xref: /openbmc/linux/block/blk-mq.c (revision 7e6f7d24)
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 	blk_add_timer(req);
785 }
786 
787 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
788 {
789 	unsigned long deadline;
790 
791 	if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
792 		return false;
793 	if (rq->rq_flags & RQF_TIMED_OUT)
794 		return false;
795 
796 	deadline = blk_rq_deadline(rq);
797 	if (time_after_eq(jiffies, deadline))
798 		return true;
799 
800 	if (*next == 0)
801 		*next = deadline;
802 	else if (time_after(*next, deadline))
803 		*next = deadline;
804 	return false;
805 }
806 
807 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
808 		struct request *rq, void *priv, bool reserved)
809 {
810 	unsigned long *next = priv;
811 
812 	/*
813 	 * Just do a quick check if it is expired before locking the request in
814 	 * so we're not unnecessarilly synchronizing across CPUs.
815 	 */
816 	if (!blk_mq_req_expired(rq, next))
817 		return;
818 
819 	/*
820 	 * We have reason to believe the request may be expired. Take a
821 	 * reference on the request to lock this request lifetime into its
822 	 * currently allocated context to prevent it from being reallocated in
823 	 * the event the completion by-passes this timeout handler.
824 	 *
825 	 * If the reference was already released, then the driver beat the
826 	 * timeout handler to posting a natural completion.
827 	 */
828 	if (!refcount_inc_not_zero(&rq->ref))
829 		return;
830 
831 	/*
832 	 * The request is now locked and cannot be reallocated underneath the
833 	 * timeout handler's processing. Re-verify this exact request is truly
834 	 * expired; if it is not expired, then the request was completed and
835 	 * reallocated as a new request.
836 	 */
837 	if (blk_mq_req_expired(rq, next))
838 		blk_mq_rq_timed_out(rq, reserved);
839 	if (refcount_dec_and_test(&rq->ref))
840 		__blk_mq_free_request(rq);
841 }
842 
843 static void blk_mq_timeout_work(struct work_struct *work)
844 {
845 	struct request_queue *q =
846 		container_of(work, struct request_queue, timeout_work);
847 	unsigned long next = 0;
848 	struct blk_mq_hw_ctx *hctx;
849 	int i;
850 
851 	/* A deadlock might occur if a request is stuck requiring a
852 	 * timeout at the same time a queue freeze is waiting
853 	 * completion, since the timeout code would not be able to
854 	 * acquire the queue reference here.
855 	 *
856 	 * That's why we don't use blk_queue_enter here; instead, we use
857 	 * percpu_ref_tryget directly, because we need to be able to
858 	 * obtain a reference even in the short window between the queue
859 	 * starting to freeze, by dropping the first reference in
860 	 * blk_freeze_queue_start, and the moment the last request is
861 	 * consumed, marked by the instant q_usage_counter reaches
862 	 * zero.
863 	 */
864 	if (!percpu_ref_tryget(&q->q_usage_counter))
865 		return;
866 
867 	blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
868 
869 	if (next != 0) {
870 		mod_timer(&q->timeout, next);
871 	} else {
872 		/*
873 		 * Request timeouts are handled as a forward rolling timer. If
874 		 * we end up here it means that no requests are pending and
875 		 * also that no request has been pending for a while. Mark
876 		 * each hctx as idle.
877 		 */
878 		queue_for_each_hw_ctx(q, hctx, i) {
879 			/* the hctx may be unmapped, so check it here */
880 			if (blk_mq_hw_queue_mapped(hctx))
881 				blk_mq_tag_idle(hctx);
882 		}
883 	}
884 	blk_queue_exit(q);
885 }
886 
887 struct flush_busy_ctx_data {
888 	struct blk_mq_hw_ctx *hctx;
889 	struct list_head *list;
890 };
891 
892 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
893 {
894 	struct flush_busy_ctx_data *flush_data = data;
895 	struct blk_mq_hw_ctx *hctx = flush_data->hctx;
896 	struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
897 
898 	spin_lock(&ctx->lock);
899 	list_splice_tail_init(&ctx->rq_list, flush_data->list);
900 	sbitmap_clear_bit(sb, bitnr);
901 	spin_unlock(&ctx->lock);
902 	return true;
903 }
904 
905 /*
906  * Process software queues that have been marked busy, splicing them
907  * to the for-dispatch
908  */
909 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
910 {
911 	struct flush_busy_ctx_data data = {
912 		.hctx = hctx,
913 		.list = list,
914 	};
915 
916 	sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
917 }
918 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
919 
920 struct dispatch_rq_data {
921 	struct blk_mq_hw_ctx *hctx;
922 	struct request *rq;
923 };
924 
925 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
926 		void *data)
927 {
928 	struct dispatch_rq_data *dispatch_data = data;
929 	struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
930 	struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
931 
932 	spin_lock(&ctx->lock);
933 	if (!list_empty(&ctx->rq_list)) {
934 		dispatch_data->rq = list_entry_rq(ctx->rq_list.next);
935 		list_del_init(&dispatch_data->rq->queuelist);
936 		if (list_empty(&ctx->rq_list))
937 			sbitmap_clear_bit(sb, bitnr);
938 	}
939 	spin_unlock(&ctx->lock);
940 
941 	return !dispatch_data->rq;
942 }
943 
944 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
945 					struct blk_mq_ctx *start)
946 {
947 	unsigned off = start ? start->index_hw : 0;
948 	struct dispatch_rq_data data = {
949 		.hctx = hctx,
950 		.rq   = NULL,
951 	};
952 
953 	__sbitmap_for_each_set(&hctx->ctx_map, off,
954 			       dispatch_rq_from_ctx, &data);
955 
956 	return data.rq;
957 }
958 
959 static inline unsigned int queued_to_index(unsigned int queued)
960 {
961 	if (!queued)
962 		return 0;
963 
964 	return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
965 }
966 
967 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
968 			   bool wait)
969 {
970 	struct blk_mq_alloc_data data = {
971 		.q = rq->q,
972 		.hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
973 		.flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
974 	};
975 
976 	might_sleep_if(wait);
977 
978 	if (rq->tag != -1)
979 		goto done;
980 
981 	if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
982 		data.flags |= BLK_MQ_REQ_RESERVED;
983 
984 	rq->tag = blk_mq_get_tag(&data);
985 	if (rq->tag >= 0) {
986 		if (blk_mq_tag_busy(data.hctx)) {
987 			rq->rq_flags |= RQF_MQ_INFLIGHT;
988 			atomic_inc(&data.hctx->nr_active);
989 		}
990 		data.hctx->tags->rqs[rq->tag] = rq;
991 	}
992 
993 done:
994 	if (hctx)
995 		*hctx = data.hctx;
996 	return rq->tag != -1;
997 }
998 
999 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1000 				int flags, void *key)
1001 {
1002 	struct blk_mq_hw_ctx *hctx;
1003 
1004 	hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1005 
1006 	list_del_init(&wait->entry);
1007 	blk_mq_run_hw_queue(hctx, true);
1008 	return 1;
1009 }
1010 
1011 /*
1012  * Mark us waiting for a tag. For shared tags, this involves hooking us into
1013  * the tag wakeups. For non-shared tags, we can simply mark us needing a
1014  * restart. For both cases, take care to check the condition again after
1015  * marking us as waiting.
1016  */
1017 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx **hctx,
1018 				 struct request *rq)
1019 {
1020 	struct blk_mq_hw_ctx *this_hctx = *hctx;
1021 	struct sbq_wait_state *ws;
1022 	wait_queue_entry_t *wait;
1023 	bool ret;
1024 
1025 	if (!(this_hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1026 		if (!test_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state))
1027 			set_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state);
1028 
1029 		/*
1030 		 * It's possible that a tag was freed in the window between the
1031 		 * allocation failure and adding the hardware queue to the wait
1032 		 * queue.
1033 		 *
1034 		 * Don't clear RESTART here, someone else could have set it.
1035 		 * At most this will cost an extra queue run.
1036 		 */
1037 		return blk_mq_get_driver_tag(rq, hctx, false);
1038 	}
1039 
1040 	wait = &this_hctx->dispatch_wait;
1041 	if (!list_empty_careful(&wait->entry))
1042 		return false;
1043 
1044 	spin_lock(&this_hctx->lock);
1045 	if (!list_empty(&wait->entry)) {
1046 		spin_unlock(&this_hctx->lock);
1047 		return false;
1048 	}
1049 
1050 	ws = bt_wait_ptr(&this_hctx->tags->bitmap_tags, this_hctx);
1051 	add_wait_queue(&ws->wait, wait);
1052 
1053 	/*
1054 	 * It's possible that a tag was freed in the window between the
1055 	 * allocation failure and adding the hardware queue to the wait
1056 	 * queue.
1057 	 */
1058 	ret = blk_mq_get_driver_tag(rq, hctx, false);
1059 	if (!ret) {
1060 		spin_unlock(&this_hctx->lock);
1061 		return false;
1062 	}
1063 
1064 	/*
1065 	 * We got a tag, remove ourselves from the wait queue to ensure
1066 	 * someone else gets the wakeup.
1067 	 */
1068 	spin_lock_irq(&ws->wait.lock);
1069 	list_del_init(&wait->entry);
1070 	spin_unlock_irq(&ws->wait.lock);
1071 	spin_unlock(&this_hctx->lock);
1072 
1073 	return true;
1074 }
1075 
1076 #define BLK_MQ_RESOURCE_DELAY	3		/* ms units */
1077 
1078 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1079 			     bool got_budget)
1080 {
1081 	struct blk_mq_hw_ctx *hctx;
1082 	struct request *rq, *nxt;
1083 	bool no_tag = false;
1084 	int errors, queued;
1085 	blk_status_t ret = BLK_STS_OK;
1086 
1087 	if (list_empty(list))
1088 		return false;
1089 
1090 	WARN_ON(!list_is_singular(list) && got_budget);
1091 
1092 	/*
1093 	 * Now process all the entries, sending them to the driver.
1094 	 */
1095 	errors = queued = 0;
1096 	do {
1097 		struct blk_mq_queue_data bd;
1098 
1099 		rq = list_first_entry(list, struct request, queuelist);
1100 
1101 		hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
1102 		if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1103 			break;
1104 
1105 		if (!blk_mq_get_driver_tag(rq, NULL, false)) {
1106 			/*
1107 			 * The initial allocation attempt failed, so we need to
1108 			 * rerun the hardware queue when a tag is freed. The
1109 			 * waitqueue takes care of that. If the queue is run
1110 			 * before we add this entry back on the dispatch list,
1111 			 * we'll re-run it below.
1112 			 */
1113 			if (!blk_mq_mark_tag_wait(&hctx, rq)) {
1114 				blk_mq_put_dispatch_budget(hctx);
1115 				/*
1116 				 * For non-shared tags, the RESTART check
1117 				 * will suffice.
1118 				 */
1119 				if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1120 					no_tag = true;
1121 				break;
1122 			}
1123 		}
1124 
1125 		list_del_init(&rq->queuelist);
1126 
1127 		bd.rq = rq;
1128 
1129 		/*
1130 		 * Flag last if we have no more requests, or if we have more
1131 		 * but can't assign a driver tag to it.
1132 		 */
1133 		if (list_empty(list))
1134 			bd.last = true;
1135 		else {
1136 			nxt = list_first_entry(list, struct request, queuelist);
1137 			bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1138 		}
1139 
1140 		ret = q->mq_ops->queue_rq(hctx, &bd);
1141 		if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1142 			/*
1143 			 * If an I/O scheduler has been configured and we got a
1144 			 * driver tag for the next request already, free it
1145 			 * again.
1146 			 */
1147 			if (!list_empty(list)) {
1148 				nxt = list_first_entry(list, struct request, queuelist);
1149 				blk_mq_put_driver_tag(nxt);
1150 			}
1151 			list_add(&rq->queuelist, list);
1152 			__blk_mq_requeue_request(rq);
1153 			break;
1154 		}
1155 
1156 		if (unlikely(ret != BLK_STS_OK)) {
1157 			errors++;
1158 			blk_mq_end_request(rq, BLK_STS_IOERR);
1159 			continue;
1160 		}
1161 
1162 		queued++;
1163 	} while (!list_empty(list));
1164 
1165 	hctx->dispatched[queued_to_index(queued)]++;
1166 
1167 	/*
1168 	 * Any items that need requeuing? Stuff them into hctx->dispatch,
1169 	 * that is where we will continue on next queue run.
1170 	 */
1171 	if (!list_empty(list)) {
1172 		bool needs_restart;
1173 
1174 		spin_lock(&hctx->lock);
1175 		list_splice_init(list, &hctx->dispatch);
1176 		spin_unlock(&hctx->lock);
1177 
1178 		/*
1179 		 * If SCHED_RESTART was set by the caller of this function and
1180 		 * it is no longer set that means that it was cleared by another
1181 		 * thread and hence that a queue rerun is needed.
1182 		 *
1183 		 * If 'no_tag' is set, that means that we failed getting
1184 		 * a driver tag with an I/O scheduler attached. If our dispatch
1185 		 * waitqueue is no longer active, ensure that we run the queue
1186 		 * AFTER adding our entries back to the list.
1187 		 *
1188 		 * If no I/O scheduler has been configured it is possible that
1189 		 * the hardware queue got stopped and restarted before requests
1190 		 * were pushed back onto the dispatch list. Rerun the queue to
1191 		 * avoid starvation. Notes:
1192 		 * - blk_mq_run_hw_queue() checks whether or not a queue has
1193 		 *   been stopped before rerunning a queue.
1194 		 * - Some but not all block drivers stop a queue before
1195 		 *   returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1196 		 *   and dm-rq.
1197 		 *
1198 		 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1199 		 * bit is set, run queue after a delay to avoid IO stalls
1200 		 * that could otherwise occur if the queue is idle.
1201 		 */
1202 		needs_restart = blk_mq_sched_needs_restart(hctx);
1203 		if (!needs_restart ||
1204 		    (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1205 			blk_mq_run_hw_queue(hctx, true);
1206 		else if (needs_restart && (ret == BLK_STS_RESOURCE))
1207 			blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1208 	}
1209 
1210 	return (queued + errors) != 0;
1211 }
1212 
1213 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1214 {
1215 	int srcu_idx;
1216 
1217 	/*
1218 	 * We should be running this queue from one of the CPUs that
1219 	 * are mapped to it.
1220 	 *
1221 	 * There are at least two related races now between setting
1222 	 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1223 	 * __blk_mq_run_hw_queue():
1224 	 *
1225 	 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1226 	 *   but later it becomes online, then this warning is harmless
1227 	 *   at all
1228 	 *
1229 	 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1230 	 *   but later it becomes offline, then the warning can't be
1231 	 *   triggered, and we depend on blk-mq timeout handler to
1232 	 *   handle dispatched requests to this hctx
1233 	 */
1234 	if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1235 		cpu_online(hctx->next_cpu)) {
1236 		printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1237 			raw_smp_processor_id(),
1238 			cpumask_empty(hctx->cpumask) ? "inactive": "active");
1239 		dump_stack();
1240 	}
1241 
1242 	/*
1243 	 * We can't run the queue inline with ints disabled. Ensure that
1244 	 * we catch bad users of this early.
1245 	 */
1246 	WARN_ON_ONCE(in_interrupt());
1247 
1248 	might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1249 
1250 	hctx_lock(hctx, &srcu_idx);
1251 	blk_mq_sched_dispatch_requests(hctx);
1252 	hctx_unlock(hctx, srcu_idx);
1253 }
1254 
1255 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1256 {
1257 	int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1258 
1259 	if (cpu >= nr_cpu_ids)
1260 		cpu = cpumask_first(hctx->cpumask);
1261 	return cpu;
1262 }
1263 
1264 /*
1265  * It'd be great if the workqueue API had a way to pass
1266  * in a mask and had some smarts for more clever placement.
1267  * For now we just round-robin here, switching for every
1268  * BLK_MQ_CPU_WORK_BATCH queued items.
1269  */
1270 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1271 {
1272 	bool tried = false;
1273 	int next_cpu = hctx->next_cpu;
1274 
1275 	if (hctx->queue->nr_hw_queues == 1)
1276 		return WORK_CPU_UNBOUND;
1277 
1278 	if (--hctx->next_cpu_batch <= 0) {
1279 select_cpu:
1280 		next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1281 				cpu_online_mask);
1282 		if (next_cpu >= nr_cpu_ids)
1283 			next_cpu = blk_mq_first_mapped_cpu(hctx);
1284 		hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1285 	}
1286 
1287 	/*
1288 	 * Do unbound schedule if we can't find a online CPU for this hctx,
1289 	 * and it should only happen in the path of handling CPU DEAD.
1290 	 */
1291 	if (!cpu_online(next_cpu)) {
1292 		if (!tried) {
1293 			tried = true;
1294 			goto select_cpu;
1295 		}
1296 
1297 		/*
1298 		 * Make sure to re-select CPU next time once after CPUs
1299 		 * in hctx->cpumask become online again.
1300 		 */
1301 		hctx->next_cpu = next_cpu;
1302 		hctx->next_cpu_batch = 1;
1303 		return WORK_CPU_UNBOUND;
1304 	}
1305 
1306 	hctx->next_cpu = next_cpu;
1307 	return next_cpu;
1308 }
1309 
1310 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1311 					unsigned long msecs)
1312 {
1313 	if (unlikely(blk_mq_hctx_stopped(hctx)))
1314 		return;
1315 
1316 	if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1317 		int cpu = get_cpu();
1318 		if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1319 			__blk_mq_run_hw_queue(hctx);
1320 			put_cpu();
1321 			return;
1322 		}
1323 
1324 		put_cpu();
1325 	}
1326 
1327 	kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1328 				    msecs_to_jiffies(msecs));
1329 }
1330 
1331 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1332 {
1333 	__blk_mq_delay_run_hw_queue(hctx, true, msecs);
1334 }
1335 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1336 
1337 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1338 {
1339 	int srcu_idx;
1340 	bool need_run;
1341 
1342 	/*
1343 	 * When queue is quiesced, we may be switching io scheduler, or
1344 	 * updating nr_hw_queues, or other things, and we can't run queue
1345 	 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1346 	 *
1347 	 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1348 	 * quiesced.
1349 	 */
1350 	hctx_lock(hctx, &srcu_idx);
1351 	need_run = !blk_queue_quiesced(hctx->queue) &&
1352 		blk_mq_hctx_has_pending(hctx);
1353 	hctx_unlock(hctx, srcu_idx);
1354 
1355 	if (need_run) {
1356 		__blk_mq_delay_run_hw_queue(hctx, async, 0);
1357 		return true;
1358 	}
1359 
1360 	return false;
1361 }
1362 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1363 
1364 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1365 {
1366 	struct blk_mq_hw_ctx *hctx;
1367 	int i;
1368 
1369 	queue_for_each_hw_ctx(q, hctx, i) {
1370 		if (blk_mq_hctx_stopped(hctx))
1371 			continue;
1372 
1373 		blk_mq_run_hw_queue(hctx, async);
1374 	}
1375 }
1376 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1377 
1378 /**
1379  * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1380  * @q: request queue.
1381  *
1382  * The caller is responsible for serializing this function against
1383  * blk_mq_{start,stop}_hw_queue().
1384  */
1385 bool blk_mq_queue_stopped(struct request_queue *q)
1386 {
1387 	struct blk_mq_hw_ctx *hctx;
1388 	int i;
1389 
1390 	queue_for_each_hw_ctx(q, hctx, i)
1391 		if (blk_mq_hctx_stopped(hctx))
1392 			return true;
1393 
1394 	return false;
1395 }
1396 EXPORT_SYMBOL(blk_mq_queue_stopped);
1397 
1398 /*
1399  * This function is often used for pausing .queue_rq() by driver when
1400  * there isn't enough resource or some conditions aren't satisfied, and
1401  * BLK_STS_RESOURCE is usually returned.
1402  *
1403  * We do not guarantee that dispatch can be drained or blocked
1404  * after blk_mq_stop_hw_queue() returns. Please use
1405  * blk_mq_quiesce_queue() for that requirement.
1406  */
1407 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1408 {
1409 	cancel_delayed_work(&hctx->run_work);
1410 
1411 	set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1412 }
1413 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1414 
1415 /*
1416  * This function is often used for pausing .queue_rq() by driver when
1417  * there isn't enough resource or some conditions aren't satisfied, and
1418  * BLK_STS_RESOURCE is usually returned.
1419  *
1420  * We do not guarantee that dispatch can be drained or blocked
1421  * after blk_mq_stop_hw_queues() returns. Please use
1422  * blk_mq_quiesce_queue() for that requirement.
1423  */
1424 void blk_mq_stop_hw_queues(struct request_queue *q)
1425 {
1426 	struct blk_mq_hw_ctx *hctx;
1427 	int i;
1428 
1429 	queue_for_each_hw_ctx(q, hctx, i)
1430 		blk_mq_stop_hw_queue(hctx);
1431 }
1432 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1433 
1434 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1435 {
1436 	clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1437 
1438 	blk_mq_run_hw_queue(hctx, false);
1439 }
1440 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1441 
1442 void blk_mq_start_hw_queues(struct request_queue *q)
1443 {
1444 	struct blk_mq_hw_ctx *hctx;
1445 	int i;
1446 
1447 	queue_for_each_hw_ctx(q, hctx, i)
1448 		blk_mq_start_hw_queue(hctx);
1449 }
1450 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1451 
1452 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1453 {
1454 	if (!blk_mq_hctx_stopped(hctx))
1455 		return;
1456 
1457 	clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1458 	blk_mq_run_hw_queue(hctx, async);
1459 }
1460 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1461 
1462 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1463 {
1464 	struct blk_mq_hw_ctx *hctx;
1465 	int i;
1466 
1467 	queue_for_each_hw_ctx(q, hctx, i)
1468 		blk_mq_start_stopped_hw_queue(hctx, async);
1469 }
1470 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1471 
1472 static void blk_mq_run_work_fn(struct work_struct *work)
1473 {
1474 	struct blk_mq_hw_ctx *hctx;
1475 
1476 	hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1477 
1478 	/*
1479 	 * If we are stopped, don't run the queue.
1480 	 */
1481 	if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1482 		return;
1483 
1484 	__blk_mq_run_hw_queue(hctx);
1485 }
1486 
1487 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1488 					    struct request *rq,
1489 					    bool at_head)
1490 {
1491 	struct blk_mq_ctx *ctx = rq->mq_ctx;
1492 
1493 	lockdep_assert_held(&ctx->lock);
1494 
1495 	trace_block_rq_insert(hctx->queue, rq);
1496 
1497 	if (at_head)
1498 		list_add(&rq->queuelist, &ctx->rq_list);
1499 	else
1500 		list_add_tail(&rq->queuelist, &ctx->rq_list);
1501 }
1502 
1503 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1504 			     bool at_head)
1505 {
1506 	struct blk_mq_ctx *ctx = rq->mq_ctx;
1507 
1508 	lockdep_assert_held(&ctx->lock);
1509 
1510 	__blk_mq_insert_req_list(hctx, rq, at_head);
1511 	blk_mq_hctx_mark_pending(hctx, ctx);
1512 }
1513 
1514 /*
1515  * Should only be used carefully, when the caller knows we want to
1516  * bypass a potential IO scheduler on the target device.
1517  */
1518 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1519 {
1520 	struct blk_mq_ctx *ctx = rq->mq_ctx;
1521 	struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1522 
1523 	spin_lock(&hctx->lock);
1524 	list_add_tail(&rq->queuelist, &hctx->dispatch);
1525 	spin_unlock(&hctx->lock);
1526 
1527 	if (run_queue)
1528 		blk_mq_run_hw_queue(hctx, false);
1529 }
1530 
1531 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1532 			    struct list_head *list)
1533 
1534 {
1535 	/*
1536 	 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1537 	 * offline now
1538 	 */
1539 	spin_lock(&ctx->lock);
1540 	while (!list_empty(list)) {
1541 		struct request *rq;
1542 
1543 		rq = list_first_entry(list, struct request, queuelist);
1544 		BUG_ON(rq->mq_ctx != ctx);
1545 		list_del_init(&rq->queuelist);
1546 		__blk_mq_insert_req_list(hctx, rq, false);
1547 	}
1548 	blk_mq_hctx_mark_pending(hctx, ctx);
1549 	spin_unlock(&ctx->lock);
1550 }
1551 
1552 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1553 {
1554 	struct request *rqa = container_of(a, struct request, queuelist);
1555 	struct request *rqb = container_of(b, struct request, queuelist);
1556 
1557 	return !(rqa->mq_ctx < rqb->mq_ctx ||
1558 		 (rqa->mq_ctx == rqb->mq_ctx &&
1559 		  blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1560 }
1561 
1562 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1563 {
1564 	struct blk_mq_ctx *this_ctx;
1565 	struct request_queue *this_q;
1566 	struct request *rq;
1567 	LIST_HEAD(list);
1568 	LIST_HEAD(ctx_list);
1569 	unsigned int depth;
1570 
1571 	list_splice_init(&plug->mq_list, &list);
1572 
1573 	list_sort(NULL, &list, plug_ctx_cmp);
1574 
1575 	this_q = NULL;
1576 	this_ctx = NULL;
1577 	depth = 0;
1578 
1579 	while (!list_empty(&list)) {
1580 		rq = list_entry_rq(list.next);
1581 		list_del_init(&rq->queuelist);
1582 		BUG_ON(!rq->q);
1583 		if (rq->mq_ctx != this_ctx) {
1584 			if (this_ctx) {
1585 				trace_block_unplug(this_q, depth, from_schedule);
1586 				blk_mq_sched_insert_requests(this_q, this_ctx,
1587 								&ctx_list,
1588 								from_schedule);
1589 			}
1590 
1591 			this_ctx = rq->mq_ctx;
1592 			this_q = rq->q;
1593 			depth = 0;
1594 		}
1595 
1596 		depth++;
1597 		list_add_tail(&rq->queuelist, &ctx_list);
1598 	}
1599 
1600 	/*
1601 	 * If 'this_ctx' is set, we know we have entries to complete
1602 	 * on 'ctx_list'. Do those.
1603 	 */
1604 	if (this_ctx) {
1605 		trace_block_unplug(this_q, depth, from_schedule);
1606 		blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1607 						from_schedule);
1608 	}
1609 }
1610 
1611 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1612 {
1613 	blk_init_request_from_bio(rq, bio);
1614 
1615 	blk_rq_set_rl(rq, blk_get_rl(rq->q, bio));
1616 
1617 	blk_account_io_start(rq, true);
1618 }
1619 
1620 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1621 {
1622 	if (rq->tag != -1)
1623 		return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1624 
1625 	return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1626 }
1627 
1628 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1629 					    struct request *rq,
1630 					    blk_qc_t *cookie)
1631 {
1632 	struct request_queue *q = rq->q;
1633 	struct blk_mq_queue_data bd = {
1634 		.rq = rq,
1635 		.last = true,
1636 	};
1637 	blk_qc_t new_cookie;
1638 	blk_status_t ret;
1639 
1640 	new_cookie = request_to_qc_t(hctx, rq);
1641 
1642 	/*
1643 	 * For OK queue, we are done. For error, caller may kill it.
1644 	 * Any other error (busy), just add it to our list as we
1645 	 * previously would have done.
1646 	 */
1647 	ret = q->mq_ops->queue_rq(hctx, &bd);
1648 	switch (ret) {
1649 	case BLK_STS_OK:
1650 		*cookie = new_cookie;
1651 		break;
1652 	case BLK_STS_RESOURCE:
1653 	case BLK_STS_DEV_RESOURCE:
1654 		__blk_mq_requeue_request(rq);
1655 		break;
1656 	default:
1657 		*cookie = BLK_QC_T_NONE;
1658 		break;
1659 	}
1660 
1661 	return ret;
1662 }
1663 
1664 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1665 						struct request *rq,
1666 						blk_qc_t *cookie,
1667 						bool bypass_insert)
1668 {
1669 	struct request_queue *q = rq->q;
1670 	bool run_queue = true;
1671 
1672 	/*
1673 	 * RCU or SRCU read lock is needed before checking quiesced flag.
1674 	 *
1675 	 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1676 	 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1677 	 * and avoid driver to try to dispatch again.
1678 	 */
1679 	if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1680 		run_queue = false;
1681 		bypass_insert = false;
1682 		goto insert;
1683 	}
1684 
1685 	if (q->elevator && !bypass_insert)
1686 		goto insert;
1687 
1688 	if (!blk_mq_get_dispatch_budget(hctx))
1689 		goto insert;
1690 
1691 	if (!blk_mq_get_driver_tag(rq, NULL, false)) {
1692 		blk_mq_put_dispatch_budget(hctx);
1693 		goto insert;
1694 	}
1695 
1696 	return __blk_mq_issue_directly(hctx, rq, cookie);
1697 insert:
1698 	if (bypass_insert)
1699 		return BLK_STS_RESOURCE;
1700 
1701 	blk_mq_sched_insert_request(rq, false, run_queue, false);
1702 	return BLK_STS_OK;
1703 }
1704 
1705 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1706 		struct request *rq, blk_qc_t *cookie)
1707 {
1708 	blk_status_t ret;
1709 	int srcu_idx;
1710 
1711 	might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1712 
1713 	hctx_lock(hctx, &srcu_idx);
1714 
1715 	ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1716 	if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1717 		blk_mq_sched_insert_request(rq, false, true, false);
1718 	else if (ret != BLK_STS_OK)
1719 		blk_mq_end_request(rq, ret);
1720 
1721 	hctx_unlock(hctx, srcu_idx);
1722 }
1723 
1724 blk_status_t blk_mq_request_issue_directly(struct request *rq)
1725 {
1726 	blk_status_t ret;
1727 	int srcu_idx;
1728 	blk_qc_t unused_cookie;
1729 	struct blk_mq_ctx *ctx = rq->mq_ctx;
1730 	struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1731 
1732 	hctx_lock(hctx, &srcu_idx);
1733 	ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true);
1734 	hctx_unlock(hctx, srcu_idx);
1735 
1736 	return ret;
1737 }
1738 
1739 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1740 {
1741 	const int is_sync = op_is_sync(bio->bi_opf);
1742 	const int is_flush_fua = op_is_flush(bio->bi_opf);
1743 	struct blk_mq_alloc_data data = { .flags = 0 };
1744 	struct request *rq;
1745 	unsigned int request_count = 0;
1746 	struct blk_plug *plug;
1747 	struct request *same_queue_rq = NULL;
1748 	blk_qc_t cookie;
1749 	unsigned int wb_acct;
1750 
1751 	blk_queue_bounce(q, &bio);
1752 
1753 	blk_queue_split(q, &bio);
1754 
1755 	if (!bio_integrity_prep(bio))
1756 		return BLK_QC_T_NONE;
1757 
1758 	if (!is_flush_fua && !blk_queue_nomerges(q) &&
1759 	    blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1760 		return BLK_QC_T_NONE;
1761 
1762 	if (blk_mq_sched_bio_merge(q, bio))
1763 		return BLK_QC_T_NONE;
1764 
1765 	wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1766 
1767 	trace_block_getrq(q, bio, bio->bi_opf);
1768 
1769 	rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1770 	if (unlikely(!rq)) {
1771 		__wbt_done(q->rq_wb, wb_acct);
1772 		if (bio->bi_opf & REQ_NOWAIT)
1773 			bio_wouldblock_error(bio);
1774 		return BLK_QC_T_NONE;
1775 	}
1776 
1777 	wbt_track(rq, wb_acct);
1778 
1779 	cookie = request_to_qc_t(data.hctx, rq);
1780 
1781 	plug = current->plug;
1782 	if (unlikely(is_flush_fua)) {
1783 		blk_mq_put_ctx(data.ctx);
1784 		blk_mq_bio_to_request(rq, bio);
1785 
1786 		/* bypass scheduler for flush rq */
1787 		blk_insert_flush(rq);
1788 		blk_mq_run_hw_queue(data.hctx, true);
1789 	} else if (plug && q->nr_hw_queues == 1) {
1790 		struct request *last = NULL;
1791 
1792 		blk_mq_put_ctx(data.ctx);
1793 		blk_mq_bio_to_request(rq, bio);
1794 
1795 		/*
1796 		 * @request_count may become stale because of schedule
1797 		 * out, so check the list again.
1798 		 */
1799 		if (list_empty(&plug->mq_list))
1800 			request_count = 0;
1801 		else if (blk_queue_nomerges(q))
1802 			request_count = blk_plug_queued_count(q);
1803 
1804 		if (!request_count)
1805 			trace_block_plug(q);
1806 		else
1807 			last = list_entry_rq(plug->mq_list.prev);
1808 
1809 		if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1810 		    blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1811 			blk_flush_plug_list(plug, false);
1812 			trace_block_plug(q);
1813 		}
1814 
1815 		list_add_tail(&rq->queuelist, &plug->mq_list);
1816 	} else if (plug && !blk_queue_nomerges(q)) {
1817 		blk_mq_bio_to_request(rq, bio);
1818 
1819 		/*
1820 		 * We do limited plugging. If the bio can be merged, do that.
1821 		 * Otherwise the existing request in the plug list will be
1822 		 * issued. So the plug list will have one request at most
1823 		 * The plug list might get flushed before this. If that happens,
1824 		 * the plug list is empty, and same_queue_rq is invalid.
1825 		 */
1826 		if (list_empty(&plug->mq_list))
1827 			same_queue_rq = NULL;
1828 		if (same_queue_rq)
1829 			list_del_init(&same_queue_rq->queuelist);
1830 		list_add_tail(&rq->queuelist, &plug->mq_list);
1831 
1832 		blk_mq_put_ctx(data.ctx);
1833 
1834 		if (same_queue_rq) {
1835 			data.hctx = blk_mq_map_queue(q,
1836 					same_queue_rq->mq_ctx->cpu);
1837 			blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1838 					&cookie);
1839 		}
1840 	} else if (q->nr_hw_queues > 1 && is_sync) {
1841 		blk_mq_put_ctx(data.ctx);
1842 		blk_mq_bio_to_request(rq, bio);
1843 		blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1844 	} else {
1845 		blk_mq_put_ctx(data.ctx);
1846 		blk_mq_bio_to_request(rq, bio);
1847 		blk_mq_sched_insert_request(rq, false, true, true);
1848 	}
1849 
1850 	return cookie;
1851 }
1852 
1853 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1854 		     unsigned int hctx_idx)
1855 {
1856 	struct page *page;
1857 
1858 	if (tags->rqs && set->ops->exit_request) {
1859 		int i;
1860 
1861 		for (i = 0; i < tags->nr_tags; i++) {
1862 			struct request *rq = tags->static_rqs[i];
1863 
1864 			if (!rq)
1865 				continue;
1866 			set->ops->exit_request(set, rq, hctx_idx);
1867 			tags->static_rqs[i] = NULL;
1868 		}
1869 	}
1870 
1871 	while (!list_empty(&tags->page_list)) {
1872 		page = list_first_entry(&tags->page_list, struct page, lru);
1873 		list_del_init(&page->lru);
1874 		/*
1875 		 * Remove kmemleak object previously allocated in
1876 		 * blk_mq_init_rq_map().
1877 		 */
1878 		kmemleak_free(page_address(page));
1879 		__free_pages(page, page->private);
1880 	}
1881 }
1882 
1883 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1884 {
1885 	kfree(tags->rqs);
1886 	tags->rqs = NULL;
1887 	kfree(tags->static_rqs);
1888 	tags->static_rqs = NULL;
1889 
1890 	blk_mq_free_tags(tags);
1891 }
1892 
1893 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1894 					unsigned int hctx_idx,
1895 					unsigned int nr_tags,
1896 					unsigned int reserved_tags)
1897 {
1898 	struct blk_mq_tags *tags;
1899 	int node;
1900 
1901 	node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1902 	if (node == NUMA_NO_NODE)
1903 		node = set->numa_node;
1904 
1905 	tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1906 				BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1907 	if (!tags)
1908 		return NULL;
1909 
1910 	tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
1911 				 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1912 				 node);
1913 	if (!tags->rqs) {
1914 		blk_mq_free_tags(tags);
1915 		return NULL;
1916 	}
1917 
1918 	tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
1919 					GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1920 					node);
1921 	if (!tags->static_rqs) {
1922 		kfree(tags->rqs);
1923 		blk_mq_free_tags(tags);
1924 		return NULL;
1925 	}
1926 
1927 	return tags;
1928 }
1929 
1930 static size_t order_to_size(unsigned int order)
1931 {
1932 	return (size_t)PAGE_SIZE << order;
1933 }
1934 
1935 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
1936 			       unsigned int hctx_idx, int node)
1937 {
1938 	int ret;
1939 
1940 	if (set->ops->init_request) {
1941 		ret = set->ops->init_request(set, rq, hctx_idx, node);
1942 		if (ret)
1943 			return ret;
1944 	}
1945 
1946 	WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1947 	return 0;
1948 }
1949 
1950 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1951 		     unsigned int hctx_idx, unsigned int depth)
1952 {
1953 	unsigned int i, j, entries_per_page, max_order = 4;
1954 	size_t rq_size, left;
1955 	int node;
1956 
1957 	node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1958 	if (node == NUMA_NO_NODE)
1959 		node = set->numa_node;
1960 
1961 	INIT_LIST_HEAD(&tags->page_list);
1962 
1963 	/*
1964 	 * rq_size is the size of the request plus driver payload, rounded
1965 	 * to the cacheline size
1966 	 */
1967 	rq_size = round_up(sizeof(struct request) + set->cmd_size,
1968 				cache_line_size());
1969 	left = rq_size * depth;
1970 
1971 	for (i = 0; i < depth; ) {
1972 		int this_order = max_order;
1973 		struct page *page;
1974 		int to_do;
1975 		void *p;
1976 
1977 		while (this_order && left < order_to_size(this_order - 1))
1978 			this_order--;
1979 
1980 		do {
1981 			page = alloc_pages_node(node,
1982 				GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1983 				this_order);
1984 			if (page)
1985 				break;
1986 			if (!this_order--)
1987 				break;
1988 			if (order_to_size(this_order) < rq_size)
1989 				break;
1990 		} while (1);
1991 
1992 		if (!page)
1993 			goto fail;
1994 
1995 		page->private = this_order;
1996 		list_add_tail(&page->lru, &tags->page_list);
1997 
1998 		p = page_address(page);
1999 		/*
2000 		 * Allow kmemleak to scan these pages as they contain pointers
2001 		 * to additional allocations like via ops->init_request().
2002 		 */
2003 		kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2004 		entries_per_page = order_to_size(this_order) / rq_size;
2005 		to_do = min(entries_per_page, depth - i);
2006 		left -= to_do * rq_size;
2007 		for (j = 0; j < to_do; j++) {
2008 			struct request *rq = p;
2009 
2010 			tags->static_rqs[i] = rq;
2011 			if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2012 				tags->static_rqs[i] = NULL;
2013 				goto fail;
2014 			}
2015 
2016 			p += rq_size;
2017 			i++;
2018 		}
2019 	}
2020 	return 0;
2021 
2022 fail:
2023 	blk_mq_free_rqs(set, tags, hctx_idx);
2024 	return -ENOMEM;
2025 }
2026 
2027 /*
2028  * 'cpu' is going away. splice any existing rq_list entries from this
2029  * software queue to the hw queue dispatch list, and ensure that it
2030  * gets run.
2031  */
2032 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2033 {
2034 	struct blk_mq_hw_ctx *hctx;
2035 	struct blk_mq_ctx *ctx;
2036 	LIST_HEAD(tmp);
2037 
2038 	hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2039 	ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2040 
2041 	spin_lock(&ctx->lock);
2042 	if (!list_empty(&ctx->rq_list)) {
2043 		list_splice_init(&ctx->rq_list, &tmp);
2044 		blk_mq_hctx_clear_pending(hctx, ctx);
2045 	}
2046 	spin_unlock(&ctx->lock);
2047 
2048 	if (list_empty(&tmp))
2049 		return 0;
2050 
2051 	spin_lock(&hctx->lock);
2052 	list_splice_tail_init(&tmp, &hctx->dispatch);
2053 	spin_unlock(&hctx->lock);
2054 
2055 	blk_mq_run_hw_queue(hctx, true);
2056 	return 0;
2057 }
2058 
2059 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2060 {
2061 	cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2062 					    &hctx->cpuhp_dead);
2063 }
2064 
2065 /* hctx->ctxs will be freed in queue's release handler */
2066 static void blk_mq_exit_hctx(struct request_queue *q,
2067 		struct blk_mq_tag_set *set,
2068 		struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2069 {
2070 	blk_mq_debugfs_unregister_hctx(hctx);
2071 
2072 	if (blk_mq_hw_queue_mapped(hctx))
2073 		blk_mq_tag_idle(hctx);
2074 
2075 	if (set->ops->exit_request)
2076 		set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2077 
2078 	blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2079 
2080 	if (set->ops->exit_hctx)
2081 		set->ops->exit_hctx(hctx, hctx_idx);
2082 
2083 	if (hctx->flags & BLK_MQ_F_BLOCKING)
2084 		cleanup_srcu_struct(hctx->srcu);
2085 
2086 	blk_mq_remove_cpuhp(hctx);
2087 	blk_free_flush_queue(hctx->fq);
2088 	sbitmap_free(&hctx->ctx_map);
2089 }
2090 
2091 static void blk_mq_exit_hw_queues(struct request_queue *q,
2092 		struct blk_mq_tag_set *set, int nr_queue)
2093 {
2094 	struct blk_mq_hw_ctx *hctx;
2095 	unsigned int i;
2096 
2097 	queue_for_each_hw_ctx(q, hctx, i) {
2098 		if (i == nr_queue)
2099 			break;
2100 		blk_mq_exit_hctx(q, set, hctx, i);
2101 	}
2102 }
2103 
2104 static int blk_mq_init_hctx(struct request_queue *q,
2105 		struct blk_mq_tag_set *set,
2106 		struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2107 {
2108 	int node;
2109 
2110 	node = hctx->numa_node;
2111 	if (node == NUMA_NO_NODE)
2112 		node = hctx->numa_node = set->numa_node;
2113 
2114 	INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2115 	spin_lock_init(&hctx->lock);
2116 	INIT_LIST_HEAD(&hctx->dispatch);
2117 	hctx->queue = q;
2118 	hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2119 
2120 	cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2121 
2122 	hctx->tags = set->tags[hctx_idx];
2123 
2124 	/*
2125 	 * Allocate space for all possible cpus to avoid allocation at
2126 	 * runtime
2127 	 */
2128 	hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2129 					GFP_KERNEL, node);
2130 	if (!hctx->ctxs)
2131 		goto unregister_cpu_notifier;
2132 
2133 	if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
2134 			      node))
2135 		goto free_ctxs;
2136 
2137 	hctx->nr_ctx = 0;
2138 
2139 	init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2140 	INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2141 
2142 	if (set->ops->init_hctx &&
2143 	    set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2144 		goto free_bitmap;
2145 
2146 	if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
2147 		goto exit_hctx;
2148 
2149 	hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
2150 	if (!hctx->fq)
2151 		goto sched_exit_hctx;
2152 
2153 	if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node))
2154 		goto free_fq;
2155 
2156 	if (hctx->flags & BLK_MQ_F_BLOCKING)
2157 		init_srcu_struct(hctx->srcu);
2158 
2159 	blk_mq_debugfs_register_hctx(q, hctx);
2160 
2161 	return 0;
2162 
2163  free_fq:
2164 	kfree(hctx->fq);
2165  sched_exit_hctx:
2166 	blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2167  exit_hctx:
2168 	if (set->ops->exit_hctx)
2169 		set->ops->exit_hctx(hctx, hctx_idx);
2170  free_bitmap:
2171 	sbitmap_free(&hctx->ctx_map);
2172  free_ctxs:
2173 	kfree(hctx->ctxs);
2174  unregister_cpu_notifier:
2175 	blk_mq_remove_cpuhp(hctx);
2176 	return -1;
2177 }
2178 
2179 static void blk_mq_init_cpu_queues(struct request_queue *q,
2180 				   unsigned int nr_hw_queues)
2181 {
2182 	unsigned int i;
2183 
2184 	for_each_possible_cpu(i) {
2185 		struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2186 		struct blk_mq_hw_ctx *hctx;
2187 
2188 		__ctx->cpu = i;
2189 		spin_lock_init(&__ctx->lock);
2190 		INIT_LIST_HEAD(&__ctx->rq_list);
2191 		__ctx->queue = q;
2192 
2193 		/*
2194 		 * Set local node, IFF we have more than one hw queue. If
2195 		 * not, we remain on the home node of the device
2196 		 */
2197 		hctx = blk_mq_map_queue(q, i);
2198 		if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2199 			hctx->numa_node = local_memory_node(cpu_to_node(i));
2200 	}
2201 }
2202 
2203 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2204 {
2205 	int ret = 0;
2206 
2207 	set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2208 					set->queue_depth, set->reserved_tags);
2209 	if (!set->tags[hctx_idx])
2210 		return false;
2211 
2212 	ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2213 				set->queue_depth);
2214 	if (!ret)
2215 		return true;
2216 
2217 	blk_mq_free_rq_map(set->tags[hctx_idx]);
2218 	set->tags[hctx_idx] = NULL;
2219 	return false;
2220 }
2221 
2222 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2223 					 unsigned int hctx_idx)
2224 {
2225 	if (set->tags[hctx_idx]) {
2226 		blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2227 		blk_mq_free_rq_map(set->tags[hctx_idx]);
2228 		set->tags[hctx_idx] = NULL;
2229 	}
2230 }
2231 
2232 static void blk_mq_map_swqueue(struct request_queue *q)
2233 {
2234 	unsigned int i, hctx_idx;
2235 	struct blk_mq_hw_ctx *hctx;
2236 	struct blk_mq_ctx *ctx;
2237 	struct blk_mq_tag_set *set = q->tag_set;
2238 
2239 	/*
2240 	 * Avoid others reading imcomplete hctx->cpumask through sysfs
2241 	 */
2242 	mutex_lock(&q->sysfs_lock);
2243 
2244 	queue_for_each_hw_ctx(q, hctx, i) {
2245 		cpumask_clear(hctx->cpumask);
2246 		hctx->nr_ctx = 0;
2247 		hctx->dispatch_from = NULL;
2248 	}
2249 
2250 	/*
2251 	 * Map software to hardware queues.
2252 	 *
2253 	 * If the cpu isn't present, the cpu is mapped to first hctx.
2254 	 */
2255 	for_each_possible_cpu(i) {
2256 		hctx_idx = q->mq_map[i];
2257 		/* unmapped hw queue can be remapped after CPU topo changed */
2258 		if (!set->tags[hctx_idx] &&
2259 		    !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2260 			/*
2261 			 * If tags initialization fail for some hctx,
2262 			 * that hctx won't be brought online.  In this
2263 			 * case, remap the current ctx to hctx[0] which
2264 			 * is guaranteed to always have tags allocated
2265 			 */
2266 			q->mq_map[i] = 0;
2267 		}
2268 
2269 		ctx = per_cpu_ptr(q->queue_ctx, i);
2270 		hctx = blk_mq_map_queue(q, i);
2271 
2272 		cpumask_set_cpu(i, hctx->cpumask);
2273 		ctx->index_hw = hctx->nr_ctx;
2274 		hctx->ctxs[hctx->nr_ctx++] = ctx;
2275 	}
2276 
2277 	mutex_unlock(&q->sysfs_lock);
2278 
2279 	queue_for_each_hw_ctx(q, hctx, i) {
2280 		/*
2281 		 * If no software queues are mapped to this hardware queue,
2282 		 * disable it and free the request entries.
2283 		 */
2284 		if (!hctx->nr_ctx) {
2285 			/* Never unmap queue 0.  We need it as a
2286 			 * fallback in case of a new remap fails
2287 			 * allocation
2288 			 */
2289 			if (i && set->tags[i])
2290 				blk_mq_free_map_and_requests(set, i);
2291 
2292 			hctx->tags = NULL;
2293 			continue;
2294 		}
2295 
2296 		hctx->tags = set->tags[i];
2297 		WARN_ON(!hctx->tags);
2298 
2299 		/*
2300 		 * Set the map size to the number of mapped software queues.
2301 		 * This is more accurate and more efficient than looping
2302 		 * over all possibly mapped software queues.
2303 		 */
2304 		sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2305 
2306 		/*
2307 		 * Initialize batch roundrobin counts
2308 		 */
2309 		hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2310 		hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2311 	}
2312 }
2313 
2314 /*
2315  * Caller needs to ensure that we're either frozen/quiesced, or that
2316  * the queue isn't live yet.
2317  */
2318 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2319 {
2320 	struct blk_mq_hw_ctx *hctx;
2321 	int i;
2322 
2323 	queue_for_each_hw_ctx(q, hctx, i) {
2324 		if (shared) {
2325 			if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2326 				atomic_inc(&q->shared_hctx_restart);
2327 			hctx->flags |= BLK_MQ_F_TAG_SHARED;
2328 		} else {
2329 			if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2330 				atomic_dec(&q->shared_hctx_restart);
2331 			hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2332 		}
2333 	}
2334 }
2335 
2336 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2337 					bool shared)
2338 {
2339 	struct request_queue *q;
2340 
2341 	lockdep_assert_held(&set->tag_list_lock);
2342 
2343 	list_for_each_entry(q, &set->tag_list, tag_set_list) {
2344 		blk_mq_freeze_queue(q);
2345 		queue_set_hctx_shared(q, shared);
2346 		blk_mq_unfreeze_queue(q);
2347 	}
2348 }
2349 
2350 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2351 {
2352 	struct blk_mq_tag_set *set = q->tag_set;
2353 
2354 	mutex_lock(&set->tag_list_lock);
2355 	list_del_rcu(&q->tag_set_list);
2356 	if (list_is_singular(&set->tag_list)) {
2357 		/* just transitioned to unshared */
2358 		set->flags &= ~BLK_MQ_F_TAG_SHARED;
2359 		/* update existing queue */
2360 		blk_mq_update_tag_set_depth(set, false);
2361 	}
2362 	mutex_unlock(&set->tag_list_lock);
2363 	synchronize_rcu();
2364 	INIT_LIST_HEAD(&q->tag_set_list);
2365 }
2366 
2367 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2368 				     struct request_queue *q)
2369 {
2370 	q->tag_set = set;
2371 
2372 	mutex_lock(&set->tag_list_lock);
2373 
2374 	/*
2375 	 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2376 	 */
2377 	if (!list_empty(&set->tag_list) &&
2378 	    !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2379 		set->flags |= BLK_MQ_F_TAG_SHARED;
2380 		/* update existing queue */
2381 		blk_mq_update_tag_set_depth(set, true);
2382 	}
2383 	if (set->flags & BLK_MQ_F_TAG_SHARED)
2384 		queue_set_hctx_shared(q, true);
2385 	list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2386 
2387 	mutex_unlock(&set->tag_list_lock);
2388 }
2389 
2390 /*
2391  * It is the actual release handler for mq, but we do it from
2392  * request queue's release handler for avoiding use-after-free
2393  * and headache because q->mq_kobj shouldn't have been introduced,
2394  * but we can't group ctx/kctx kobj without it.
2395  */
2396 void blk_mq_release(struct request_queue *q)
2397 {
2398 	struct blk_mq_hw_ctx *hctx;
2399 	unsigned int i;
2400 
2401 	/* hctx kobj stays in hctx */
2402 	queue_for_each_hw_ctx(q, hctx, i) {
2403 		if (!hctx)
2404 			continue;
2405 		kobject_put(&hctx->kobj);
2406 	}
2407 
2408 	q->mq_map = NULL;
2409 
2410 	kfree(q->queue_hw_ctx);
2411 
2412 	/*
2413 	 * release .mq_kobj and sw queue's kobject now because
2414 	 * both share lifetime with request queue.
2415 	 */
2416 	blk_mq_sysfs_deinit(q);
2417 
2418 	free_percpu(q->queue_ctx);
2419 }
2420 
2421 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2422 {
2423 	struct request_queue *uninit_q, *q;
2424 
2425 	uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node, NULL);
2426 	if (!uninit_q)
2427 		return ERR_PTR(-ENOMEM);
2428 
2429 	q = blk_mq_init_allocated_queue(set, uninit_q);
2430 	if (IS_ERR(q))
2431 		blk_cleanup_queue(uninit_q);
2432 
2433 	return q;
2434 }
2435 EXPORT_SYMBOL(blk_mq_init_queue);
2436 
2437 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2438 {
2439 	int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2440 
2441 	BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2442 			   __alignof__(struct blk_mq_hw_ctx)) !=
2443 		     sizeof(struct blk_mq_hw_ctx));
2444 
2445 	if (tag_set->flags & BLK_MQ_F_BLOCKING)
2446 		hw_ctx_size += sizeof(struct srcu_struct);
2447 
2448 	return hw_ctx_size;
2449 }
2450 
2451 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2452 						struct request_queue *q)
2453 {
2454 	int i, j;
2455 	struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2456 
2457 	blk_mq_sysfs_unregister(q);
2458 
2459 	/* protect against switching io scheduler  */
2460 	mutex_lock(&q->sysfs_lock);
2461 	for (i = 0; i < set->nr_hw_queues; i++) {
2462 		int node;
2463 
2464 		if (hctxs[i])
2465 			continue;
2466 
2467 		node = blk_mq_hw_queue_to_node(q->mq_map, i);
2468 		hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set),
2469 					GFP_KERNEL, node);
2470 		if (!hctxs[i])
2471 			break;
2472 
2473 		if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2474 						node)) {
2475 			kfree(hctxs[i]);
2476 			hctxs[i] = NULL;
2477 			break;
2478 		}
2479 
2480 		atomic_set(&hctxs[i]->nr_active, 0);
2481 		hctxs[i]->numa_node = node;
2482 		hctxs[i]->queue_num = i;
2483 
2484 		if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2485 			free_cpumask_var(hctxs[i]->cpumask);
2486 			kfree(hctxs[i]);
2487 			hctxs[i] = NULL;
2488 			break;
2489 		}
2490 		blk_mq_hctx_kobj_init(hctxs[i]);
2491 	}
2492 	for (j = i; j < q->nr_hw_queues; j++) {
2493 		struct blk_mq_hw_ctx *hctx = hctxs[j];
2494 
2495 		if (hctx) {
2496 			if (hctx->tags)
2497 				blk_mq_free_map_and_requests(set, j);
2498 			blk_mq_exit_hctx(q, set, hctx, j);
2499 			kobject_put(&hctx->kobj);
2500 			hctxs[j] = NULL;
2501 
2502 		}
2503 	}
2504 	q->nr_hw_queues = i;
2505 	mutex_unlock(&q->sysfs_lock);
2506 	blk_mq_sysfs_register(q);
2507 }
2508 
2509 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2510 						  struct request_queue *q)
2511 {
2512 	/* mark the queue as mq asap */
2513 	q->mq_ops = set->ops;
2514 
2515 	q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2516 					     blk_mq_poll_stats_bkt,
2517 					     BLK_MQ_POLL_STATS_BKTS, q);
2518 	if (!q->poll_cb)
2519 		goto err_exit;
2520 
2521 	q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2522 	if (!q->queue_ctx)
2523 		goto err_exit;
2524 
2525 	/* init q->mq_kobj and sw queues' kobjects */
2526 	blk_mq_sysfs_init(q);
2527 
2528 	q->queue_hw_ctx = kcalloc_node(nr_cpu_ids, sizeof(*(q->queue_hw_ctx)),
2529 						GFP_KERNEL, set->numa_node);
2530 	if (!q->queue_hw_ctx)
2531 		goto err_percpu;
2532 
2533 	q->mq_map = set->mq_map;
2534 
2535 	blk_mq_realloc_hw_ctxs(set, q);
2536 	if (!q->nr_hw_queues)
2537 		goto err_hctxs;
2538 
2539 	INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2540 	blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2541 
2542 	q->nr_queues = nr_cpu_ids;
2543 
2544 	q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2545 
2546 	if (!(set->flags & BLK_MQ_F_SG_MERGE))
2547 		queue_flag_set_unlocked(QUEUE_FLAG_NO_SG_MERGE, q);
2548 
2549 	q->sg_reserved_size = INT_MAX;
2550 
2551 	INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2552 	INIT_LIST_HEAD(&q->requeue_list);
2553 	spin_lock_init(&q->requeue_lock);
2554 
2555 	blk_queue_make_request(q, blk_mq_make_request);
2556 	if (q->mq_ops->poll)
2557 		q->poll_fn = blk_mq_poll;
2558 
2559 	/*
2560 	 * Do this after blk_queue_make_request() overrides it...
2561 	 */
2562 	q->nr_requests = set->queue_depth;
2563 
2564 	/*
2565 	 * Default to classic polling
2566 	 */
2567 	q->poll_nsec = -1;
2568 
2569 	if (set->ops->complete)
2570 		blk_queue_softirq_done(q, set->ops->complete);
2571 
2572 	blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2573 	blk_mq_add_queue_tag_set(set, q);
2574 	blk_mq_map_swqueue(q);
2575 
2576 	if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2577 		int ret;
2578 
2579 		ret = elevator_init_mq(q);
2580 		if (ret)
2581 			return ERR_PTR(ret);
2582 	}
2583 
2584 	return q;
2585 
2586 err_hctxs:
2587 	kfree(q->queue_hw_ctx);
2588 err_percpu:
2589 	free_percpu(q->queue_ctx);
2590 err_exit:
2591 	q->mq_ops = NULL;
2592 	return ERR_PTR(-ENOMEM);
2593 }
2594 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2595 
2596 void blk_mq_free_queue(struct request_queue *q)
2597 {
2598 	struct blk_mq_tag_set	*set = q->tag_set;
2599 
2600 	blk_mq_del_queue_tag_set(q);
2601 	blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2602 }
2603 
2604 /* Basically redo blk_mq_init_queue with queue frozen */
2605 static void blk_mq_queue_reinit(struct request_queue *q)
2606 {
2607 	WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2608 
2609 	blk_mq_debugfs_unregister_hctxs(q);
2610 	blk_mq_sysfs_unregister(q);
2611 
2612 	/*
2613 	 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2614 	 * we should change hctx numa_node according to the new topology (this
2615 	 * involves freeing and re-allocating memory, worth doing?)
2616 	 */
2617 	blk_mq_map_swqueue(q);
2618 
2619 	blk_mq_sysfs_register(q);
2620 	blk_mq_debugfs_register_hctxs(q);
2621 }
2622 
2623 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2624 {
2625 	int i;
2626 
2627 	for (i = 0; i < set->nr_hw_queues; i++)
2628 		if (!__blk_mq_alloc_rq_map(set, i))
2629 			goto out_unwind;
2630 
2631 	return 0;
2632 
2633 out_unwind:
2634 	while (--i >= 0)
2635 		blk_mq_free_rq_map(set->tags[i]);
2636 
2637 	return -ENOMEM;
2638 }
2639 
2640 /*
2641  * Allocate the request maps associated with this tag_set. Note that this
2642  * may reduce the depth asked for, if memory is tight. set->queue_depth
2643  * will be updated to reflect the allocated depth.
2644  */
2645 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2646 {
2647 	unsigned int depth;
2648 	int err;
2649 
2650 	depth = set->queue_depth;
2651 	do {
2652 		err = __blk_mq_alloc_rq_maps(set);
2653 		if (!err)
2654 			break;
2655 
2656 		set->queue_depth >>= 1;
2657 		if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2658 			err = -ENOMEM;
2659 			break;
2660 		}
2661 	} while (set->queue_depth);
2662 
2663 	if (!set->queue_depth || err) {
2664 		pr_err("blk-mq: failed to allocate request map\n");
2665 		return -ENOMEM;
2666 	}
2667 
2668 	if (depth != set->queue_depth)
2669 		pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2670 						depth, set->queue_depth);
2671 
2672 	return 0;
2673 }
2674 
2675 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2676 {
2677 	if (set->ops->map_queues) {
2678 		int cpu;
2679 		/*
2680 		 * transport .map_queues is usually done in the following
2681 		 * way:
2682 		 *
2683 		 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2684 		 * 	mask = get_cpu_mask(queue)
2685 		 * 	for_each_cpu(cpu, mask)
2686 		 * 		set->mq_map[cpu] = queue;
2687 		 * }
2688 		 *
2689 		 * When we need to remap, the table has to be cleared for
2690 		 * killing stale mapping since one CPU may not be mapped
2691 		 * to any hw queue.
2692 		 */
2693 		for_each_possible_cpu(cpu)
2694 			set->mq_map[cpu] = 0;
2695 
2696 		return set->ops->map_queues(set);
2697 	} else
2698 		return blk_mq_map_queues(set);
2699 }
2700 
2701 /*
2702  * Alloc a tag set to be associated with one or more request queues.
2703  * May fail with EINVAL for various error conditions. May adjust the
2704  * requested depth down, if if it too large. In that case, the set
2705  * value will be stored in set->queue_depth.
2706  */
2707 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2708 {
2709 	int ret;
2710 
2711 	BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2712 
2713 	if (!set->nr_hw_queues)
2714 		return -EINVAL;
2715 	if (!set->queue_depth)
2716 		return -EINVAL;
2717 	if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2718 		return -EINVAL;
2719 
2720 	if (!set->ops->queue_rq)
2721 		return -EINVAL;
2722 
2723 	if (!set->ops->get_budget ^ !set->ops->put_budget)
2724 		return -EINVAL;
2725 
2726 	if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2727 		pr_info("blk-mq: reduced tag depth to %u\n",
2728 			BLK_MQ_MAX_DEPTH);
2729 		set->queue_depth = BLK_MQ_MAX_DEPTH;
2730 	}
2731 
2732 	/*
2733 	 * If a crashdump is active, then we are potentially in a very
2734 	 * memory constrained environment. Limit us to 1 queue and
2735 	 * 64 tags to prevent using too much memory.
2736 	 */
2737 	if (is_kdump_kernel()) {
2738 		set->nr_hw_queues = 1;
2739 		set->queue_depth = min(64U, set->queue_depth);
2740 	}
2741 	/*
2742 	 * There is no use for more h/w queues than cpus.
2743 	 */
2744 	if (set->nr_hw_queues > nr_cpu_ids)
2745 		set->nr_hw_queues = nr_cpu_ids;
2746 
2747 	set->tags = kcalloc_node(nr_cpu_ids, sizeof(struct blk_mq_tags *),
2748 				 GFP_KERNEL, set->numa_node);
2749 	if (!set->tags)
2750 		return -ENOMEM;
2751 
2752 	ret = -ENOMEM;
2753 	set->mq_map = kcalloc_node(nr_cpu_ids, sizeof(*set->mq_map),
2754 				   GFP_KERNEL, set->numa_node);
2755 	if (!set->mq_map)
2756 		goto out_free_tags;
2757 
2758 	ret = blk_mq_update_queue_map(set);
2759 	if (ret)
2760 		goto out_free_mq_map;
2761 
2762 	ret = blk_mq_alloc_rq_maps(set);
2763 	if (ret)
2764 		goto out_free_mq_map;
2765 
2766 	mutex_init(&set->tag_list_lock);
2767 	INIT_LIST_HEAD(&set->tag_list);
2768 
2769 	return 0;
2770 
2771 out_free_mq_map:
2772 	kfree(set->mq_map);
2773 	set->mq_map = NULL;
2774 out_free_tags:
2775 	kfree(set->tags);
2776 	set->tags = NULL;
2777 	return ret;
2778 }
2779 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2780 
2781 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2782 {
2783 	int i;
2784 
2785 	for (i = 0; i < nr_cpu_ids; i++)
2786 		blk_mq_free_map_and_requests(set, i);
2787 
2788 	kfree(set->mq_map);
2789 	set->mq_map = NULL;
2790 
2791 	kfree(set->tags);
2792 	set->tags = NULL;
2793 }
2794 EXPORT_SYMBOL(blk_mq_free_tag_set);
2795 
2796 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2797 {
2798 	struct blk_mq_tag_set *set = q->tag_set;
2799 	struct blk_mq_hw_ctx *hctx;
2800 	int i, ret;
2801 
2802 	if (!set)
2803 		return -EINVAL;
2804 
2805 	blk_mq_freeze_queue(q);
2806 	blk_mq_quiesce_queue(q);
2807 
2808 	ret = 0;
2809 	queue_for_each_hw_ctx(q, hctx, i) {
2810 		if (!hctx->tags)
2811 			continue;
2812 		/*
2813 		 * If we're using an MQ scheduler, just update the scheduler
2814 		 * queue depth. This is similar to what the old code would do.
2815 		 */
2816 		if (!hctx->sched_tags) {
2817 			ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
2818 							false);
2819 		} else {
2820 			ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2821 							nr, true);
2822 		}
2823 		if (ret)
2824 			break;
2825 	}
2826 
2827 	if (!ret)
2828 		q->nr_requests = nr;
2829 
2830 	blk_mq_unquiesce_queue(q);
2831 	blk_mq_unfreeze_queue(q);
2832 
2833 	return ret;
2834 }
2835 
2836 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2837 							int nr_hw_queues)
2838 {
2839 	struct request_queue *q;
2840 
2841 	lockdep_assert_held(&set->tag_list_lock);
2842 
2843 	if (nr_hw_queues > nr_cpu_ids)
2844 		nr_hw_queues = nr_cpu_ids;
2845 	if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2846 		return;
2847 
2848 	list_for_each_entry(q, &set->tag_list, tag_set_list)
2849 		blk_mq_freeze_queue(q);
2850 
2851 	set->nr_hw_queues = nr_hw_queues;
2852 	blk_mq_update_queue_map(set);
2853 	list_for_each_entry(q, &set->tag_list, tag_set_list) {
2854 		blk_mq_realloc_hw_ctxs(set, q);
2855 		blk_mq_queue_reinit(q);
2856 	}
2857 
2858 	list_for_each_entry(q, &set->tag_list, tag_set_list)
2859 		blk_mq_unfreeze_queue(q);
2860 }
2861 
2862 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2863 {
2864 	mutex_lock(&set->tag_list_lock);
2865 	__blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2866 	mutex_unlock(&set->tag_list_lock);
2867 }
2868 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2869 
2870 /* Enable polling stats and return whether they were already enabled. */
2871 static bool blk_poll_stats_enable(struct request_queue *q)
2872 {
2873 	if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2874 	    blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
2875 		return true;
2876 	blk_stat_add_callback(q, q->poll_cb);
2877 	return false;
2878 }
2879 
2880 static void blk_mq_poll_stats_start(struct request_queue *q)
2881 {
2882 	/*
2883 	 * We don't arm the callback if polling stats are not enabled or the
2884 	 * callback is already active.
2885 	 */
2886 	if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2887 	    blk_stat_is_active(q->poll_cb))
2888 		return;
2889 
2890 	blk_stat_activate_msecs(q->poll_cb, 100);
2891 }
2892 
2893 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2894 {
2895 	struct request_queue *q = cb->data;
2896 	int bucket;
2897 
2898 	for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
2899 		if (cb->stat[bucket].nr_samples)
2900 			q->poll_stat[bucket] = cb->stat[bucket];
2901 	}
2902 }
2903 
2904 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2905 				       struct blk_mq_hw_ctx *hctx,
2906 				       struct request *rq)
2907 {
2908 	unsigned long ret = 0;
2909 	int bucket;
2910 
2911 	/*
2912 	 * If stats collection isn't on, don't sleep but turn it on for
2913 	 * future users
2914 	 */
2915 	if (!blk_poll_stats_enable(q))
2916 		return 0;
2917 
2918 	/*
2919 	 * As an optimistic guess, use half of the mean service time
2920 	 * for this type of request. We can (and should) make this smarter.
2921 	 * For instance, if the completion latencies are tight, we can
2922 	 * get closer than just half the mean. This is especially
2923 	 * important on devices where the completion latencies are longer
2924 	 * than ~10 usec. We do use the stats for the relevant IO size
2925 	 * if available which does lead to better estimates.
2926 	 */
2927 	bucket = blk_mq_poll_stats_bkt(rq);
2928 	if (bucket < 0)
2929 		return ret;
2930 
2931 	if (q->poll_stat[bucket].nr_samples)
2932 		ret = (q->poll_stat[bucket].mean + 1) / 2;
2933 
2934 	return ret;
2935 }
2936 
2937 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2938 				     struct blk_mq_hw_ctx *hctx,
2939 				     struct request *rq)
2940 {
2941 	struct hrtimer_sleeper hs;
2942 	enum hrtimer_mode mode;
2943 	unsigned int nsecs;
2944 	ktime_t kt;
2945 
2946 	if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
2947 		return false;
2948 
2949 	/*
2950 	 * poll_nsec can be:
2951 	 *
2952 	 * -1:	don't ever hybrid sleep
2953 	 *  0:	use half of prev avg
2954 	 * >0:	use this specific value
2955 	 */
2956 	if (q->poll_nsec == -1)
2957 		return false;
2958 	else if (q->poll_nsec > 0)
2959 		nsecs = q->poll_nsec;
2960 	else
2961 		nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2962 
2963 	if (!nsecs)
2964 		return false;
2965 
2966 	rq->rq_flags |= RQF_MQ_POLL_SLEPT;
2967 
2968 	/*
2969 	 * This will be replaced with the stats tracking code, using
2970 	 * 'avg_completion_time / 2' as the pre-sleep target.
2971 	 */
2972 	kt = nsecs;
2973 
2974 	mode = HRTIMER_MODE_REL;
2975 	hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2976 	hrtimer_set_expires(&hs.timer, kt);
2977 
2978 	hrtimer_init_sleeper(&hs, current);
2979 	do {
2980 		if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
2981 			break;
2982 		set_current_state(TASK_UNINTERRUPTIBLE);
2983 		hrtimer_start_expires(&hs.timer, mode);
2984 		if (hs.task)
2985 			io_schedule();
2986 		hrtimer_cancel(&hs.timer);
2987 		mode = HRTIMER_MODE_ABS;
2988 	} while (hs.task && !signal_pending(current));
2989 
2990 	__set_current_state(TASK_RUNNING);
2991 	destroy_hrtimer_on_stack(&hs.timer);
2992 	return true;
2993 }
2994 
2995 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2996 {
2997 	struct request_queue *q = hctx->queue;
2998 	long state;
2999 
3000 	/*
3001 	 * If we sleep, have the caller restart the poll loop to reset
3002 	 * the state. Like for the other success return cases, the
3003 	 * caller is responsible for checking if the IO completed. If
3004 	 * the IO isn't complete, we'll get called again and will go
3005 	 * straight to the busy poll loop.
3006 	 */
3007 	if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
3008 		return true;
3009 
3010 	hctx->poll_considered++;
3011 
3012 	state = current->state;
3013 	while (!need_resched()) {
3014 		int ret;
3015 
3016 		hctx->poll_invoked++;
3017 
3018 		ret = q->mq_ops->poll(hctx, rq->tag);
3019 		if (ret > 0) {
3020 			hctx->poll_success++;
3021 			set_current_state(TASK_RUNNING);
3022 			return true;
3023 		}
3024 
3025 		if (signal_pending_state(state, current))
3026 			set_current_state(TASK_RUNNING);
3027 
3028 		if (current->state == TASK_RUNNING)
3029 			return true;
3030 		if (ret < 0)
3031 			break;
3032 		cpu_relax();
3033 	}
3034 
3035 	__set_current_state(TASK_RUNNING);
3036 	return false;
3037 }
3038 
3039 static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
3040 {
3041 	struct blk_mq_hw_ctx *hctx;
3042 	struct request *rq;
3043 
3044 	if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3045 		return false;
3046 
3047 	hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3048 	if (!blk_qc_t_is_internal(cookie))
3049 		rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3050 	else {
3051 		rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3052 		/*
3053 		 * With scheduling, if the request has completed, we'll
3054 		 * get a NULL return here, as we clear the sched tag when
3055 		 * that happens. The request still remains valid, like always,
3056 		 * so we should be safe with just the NULL check.
3057 		 */
3058 		if (!rq)
3059 			return false;
3060 	}
3061 
3062 	return __blk_mq_poll(hctx, rq);
3063 }
3064 
3065 static int __init blk_mq_init(void)
3066 {
3067 	cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3068 				blk_mq_hctx_notify_dead);
3069 	return 0;
3070 }
3071 subsys_initcall(blk_mq_init);
3072