct *part,
110 		      unsigned int inflight[2])
111 {
112 	struct mq_inflight mi = { .part = part, .inflight = inflight, };
113 
114 	inflight[0] = inflight[1] = 0;
115 	blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
116 }
117 
118 static void blk_mq_check_inflight_rw(struct blk_mq_hw_ctx *hctx,
119 				     struct request *rq, void *priv,
120 				     bool reserved)
121 {
122 	struct mq_inflight *mi = priv;
123 
124 	if (rq->part == mi->part)
125 		mi->inflight[rq_data_dir(rq)]++;
126 }
127 
128 void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
129 			 unsigned int inflight[2])
130 {
131 	struct mq_inflight mi = { .part = part, .inflight = inflight, };
132 
133 	inflight[0] = inflight[1] = 0;
134 	blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight_rw, &mi);
135 }
136 
137 void blk_freeze_queue_start(struct request_queue *q)
138 {
139 	int freeze_depth;
140 
141 	freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
142 	if (freeze_depth == 1) {
143 		percpu_ref_kill(&q->q_usage_counter);
144 		if (q->mq_ops)
145 			blk_mq_run_hw_queues(q, false);
146 	}
147 }
148 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
149 
150 void blk_mq_freeze_queue_wait(struct request_queue *q)
151 {
152 	wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
153 }
154 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
155 
156 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
157 				     unsigned long timeout)
158 {
159 	return wait_event_timeout(q->mq_freeze_wq,
160 					percpu_ref_is_zero(&q->q_usage_counter),
161 					timeout);
162 }
163 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
164 
165 /*
166  * Guarantee no request is in use, so we can change any data structure of
167  * the queue afterward.
168  */
169 void blk_freeze_queue(struct request_queue *q)
170 {
171 	/*
172 	 * In the !blk_mq case we are only calling this to kill the
173 	 * q_usage_counter, otherwise this increases the freeze depth
174 	 * and waits for it to return to zero.  For this reason there is
175 	 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
176 	 * exported to drivers as the only user for unfreeze is blk_mq.
177 	 */
178 	blk_freeze_queue_start(q);
179 	if (!q->mq_ops)
180 		blk_drain_queue(q);
181 	blk_mq_freeze_queue_wait(q);
182 }
183 
184 void blk_mq_freeze_queue(struct request_queue *q)
185 {
186 	/*
187 	 * ...just an alias to keep freeze and unfreeze actions balanced
188 	 * in the blk_mq_* namespace
189 	 */
190 	blk_freeze_queue(q);
191 }
192 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
193 
194 void blk_mq_unfreeze_queue(struct request_queue *q)
195 {
196 	int freeze_depth;
197 
198 	freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
199 	WARN_ON_ONCE(freeze_depth < 0);
200 	if (!freeze_depth) {
201 		percpu_ref_reinit(&q->q_usage_counter);
202 		wake_up_all(&q->mq_freeze_wq);
203 	}
204 }
205 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
206 
207 /*
208  * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
209  * mpt3sas driver such that this function can be removed.
210  */
211 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
212 {
213 	blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
214 }
215 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
216 
217 /**
218  * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
219  * @q: request queue.
220  *
221  * Note: this function does not prevent that the struct request end_io()
222  * callback function is invoked. Once this function is returned, we make
223  * sure no dispatch can happen until the queue is unquiesced via
224  * blk_mq_unquiesce_queue().
225  */
226 void blk_mq_quiesce_queue(struct request_queue *q)
227 {
228 	struct blk_mq_hw_ctx *hctx;
229 	unsigned int i;
230 	bool rcu = false;
231 
232 	blk_mq_quiesce_queue_nowait(q);
233 
234 	queue_for_each_hw_ctx(q, hctx, i) {
235 		if (hctx->flags & BLK_MQ_F_BLOCKING)
236 			synchronize_srcu(hctx->srcu);
237 		else
238 			rcu = true;
239 	}
240 	if (rcu)
241 		synchronize_rcu();
242 }
243 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
244 
245 /*
246  * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
247  * @q: request queue.
248  *
249  * This function recovers queue into the state before quiescing
250  * which is done by blk_mq_quiesce_queue.
251  */
252 void blk_mq_unquiesce_queue(struct request_queue *q)
253 {
254 	blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
255 
256 	/* dispatch requests which are inserted during quiescing */
257 	blk_mq_run_hw_queues(q, true);
258 }
259 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
260 
261 void blk_mq_wake_waiters(struct request_queue *q)
262 {
263 	struct blk_mq_hw_ctx *hctx;
264 	unsigned int i;
265 
266 	queue_for_each_hw_ctx(q, hctx, i)
267 		if (blk_mq_hw_queue_mapped(hctx))
268 			blk_mq_tag_wakeup_all(hctx->tags, true);
269 }
270 
271 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
272 {
273 	return blk_mq_has_free_tags(hctx->tags);
274 }
275 EXPORT_SYMBOL(blk_mq_can_queue);
276 
277 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
278 		unsigned int tag, unsigned int op)
279 {
280 	struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
281 	struct request *rq = tags->static_rqs[tag];
282 	req_flags_t rq_flags = 0;
283 
284 	if (data->flags & BLK_MQ_REQ_INTERNAL) {
285 		rq->tag = -1;
286 		rq->internal_tag = tag;
287 	} else {
288 		if (blk_mq_tag_busy(data->hctx)) {
289 			rq_flags = RQF_MQ_INFLIGHT;
290 			atomic_inc(&data->hctx->nr_active);
291 		}
292 		rq->tag = tag;
293 		rq->internal_tag = -1;
294 		data->hctx->tags->rqs[rq->tag] = rq;
295 	}
296 
297 	/* csd/requeue_work/fifo_time is initialized before use */
298 	rq->q = data->q;
299 	rq->mq_ctx = data->ctx;
300 	rq->rq_flags = rq_flags;
301 	rq->cpu = -1;
302 	rq->cmd_flags = op;
303 	if (data->flags & BLK_MQ_REQ_PREEMPT)
304 		rq->rq_flags |= RQF_PREEMPT;
305 	if (blk_queue_io_stat(data->q))
306 		rq->rq_flags |= RQF_IO_STAT;
307 	INIT_LIST_HEAD(&rq->queuelist);
308 	INIT_HLIST_NODE(&rq->hash);
309 	RB_CLEAR_NODE(&rq->rb_node);
310 	rq->rq_disk = NULL;
311 	rq->part = NULL;
312 	rq->start_time_ns = ktime_get_ns();
313 	rq->io_start_time_ns = 0;
314 	rq->nr_phys_segments = 0;
315 #if defined(CONFIG_BLK_DEV_INTEGRITY)
316 	rq->nr_integrity_segments = 0;
317 #endif
318 	rq->special = NULL;
319 	/* tag was already set */
320 	rq->extra_len = 0;
321 	rq->__deadline = 0;
322 
323 	INIT_LIST_HEAD(&rq->timeout_list);
324 	rq->timeout = 0;
325 
326 	rq->end_io = NULL;
327 	rq->end_io_data = NULL;
328 	rq->next_rq = NULL;
329 
330 #ifdef CONFIG_BLK_CGROUP
331 	rq->rl = NULL;
332 #endif
333 
334 	data->ctx->rq_dispatched[op_is_sync(op)]++;
335 	refcount_set(&rq->ref, 1);
336 	return rq;
337 }
338 
339 static struct request *blk_mq_get_request(struct request_queue *q,
340 		struct bio *bio, unsigned int op,
341 		struct blk_mq_alloc_data *data)
342 {
343 	struct elevator_queue *e = q->elevator;
344 	struct request *rq;
345 	unsigned int tag;
346 	bool put_ctx_on_error = false;
347 
348 	blk_queue_enter_live(q);
349 	data->q = q;
350 	if (likely(!data->ctx)) {
351 		data->ctx = blk_mq_get_ctx(q);
352 		put_ctx_on_error = true;
353 	}
354 	if (likely(!data->hctx))
355 		data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
356 	if (op & REQ_NOWAIT)
357 		data->flags |= BLK_MQ_REQ_NOWAIT;
358 
359 	if (e) {
360 		data->flags |= BLK_MQ_REQ_INTERNAL;
361 
362 		/*
363 		 * Flush requests are special and go directly to the
364 		 * dispatch list. Don't include reserved tags in the
365 		 * limiting, as it isn't useful.
366 		 */
367 		if (!op_is_flush(op) && e->type->ops.mq.limit_depth &&
368 		    !(data->flags & BLK_MQ_REQ_RESERVED))
369 			e->type->ops.mq.limit_depth(op, data);
370 	}
371 
372 	tag = blk_mq_get_tag(data);
373 	if (tag == BLK_MQ_TAG_FAIL) {
374 		if (put_ctx_on_error) {
375 			blk_mq_put_ctx(data->ctx);
376 			data->ctx = NULL;
377 		}
378 		blk_queue_exit(q);
379 		return NULL;
380 	}
381 
382 	rq = blk_mq_rq_ctx_init(data, tag, op);
383 	if (!op_is_flush(op)) {
384 		rq->elv.icq = NULL;
385 		if (e && e->type->ops.mq.prepare_request) {
386 			if (e->type->icq_cache && rq_ioc(bio))
387 				blk_mq_sched_assign_ioc(rq, bio);
388 
389 			e->type->ops.mq.prepare_request(rq, bio);
390 			rq->rq_flags |= RQF_ELVPRIV;
391 		}
392 	}
393 	data->hctx->queued++;
394 	return rq;
395 }
396 
397 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
398 		blk_mq_req_flags_t flags)
399 {
400 	struct blk_mq_alloc_data alloc_data = { .flags = flags };
401 	struct request *rq;
402 	int ret;
403 
404 	ret = blk_queue_enter(q, flags);
405 	if (ret)
406 		return ERR_PTR(ret);
407 
408 	rq = blk_mq_get_request(q, NULL, op, &alloc_data);
409 	blk_queue_exit(q);
410 
411 	if (!rq)
412 		return ERR_PTR(-EWOULDBLOCK);
413 
414 	blk_mq_put_ctx(alloc_data.ctx);
415 
416 	rq->__data_len = 0;
417 	rq->__sector = (sector_t) -1;
418 	rq->bio = rq->biotail = NULL;
419 	return rq;
420 }
421 EXPORT_SYMBOL(blk_mq_alloc_request);
422 
423 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
424 	unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
425 {
426 	struct blk_mq_alloc_data alloc_data = { .flags = flags };
427 	struct request *rq;
428 	unsigned int cpu;
429 	int ret;
430 
431 	/*
432 	 * If the tag allocator sleeps we could get an allocation for a
433 	 * different hardware context.  No need to complicate the low level
434 	 * allocator for this for the rare use case of a command tied to
435 	 * a specific queue.
436 	 */
437 	if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
438 		return ERR_PTR(-EINVAL);
439 
440 	if (hctx_idx >= q->nr_hw_queues)
441 		return ERR_PTR(-EIO);
442 
443 	ret = blk_queue_enter(q, flags);
444 	if (ret)
445 		return ERR_PTR(ret);
446 
447 	/*
448 	 * Check if the hardware context is actually mapped to anything.
449 	 * If not tell the caller that it should skip this queue.
450 	 */
451 	alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
452 	if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
453 		blk_queue_exit(q);
454 		return ERR_PTR(-EXDEV);
455 	}
456 	cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
457 	alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
458 
459 	rq = blk_mq_get_request(q, NULL, op, &alloc_data);
460 	blk_queue_exit(q);
461 
462 	if (!rq)
463 		return ERR_PTR(-EWOULDBLOCK);
464 
465 	return rq;
466 }
467 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
468 
469 static void __blk_mq_free_request(struct request *rq)
470 {
471 	struct request_queue *q = rq->q;
472 	struct blk_mq_ctx *ctx = rq->mq_ctx;
473 	struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
474 	const int sched_tag = rq->internal_tag;
475 
476 	if (rq->tag != -1)
477 		blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
478 	if (sched_tag != -1)
479 		blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
480 	blk_mq_sched_restart(hctx);
481 	blk_queue_exit(q);
482 }
483 
484 void blk_mq_free_request(struct request *rq)
485 {
486 	struct request_queue *q = rq->q;
487 	struct elevator_queue *e = q->elevator;
488 	struct blk_mq_ctx *ctx = rq->mq_ctx;
489 	struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
490 
491 	if (rq->rq_flags & RQF_ELVPRIV) {
492 		if (e && e->type->ops.mq.finish_request)
493 			e->type->ops.mq.finish_request(rq);
494 		if (rq->elv.icq) {
495 			put_io_context(rq->elv.icq->ioc);
496 			rq->elv.icq = NULL;
497 		}
498 	}
499 
500 	ctx->rq_completed[rq_is_sync(rq)]++;
501 	if (rq->rq_flags & RQF_MQ_INFLIGHT)
502 		atomic_dec(&hctx->nr_active);
503 
504 	if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
505 		laptop_io_completion(q->backing_dev_info);
506 
507 	wbt_done(q->rq_wb, rq);
508 
509 	if (blk_rq_rl(rq))
510 		blk_put_rl(blk_rq_rl(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 = ktime_get_ns();
521 
522 	if (rq->rq_flags & RQF_STATS) {
523 		blk_mq_poll_stats_start(rq->q);
524 		blk_stat_add(rq, now);
525 	}
526 
527 	blk_account_io_done(rq, now);
528 
529 	if (rq->end_io) {
530 		wbt_done(rq->q->rq_wb, rq);
531 		rq->end_io(rq, error);
532 	} else {
533 		if (unlikely(blk_bidi_rq(rq)))
534 			blk_mq_free_request(rq->next_rq);
535 		blk_mq_free_request(rq);
536 	}
537 }
538 EXPORT_SYMBOL(__blk_mq_end_request);
539 
540 void blk_mq_end_request(struct request *rq, blk_status_t error)
541 {
542 	if (blk_update_request(rq, error, blk_rq_bytes(rq)))
543 		BUG();
544 	__blk_mq_end_request(rq, error);
545 }
546 EXPORT_SYMBOL(blk_mq_end_request);
547 
548 static void __blk_mq_complete_request_remote(void *data)
549 {
550 	struct request *rq = data;
551 
552 	rq->q->softirq_done_fn(rq);
553 }
554 
555 static void __blk_mq_complete_request(struct request *rq)
556 {
557 	struct blk_mq_ctx *ctx = rq->mq_ctx;
558 	bool shared = false;
559 	int cpu;
560 
561 	if (cmpxchg(&rq->state, MQ_RQ_IN_FLIGHT, MQ_RQ_COMPLETE) !=
562 			MQ_RQ_IN_FLIGHT)
563 		return;
564 
565 	if (rq->internal_tag != -1)
566 		blk_mq_sched_completed_request(rq);
567 
568 	if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
569 		rq->q->softirq_done_fn(rq);
570 		return;
571 	}
572 
573 	cpu = get_cpu();
574 	if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
575 		shared = cpus_share_cache(cpu, ctx->cpu);
576 
577 	if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
578 		rq->csd.func = __blk_mq_complete_request_remote;
579 		rq->csd.info = rq;
580 		rq->csd.flags = 0;
581 		smp_call_function_single_async(ctx->cpu, &rq->csd);
582 	} else {
583 		rq->q->softirq_done_fn(rq);
584 	}
585 	put_cpu();
586 }
587 
588 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
589 	__releases(hctx->srcu)
590 {
591 	if (!(hctx->flags & BLK_MQ_F_BLOCKING))
592 		rcu_read_unlock();
593 	else
594 		srcu_read_unlock(hctx->srcu, srcu_idx);
595 }
596 
597 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
598 	__acquires(hctx->srcu)
599 {
600 	if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
601 		/* shut up gcc false positive */
602 		*srcu_idx = 0;
603 		rcu_read_lock();
604 	} else
605 		*srcu_idx = srcu_read_lock(hctx->srcu);
606 }
607 
608 /**
609  * blk_mq_complete_request - end I/O on a request
610  * @rq:		the request being processed
611  *
612  * Description:
613  *	Ends all I/O on a request. It does not handle partial completions.
614  *	The actual completion happens out-of-order, through a IPI handler.
615  **/
616 void blk_mq_complete_request(struct request *rq)
617 {
618 	if (unlikely(blk_should_fake_timeout(rq->q)))
619 		return;
620 	__blk_mq_complete_request(rq);
621 }
622 EXPORT_SYMBOL(blk_mq_complete_request);
623 
624 int blk_mq_request_started(struct request *rq)
625 {
626 	return blk_mq_rq_state(rq) != MQ_RQ_IDLE;
627 }
628 EXPORT_SYMBOL_GPL(blk_mq_request_started);
629 
630 void blk_mq_start_request(struct request *rq)
631 {
632 	struct request_queue *q = rq->q;
633 
634 	blk_mq_sched_started_request(rq);
635 
636 	trace_block_rq_issue(q, rq);
637 
638 	if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
639 		rq->io_start_time_ns = ktime_get_ns();
640 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
641 		rq->throtl_size = blk_rq_sectors(rq);
642 #endif
643 		rq->rq_flags |= RQF_STATS;
644 		wbt_issue(q->rq_wb, rq);
645 	}
646 
647 	WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
648 
649 	blk_add_timer(rq);
650 	WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
651 
652 	if (q->dma_drain_size && blk_rq_bytes(rq)) {
653 		/*
654 		 * Make sure space for the drain appears.  We know we can do
655 		 * this because max_hw_segments has been adjusted to be one
656 		 * fewer than the device can handle.
657 		 */
658 		rq->nr_phys_segments++;
659 	}
660 }
661 EXPORT_SYMBOL(blk_mq_start_request);
662 
663 static void __blk_mq_requeue_request(struct request *rq)
664 {
665 	struct request_queue *q = rq->q;
666 
667 	blk_mq_put_driver_tag(rq);
668 
669 	trace_block_rq_requeue(q, rq);
670 	wbt_requeue(q->rq_wb, rq);
671 
672 	if (blk_mq_request_started(rq)) {
673 		WRITE_ONCE(rq->state, MQ_RQ_IDLE);
674 		rq->rq_flags &= ~RQF_TIMED_OUT;
675 		if (q->dma_drain_size && blk_rq_bytes(rq))
676 			rq->nr_phys_segments--;
677 	}
678 }
679 
680 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
681 {
682 	__blk_mq_requeue_request(rq);
683 
684 	/* this request will be re-inserted to io scheduler queue */
685 	blk_mq_sched_requeue_request(rq);
686 
687 	BUG_ON(blk_queued_rq(rq));
688 	blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
689 }
690 EXPORT_SYMBOL(blk_mq_requeue_request);
691 
692 static void blk_mq_requeue_work(struct work_struct *work)
693 {
694 	struct request_queue *q =
695 		container_of(work, struct request_queue, requeue_work.work);
696 	LIST_HEAD(rq_list);
697 	struct request *rq, *next;
698 
699 	spin_lock_irq(&q->requeue_lock);
700 	list_splice_init(&q->requeue_list, &rq_list);
701 	spin_unlock_irq(&q->requeue_lock);
702 
703 	list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
704 		if (!(rq->rq_flags & RQF_SOFTBARRIER))
705 			continue;
706 
707 		rq->rq_flags &= ~RQF_SOFTBARRIER;
708 		list_del_init(&rq->queuelist);
709 		blk_mq_sched_insert_request(rq, true, false, false);
710 	}
711 
712 	while (!list_empty(&rq_list)) {
713 		rq = list_entry(rq_list.next, struct request, queuelist);
714 		list_del_init(&rq->queuelist);
715 		blk_mq_sched_insert_request(rq, false, false, false);
716 	}
717 
718 	blk_mq_run_hw_queues(q, false);
719 }
720 
721 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
722 				bool kick_requeue_list)
723 {
724 	struct request_queue *q = rq->q;
725 	unsigned long flags;
726 
727 	/*
728 	 * We abuse this flag that is otherwise used by the I/O scheduler to
729 	 * request head insertion from the workqueue.
730 	 */
731 	BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
732 
733 	spin_lock_irqsave(&q->requeue_lock, flags);
734 	if (at_head) {
735 		rq->rq_flags |= RQF_SOFTBARRIER;
736 		list_add(&rq->queuelist, &q->requeue_list);
737 	} else {
738 		list_add_tail(&rq->queuelist, &q->requeue_list);
739 	}
740 	spin_unlock_irqrestore(&q->requeue_lock, flags);
741 
742 	if (kick_requeue_list)
743 		blk_mq_kick_requeue_list(q);
744 }
745 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
746 
747 void blk_mq_kick_requeue_list(struct request_queue *q)
748 {
749 	kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
750 }
751 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
752 
753 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
754 				    unsigned long msecs)
755 {
756 	kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
757 				    msecs_to_jiffies(msecs));
758 }
759 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
760 
761 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
762 {
763 	if (tag < tags->nr_tags) {
764 		prefetch(tags->rqs[tag]);
765 		return tags->rqs[tag];
766 	}
767 
768 	return NULL;
769 }
770 EXPORT_SYMBOL(blk_mq_tag_to_rq);
771 
772 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
773 {
774 	req->rq_flags |= RQF_TIMED_OUT;
775 	if (req->q->mq_ops->timeout) {
776 		enum blk_eh_timer_return ret;
777 
778 		ret = req->q->mq_ops->timeout(req, reserved);
779 		if (ret == BLK_EH_DONE)
780 			return;
781 		WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
782 	}
783 
784 	req->rq_flags &= ~RQF_TIMED_OUT;
785 	blk_add_timer(req);
786 }
787 
788 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
789 {
790 	unsigned long deadline;
791 
792 	if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
793 		return false;
794 	if (rq->rq_flags & RQF_TIMED_OUT)
795 		return false;
796 
797 	deadline = blk_rq_deadline(rq);
798 	if (time_after_eq(jiffies, deadline))
799 		return true;
800 
801 	if (*next == 0)
802 		*next = deadline;
803 	else if (time_after(*next, deadline))
804 		*next = deadline;
805 	return false;
806 }
807 
808 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
809 		struct request *rq, void *priv, bool reserved)
810 {
811 	unsigned long *next = priv;
812 
813 	/*
814 	 * Just do a quick check if it is expired before locking the request in
815 	 * so we're not unnecessarilly synchronizing across CPUs.
816 	 */
817 	if (!blk_mq_req_expired(rq, next))
818 		return;
819 
820 	/*
821 	 * We have reason to believe the request may be expired. Take a
822 	 * reference on the request to lock this request lifetime into its
823 	 * currently allocated context to prevent it from being reallocated in
824 	 * the event the completion by-passes this timeout handler.
825 	 *
826 	 * If the reference was already released, then the driver beat the
827 	 * timeout handler to posting a natural completion.
828 	 */
829 	if (!refcount_inc_not_zero(&rq->ref))
830 		return;
831 
832 	/*
833 	 * The request is now locked and cannot be reallocated underneath the
834 	 * timeout handler's processing. Re-verify this exact request is truly
835 	 * expired; if it is not expired, then the request was completed and
836 	 * reallocated as a new request.
837 	 */
838 	if (blk_mq_req_expired(rq, next))
839 		blk_mq_rq_timed_out(rq, reserved);
840 	if (refcount_dec_and_test(&rq->ref))
841 		__blk_mq_free_request(rq);
842 }
843 
844 static void blk_mq_timeout_work(struct work_struct *work)
845 {
846 	struct request_queue *q =
847 		container_of(work, struct request_queue, timeout_work);
848 	unsigned long next = 0;
849 	struct blk_mq_hw_ctx *hctx;
850 	int i;
851 
852 	/* A deadlock might occur if a request is stuck requiring a
853 	 * timeout at the same time a queue freeze is waiting
854 	 * completion, since the timeout code would not be able to
855 	 * acquire the queue reference here.
856 	 *
857 	 * That's why we don't use blk_queue_enter here; instead, we use
858 	 * percpu_ref_tryget directly, because we need to be able to
859 	 * obtain a reference even in the short window between the queue
860 	 * starting to freeze, by dropping the first reference in
861 	 * blk_freeze_queue_start, and the moment the last request is
862 	 * consumed, marked by the instant q_usage_counter reaches
863 	 * zero.
864 	 */
865 	if (!percpu_ref_tryget(&q->q_usage_counter))
866 		return;
867 
868 	blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
869 
870 	if (next != 0) {
871 		mod_timer(&q->timeout, next);
872 	} else {
873 		/*
874 		 * Request timeouts are handled as a forward rolling timer. If
875 		 * we end up here it means that no requests are pending and
876 		 * also that no request has been pending for a while. Mark
877 		 * each hctx as idle.
878 		 */
879 		queue_for_each_hw_ctx(q, hctx, i) {
880 			/* the hctx may be unmapped, so check it here */
881 			if (blk_mq_hw_queue_mapped(hctx))
882 				blk_mq_tag_idle(hctx);
883 		}
884 	}
885 	blk_queue_exit(q);
886 }
887 
888 struct flush_busy_ctx_data {
889 	struct blk_mq_hw_ctx *hctx;
890 	struct list_head *list;
891 };
892 
893 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
894 {
895 	struct flush_busy_ctx_data *flush_data = data;
896 	struct blk_mq_hw_ctx *hctx = flush_data->hctx;
897 	struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
898 
899 	spin_lock(&ctx->lock);
900 	list_splice_tail_init(&ctx->rq_list, flush_data->list);
901 	sbitmap_clear_bit(sb, bitnr);
902 	spin_unlock(&ctx->lock);
903 	return true;
904 }
905 
906 /*
907  * Process software queues that have been marked busy, splicing them
908  * to the for-dispatch
909  */
910 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
911 {
912 	struct flush_busy_ctx_data data = {
913 		.hctx = hctx,
914 		.list = list,
915 	};
916 
917 	sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
918 }
919 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
920 
921 struct dispatch_rq_data {
922 	struct blk_mq_hw_ctx *hctx;
923 	struct request *rq;
924 };
925 
926 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
927 		void *data)
928 {
929 	struct dispatch_rq_data *dispatch_data = data;
930 	struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
931 	struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
932 
933 	spin_lock(&ctx->lock);
934 	if (!list_empty(&ctx->rq_list)) {
935 		dispatch_data->rq = list_entry_rq(ctx->rq_list.next);
936 		list_del_init(&dispatch_data->rq->queuelist);
937 		if (list_empty(&ctx->rq_list))
938 			sbitmap_clear_bit(sb, bitnr);
939 	}
940 	spin_unlock(&ctx->lock);
941 
942 	return !dispatch_data->rq;
943 }
944 
945 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
946 					struct blk_mq_ctx *start)
947 {
948 	unsigned off = start ? start->index_hw : 0;
949 	struct dispatch_rq_data data = {
950 		.hctx = hctx,
951 		.rq   = NULL,
952 	};
953 
954 	__sbitmap_for_each_set(&hctx->ctx_map, off,
955 			       dispatch_rq_from_ctx, &data);
956 
957 	return data.rq;
958 }
959 
960 static inline unsigned int queued_to_index(unsigned int queued)
961 {
962 	if (!queued)
963 		return 0;
964 
965 	return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
966 }
967 
968 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
969 			   bool wait)
970 {
971 	struct blk_mq_alloc_data data = {
972 		.q = rq->q,
973 		.hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
974 		.flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
975 	};
976 
977 	might_sleep_if(wait);
978 
979 	if (rq->tag != -1)
980 		goto done;
981 
982 	if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
983 		data.flags |= BLK_MQ_REQ_RESERVED;
984 
985 	rq->tag = blk_mq_get_tag(&data);
986 	if (rq->tag >= 0) {
987 		if (blk_mq_tag_busy(data.hctx)) {
988 			rq->rq_flags |= RQF_MQ_INFLIGHT;
989 			atomic_inc(&data.hctx->nr_active);
990 		}
991 		data.hctx->tags->rqs[rq->tag] = rq;
992 	}
993 
994 done:
995 	if (hctx)
996 		*hctx = data.hctx;
997 	return rq->tag != -1;
998 }
999 
1000 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1001 				int flags, void *key)
1002 {
1003 	struct blk_mq_hw_ctx *hctx;
1004 
1005 	hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1006 
1007 	list_del_init(&wait->entry);
1008 	blk_mq_run_hw_queue(hctx, true);
1009 	return 1;
1010 }
1011 
1012 /*
1013  * Mark us waiting for a tag. For shared tags, this involves hooking us into
1014  * the tag wakeups. For non-shared tags, we can simply mark us needing a
1015  * restart. For both cases, take care to check the condition again after
1016  * marking us as waiting.
1017  */
1018 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx **hctx,
1019 				 struct request *rq)
1020 {
1021 	struct blk_mq_hw_ctx *this_hctx = *hctx;
1022 	struct sbq_wait_state *ws;
1023 	wait_queue_entry_t *wait;
1024 	bool ret;
1025 
1026 	if (!(this_hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1027 		if (!test_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state))
1028 			set_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state);
1029 
1030 		/*
1031 		 * It's possible that a tag was freed in the window between the
1032 		 * allocation failure and adding the hardware queue to the wait
1033 		 * queue.
1034 		 *
1035 		 * Don't clear RESTART here, someone else could have set it.
1036 		 * At most this will cost an extra queue run.
1037 		 */
1038 		return blk_mq_get_driver_tag(rq, hctx, false);
1039 	}
1040 
1041 	wait = &this_hctx->dispatch_wait;
1042 	if (!list_empty_careful(&wait->entry))
1043 		return false;
1044 
1045 	spin_lock(&this_hctx->lock);
1046 	if (!list_empty(&wait->entry)) {
1047 		spin_unlock(&this_hctx->lock);
1048 		return false;
1049 	}
1050 
1051 	ws = bt_wait_ptr(&this_hctx->tags->bitmap_tags, this_hctx);
1052 	add_wait_queue(&ws->wait, wait);
1053 
1054 	/*
1055 	 * It's possible that a tag was freed in the window between the
1056 	 * allocation failure and adding the hardware queue to the wait
1057 	 * queue.
1058 	 */
1059 	ret = blk_mq_get_driver_tag(rq, hctx, false);
1060 	if (!ret) {
1061 		spin_unlock(&this_hctx->lock);
1062 		return false;
1063 	}
1064 
1065 	/*
1066 	 * We got a tag, remove ourselves from the wait queue to ensure
1067 	 * someone else gets the wakeup.
1068 	 */
1069 	spin_lock_irq(&ws->wait.lock);
1070 	list_del_init(&wait->entry);
1071 	spin_unlock_irq(&ws->wait.lock);
1072 	spin_unlock(&this_hctx->lock);
1073 
1074 	return true;
1075 }
1076 
1077 #define BLK_MQ_RESOURCE_DELAY	3		/* ms units */
1078 
1079 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1080 			     bool got_budget)
1081 {
1082 	struct blk_mq_hw_ctx *hctx;
1083 	struct request *rq, *nxt;
1084 	bool no_tag = false;
1085 	int errors, queued;
1086 	blk_status_t ret = BLK_STS_OK;
1087 
1088 	if (list_empty(list))
1089 		return false;
1090 
1091 	WARN_ON(!list_is_singular(list) && got_budget);
1092 
1093 	/*
1094 	 * Now process all the entries, sending them to the driver.
1095 	 */
1096 	errors = queued = 0;
1097 	do {
1098 		struct blk_mq_queue_data bd;
1099 
1100 		rq = list_first_entry(list, struct request, queuelist);
1101 
1102 		hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
1103 		if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1104 			break;
1105 
1106 		if (!blk_mq_get_driver_tag(rq, NULL, false)) {
1107 			/*
1108 			 * The initial allocation attempt failed, so we need to
1109 			 * rerun the hardware queue when a tag is freed. The
1110 			 * waitqueue takes care of that. If the queue is run
1111 			 * before we add this entry back on the dispatch list,
1112 			 * we'll re-run it below.
1113 			 */
1114 			if (!blk_mq_mark_tag_wait(&hctx, rq)) {
1115 				blk_mq_put_dispatch_budget(hctx);
1116 				/*
1117 				 * For non-shared tags, the RESTART check
1118 				 * will suffice.
1119 				 */
1120 				if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1121 					no_tag = true;
1122 				break;
1123 			}
1124 		}
1125 
1126 		list_del_init(&rq->queuelist);
1127 
1128 		bd.rq = rq;
1129 
1130 		/*
1131 		 * Flag last if we have no more requests, or if we have more
1132 		 * but can't assign a driver tag to it.
1133 		 */
1134 		if (list_empty(list))
1135 			bd.last = true;
1136 		else {
1137 			nxt = list_first_entry(list, struct request, queuelist);
1138 			bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1139 		}
1140 
1141 		ret = q->mq_ops->queue_rq(hctx, &bd);
1142 		if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1143 			/*
1144 			 * If an I/O scheduler has been configured and we got a
1145 			 * driver tag for the next request already, free it
1146 			 * again.
1147 			 */
1148 			if (!list_empty(list)) {
1149 				nxt = list_first_entry(list, struct request, queuelist);
1150 				blk_mq_put_driver_tag(nxt);
1151 			}
1152 			list_add(&rq->queuelist, list);
1153 			__blk_mq_requeue_request(rq);
1154 			break;
1155 		}
1156 
1157 		if (unlikely(ret != BLK_STS_OK)) {
1158 			errors++;
1159 			blk_mq_end_request(rq, BLK_STS_IOERR);
1160 			continue;
1161 		}
1162 
1163 		queued++;
1164 	} while (!list_empty(list));
1165 
1166 	hctx->dispatched[queued_to_index(queued)]++;
1167 
1168 	/*
1169 	 * Any items that need requeuing? Stuff them into hctx->dispatch,
1170 	 * that is where we will continue on next queue run.
1171 	 */
1172 	if (!list_empty(list)) {
1173 		bool needs_restart;
1174 
1175 		spin_lock(&hctx->lock);
1176 		list_splice_init(list, &hctx->dispatch);
1177 		spin_unlock(&hctx->lock);
1178 
1179 		/*
1180 		 * If SCHED_RESTART was set by the caller of this function and
1181 		 * it is no longer set that means that it was cleared by another
1182 		 * thread and hence that a queue rerun is needed.
1183 		 *
1184 		 * If 'no_tag' is set, that means that we failed getting
1185 		 * a driver tag with an I/O scheduler attached. If our dispatch
1186 		 * waitqueue is no longer active, ensure that we run the queue
1187 		 * AFTER adding our entries back to the list.
1188 		 *
1189 		 * If no I/O scheduler has been configured it is possible that
1190 		 * the hardware queue got stopped and restarted before requests
1191 		 * were pushed back onto the dispatch list. Rerun the queue to
1192 		 * avoid starvation. Notes:
1193 		 * - blk_mq_run_hw_queue() checks whether or not a queue has
1194 		 *   been stopped before rerunning a queue.
1195 		 * - Some but not all block drivers stop a queue before
1196 		 *   returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1197 		 *   and dm-rq.
1198 		 *
1199 		 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1200 		 * bit is set, run queue after a delay to avoid IO stalls
1201 		 * that could otherwise occur if the queue is idle.
1202 		 */
1203 		needs_restart = blk_mq_sched_needs_restart(hctx);
1204 		if (!needs_restart ||
1205 		    (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1206 			blk_mq_run_hw_queue(hctx, true);
1207 		else if (needs_restart && (ret == BLK_STS_RESOURCE))
1208 			blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1209 	}
1210 
1211 	return (queued + errors) != 0;
1212 }
1213 
1214 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1215 {
1216 	int srcu_idx;
1217 
1218 	/*
1219 	 * We should be running this queue from one of the CPUs that
1220 	 * are mapped to it.
1221 	 *
1222 	 * There are at least two related races now between setting
1223 	 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1224 	 * __blk_mq_run_hw_queue():
1225 	 *
1226 	 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1227 	 *   but later it becomes online, then this warning is harmless
1228 	 *   at all
1229 	 *
1230 	 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1231 	 *   but later it becomes offline, then the warning can't be
1232 	 *   triggered, and we depend on blk-mq timeout handler to
1233 	 *   handle dispatched requests to this hctx
1234 	 */
1235 	if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1236 		cpu_online(hctx->next_cpu)) {
1237 		printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1238 			raw_smp_processor_id(),
1239 			cpumask_empty(hctx->cpumask) ? "inactive": "active");
1240 		dump_stack();
1241 	}
1242 
1243 	/*
1244 	 * We can't run the queue inline with ints disabled. Ensure that
1245 	 * we catch bad users of this early.
1246 	 */
1247 	WARN_ON_ONCE(in_interrupt());
1248 
1249 	might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1250 
1251 	hctx_lock(hctx, &srcu_idx);
1252 	blk_mq_sched_dispatch_requests(hctx);
1253 	hctx_unlock(hctx, srcu_idx);
1254 }
1255 
1256 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1257 {
1258 	int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1259 
1260 	if (cpu >= nr_cpu_ids)
1261 		cpu = cpumask_first(hctx->cpumask);
1262 	return cpu;
1263 }
1264 
1265 /*
1266  * It'd be great if the workqueue API had a way to pass
1267  * in a mask and had some smarts for more clever placement.
1268  * For now we just round-robin here, switching for every
1269  * BLK_MQ_CPU_WORK_BATCH queued items.
1270  */
1271 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1272 {
1273 	bool tried = false;
1274 	int next_cpu = hctx->next_cpu;
1275 
1276 	if (hctx->queue->nr_hw_queues == 1)
1277 		return WORK_CPU_UNBOUND;
1278 
1279 	if (--hctx->next_cpu_batch <= 0) {
1280 select_cpu:
1281 		next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1282 				cpu_online_mask);
1283 		if (next_cpu >= nr_cpu_ids)
1284 			next_cpu = blk_mq_first_mapped_cpu(hctx);
1285 		hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1286 	}
1287 
1288 	/*
1289 	 * Do unbound schedule if we can't find a online CPU for this hctx,
1290 	 * and it should only happen in the path of handling CPU DEAD.
1291 	 */
1292 	if (!cpu_online(next_cpu)) {
1293 		if (!tried) {
1294 			tried = true;
1295 			goto select_cpu;
1296 		}
1297 
1298 		/*
1299 		 * Make sure to re-select CPU next time once after CPUs
1300 		 * in hctx->cpumask become online again.
1301 		 */
1302 		hctx->next_cpu = next_cpu;
1303 		hctx->next_cpu_batch = 1;
1304 		return WORK_CPU_UNBOUND;
1305 	}
1306 
1307 	hctx->next_cpu = next_cpu;
1308 	return next_cpu;
1309 }
1310 
1311 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1312 					unsigned long msecs)
1313 {
1314 	if (unlikely(blk_mq_hctx_stopped(hctx)))
1315 		return;
1316 
1317 	if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1318 		int cpu = get_cpu();
1319 		if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1320 			__blk_mq_run_hw_queue(hctx);
1321 			put_cpu();
1322 			return;
1323 		}
1324 
1325 		put_cpu();
1326 	}
1327 
1328 	kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1329 				    msecs_to_jiffies(msecs));
1330 }
1331 
1332 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1333 {
1334 	__blk_mq_delay_run_hw_queue(hctx, true, msecs);
1335 }
1336 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1337 
1338 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1339 {
1340 	int srcu_idx;
1341 	bool need_run;
1342 
1343 	/*
1344 	 * When queue is quiesced, we may be switching io scheduler, or
1345 	 * updating nr_hw_queues, or other things, and we can't run queue
1346 	 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1347 	 *
1348 	 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1349 	 * quiesced.
1350 	 */
1351 	hctx_lock(hctx, &srcu_idx);
1352 	need_run = !blk_queue_quiesced(hctx->queue) &&
1353 		blk_mq_hctx_has_pending(hctx);
1354 	hctx_unlock(hctx, srcu_idx);
1355 
1356 	if (need_run) {
1357 		__blk_mq_delay_run_hw_queue(hctx, async, 0);
1358 		return true;
1359 	}
1360 
1361 	return false;
1362 }
1363 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1364 
1365 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1366 {
1367 	struct blk_mq_hw_ctx *hctx;
1368 	int i;
1369 
1370 	queue_for_each_hw_ctx(q, hctx, i) {
1371 		if (blk_mq_hctx_stopped(hctx))
1372 			continue;
1373 
1374 		blk_mq_run_hw_queue(hctx, async);
1375 	}
1376 }
1377 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1378 
1379 /**
1380  * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1381  * @q: request queue.
1382  *
1383  * The caller is responsible for serializing this function against
1384  * blk_mq_{start,stop}_hw_queue().
1385  */
1386 bool blk_mq_queue_stopped(struct request_queue *q)
1387 {
1388 	struct blk_mq_hw_ctx *hctx;
1389 	int i;
1390 
1391 	queue_for_each_hw_ctx(q, hctx, i)
1392 		if (blk_mq_hctx_stopped(hctx))
1393 			return true;
1394 
1395 	return false;
1396 }
1397 EXPORT_SYMBOL(blk_mq_queue_stopped);
1398 
1399 /*
1400  * This function is often used for pausing .queue_rq() by driver when
1401  * there isn't enough resource or some conditions aren't satisfied, and
1402  * BLK_STS_RESOURCE is usually returned.
1403  *
1404  * We do not guarantee that dispatch can be drained or blocked
1405  * after blk_mq_stop_hw_queue() returns. Please use
1406  * blk_mq_quiesce_queue() for that requirement.
1407  */
1408 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1409 {
1410 	cancel_delayed_work(&hctx->run_work);
1411 
1412 	set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1413 }
1414 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1415 
1416 /*
1417  * This function is often used for pausing .queue_rq() by driver when
1418  * there isn't enough resource or some conditions aren't satisfied, and
1419  * BLK_STS_RESOURCE is usually returned.
1420  *
1421  * We do not guarantee that dispatch can be drained or blocked
1422  * after blk_mq_stop_hw_queues() returns. Please use
1423  * blk_mq_quiesce_queue() for that requirement.
1424  */
1425 void blk_mq_stop_hw_queues(struct request_queue *q)
1426 {
1427 	struct blk_mq_hw_ctx *hctx;
1428 	int i;
1429 
1430 	queue_for_each_hw_ctx(q, hctx, i)
1431 		blk_mq_stop_hw_queue(hctx);
1432 }
1433 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1434 
1435 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1436 {
1437 	clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1438 
1439 	blk_mq_run_hw_queue(hctx, false);
1440 }
1441 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1442 
1443 void blk_mq_start_hw_queues(struct request_queue *q)
1444 {
1445 	struct blk_mq_hw_ctx *hctx;
1446 	int i;
1447 
1448 	queue_for_each_hw_ctx(q, hctx, i)
1449 		blk_mq_start_hw_queue(hctx);
1450 }
1451 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1452 
1453 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1454 {
1455 	if (!blk_mq_hctx_stopped(hctx))
1456 		return;
1457 
1458 	clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1459 	blk_mq_run_hw_queue(hctx, async);
1460 }
1461 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1462 
1463 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1464 {
1465 	struct blk_mq_hw_ctx *hctx;
1466 	int i;
1467 
1468 	queue_for_each_hw_ctx(q, hctx, i)
1469 		blk_mq_start_stopped_hw_queue(hctx, async);
1470 }
1471 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1472 
1473 static void blk_mq_run_work_fn(struct work_struct *work)
1474 {
1475 	struct blk_mq_hw_ctx *hctx;
1476 
1477 	hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1478 
1479 	/*
1480 	 * If we are stopped, don't run the queue.
1481 	 */
1482 	if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1483 		return;
1484 
1485 	__blk_mq_run_hw_queue(hctx);
1486 }
1487 
1488 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1489 					    struct request *rq,
1490 					    bool at_head)
1491 {
1492 	struct blk_mq_ctx *ctx = rq->mq_ctx;
1493 
1494 	lockdep_assert_held(&ctx->lock);
1495 
1496 	trace_block_rq_insert(hctx->queue, rq);
1497 
1498 	if (at_head)
1499 		list_add(&rq->queuelist, &ctx->rq_list);
1500 	else
1501 		list_add_tail(&rq->queuelist, &ctx->rq_list);
1502 }
1503 
1504 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1505 			     bool at_head)
1506 {
1507 	struct blk_mq_ctx *ctx = rq->mq_ctx;
1508 
1509 	lockdep_assert_held(&ctx->lock);
1510 
1511 	__blk_mq_insert_req_list(hctx, rq, at_head);
1512 	blk_mq_hctx_mark_pending(hctx, ctx);
1513 }
1514 
1515 /*
1516  * Should only be used carefully, when the caller knows we want to
1517  * bypass a potential IO scheduler on the target device.
1518  */
1519 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1520 {
1521 	struct blk_mq_ctx *ctx = rq->mq_ctx;
1522 	struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1523 
1524 	spin_lock(&hctx->lock);
1525 	list_add_tail(&rq->queuelist, &hctx->dispatch);
1526 	spin_unlock(&hctx->lock);
1527 
1528 	if (run_queue)
1529 		blk_mq_run_hw_queue(hctx, false);
1530 }
1531 
1532 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1533 			    struct list_head *list)
1534 
1535 {
1536 	/*
1537 	 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1538 	 * offline now
1539 	 */
1540 	spin_lock(&ctx->lock);
1541 	while (!list_empty(list)) {
1542 		struct request *rq;
1543 
1544 		rq = list_first_entry(list, struct request, queuelist);
1545 		BUG_ON(rq->mq_ctx != ctx);
1546 		list_del_init(&rq->queuelist);
1547 		__blk_mq_insert_req_list(hctx, rq, false);
1548 	}
1549 	blk_mq_hctx_mark_pending(hctx, ctx);
1550 	spin_unlock(&ctx->lock);
1551 }
1552 
1553 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1554 {
1555 	struct request *rqa = container_of(a, struct request, queuelist);
1556 	struct request *rqb = container_of(b, struct request, queuelist);
1557 
1558 	return !(rqa->mq_ctx < rqb->mq_ctx ||
1559 		 (rqa->mq_ctx == rqb->mq_ctx &&
1560 		  blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1561 }
1562 
1563 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1564 {
1565 	struct blk_mq_ctx *this_ctx;
1566 	struct request_queue *this_q;
1567 	struct request *rq;
1568 	LIST_HEAD(list);
1569 	LIST_HEAD(ctx_list);
1570 	unsigned int depth;
1571 
1572 	list_splice_init(&plug->mq_list, &list);
1573 
1574 	list_sort(NULL, &list, plug_ctx_cmp);
1575 
1576 	this_q = NULL;
1577 	this_ctx = NULL;
1578 	depth = 0;
1579 
1580 	while (!list_empty(&list)) {
1581 		rq = list_entry_rq(list.next);
1582 		list_del_init(&rq->queuelist);
1583 		BUG_ON(!rq->q);
1584 		if (rq->mq_ctx != this_ctx) {
1585 			if (this_ctx) {
1586 				trace_block_unplug(this_q, depth, from_schedule);
1587 				blk_mq_sched_insert_requests(this_q, this_ctx,
1588 								&ctx_list,
1589 								from_schedule);
1590 			}
1591 
1592 			this_ctx = rq->mq_ctx;
1593 			this_q = rq->q;
1594 			depth = 0;
1595 		}
1596 
1597 		depth++;
1598 		list_add_tail(&rq->queuelist, &ctx_list);
1599 	}
1600 
1601 	/*
1602 	 * If 'this_ctx' is set, we know we have entries to complete
1603 	 * on 'ctx_list'. Do those.
1604 	 */
1605 	if (this_ctx) {
1606 		trace_block_unplug(this_q, depth, from_schedule);
1607 		blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1608 						from_schedule);
1609 	}
1610 }
1611 
1612 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1613 {
1614 	blk_init_request_from_bio(rq, bio);
1615 
1616 	blk_rq_set_rl(rq, blk_get_rl(rq->q, bio));
1617 
1618 	blk_account_io_start(rq, true);
1619 }
1620 
1621 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1622 {
1623 	if (rq->tag != -1)
1624 		return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1625 
1626 	return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1627 }
1628 
1629 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1630 					    struct request *rq,
1631 					    blk_qc_t *cookie)
1632 {
1633 	struct request_queue *q = rq->q;
1634 	struct blk_mq_queue_data bd = {
1635 		.rq = rq,
1636 		.last = true,
1637 	};
1638 	blk_qc_t new_cookie;
1639 	blk_status_t ret;
1640 
1641 	new_cookie = request_to_qc_t(hctx, rq);
1642 
1643 	/*
1644 	 * For OK queue, we are done. For error, caller may kill it.
1645 	 * Any other error (busy), just add it to our list as we
1646 	 * previously would have done.
1647 	 */
1648 	ret = q->mq_ops->queue_rq(hctx, &bd);
1649 	switch (ret) {
1650 	case BLK_STS_OK:
1651 		*cookie = new_cookie;
1652 		break;
1653 	case BLK_STS_RESOURCE:
1654 	case BLK_STS_DEV_RESOURCE:
1655 		__blk_mq_requeue_request(rq);
1656 		break;
1657 	default:
1658 		*cookie = BLK_QC_T_NONE;
1659 		break;
1660 	}
1661 
1662 	return ret;
1663 }
1664 
1665 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1666 						struct request *rq,
1667 						blk_qc_t *cookie,
1668 						bool bypass_insert)
1669 {
1670 	struct request_queue *q = rq->q;
1671 	bool run_queue = true;
1672 
1673 	/*
1674 	 * RCU or SRCU read lock is needed before checking quiesced flag.
1675 	 *
1676 	 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1677 	 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1678 	 * and avoid driver to try to dispatch again.
1679 	 */
1680 	if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1681 		run_queue = false;
1682 		bypass_insert = false;
1683 		goto insert;
1684 	}
1685 
1686 	if (q->elevator && !bypass_insert)
1687 		goto insert;
1688 
1689 	if (!blk_mq_get_dispatch_budget(hctx))
1690 		goto insert;
1691 
1692 	if (!blk_mq_get_driver_tag(rq, NULL, false)) {
1693 		blk_mq_put_dispatch_budget(hctx);
1694 		goto insert;
1695 	}
1696 
1697 	return __blk_mq_issue_directly(hctx, rq, cookie);
1698 insert:
1699 	if (bypass_insert)
1700 		return BLK_STS_RESOURCE;
1701 
1702 	blk_mq_sched_insert_request(rq, false, run_queue, false);
1703 	return BLK_STS_OK;
1704 }
1705 
1706 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1707 		struct request *rq, blk_qc_t *cookie)
1708 {
1709 	blk_status_t ret;
1710 	int srcu_idx;
1711 
1712 	might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1713 
1714 	hctx_lock(hctx, &srcu_idx);
1715 
1716 	ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1717 	if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1718 		blk_mq_sched_insert_request(rq, false, true, false);
1719 	else if (ret != BLK_STS_OK)
1720 		blk_mq_end_request(rq, ret);
1721 
1722 	hctx_unlock(hctx, srcu_idx);
1723 }
1724 
1725 blk_status_t blk_mq_request_issue_directly(struct request *rq)
1726 {
1727 	blk_status_t ret;
1728 	int srcu_idx;
1729 	blk_qc_t unused_cookie;
1730 	struct blk_mq_ctx *ctx = rq->mq_ctx;
1731 	struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1732 
1733 	hctx_lock(hctx, &srcu_idx);
1734 	ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true);
1735 	hctx_unlock(hctx, srcu_idx);
1736 
1737 	return ret;
1738 }
1739 
1740 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1741 {
1742 	const int is_sync = op_is_sync(bio->bi_opf);
1743 	const int is_flush_fua = op_is_flush(bio->bi_opf);
1744 	struct blk_mq_alloc_data data = { .flags = 0 };
1745 	struct request *rq;
1746 	unsigned int request_count = 0;
1747 	struct blk_plug *plug;
1748 	struct request *same_queue_rq = NULL;
1749 	blk_qc_t cookie;
1750 	unsigned int wb_acct;
1751 
1752 	blk_queue_bounce(q, &bio);
1753 
1754 	blk_queue_split(q, &bio);
1755 
1756 	if (!bio_integrity_prep(bio))
1757 		return BLK_QC_T_NONE;
1758 
1759 	if (!is_flush_fua && !blk_queue_nomerges(q) &&
1760 	    blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1761 		return BLK_QC_T_NONE;
1762 
1763 	if (blk_mq_sched_bio_merge(q, bio))
1764 		return BLK_QC_T_NONE;
1765 
1766 	wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1767 
1768 	trace_block_getrq(q, bio, bio->bi_opf);
1769 
1770 	rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1771 	if (unlikely(!rq)) {
1772 		__wbt_done(q->rq_wb, wb_acct);
1773 		if (bio->bi_opf & REQ_NOWAIT)
1774 			bio_wouldblock_error(bio);
1775 		return BLK_QC_T_NONE;
1776 	}
1777 
1778 	wbt_track(rq, wb_acct);
1779 
1780 	cookie = request_to_qc_t(data.hctx, rq);
1781 
1782 	plug = current->plug;
1783 	if (unlikely(is_flush_fua)) {
1784 		blk_mq_put_ctx(data.ctx);
1785 		blk_mq_bio_to_request(rq, bio);
1786 
1787 		/* bypass scheduler for flush rq */
1788 		blk_insert_flush(rq);
1789 		blk_mq_run_hw_queue(data.hctx, true);
1790 	} else if (plug && q->nr_hw_queues == 1) {
1791 		struct request *last = NULL;
1792 
1793 		blk_mq_put_ctx(data.ctx);
1794 		blk_mq_bio_to_request(rq, bio);
1795 
1796 		/*
1797 		 * @request_count may become stale because of schedule
1798 		 * out, so check the list again.
1799 		 */
1800 		if (list_empty(&plug->mq_list))
1801 			request_count = 0;
1802 		else if (blk_queue_nomerges(q))
1803 			request_count = blk_plug_queued_count(q);
1804 
1805 		if (!request_count)
1806 			trace_block_plug(q);
1807 		else
1808 			last = list_entry_rq(plug->mq_list.prev);
1809 
1810 		if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1811 		    blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1812 			blk_flush_plug_list(plug, false);
1813 			trace_block_plug(q);
1814 		}
1815 
1816 		list_add_tail(&rq->queuelist, &plug->mq_list);
1817 	} else if (plug && !blk_queue_nomerges(q)) {
1818 		blk_mq_bio_to_request(rq, bio);
1819 
1820 		/*
1821 		 * We do limited plugging. If the bio can be merged, do that.
1822 		 * Otherwise the existing request in the plug list will be
1823 		 * issued. So the plug list will have one request at most
1824 		 * The plug list might get flushed before this. If that happens,
1825 		 * the plug list is empty, and same_queue_rq is invalid.
1826 		 */
1827 		if (list_empty(&plug->mq_list))
1828 			same_queue_rq = NULL;
1829 		if (same_queue_rq)
1830 			list_del_init(&same_queue_rq->queuelist);
1831 		list_add_tail(&rq->queuelist, &plug->mq_list);
1832 
1833 		blk_mq_put_ctx(data.ctx);
1834 
1835 		if (same_queue_rq) {
1836 			data.hctx = blk_mq_map_queue(q,
1837 					same_queue_rq->mq_ctx->cpu);
1838 			blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1839 					&cookie);
1840 		}
1841 	} else if (q->nr_hw_queues > 1 && is_sync) {
1842 		blk_mq_put_ctx(data.ctx);
1843 		blk_mq_bio_to_request(rq, bio);
1844 		blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1845 	} else {
1846 		blk_mq_put_ctx(data.ctx);
1847 		blk_mq_bio_to_request(rq, bio);
1848 		blk_mq_sched_insert_request(rq, false, true, true);
1849 	}
1850 
1851 	return cookie;
1852 }
1853 
1854 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1855 		     unsigned int hctx_idx)
1856 {
1857 	struct page *page;
1858 
1859 	if (tags->rqs && set->ops->exit_request) {
1860 		int i;
1861 
1862 		for (i = 0; i < tags->nr_tags; i++) {
1863 			struct request *rq = tags->static_rqs[i];
1864 
1865 			if (!rq)
1866 				continue;
1867 			set->ops->exit_request(set, rq, hctx_idx);
1868 			tags->static_rqs[i] = NULL;
1869 		}
1870 	}
1871 
1872 	while (!list_empty(&tags->page_list)) {
1873 		page = list_first_entry(&tags->page_list, struct page, lru);
1874 		list_del_init(&page->lru);
1875 		/*
1876 		 * Remove kmemleak object previously allocated in
1877 		 * blk_mq_init_rq_map().
1878 		 */
1879 		kmemleak_free(page_address(page));
1880 		__free_pages(page, page->private);
1881 	}
1882 }
1883 
1884 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1885 {
1886 	kfree(tags->rqs);
1887 	tags->rqs = NULL;
1888 	kfree(tags->static_rqs);
1889 	tags->static_rqs = NULL;
1890 
1891 	blk_mq_free_tags(tags);
1892 }
1893 
1894 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1895 					unsigned int hctx_idx,
1896 					unsigned int nr_tags,
1897 					unsigned int reserved_tags)
1898 {
1899 	struct blk_mq_tags *tags;
1900 	int node;
1901 
1902 	node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1903 	if (node == NUMA_NO_NODE)
1904 		node = set->numa_node;
1905 
1906 	tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1907 				BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1908 	if (!tags)
1909 		return NULL;
1910 
1911 	tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
1912 				 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1913 				 node);
1914 	if (!tags->rqs) {
1915 		blk_mq_free_tags(tags);
1916 		return NULL;
1917 	}
1918 
1919 	tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
1920 					GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1921 					node);
1922 	if (!tags->static_rqs) {
1923 		kfree(tags->rqs);
1924 		blk_mq_free_tags(tags);
1925 		return NULL;
1926 	}
1927 
1928 	return tags;
1929 }
1930 
1931 static size_t order_to_size(unsigned int order)
1932 {
1933 	return (size_t)PAGE_SIZE << order;
1934 }
1935 
1936 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
1937 			       unsigned int hctx_idx, int node)
1938 {
1939 	int ret;
1940 
1941 	if (set->ops->init_request) {
1942 		ret = set->ops->init_request(set, rq, hctx_idx, node);
1943 		if (ret)
1944 			return ret;
1945 	}
1946 
1947 	WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1948 	return 0;
1949 }
1950 
1951 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1952 		     unsigned int hctx_idx, unsigned int depth)
1953 {
1954 	unsigned int i, j, entries_per_page, max_order = 4;
1955 	size_t rq_size, left;
1956 	int node;
1957 
1958 	node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1959 	if (node == NUMA_NO_NODE)
1960 		node = set->numa_node;
1961 
1962 	INIT_LIST_HEAD(&tags->page_list);
1963 
1964 	/*
1965 	 * rq_size is the size of the request plus driver payload, rounded
1966 	 * to the cacheline size
1967 	 */
1968 	rq_size = round_up(sizeof(struct request) + set->cmd_size,
1969 				cache_line_size());
1970 	left = rq_size * depth;
1971 
1972 	for (i = 0; i < depth; ) {
1973 		int this_order = max_order;
1974 		struct page *page;
1975 		int to_do;
1976 		void *p;
1977 
1978 		while (this_order && left < order_to_size(this_order - 1))
1979 			this_order--;
1980 
1981 		do {
1982 			page = alloc_pages_node(node,
1983 				GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1984 				this_order);
1985 			if (page)
1986 				break;
1987 			if (!this_order--)
1988 				break;
1989 			if (order_to_size(this_order) < rq_size)
1990 				break;
1991 		} while (1);
1992 
1993 		if (!page)
1994 			goto fail;
1995 
1996 		page->private = this_order;
1997 		list_add_tail(&page->lru, &tags->page_list);
1998 
1999 		p = page_address(page);
2000 		/*
2001 		 * Allow kmemleak to scan these pages as they contain pointers
2002 		 * to additional allocations like via ops->init_request().
2003 		 */
2004 		kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2005 		entries_per_page = order_to_size(this_order) / rq_size;
2006 		to_do = min(entries_per_page, depth - i);
2007 		left -= to_do * rq_size;
2008 		for (j = 0; j < to_do; j++) {
2009 			struct request *rq = p;
2010 
2011 			tags->static_rqs[i] = rq;
2012 			if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2013 				tags->static_rqs[i] = NULL;
2014 				goto fail;
2015 			}
2016 
2017 			p += rq_size;
2018 			i++;
2019 		}
2020 	}
2021 	return 0;
2022 
2023 fail:
2024 	blk_mq_free_rqs(set, tags, hctx_idx);
2025 	return -ENOMEM;
2026 }
2027 
2028 /*
2029  * 'cpu' is going away. splice any existing rq_list entries from this
2030  * software queue to the hw queue dispatch list, and ensure that it
2031  * gets run.
2032  */
2033 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2034 {
2035 	struct blk_mq_hw_ctx *hctx;
2036 	struct blk_mq_ctx *ctx;
2037 	LIST_HEAD(tmp);
2038 
2039 	hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2040 	ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2041 
2042 	spin_lock(&ctx->lock);
2043 	if (!list_empty(&ctx->rq_list)) {
2044 		list_splice_init(&ctx->rq_list, &tmp);
2045 		blk_mq_hctx_clear_pending(hctx, ctx);
2046 	}
2047 	spin_unlock(&ctx->lock);
2048 
2049 	if (list_empty(&tmp))
2050 		return 0;
2051 
2052 	spin_lock(&hctx->lock);
2053 	list_splice_tail_init(&tmp, &hctx->dispatch);
2054 	spin_unlock(&hctx->lock);
2055 
2056 	blk_mq_run_hw_queue(hctx, true);
2057 	return 0;
2058 }
2059 
2060 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2061 {
2062 	cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2063 					    &hctx->cpuhp_dead);
2064 }
2065 
2066 /* hctx->ctxs will be freed in queue's release handler */
2067 static void blk_mq_exit_hctx(struct request_queue *q,
2068 		struct blk_mq_tag_set *set,
2069 		struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2070 {
2071 	blk_mq_debugfs_unregister_hctx(hctx);
2072 
2073 	if (blk_mq_hw_queue_mapped(hctx))
2074 		blk_mq_tag_idle(hctx);
2075 
2076 	if (set->ops->exit_request)
2077 		set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2078 
2079 	blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2080 
2081 	if (set->ops->exit_hctx)
2082 		set->ops->exit_hctx(hctx, hctx_idx);
2083 
2084 	if (hctx->flags & BLK_MQ_F_BLOCKING)
2085 		cleanup_srcu_struct(hctx->srcu);
2086 
2087 	blk_mq_remove_cpuhp(hctx);
2088 	blk_free_flush_queue(hctx->fq);
2089 	sbitmap_free(&hctx->ctx_map);
2090 }
2091 
2092 static void blk_mq_exit_hw_queues(struct request_queue *q,
2093 		struct blk_mq_tag_set *set, int nr_queue)
2094 {
2095 	struct blk_mq_hw_ctx *hctx;
2096 	unsigned int i;
2097 
2098 	queue_for_each_hw_ctx(q, hctx, i) {
2099 		if (i == nr_queue)
2100 			break;
2101 		blk_mq_exit_hctx(q, set, hctx, i);
2102 	}
2103 }
2104 
2105 static int blk_mq_init_hctx(struct request_queue *q,
2106 		struct blk_mq_tag_set *set,
2107 		struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2108 {
2109 	int node;
2110 
2111 	node = hctx->numa_node;
2112 	if (node == NUMA_NO_NODE)
2113 		node = hctx->numa_node = set->numa_node;
2114 
2115 	INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2116 	spin_lock_init(&hctx->lock);
2117 	INIT_LIST_HEAD(&hctx->dispatch);
2118 	hctx->queue = q;
2119 	hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2120 
2121 	cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2122 
2123 	hctx->tags = set->tags[hctx_idx];
2124 
2125 	/*
2126 	 * Allocate space for all possible cpus to avoid allocation at
2127 	 * runtime
2128 	 */
2129 	hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2130 					GFP_KERNEL, node);
2131 	if (!hctx->ctxs)
2132 		goto unregister_cpu_notifier;
2133 
2134 	if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
2135 			      node))
2136 		goto free_ctxs;
2137 
2138 	hctx->nr_ctx = 0;
2139 
2140 	init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2141 	INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2142 
2143 	if (set->ops->init_hctx &&
2144 	    set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2145 		goto free_bitmap;
2146 
2147 	if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
2148 		goto exit_hctx;
2149 
2150 	hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
2151 	if (!hctx->fq)
2152 		goto sched_exit_hctx;
2153 
2154 	if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node))
2155 		goto free_fq;
2156 
2157 	if (hctx->flags & BLK_MQ_F_BLOCKING)
2158 		init_srcu_struct(hctx->srcu);
2159 
2160 	blk_mq_debugfs_register_hctx(q, hctx);
2161 
2162 	return 0;
2163 
2164  free_fq:
2165 	kfree(hctx->fq);
2166  sched_exit_hctx:
2167 	blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2168  exit_hctx:
2169 	if (set->ops->exit_hctx)
2170 		set->ops->exit_hctx(hctx, hctx_idx);
2171  free_bitmap:
2172 	sbitmap_free(&hctx->ctx_map);
2173  free_ctxs:
2174 	kfree(hctx->ctxs);
2175  unregister_cpu_notifier:
2176 	blk_mq_remove_cpuhp(hctx);
2177 	return -1;
2178 }
2179 
2180 static void blk_mq_init_cpu_queues(struct request_queue *q,
2181 				   unsigned int nr_hw_queues)
2182 {
2183 	unsigned int i;
2184 
2185 	for_each_possible_cpu(i) {
2186 		struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2187 		struct blk_mq_hw_ctx *hctx;
2188 
2189 		__ctx->cpu = i;
2190 		spin_lock_init(&__ctx->lock);
2191 		INIT_LIST_HEAD(&__ctx->rq_list);
2192 		__ctx->queue = q;
2193 
2194 		/*
2195 		 * Set local node, IFF we have more than one hw queue. If
2196 		 * not, we remain on the home node of the device
2197 		 */
2198 		hctx = blk_mq_map_queue(q, i);
2199 		if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2200 			hctx->numa_node = local_memory_node(cpu_to_node(i));
2201 	}
2202 }
2203 
2204 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2205 {
2206 	int ret = 0;
2207 
2208 	set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2209 					set->queue_depth, set->reserved_tags);
2210 	if (!set->tags[hctx_idx])
2211 		return false;
2212 
2213 	ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2214 				set->queue_depth);
2215 	if (!ret)
2216 		return true;
2217 
2218 	blk_mq_free_rq_map(set->tags[hctx_idx]);
2219 	set->tags[hctx_idx] = NULL;
2220 	return false;
2221 }
2222 
2223 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2224 					 unsigned int hctx_idx)
2225 {
2226 	if (set->tags[hctx_idx]) {
2227 		blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2228 		blk_mq_free_rq_map(set->tags[hctx_idx]);
2229 		set->tags[hctx_idx] = NULL;
2230 	}
2231 }
2232 
2233 static void blk_mq_map_swqueue(struct request_queue *q)
2234 {
2235 	unsigned int i, hctx_idx;
2236 	struct blk_mq_hw_ctx *hctx;
2237 	struct blk_mq_ctx *ctx;
2238 	struct blk_mq_tag_set *set = q->tag_set;
2239 
2240 	/*
2241 	 * Avoid others reading imcomplete hctx->cpumask through sysfs
2242 	 */
2243 	mutex_lock(&q->sysfs_lock);
2244 
2245 	queue_for_each_hw_ctx(q, hctx, i) {
2246 		cpumask_clear(hctx->cpumask);
2247 		hctx->nr_ctx = 0;
2248 		hctx->dispatch_from = NULL;
2249 	}
2250 
2251 	/*
2252 	 * Map software to hardware queues.
2253 	 *
2254 	 * If the cpu isn't present, the cpu is mapped to first hctx.
2255 	 */
2256 	for_each_possible_cpu(i) {
2257 		hctx_idx = q->mq_map[i];
2258 		/* unmapped hw queue can be remapped after CPU topo changed */
2259 		if (!set->tags[hctx_idx] &&
2260 		    !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2261 			/*
2262 			 * If tags initialization fail for some hctx,
2263 			 * that hctx won't be brought online.  In this
2264 			 * case, remap the current ctx to hctx[0] which
2265 			 * is guaranteed to always have tags allocated
2266 			 */
2267 			q->mq_map[i] = 0;
2268 		}
2269 
2270 		ctx = per_cpu_ptr(q->queue_ctx, i);
2271 		hctx = blk_mq_map_queue(q, i);
2272 
2273 		cpumask_set_cpu(i, hctx->cpumask);
2274 		ctx->index_hw = hctx->nr_ctx;
2275 		hctx->ctxs[hctx->nr_ctx++] = ctx;
2276 	}
2277 
2278 	mutex_unlock(&q->sysfs_lock);
2279 
2280 	queue_for_each_hw_ctx(q, hctx, i) {
2281 		/*
2282 		 * If no software queues are mapped to this hardware queue,
2283 		 * disable it and free the request entries.
2284 		 */
2285 		if (!hctx->nr_ctx) {
2286 			/* Never unmap queue 0.  We need it as a
2287 			 * fallback in case of a new remap fails
2288 			 * allocation
2289 			 */
2290 			if (i && set->tags[i])
2291 				blk_mq_free_map_and_requests(set, i);
2292 
2293 			hctx->tags = NULL;
2294 			continue;
2295 		}
2296 
2297 		hctx->tags = set->tags[i];
2298 		WARN_ON(!hctx->tags);
2299 
2300 		/*
2301 		 * Set the map size to the number of mapped software queues.
2302 		 * This is more accurate and more efficient than looping
2303 		 * over all possibly mapped software queues.
2304 		 */
2305 		sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2306 
2307 		/*
2308 		 * Initialize batch roundrobin counts
2309 		 */
2310 		hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2311 		hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2312 	}
2313 }
2314 
2315 /*
2316  * Caller needs to ensure that we're either frozen/quiesced, or that
2317  * the queue isn't live yet.
2318  */
2319 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2320 {
2321 	struct blk_mq_hw_ctx *hctx;
2322 	int i;
2323 
2324 	queue_for_each_hw_ctx(q, hctx, i) {
2325 		if (shared) {
2326 			if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2327 				atomic_inc(&q->shared_hctx_restart);
2328 			hctx->flags |= BLK_MQ_F_TAG_SHARED;
2329 		} else {
2330 			if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2331 				atomic_dec(&q->shared_hctx_restart);
2332 			hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2333 		}
2334 	}
2335 }
2336 
2337 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2338 					bool shared)
2339 {
2340 	struct request_queue *q;
2341 
2342 	lockdep_assert_held(&set->tag_list_lock);
2343 
2344 	list_for_each_entry(q, &set->tag_list, tag_set_list) {
2345 		blk_mq_freeze_queue(q);
2346 		queue_set_hctx_shared(q, shared);
2347 		blk_mq_unfreeze_queue(q);
2348 	}
2349 }
2350 
2351 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2352 {
2353 	struct blk_mq_tag_set *set = q->tag_set;
2354 
2355 	mutex_lock(&set->tag_list_lock);
2356 	list_del_rcu(&q->tag_set_list);
2357 	if (list_is_singular(&set->tag_list)) {
2358 		/* just transitioned to unshared */
2359 		set->flags &= ~BLK_MQ_F_TAG_SHARED;
2360 		/* update existing queue */
2361 		blk_mq_update_tag_set_depth(set, false);
2362 	}
2363 	mutex_unlock(&set->tag_list_lock);
2364 	synchronize_rcu();
2365 	INIT_LIST_HEAD(&q->tag_set_list);
2366 }
2367 
2368 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2369 				     struct request_queue *q)
2370 {
2371 	q->tag_set = set;
2372 
2373 	mutex_lock(&set->tag_list_lock);
2374 
2375 	/*
2376 	 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2377 	 */
2378 	if (!list_empty(&set->tag_list) &&
2379 	    !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2380 		set->flags |= BLK_MQ_F_TAG_SHARED;
2381 		/* update existing queue */
2382 		blk_mq_update_tag_set_depth(set, true);
2383 	}
2384 	if (set->flags & BLK_MQ_F_TAG_SHARED)
2385 		queue_set_hctx_shared(q, true);
2386 	list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2387 
2388 	mutex_unlock(&set->tag_list_lock);
2389 }
2390 
2391 /*
2392  * It is the actual release handler for mq, but we do it from
2393  * request queue's release handler for avoiding use-after-free
2394  * and headache because q->mq_kobj shouldn't have been introduced,
2395  * but we can't group ctx/kctx kobj without it.
2396  */
2397 void blk_mq_release(struct request_queue *q)
2398 {
2399 	struct blk_mq_hw_ctx *hctx;
2400 	unsigned int i;
2401 
2402 	/* hctx kobj stays in hctx */
2403 	queue_for_each_hw_ctx(q, hctx, i) {
2404 		if (!hctx)
2405 			continue;
2406 		kobject_put(&hctx->kobj);
2407 	}
2408 
2409 	q->mq_map = NULL;
2410 
2411 	kfree(q->queue_hw_ctx);
2412 
2413 	/*
2414 	 * release .mq_kobj and sw queue's kobject now because
2415 	 * both share lifetime with request queue.
2416 	 */
2417 	blk_mq_sysfs_deinit(q);
2418 
2419 	free_percpu(q->queue_ctx);
2420 }
2421 
2422 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2423 {
2424 	struct request_queue *uninit_q, *q;
2425 
2426 	uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node, NULL);
2427 	if (!uninit_q)
2428 		return ERR_PTR(-ENOMEM);
2429 
2430 	q = blk_mq_init_allocated_queue(set, uninit_q);
2431 	if (IS_ERR(q))
2432 		blk_cleanup_queue(uninit_q);
2433 
2434 	return q;
2435 }
2436 EXPORT_SYMBOL(blk_mq_init_queue);
2437 
2438 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2439 {
2440 	int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2441 
2442 	BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2443 			   __alignof__(struct blk_mq_hw_ctx)) !=
2444 		     sizeof(struct blk_mq_hw_ctx));
2445 
2446 	if (tag_set->flags & BLK_MQ_F_BLOCKING)
2447 		hw_ctx_size += sizeof(struct srcu_struct);
2448 
2449 	return hw_ctx_size;
2450 }
2451 
2452 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2453 						struct request_queue *q)
2454 {
2455 	int i, j;
2456 	struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2457 
2458 	blk_mq_sysfs_unregister(q);
2459 
2460 	/* protect against switching io scheduler  */
2461 	mutex_lock(&q->sysfs_lock);
2462 	for (i = 0; i < set->nr_hw_queues; i++) {
2463 		int node;
2464 
2465 		if (hctxs[i])
2466 			continue;
2467 
2468 		node = blk_mq_hw_queue_to_node(q->mq_map, i);
2469 		hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set),
2470 					GFP_KERNEL, node);
2471 		if (!hctxs[i])
2472 			break;
2473 
2474 		if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2475 						node)) {
2476 			kfree(hctxs[i]);
2477 			hctxs[i] = NULL;
2478 			break;
2479 		}
2480 
2481 		atomic_set(&hctxs[i]->nr_active, 0);
2482 		hctxs[i]->numa_node = node;
2483 		hctxs[i]->queue_num = i;
2484 
2485 		if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2486 			free_cpumask_var(hctxs[i]->cpumask);
2487 			kfree(hctxs[i]);
2488 			hctxs[i] = NULL;
2489 			break;
2490 		}
2491 		blk_mq_hctx_kobj_init(hctxs[i]);
2492 	}
2493 	for (j = i; j < q->nr_hw_queues; j++) {
2494 		struct blk_mq_hw_ctx *hctx = hctxs[j];
2495 
2496 		if (hctx) {
2497 			if (hctx->tags)
2498 				blk_mq_free_map_and_requests(set, j);
2499 			blk_mq_exit_hctx(q, set, hctx, j);
2500 			kobject_put(&hctx->kobj);
2501 			hctxs[j] = NULL;
2502 
2503 		}
2504 	}
2505 	q->nr_hw_queues = i;
2506 	mutex_unlock(&q->sysfs_lock);
2507 	blk_mq_sysfs_register(q);
2508 }
2509 
2510 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2511 						  struct request_queue *q)
2512 {
2513 	/* mark the queue as mq asap */
2514 	q->mq_ops = set->ops;
2515 
2516 	q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2517 					     blk_mq_poll_stats_bkt,
2518 					     BLK_MQ_POLL_STATS_BKTS, q);
2519 	if (!q->poll_cb)
2520 		goto err_exit;
2521 
2522 	q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2523 	if (!q->queue_ctx)
2524 		goto err_exit;
2525 
2526 	/* init q->mq_kobj and sw queues' kobjects */
2527 	blk_mq_sysfs_init(q);
2528 
2529 	q->queue_hw_ctx = kcalloc_node(nr_cpu_ids, sizeof(*(q->queue_hw_ctx)),
2530 						GFP_KERNEL, set->numa_node);
2531 	if (!q->queue_hw_ctx)
2532 		goto err_percpu;
2533 
2534 	q->mq_map = set->mq_map;
2535 
2536 	blk_mq_realloc_hw_ctxs(set, q);
2537 	if (!q->nr_hw_queues)
2538 		goto err_hctxs;
2539 
2540 	INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2541 	blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2542 
2543 	q->nr_queues = nr_cpu_ids;
2544 
2545 	q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2546 
2547 	if (!(set->flags & BLK_MQ_F_SG_MERGE))
2548 		queue_flag_set_unlocked(QUEUE_FLAG_NO_SG_MERGE, q);
2549 
2550 	q->sg_reserved_size = INT_MAX;
2551 
2552 	INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2553 	INIT_LIST_HEAD(&q->requeue_list);
2554 	spin_lock_init(&q->requeue_lock);
2555 
2556 	blk_queue_make_request(q, blk_mq_make_request);
2557 	if (q->mq_ops->poll)
2558 		q->poll_fn = blk_mq_poll;
2559 
2560 	/*
2561 	 * Do this after blk_queue_make_request() overrides it...
2562 	 */
2563 	q->nr_requests = set->queue_depth;
2564 
2565 	/*
2566 	 * Default to classic polling
2567 	 */
2568 	q->poll_nsec = -1;
2569 
2570 	if (set->ops->complete)
2571 		blk_queue_softirq_done(q, set->ops->complete);
2572 
2573 	blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2574 	blk_mq_add_queue_tag_set(set, q);
2575 	blk_mq_map_swqueue(q);
2576 
2577 	if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2578 		int ret;
2579 
2580 		ret = elevator_init_mq(q);
2581 		if (ret)
2582 			return ERR_PTR(ret);
2583 	}
2584 
2585 	return q;
2586 
2587 err_hctxs:
2588 	kfree(q->queue_hw_ctx);
2589 err_percpu:
2590 	free_percpu(q->queue_ctx);
2591 err_exit:
2592 	q->mq_ops = NULL;
2593 	return ERR_PTR(-ENOMEM);
2594 }
2595 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2596 
2597 void blk_mq_free_queue(struct request_queue *q)
2598 {
2599 	struct blk_mq_tag_set	*set = q->tag_set;
2600 
2601 	blk_mq_del_queue_tag_set(q);
2602 	blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2603 }
2604 
2605 /* Basically redo blk_mq_init_queue with queue frozen */
2606 static void blk_mq_queue_reinit(struct request_queue *q)
2607 {
2608 	WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2609 
2610 	blk_mq_debugfs_unregister_hctxs(q);
2611 	blk_mq_sysfs_unregister(q);
2612 
2613 	/*
2614 	 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2615 	 * we should change hctx numa_node according to the new topology (this
2616 	 * involves freeing and re-allocating memory, worth doing?)
2617 	 */
2618 	blk_mq_map_swqueue(q);
2619 
2620 	blk_mq_sysfs_register(q);
2621 	blk_mq_debugfs_register_hctxs(q);
2622 }
2623 
2624 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2625 {
2626 	int i;
2627 
2628 	for (i = 0; i < set->nr_hw_queues; i++)
2629 		if (!__blk_mq_alloc_rq_map(set, i))
2630 			goto out_unwind;
2631 
2632 	return 0;
2633 
2634 out_unwind:
2635 	while (--i >= 0)
2636 		blk_mq_free_rq_map(set->tags[i]);
2637 
2638 	return -ENOMEM;
2639 }
2640 
2641 /*
2642  * Allocate the request maps associated with this tag_set. Note that this
2643  * may reduce the depth asked for, if memory is tight. set->queue_depth
2644  * will be updated to reflect the allocated depth.
2645  */
2646 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2647 {
2648 	unsigned int depth;
2649 	int err;
2650 
2651 	depth = set->queue_depth;
2652 	do {
2653 		err = __blk_mq_alloc_rq_maps(set);
2654 		if (!err)
2655 			break;
2656 
2657 		set->queue_depth >>= 1;
2658 		if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2659 			err = -ENOMEM;
2660 			break;
2661 		}
2662 	} while (set->queue_depth);
2663 
2664 	if (!set->queue_depth || err) {
2665 		pr_err("blk-mq: failed to allocate request map\n");
2666 		return -ENOMEM;
2667 	}
2668 
2669 	if (depth != set->queue_depth)
2670 		pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2671 						depth, set->queue_depth);
2672 
2673 	return 0;
2674 }
2675 
2676 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2677 {
2678 	if (set->ops->map_queues) {
2679 		int cpu;
2680 		/*
2681 		 * transport .map_queues is usually done in the following
2682 		 * way:
2683 		 *
2684 		 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2685 		 * 	mask = get_cpu_mask(queue)
2686 		 * 	for_each_cpu(cpu, mask)
2687 		 * 		set->mq_map[cpu] = queue;
2688 		 * }
2689 		 *
2690 		 * When we need to remap, the table has to be cleared for
2691 		 * killing stale mapping since one CPU may not be mapped
2692 		 * to any hw queue.
2693 		 */
2694 		for_each_possible_cpu(cpu)
2695 			set->mq_map[cpu] = 0;
2696 
2697 		return set->ops->map_queues(set);
2698 	} else
2699 		return blk_mq_map_queues(set);
2700 }
2701 
2702 /*
2703  * Alloc a tag set to be associated with one or more request queues.
2704  * May fail with EINVAL for various error conditions. May adjust the
2705  * requested depth down, if if it too large. In that case, the set
2706  * value will be stored in set->queue_depth.
2707  */
2708 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2709 {
2710 	int ret;
2711 
2712 	BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2713 
2714 	if (!set->nr_hw_queues)
2715 		return -EINVAL;
2716 	if (!set->queue_depth)
2717 		return -EINVAL;
2718 	if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2719 		return -EINVAL;
2720 
2721 	if (!set->ops->queue_rq)
2722 		return -EINVAL;
2723 
2724 	if (!set->ops->get_budget ^ !set->ops->put_budget)
2725 		return -EINVAL;
2726 
2727 	if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2728 		pr_info("blk-mq: reduced tag depth to %u\n",
2729 			BLK_MQ_MAX_DEPTH);
2730 		set->queue_depth = BLK_MQ_MAX_DEPTH;
2731 	}
2732 
2733 	/*
2734 	 * If a crashdump is active, then we are potentially in a very
2735 	 * memory constrained environment. Limit us to 1 queue and
2736 	 * 64 tags to prevent using too much memory.
2737 	 */
2738 	if (is_kdump_kernel()) {
2739 		set->nr_hw_queues = 1;
2740 		set->queue_depth = min(64U, set->queue_depth);
2741 	}
2742 	/*
2743 	 * There is no use for more h/w queues than cpus.
2744 	 */
2745 	if (set->nr_hw_queues > nr_cpu_ids)
2746 		set->nr_hw_queues = nr_cpu_ids;
2747 
2748 	set->tags = kcalloc_node(nr_cpu_ids, sizeof(struct blk_mq_tags *),
2749 				 GFP_KERNEL, set->numa_node);
2750 	if (!set->tags)
2751 		return -ENOMEM;
2752 
2753 	ret = -ENOMEM;
2754 	set->mq_map = kcalloc_node(nr_cpu_ids, sizeof(*set->mq_map),
2755 				   GFP_KERNEL, set->numa_node);
2756 	if (!set->mq_map)
2757 		goto out_free_tags;
2758 
2759 	ret = blk_mq_update_queue_map(set);
2760 	if (ret)
2761 		goto out_free_mq_map;
2762 
2763 	ret = blk_mq_alloc_rq_maps(set);
2764 	if (ret)
2765 		goto out_free_mq_map;
2766 
2767 	mutex_init(&set->tag_list_lock);
2768 	INIT_LIST_HEAD(&set->tag_list);
2769 
2770 	return 0;
2771 
2772 out_free_mq_map:
2773 	kfree(set->mq_map);
2774 	set->mq_map = NULL;
2775 out_free_tags:
2776 	kfree(set->tags);
2777 	set->tags = NULL;
2778 	return ret;
2779 }
2780 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2781 
2782 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2783 {
2784 	int i;
2785 
2786 	for (i = 0; i < nr_cpu_ids; i++)
2787 		blk_mq_free_map_and_requests(set, i);
2788 
2789 	kfree(set->mq_map);
2790 	set->mq_map = NULL;
2791 
2792 	kfree(set->tags);
2793 	set->tags = NULL;
2794 }
2795 EXPORT_SYMBOL(blk_mq_free_tag_set);
2796 
2797 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2798 {
2799 	struct blk_mq_tag_set *set = q->tag_set;
2800 	struct blk_mq_hw_ctx *hctx;
2801 	int i, ret;
2802 
2803 	if (!set)
2804 		return -EINVAL;
2805 
2806 	blk_mq_freeze_queue(q);
2807 	blk_mq_quiesce_queue(q);
2808 
2809 	ret = 0;
2810 	queue_for_each_hw_ctx(q, hctx, i) {
2811 		if (!hctx->tags)
2812 			continue;
2813 		/*
2814 		 * If we're using an MQ scheduler, just update the scheduler
2815 		 * queue depth. This is similar to what the old code would do.
2816 		 */
2817 		if (!hctx->sched_tags) {
2818 			ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
2819 							false);
2820 		} else {
2821 			ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2822 							nr, true);
2823 		}
2824 		if (ret)
2825 			break;
2826 	}
2827 
2828 	if (!ret)
2829 		q->nr_requests = nr;
2830 
2831 	blk_mq_unquiesce_queue(q);
2832 	blk_mq_unfreeze_queue(q);
2833 
2834 	return ret;
2835 }
2836 
2837 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2838 							int nr_hw_queues)
2839 {
2840 	struct request_queue *q;
2841 
2842 	lockdep_assert_held(&set->tag_list_lock);
2843 
2844 	if (nr_hw_queues > nr_cpu_ids)
2845 		nr_hw_queues = nr_cpu_ids;
2846 	if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2847 		return;
2848 
2849 	list_for_each_entry(q, &set->tag_list, tag_set_list)
2850 		blk_mq_freeze_queue(q);
2851 
2852 	set->nr_hw_queues = nr_hw_queues;
2853 	blk_mq_update_queue_map(set);
2854 	list_for_each_entry(q, &set->tag_list, tag_set_list) {
2855 		blk_mq_realloc_hw_ctxs(set, q);
2856 		blk_mq_queue_reinit(q);
2857 	}
2858 
2859 	list_for_each_entry(q, &set->tag_list, tag_set_list)
2860 		blk_mq_unfreeze_queue(q);
2861 }
2862 
2863 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2864 {
2865 	mutex_lock(&set->tag_list_lock);
2866 	__blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2867 	mutex_unlock(&set->tag_list_lock);
2868 }
2869 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2870 
2871 /* Enable polling stats and return whether they were already enabled. */
2872 static bool blk_poll_stats_enable(struct request_queue *q)
2873 {
2874 	if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2875 	    blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
2876 		return true;
2877 	blk_stat_add_callback(q, q->poll_cb);
2878 	return false;
2879 }
2880 
2881 static void blk_mq_poll_stats_start(struct request_queue *q)
2882 {
2883 	/*
2884 	 * We don't arm the callback if polling stats are not enabled or the
2885 	 * callback is already active.
2886 	 */
2887 	if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2888 	    blk_stat_is_active(q->poll_cb))
2889 		return;
2890 
2891 	blk_stat_activate_msecs(q->poll_cb, 100);
2892 }
2893 
2894 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2895 {
2896 	struct request_queue *q = cb->data;
2897 	int bucket;
2898 
2899 	for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
2900 		if (cb->stat[bucket].nr_samples)
2901 			q->poll_stat[bucket] = cb->stat[bucket];
2902 	}
2903 }
2904 
2905 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2906 				       struct blk_mq_hw_ctx *hctx,
2907 				       struct request *rq)
2908 {
2909 	unsigned long ret = 0;
2910 	int bucket;
2911 
2912 	/*
2913 	 * If stats collection isn't on, don't sleep but turn it on for
2914 	 * future users
2915 	 */
2916 	if (!blk_poll_stats_enable(q))
2917 		return 0;
2918 
2919 	/*
2920 	 * As an optimistic guess, use half of the mean service time
2921 	 * for this type of request. We can (and should) make this smarter.
2922 	 * For instance, if the completion latencies are tight, we can
2923 	 * get closer than just half the mean. This is especially
2924 	 * important on devices where the completion latencies are longer
2925 	 * than ~10 usec. We do use the stats for the relevant IO size
2926 	 * if available which does lead to better estimates.
2927 	 */
2928 	bucket = blk_mq_poll_stats_bkt(rq);
2929 	if (bucket < 0)
2930 		return ret;
2931 
2932 	if (q->poll_stat[bucket].nr_samples)
2933 		ret = (q->poll_stat[bucket].mean + 1) / 2;
2934 
2935 	return ret;
2936 }
2937 
2938 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2939 				     struct blk_mq_hw_ctx *hctx,
2940 				     struct request *rq)
2941 {
2942 	struct hrtimer_sleeper hs;
2943 	enum hrtimer_mode mode;
2944 	unsigned int nsecs;
2945 	ktime_t kt;
2946 
2947 	if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
2948 		return false;
2949 
2950 	/*
2951 	 * poll_nsec can be:
2952 	 *
2953 	 * -1:	don't ever hybrid sleep
2954 	 *  0:	use half of prev avg
2955 	 * >0:	use this specific value
2956 	 */
2957 	if (q->poll_nsec == -1)
2958 		return false;
2959 	else if (q->poll_nsec > 0)
2960 		nsecs = q->poll_nsec;
2961 	else
2962 		nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2963 
2964 	if (!nsecs)
2965 		return false;
2966 
2967 	rq->rq_flags |= RQF_MQ_POLL_SLEPT;
2968 
2969 	/*
2970 	 * This will be replaced with the stats tracking code, using
2971 	 * 'avg_completion_time / 2' as the pre-sleep target.
2972 	 */
2973 	kt = nsecs;
2974 
2975 	mode = HRTIMER_MODE_REL;
2976 	hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2977 	hrtimer_set_expires(&hs.timer, kt);
2978 
2979 	hrtimer_init_sleeper(&hs, current);
2980 	do {
2981 		if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
2982 			break;
2983 		set_current_state(TASK_UNINTERRUPTIBLE);
2984 		hrtimer_start_expires(&hs.timer, mode);
2985 		if (hs.task)
2986 			io_schedule();
2987 		hrtimer_cancel(&hs.timer);
2988 		mode = HRTIMER_MODE_ABS;
2989 	} while (hs.task && !signal_pending(current));
2990 
2991 	__set_current_state(TASK_RUNNING);
2992 	destroy_hrtimer_on_stack(&hs.timer);
2993 	return true;
2994 }
2995 
2996 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2997 {
2998 	struct request_queue *q = hctx->queue;
2999 	long state;
3000 
3001 	/*
3002 	 * If we sleep, have the caller restart the poll loop to reset
3003 	 * the state. Like for the other success return cases, the
3004 	 * caller is responsible for checking if the IO completed. If
3005 	 * the IO isn't complete, we'll get called again and will go
3006 	 * straight to the busy poll loop.
3007 	 */
3008 	if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
3009 		return true;
3010 
3011 	hctx->poll_considered++;
3012 
3013 	state = current->state;
3014 	while (!need_resched()) {
3015 		int ret;
3016 
3017 		hctx->poll_invoked++;
3018 
3019 		ret = q->mq_ops->poll(hctx, rq->tag);
3020 		if (ret > 0) {
3021 			hctx->poll_success++;
3022 			set_current_state(TASK_RUNNING);
3023 			return true;
3024 		}
3025 
3026 		if (signal_pending_state(state, current))
3027 			set_current_state(TASK_RUNNING);
3028 
3029 		if (current->state == TASK_RUNNING)
3030 			return true;
3031 		if (ret < 0)
3032 			break;
3033 		cpu_relax();
3034 	}
3035 
3036 	__set_current_state(TASK_RUNNING);
3037 	return false;
3038 }
3039 
3040 static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
3041 {
3042 	struct blk_mq_hw_ctx *hctx;
3043 	struct request *rq;
3044 
3045 	if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3046 		return false;
3047 
3048 	hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3049 	if (!blk_qc_t_is_internal(cookie))
3050 		rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3051 	else {
3052 		rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3053 		/*
3054 		 * With scheduling, if the request has completed, we'll
3055 		 * get a NULL return here, as we clear the sched tag when
3056 		 * that happens. The request still remains valid, like always,
3057 		 * so we should be safe with just the NULL check.
3058 		 */
3059 		if (!rq)
3060 			return false;
3061 	}
3062 
3063 	return __blk_mq_poll(hctx, rq);
3064 }
3065 
3066 static int __init blk_mq_init(void)
3067 {
3068 	cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3069 				blk_mq_hctx_notify_dead);
3070 	return 0;
3071 }
3072 subsys_initcall(blk_mq_init);
3073