1	// SPDX-License-Identifier: GPL-2.0
2	/*
3	* Block multiqueue core code
4	*
5	* Copyright (C) 2013-2014 Jens Axboe
6	* Copyright (C) 2013-2014 Christoph Hellwig
7	*/
8	#include <linux/kernel.h>
9	#include <linux/module.h>
10	#include <linux/backing-dev.h>
11	#include <linux/bio.h>
12	#include <linux/blkdev.h>
13	#include <linux/blk-integrity.h>
14	#include <linux/kmemleak.h>
15	#include <linux/mm.h>
16	#include <linux/init.h>
17	#include <linux/slab.h>
18	#include <linux/workqueue.h>
19	#include <linux/smp.h>
20	#include <linux/interrupt.h>
21	#include <linux/llist.h>
22	#include <linux/cpu.h>
23	#include <linux/cache.h>
24	#include <linux/sched/sysctl.h>
25	#include <linux/sched/topology.h>
26	#include <linux/sched/signal.h>
27	#include <linux/delay.h>
28	#include <linux/crash_dump.h>
29	#include <linux/prefetch.h>
30	#include <linux/blk-crypto.h>
31	#include <linux/part_stat.h>
32
33	#include <trace/events/block.h>
34
35	#include <linux/t10-pi.h>
36	#include "blk.h"
37	#include "blk-mq.h"
38	#include "blk-mq-debugfs.h"
39	#include "blk-pm.h"
40	#include "blk-stat.h"
41	#include "blk-mq-sched.h"
42	#include "blk-rq-qos.h"
43	#include "blk-ioprio.h"
44
45	static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
46	static DEFINE_PER_CPU(call_single_data_t, blk_cpu_csd);
47
48	static void blk_mq_insert_request(struct request *rq, blk_insert_t flags);
49	static void blk_mq_request_bypass_insert(struct request *rq,
50	blk_insert_t flags);
51	static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
52	struct list_head *list);
53	static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
54	struct io_comp_batch *iob, unsigned int flags);
55
56	/*
57	* Check if any of the ctx, dispatch list or elevator
58	* have pending work in this hardware queue.
59	*/
60	static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
61	{
62	return !list_empty_careful(head: &hctx->dispatch) ||
63	sbitmap_any_bit_set(sb: &hctx->ctx_map) ||
64	blk_mq_sched_has_work(hctx);
65	}
66
67	/*
68	* Mark this ctx as having pending work in this hardware queue
69	*/
70	static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
71	struct blk_mq_ctx *ctx)
72	{
73	const int bit = ctx->index_hw[hctx->type];
74
75	if (!sbitmap_test_bit(sb: &hctx->ctx_map, bitnr: bit))
76	sbitmap_set_bit(sb: &hctx->ctx_map, bitnr: bit);
77	}
78
79	static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
80	struct blk_mq_ctx *ctx)
81	{
82	const int bit = ctx->index_hw[hctx->type];
83
84	sbitmap_clear_bit(sb: &hctx->ctx_map, bitnr: bit);
85	}
86
87	struct mq_inflight {
88	struct block_device *part;
89	unsigned int inflight[2];
90	};
91
92	static bool blk_mq_check_inflight(struct request *rq, void *priv)
93	{
94	struct mq_inflight *mi = priv;
95
96	if (rq->part && blk_do_io_stat(rq) &&
97	(!mi->part->bd_partno || rq->part == mi->part) &&
98	blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
99	mi->inflight[rq_data_dir(rq)]++;
100
101	return true;
102	}
103
104	unsigned int blk_mq_in_flight(struct request_queue *q,
105	struct block_device *part)
106	{
107	struct mq_inflight mi = { .part = part };
108
109	blk_mq_queue_tag_busy_iter(q, fn: blk_mq_check_inflight, priv: &mi);
110
111	return mi.inflight[0] + mi.inflight[1];
112	}
113
114	void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
115	unsigned int inflight[2])
116	{
117	struct mq_inflight mi = { .part = part };
118
119	blk_mq_queue_tag_busy_iter(q, fn: blk_mq_check_inflight, priv: &mi);
120	inflight[0] = mi.inflight[0];
121	inflight[1] = mi.inflight[1];
122	}
123
124	void blk_freeze_queue_start(struct request_queue *q)
125	{
126	mutex_lock(&q->mq_freeze_lock);
127	if (++q->mq_freeze_depth == 1) {
128	percpu_ref_kill(ref: &q->q_usage_counter);
129	mutex_unlock(lock: &q->mq_freeze_lock);
130	if (queue_is_mq(q))
131	blk_mq_run_hw_queues(q, async: false);
132	} else {
133	mutex_unlock(lock: &q->mq_freeze_lock);
134	}
135	}
136	EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
137
138	void blk_mq_freeze_queue_wait(struct request_queue *q)
139	{
140	wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
141	}
142	EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
143
144	int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
145	unsigned long timeout)
146	{
147	return wait_event_timeout(q->mq_freeze_wq,
148	percpu_ref_is_zero(&q->q_usage_counter),
149	timeout);
150	}
151	EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
152
153	/*
154	* Guarantee no request is in use, so we can change any data structure of
155	* the queue afterward.
156	*/
157	void blk_freeze_queue(struct request_queue *q)
158	{
159	/*
160	* In the !blk_mq case we are only calling this to kill the
161	* q_usage_counter, otherwise this increases the freeze depth
162	* and waits for it to return to zero. For this reason there is
163	* no blk_unfreeze_queue(), and blk_freeze_queue() is not
164	* exported to drivers as the only user for unfreeze is blk_mq.
165	*/
166	blk_freeze_queue_start(q);
167	blk_mq_freeze_queue_wait(q);
168	}
169
170	void blk_mq_freeze_queue(struct request_queue *q)
171	{
172	/*
173	* ...just an alias to keep freeze and unfreeze actions balanced
174	* in the blk_mq_* namespace
175	*/
176	blk_freeze_queue(q);
177	}
178	EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
179
180	void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
181	{
182	mutex_lock(&q->mq_freeze_lock);
183	if (force_atomic)
184	q->q_usage_counter.data->force_atomic = true;
185	q->mq_freeze_depth--;
186	WARN_ON_ONCE(q->mq_freeze_depth < 0);
187	if (!q->mq_freeze_depth) {
188	percpu_ref_resurrect(ref: &q->q_usage_counter);
189	wake_up_all(&q->mq_freeze_wq);
190	}
191	mutex_unlock(lock: &q->mq_freeze_lock);
192	}
193
194	void blk_mq_unfreeze_queue(struct request_queue *q)
195	{
196	__blk_mq_unfreeze_queue(q, force_atomic: false);
197	}
198	EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
199
200	/*
201	* FIXME: replace the scsi_internal_device_*block_nowait() calls in the
202	* mpt3sas driver such that this function can be removed.
203	*/
204	void blk_mq_quiesce_queue_nowait(struct request_queue *q)
205	{
206	unsigned long flags;
207
208	spin_lock_irqsave(&q->queue_lock, flags);
209	if (!q->quiesce_depth++)
210	blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
211	spin_unlock_irqrestore(lock: &q->queue_lock, flags);
212	}
213	EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
214
215	/**
216	* blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
217	* @set: tag_set to wait on
218	*
219	* Note: it is driver's responsibility for making sure that quiesce has
220	* been started on or more of the request_queues of the tag_set. This
221	* function only waits for the quiesce on those request_queues that had
222	* the quiesce flag set using blk_mq_quiesce_queue_nowait.
223	*/
224	void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set)
225	{
226	if (set->flags & BLK_MQ_F_BLOCKING)
227	synchronize_srcu(ssp: set->srcu);
228	else
229	synchronize_rcu();
230	}
231	EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
232
233	/**
234	* blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
235	* @q: request queue.
236	*
237	* Note: this function does not prevent that the struct request end_io()
238	* callback function is invoked. Once this function is returned, we make
239	* sure no dispatch can happen until the queue is unquiesced via
240	* blk_mq_unquiesce_queue().
241	*/
242	void blk_mq_quiesce_queue(struct request_queue *q)
243	{
244	blk_mq_quiesce_queue_nowait(q);
245	/* nothing to wait for non-mq queues */
246	if (queue_is_mq(q))
247	blk_mq_wait_quiesce_done(q->tag_set);
248	}
249	EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
250
251	/*
252	* blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
253	* @q: request queue.
254	*
255	* This function recovers queue into the state before quiescing
256	* which is done by blk_mq_quiesce_queue.
257	*/
258	void blk_mq_unquiesce_queue(struct request_queue *q)
259	{
260	unsigned long flags;
261	bool run_queue = false;
262
263	spin_lock_irqsave(&q->queue_lock, flags);
264	if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
265	;
266	} else if (!--q->quiesce_depth) {
267	blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
268	run_queue = true;
269	}
270	spin_unlock_irqrestore(lock: &q->queue_lock, flags);
271
272	/* dispatch requests which are inserted during quiescing */
273	if (run_queue)
274	blk_mq_run_hw_queues(q, async: true);
275	}
276	EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
277
278	void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set)
279	{
280	struct request_queue *q;
281
282	mutex_lock(&set->tag_list_lock);
283	list_for_each_entry(q, &set->tag_list, tag_set_list) {
284	if (!blk_queue_skip_tagset_quiesce(q))
285	blk_mq_quiesce_queue_nowait(q);
286	}
287	blk_mq_wait_quiesce_done(set);
288	mutex_unlock(lock: &set->tag_list_lock);
289	}
290	EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset);
291
292	void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set)
293	{
294	struct request_queue *q;
295
296	mutex_lock(&set->tag_list_lock);
297	list_for_each_entry(q, &set->tag_list, tag_set_list) {
298	if (!blk_queue_skip_tagset_quiesce(q))
299	blk_mq_unquiesce_queue(q);
300	}
301	mutex_unlock(lock: &set->tag_list_lock);
302	}
303	EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset);
304
305	void blk_mq_wake_waiters(struct request_queue *q)
306	{
307	struct blk_mq_hw_ctx *hctx;
308	unsigned long i;
309
310	queue_for_each_hw_ctx(q, hctx, i)
311	if (blk_mq_hw_queue_mapped(hctx))
312	blk_mq_tag_wakeup_all(tags: hctx->tags, true);
313	}
314
315	void blk_rq_init(struct request_queue *q, struct request *rq)
316	{
317	memset(rq, 0, sizeof(*rq));
318
319	INIT_LIST_HEAD(list: &rq->queuelist);
320	rq->q = q;
321	rq->__sector = (sector_t) -1;
322	INIT_HLIST_NODE(h: &rq->hash);
323	RB_CLEAR_NODE(&rq->rb_node);
324	rq->tag = BLK_MQ_NO_TAG;
325	rq->internal_tag = BLK_MQ_NO_TAG;
326	rq->start_time_ns = ktime_get_ns();
327	rq->part = NULL;
328	blk_crypto_rq_set_defaults(rq);
329	}
330	EXPORT_SYMBOL(blk_rq_init);
331
332	/* Set start and alloc time when the allocated request is actually used */
333	static inline void blk_mq_rq_time_init(struct request *rq, u64 alloc_time_ns)
334	{
335	if (blk_mq_need_time_stamp(rq))
336	rq->start_time_ns = ktime_get_ns();
337	else
338	rq->start_time_ns = 0;
339
340	#ifdef CONFIG_BLK_RQ_ALLOC_TIME
341	if (blk_queue_rq_alloc_time(rq->q))
342	rq->alloc_time_ns = alloc_time_ns ?: rq->start_time_ns;
343	else
344	rq->alloc_time_ns = 0;
345	#endif
346	}
347
348	static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
349	struct blk_mq_tags *tags, unsigned int tag)
350	{
351	struct blk_mq_ctx *ctx = data->ctx;
352	struct blk_mq_hw_ctx *hctx = data->hctx;
353	struct request_queue *q = data->q;
354	struct request *rq = tags->static_rqs[tag];
355
356	rq->q = q;
357	rq->mq_ctx = ctx;
358	rq->mq_hctx = hctx;
359	rq->cmd_flags = data->cmd_flags;
360
361	if (data->flags & BLK_MQ_REQ_PM)
362	data->rq_flags |= RQF_PM;
363	if (blk_queue_io_stat(q))
364	data->rq_flags |= RQF_IO_STAT;
365	rq->rq_flags = data->rq_flags;
366
367	if (data->rq_flags & RQF_SCHED_TAGS) {
368	rq->tag = BLK_MQ_NO_TAG;
369	rq->internal_tag = tag;
370	} else {
371	rq->tag = tag;
372	rq->internal_tag = BLK_MQ_NO_TAG;
373	}
374	rq->timeout = 0;
375
376	rq->part = NULL;
377	rq->io_start_time_ns = 0;
378	rq->stats_sectors = 0;
379	rq->nr_phys_segments = 0;
380	#if defined(CONFIG_BLK_DEV_INTEGRITY)
381	rq->nr_integrity_segments = 0;
382	#endif
383	rq->end_io = NULL;
384	rq->end_io_data = NULL;
385
386	blk_crypto_rq_set_defaults(rq);
387	INIT_LIST_HEAD(list: &rq->queuelist);
388	/* tag was already set */
389	WRITE_ONCE(rq->deadline, 0);
390	req_ref_set(req: rq, value: 1);
391
392	if (rq->rq_flags & RQF_USE_SCHED) {
393	struct elevator_queue *e = data->q->elevator;
394
395	INIT_HLIST_NODE(h: &rq->hash);
396	RB_CLEAR_NODE(&rq->rb_node);
397
398	if (e->type->ops.prepare_request)
399	e->type->ops.prepare_request(rq);
400	}
401
402	return rq;
403	}
404
405	static inline struct request *
406	__blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data)
407	{
408	unsigned int tag, tag_offset;
409	struct blk_mq_tags *tags;
410	struct request *rq;
411	unsigned long tag_mask;
412	int i, nr = 0;
413
414	tag_mask = blk_mq_get_tags(data, nr_tags: data->nr_tags, offset: &tag_offset);
415	if (unlikely(!tag_mask))
416	return NULL;
417
418	tags = blk_mq_tags_from_data(data);
419	for (i = 0; tag_mask; i++) {
420	if (!(tag_mask & (1UL << i)))
421	continue;
422	tag = tag_offset + i;
423	prefetch(tags->static_rqs[tag]);
424	tag_mask &= ~(1UL << i);
425	rq = blk_mq_rq_ctx_init(data, tags, tag);
426	rq_list_add(data->cached_rq, rq);
427	nr++;
428	}
429	if (!(data->rq_flags & RQF_SCHED_TAGS))
430	blk_mq_add_active_requests(hctx: data->hctx, val: nr);
431	/* caller already holds a reference, add for remainder */
432	percpu_ref_get_many(ref: &data->q->q_usage_counter, nr: nr - 1);
433	data->nr_tags -= nr;
434
435	return rq_list_pop(data->cached_rq);
436	}
437
438	static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
439	{
440	struct request_queue *q = data->q;
441	u64 alloc_time_ns = 0;
442	struct request *rq;
443	unsigned int tag;
444
445	/* alloc_time includes depth and tag waits */
446	if (blk_queue_rq_alloc_time(q))
447	alloc_time_ns = ktime_get_ns();
448
449	if (data->cmd_flags & REQ_NOWAIT)
450	data->flags |= BLK_MQ_REQ_NOWAIT;
451
452	if (q->elevator) {
453	/*
454	* All requests use scheduler tags when an I/O scheduler is
455	* enabled for the queue.
456	*/
457	data->rq_flags |= RQF_SCHED_TAGS;
458
459	/*
460	* Flush/passthrough requests are special and go directly to the
461	* dispatch list.
462	*/
463	if ((data->cmd_flags & REQ_OP_MASK) != REQ_OP_FLUSH &&
464	!blk_op_is_passthrough(op: data->cmd_flags)) {
465	struct elevator_mq_ops *ops = &q->elevator->type->ops;
466
467	WARN_ON_ONCE(data->flags & BLK_MQ_REQ_RESERVED);
468
469	data->rq_flags |= RQF_USE_SCHED;
470	if (ops->limit_depth)
471	ops->limit_depth(data->cmd_flags, data);
472	}
473	}
474
475	retry:
476	data->ctx = blk_mq_get_ctx(q);
477	data->hctx = blk_mq_map_queue(q, opf: data->cmd_flags, ctx: data->ctx);
478	if (!(data->rq_flags & RQF_SCHED_TAGS))
479	blk_mq_tag_busy(hctx: data->hctx);
480
481	if (data->flags & BLK_MQ_REQ_RESERVED)
482	data->rq_flags |= RQF_RESV;
483
484	/*
485	* Try batched alloc if we want more than 1 tag.
486	*/
487	if (data->nr_tags > 1) {
488	rq = __blk_mq_alloc_requests_batch(data);
489	if (rq) {
490	blk_mq_rq_time_init(rq, alloc_time_ns);
491	return rq;
492	}
493	data->nr_tags = 1;
494	}
495
496	/*
497	* Waiting allocations only fail because of an inactive hctx. In that
498	* case just retry the hctx assignment and tag allocation as CPU hotplug
499	* should have migrated us to an online CPU by now.
500	*/
501	tag = blk_mq_get_tag(data);
502	if (tag == BLK_MQ_NO_TAG) {
503	if (data->flags & BLK_MQ_REQ_NOWAIT)
504	return NULL;
505	/*
506	* Give up the CPU and sleep for a random short time to
507	* ensure that thread using a realtime scheduling class
508	* are migrated off the CPU, and thus off the hctx that
509	* is going away.
510	*/
511	msleep(msecs: 3);
512	goto retry;
513	}
514
515	if (!(data->rq_flags & RQF_SCHED_TAGS))
516	blk_mq_inc_active_requests(hctx: data->hctx);
517	rq = blk_mq_rq_ctx_init(data, tags: blk_mq_tags_from_data(data), tag);
518	blk_mq_rq_time_init(rq, alloc_time_ns);
519	return rq;
520	}
521
522	static struct request *blk_mq_rq_cache_fill(struct request_queue *q,
523	struct blk_plug *plug,
524	blk_opf_t opf,
525	blk_mq_req_flags_t flags)
526	{
527	struct blk_mq_alloc_data data = {
528	.q = q,
529	.flags = flags,
530	.cmd_flags = opf,
531	.nr_tags = plug->nr_ios,
532	.cached_rq = &plug->cached_rq,
533	};
534	struct request *rq;
535
536	if (blk_queue_enter(q, flags))
537	return NULL;
538
539	plug->nr_ios = 1;
540
541	rq = __blk_mq_alloc_requests(data: &data);
542	if (unlikely(!rq))
543	blk_queue_exit(q);
544	return rq;
545	}
546
547	static struct request *blk_mq_alloc_cached_request(struct request_queue *q,
548	blk_opf_t opf,
549	blk_mq_req_flags_t flags)
550	{
551	struct blk_plug *plug = current->plug;
552	struct request *rq;
553
554	if (!plug)
555	return NULL;
556
557	if (rq_list_empty(plug->cached_rq)) {
558	if (plug->nr_ios == 1)
559	return NULL;
560	rq = blk_mq_rq_cache_fill(q, plug, opf, flags);
561	if (!rq)
562	return NULL;
563	} else {
564	rq = rq_list_peek(&plug->cached_rq);
565	if (!rq || rq->q != q)
566	return NULL;
567
568	if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type)
569	return NULL;
570	if (op_is_flush(op: rq->cmd_flags) != op_is_flush(op: opf))
571	return NULL;
572
573	plug->cached_rq = rq_list_next(rq);
574	blk_mq_rq_time_init(rq, alloc_time_ns: 0);
575	}
576
577	rq->cmd_flags = opf;
578	INIT_LIST_HEAD(list: &rq->queuelist);
579	return rq;
580	}
581
582	struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf,
583	blk_mq_req_flags_t flags)
584	{
585	struct request *rq;
586
587	rq = blk_mq_alloc_cached_request(q, opf, flags);
588	if (!rq) {
589	struct blk_mq_alloc_data data = {
590	.q = q,
591	.flags = flags,
592	.cmd_flags = opf,
593	.nr_tags = 1,
594	};
595	int ret;
596
597	ret = blk_queue_enter(q, flags);
598	if (ret)
599	return ERR_PTR(error: ret);
600
601	rq = __blk_mq_alloc_requests(data: &data);
602	if (!rq)
603	goto out_queue_exit;
604	}
605	rq->__data_len = 0;
606	rq->__sector = (sector_t) -1;
607	rq->bio = rq->biotail = NULL;
608	return rq;
609	out_queue_exit:
610	blk_queue_exit(q);
611	return ERR_PTR(error: -EWOULDBLOCK);
612	}
613	EXPORT_SYMBOL(blk_mq_alloc_request);
614
615	struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
616	blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx)
617	{
618	struct blk_mq_alloc_data data = {
619	.q = q,
620	.flags = flags,
621	.cmd_flags = opf,
622	.nr_tags = 1,
623	};
624	u64 alloc_time_ns = 0;
625	struct request *rq;
626	unsigned int cpu;
627	unsigned int tag;
628	int ret;
629
630	/* alloc_time includes depth and tag waits */
631	if (blk_queue_rq_alloc_time(q))
632	alloc_time_ns = ktime_get_ns();
633
634	/*
635	* If the tag allocator sleeps we could get an allocation for a
636	* different hardware context. No need to complicate the low level
637	* allocator for this for the rare use case of a command tied to
638	* a specific queue.
639	*/
640	if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) ||
641	WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED)))
642	return ERR_PTR(error: -EINVAL);
643
644	if (hctx_idx >= q->nr_hw_queues)
645	return ERR_PTR(error: -EIO);
646
647	ret = blk_queue_enter(q, flags);
648	if (ret)
649	return ERR_PTR(error: ret);
650
651	/*
652	* Check if the hardware context is actually mapped to anything.
653	* If not tell the caller that it should skip this queue.
654	*/
655	ret = -EXDEV;
656	data.hctx = xa_load(&q->hctx_table, index: hctx_idx);
657	if (!blk_mq_hw_queue_mapped(hctx: data.hctx))
658	goto out_queue_exit;
659	cpu = cpumask_first_and(srcp1: data.hctx->cpumask, cpu_online_mask);
660	if (cpu >= nr_cpu_ids)
661	goto out_queue_exit;
662	data.ctx = __blk_mq_get_ctx(q, cpu);
663
664	if (q->elevator)
665	data.rq_flags |= RQF_SCHED_TAGS;
666	else
667	blk_mq_tag_busy(hctx: data.hctx);
668
669	if (flags & BLK_MQ_REQ_RESERVED)
670	data.rq_flags |= RQF_RESV;
671
672	ret = -EWOULDBLOCK;
673	tag = blk_mq_get_tag(data: &data);
674	if (tag == BLK_MQ_NO_TAG)
675	goto out_queue_exit;
676	if (!(data.rq_flags & RQF_SCHED_TAGS))
677	blk_mq_inc_active_requests(hctx: data.hctx);
678	rq = blk_mq_rq_ctx_init(data: &data, tags: blk_mq_tags_from_data(data: &data), tag);
679	blk_mq_rq_time_init(rq, alloc_time_ns);
680	rq->__data_len = 0;
681	rq->__sector = (sector_t) -1;
682	rq->bio = rq->biotail = NULL;
683	return rq;
684
685	out_queue_exit:
686	blk_queue_exit(q);
687	return ERR_PTR(error: ret);
688	}
689	EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
690
691	static void blk_mq_finish_request(struct request *rq)
692	{
693	struct request_queue *q = rq->q;
694
695	if (rq->rq_flags & RQF_USE_SCHED) {
696	q->elevator->type->ops.finish_request(rq);
697	/*
698	* For postflush request that may need to be
699	* completed twice, we should clear this flag
700	* to avoid double finish_request() on the rq.
701	*/
702	rq->rq_flags &= ~RQF_USE_SCHED;
703	}
704	}
705
706	static void __blk_mq_free_request(struct request *rq)
707	{
708	struct request_queue *q = rq->q;
709	struct blk_mq_ctx *ctx = rq->mq_ctx;
710	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
711	const int sched_tag = rq->internal_tag;
712
713	blk_crypto_free_request(rq);
714	blk_pm_mark_last_busy(rq);
715	rq->mq_hctx = NULL;
716
717	if (rq->tag != BLK_MQ_NO_TAG) {
718	blk_mq_dec_active_requests(hctx);
719	blk_mq_put_tag(tags: hctx->tags, ctx, tag: rq->tag);
720	}
721	if (sched_tag != BLK_MQ_NO_TAG)
722	blk_mq_put_tag(tags: hctx->sched_tags, ctx, tag: sched_tag);
723	blk_mq_sched_restart(hctx);
724	blk_queue_exit(q);
725	}
726
727	void blk_mq_free_request(struct request *rq)
728	{
729	struct request_queue *q = rq->q;
730
731	blk_mq_finish_request(rq);
732
733	if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
734	laptop_io_completion(info: q->disk->bdi);
735
736	rq_qos_done(q, rq);
737
738	WRITE_ONCE(rq->state, MQ_RQ_IDLE);
739	if (req_ref_put_and_test(req: rq))
740	__blk_mq_free_request(rq);
741	}
742	EXPORT_SYMBOL_GPL(blk_mq_free_request);
743
744	void blk_mq_free_plug_rqs(struct blk_plug *plug)
745	{
746	struct request *rq;
747
748	while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
749	blk_mq_free_request(rq);
750	}
751
752	void blk_dump_rq_flags(struct request *rq, char *msg)
753	{
754	printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
755	rq->q->disk ? rq->q->disk->disk_name : "?",
756	(__force unsigned long long) rq->cmd_flags);
757
758	printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
759	(unsigned long long)blk_rq_pos(rq),
760	blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
761	printk(KERN_INFO " bio %p, biotail %p, len %u\n",
762	rq->bio, rq->biotail, blk_rq_bytes(rq));
763	}
764	EXPORT_SYMBOL(blk_dump_rq_flags);
765
766	static void req_bio_endio(struct request *rq, struct bio *bio,
767	unsigned int nbytes, blk_status_t error)
768	{
769	if (unlikely(error)) {
770	bio->bi_status = error;
771	} else if (req_op(req: rq) == REQ_OP_ZONE_APPEND) {
772	/*
773	* Partial zone append completions cannot be supported as the
774	* BIO fragments may end up not being written sequentially.
775	*/
776	if (bio->bi_iter.bi_size != nbytes)
777	bio->bi_status = BLK_STS_IOERR;
778	else
779	bio->bi_iter.bi_sector = rq->__sector;
780	}
781
782	bio_advance(bio, nbytes);
783
784	if (unlikely(rq->rq_flags & RQF_QUIET))
785	bio_set_flag(bio, bit: BIO_QUIET);
786	/* don't actually finish bio if it's part of flush sequence */
787	if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
788	bio_endio(bio);
789	}
790
791	static void blk_account_io_completion(struct request *req, unsigned int bytes)
792	{
793	if (req->part && blk_do_io_stat(rq: req)) {
794	const int sgrp = op_stat_group(op: req_op(req));
795
796	part_stat_lock();
797	part_stat_add(req->part, sectors[sgrp], bytes >> 9);
798	part_stat_unlock();
799	}
800	}
801
802	static void blk_print_req_error(struct request *req, blk_status_t status)
803	{
804	printk_ratelimited(KERN_ERR
805	"%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
806	"phys_seg %u prio class %u\n",
807	blk_status_to_str(status),
808	req->q->disk ? req->q->disk->disk_name : "?",
809	blk_rq_pos(req), (__force u32)req_op(req),
810	blk_op_str(req_op(req)),
811	(__force u32)(req->cmd_flags & ~REQ_OP_MASK),
812	req->nr_phys_segments,
813	IOPRIO_PRIO_CLASS(req->ioprio));
814	}
815
816	/*
817	* Fully end IO on a request. Does not support partial completions, or
818	* errors.
819	*/
820	static void blk_complete_request(struct request *req)
821	{
822	const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
823	int total_bytes = blk_rq_bytes(rq: req);
824	struct bio *bio = req->bio;
825
826	trace_block_rq_complete(rq: req, BLK_STS_OK, nr_bytes: total_bytes);
827
828	if (!bio)
829	return;
830
831	#ifdef CONFIG_BLK_DEV_INTEGRITY
832	if (blk_integrity_rq(rq: req) && req_op(req) == REQ_OP_READ)
833	req->q->integrity.profile->complete_fn(req, total_bytes);
834	#endif
835
836	/*
837	* Upper layers may call blk_crypto_evict_key() anytime after the last
838	* bio_endio(). Therefore, the keyslot must be released before that.
839	*/
840	blk_crypto_rq_put_keyslot(rq: req);
841
842	blk_account_io_completion(req, bytes: total_bytes);
843
844	do {
845	struct bio *next = bio->bi_next;
846
847	/* Completion has already been traced */
848	bio_clear_flag(bio, bit: BIO_TRACE_COMPLETION);
849
850	if (req_op(req) == REQ_OP_ZONE_APPEND)
851	bio->bi_iter.bi_sector = req->__sector;
852
853	if (!is_flush)
854	bio_endio(bio);
855	bio = next;
856	} while (bio);
857
858	/*
859	* Reset counters so that the request stacking driver
860	* can find how many bytes remain in the request
861	* later.
862	*/
863	if (!req->end_io) {
864	req->bio = NULL;
865	req->__data_len = 0;
866	}
867	}
868
869	/**
870	* blk_update_request - Complete multiple bytes without completing the request
871	* @req: the request being processed
872	* @error: block status code
873	* @nr_bytes: number of bytes to complete for @req
874	*
875	* Description:
876	* Ends I/O on a number of bytes attached to @req, but doesn't complete
877	* the request structure even if @req doesn't have leftover.
878	* If @req has leftover, sets it up for the next range of segments.
879	*
880	* Passing the result of blk_rq_bytes() as @nr_bytes guarantees
881	* %false return from this function.
882	*
883	* Note:
884	* The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
885	* except in the consistency check at the end of this function.
886	*
887	* Return:
888	* %false - this request doesn't have any more data
889	* %true - this request has more data
890	**/
891	bool blk_update_request(struct request *req, blk_status_t error,
892	unsigned int nr_bytes)
893	{
894	int total_bytes;
895
896	trace_block_rq_complete(rq: req, error, nr_bytes);
897
898	if (!req->bio)
899	return false;
900
901	#ifdef CONFIG_BLK_DEV_INTEGRITY
902	if (blk_integrity_rq(rq: req) && req_op(req) == REQ_OP_READ &&
903	error == BLK_STS_OK)
904	req->q->integrity.profile->complete_fn(req, nr_bytes);
905	#endif
906
907	/*
908	* Upper layers may call blk_crypto_evict_key() anytime after the last
909	* bio_endio(). Therefore, the keyslot must be released before that.
910	*/
911	if (blk_crypto_rq_has_keyslot(rq: req) && nr_bytes >= blk_rq_bytes(rq: req))
912	__blk_crypto_rq_put_keyslot(rq: req);
913
914	if (unlikely(error && !blk_rq_is_passthrough(req) &&
915	!(req->rq_flags & RQF_QUIET)) &&
916	!test_bit(GD_DEAD, &req->q->disk->state)) {
917	blk_print_req_error(req, status: error);
918	trace_block_rq_error(rq: req, error, nr_bytes);
919	}
920
921	blk_account_io_completion(req, bytes: nr_bytes);
922
923	total_bytes = 0;
924	while (req->bio) {
925	struct bio *bio = req->bio;
926	unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
927
928	if (bio_bytes == bio->bi_iter.bi_size)
929	req->bio = bio->bi_next;
930
931	/* Completion has already been traced */
932	bio_clear_flag(bio, bit: BIO_TRACE_COMPLETION);
933	req_bio_endio(rq: req, bio, nbytes: bio_bytes, error);
934
935	total_bytes += bio_bytes;
936	nr_bytes -= bio_bytes;
937
938	if (!nr_bytes)
939	break;
940	}
941
942	/*
943	* completely done
944	*/
945	if (!req->bio) {
946	/*
947	* Reset counters so that the request stacking driver
948	* can find how many bytes remain in the request
949	* later.
950	*/
951	req->__data_len = 0;
952	return false;
953	}
954
955	req->__data_len -= total_bytes;
956
957	/* update sector only for requests with clear definition of sector */
958	if (!blk_rq_is_passthrough(rq: req))
959	req->__sector += total_bytes >> 9;
960
961	/* mixed attributes always follow the first bio */
962	if (req->rq_flags & RQF_MIXED_MERGE) {
963	req->cmd_flags &= ~REQ_FAILFAST_MASK;
964	req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
965	}
966
967	if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
968	/*
969	* If total number of sectors is less than the first segment
970	* size, something has gone terribly wrong.
971	*/
972	if (blk_rq_bytes(rq: req) < blk_rq_cur_bytes(rq: req)) {
973	blk_dump_rq_flags(req, "request botched");
974	req->__data_len = blk_rq_cur_bytes(rq: req);
975	}
976
977	/* recalculate the number of segments */
978	req->nr_phys_segments = blk_recalc_rq_segments(rq: req);
979	}
980
981	return true;
982	}
983	EXPORT_SYMBOL_GPL(blk_update_request);
984
985	static inline void blk_account_io_done(struct request *req, u64 now)
986	{
987	trace_block_io_done(rq: req);
988
989	/*
990	* Account IO completion. flush_rq isn't accounted as a
991	* normal IO on queueing nor completion. Accounting the
992	* containing request is enough.
993	*/
994	if (blk_do_io_stat(rq: req) && req->part &&
995	!(req->rq_flags & RQF_FLUSH_SEQ)) {
996	const int sgrp = op_stat_group(op: req_op(req));
997
998	part_stat_lock();
999	update_io_ticks(part: req->part, now: jiffies, end: true);
1000	part_stat_inc(req->part, ios[sgrp]);
1001	part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
1002	part_stat_unlock();
1003	}
1004	}
1005
1006	static inline void blk_account_io_start(struct request *req)
1007	{
1008	trace_block_io_start(rq: req);
1009
1010	if (blk_do_io_stat(rq: req)) {
1011	/*
1012	* All non-passthrough requests are created from a bio with one
1013	* exception: when a flush command that is part of a flush sequence
1014	* generated by the state machine in blk-flush.c is cloned onto the
1015	* lower device by dm-multipath we can get here without a bio.
1016	*/
1017	if (req->bio)
1018	req->part = req->bio->bi_bdev;
1019	else
1020	req->part = req->q->disk->part0;
1021
1022	part_stat_lock();
1023	update_io_ticks(part: req->part, now: jiffies, end: false);
1024	part_stat_unlock();
1025	}
1026	}
1027
1028	static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
1029	{
1030	if (rq->rq_flags & RQF_STATS)
1031	blk_stat_add(rq, now);
1032
1033	blk_mq_sched_completed_request(rq, now);
1034	blk_account_io_done(req: rq, now);
1035	}
1036
1037	inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
1038	{
1039	if (blk_mq_need_time_stamp(rq))
1040	__blk_mq_end_request_acct(rq, now: ktime_get_ns());
1041
1042	blk_mq_finish_request(rq);
1043
1044	if (rq->end_io) {
1045	rq_qos_done(q: rq->q, rq);
1046	if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1047	blk_mq_free_request(rq);
1048	} else {
1049	blk_mq_free_request(rq);
1050	}
1051	}
1052	EXPORT_SYMBOL(__blk_mq_end_request);
1053
1054	void blk_mq_end_request(struct request *rq, blk_status_t error)
1055	{
1056	if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1057	BUG();
1058	__blk_mq_end_request(rq, error);
1059	}
1060	EXPORT_SYMBOL(blk_mq_end_request);
1061
1062	#define TAG_COMP_BATCH 32
1063
1064	static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1065	int *tag_array, int nr_tags)
1066	{
1067	struct request_queue *q = hctx->queue;
1068
1069	blk_mq_sub_active_requests(hctx, val: nr_tags);
1070
1071	blk_mq_put_tags(tags: hctx->tags, tag_array, nr_tags);
1072	percpu_ref_put_many(ref: &q->q_usage_counter, nr: nr_tags);
1073	}
1074
1075	void blk_mq_end_request_batch(struct io_comp_batch *iob)
1076	{
1077	int tags[TAG_COMP_BATCH], nr_tags = 0;
1078	struct blk_mq_hw_ctx *cur_hctx = NULL;
1079	struct request *rq;
1080	u64 now = 0;
1081
1082	if (iob->need_ts)
1083	now = ktime_get_ns();
1084
1085	while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1086	prefetch(rq->bio);
1087	prefetch(rq->rq_next);
1088
1089	blk_complete_request(req: rq);
1090	if (iob->need_ts)
1091	__blk_mq_end_request_acct(rq, now);
1092
1093	blk_mq_finish_request(rq);
1094
1095	rq_qos_done(q: rq->q, rq);
1096
1097	/*
1098	* If end_io handler returns NONE, then it still has
1099	* ownership of the request.
1100	*/
1101	if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
1102	continue;
1103
1104	WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1105	if (!req_ref_put_and_test(req: rq))
1106	continue;
1107
1108	blk_crypto_free_request(rq);
1109	blk_pm_mark_last_busy(rq);
1110
1111	if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1112	if (cur_hctx)
1113	blk_mq_flush_tag_batch(hctx: cur_hctx, tag_array: tags, nr_tags);
1114	nr_tags = 0;
1115	cur_hctx = rq->mq_hctx;
1116	}
1117	tags[nr_tags++] = rq->tag;
1118	}
1119
1120	if (nr_tags)
1121	blk_mq_flush_tag_batch(hctx: cur_hctx, tag_array: tags, nr_tags);
1122	}
1123	EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1124
1125	static void blk_complete_reqs(struct llist_head *list)
1126	{
1127	struct llist_node *entry = llist_reverse_order(head: llist_del_all(head: list));
1128	struct request *rq, *next;
1129
1130	llist_for_each_entry_safe(rq, next, entry, ipi_list)
1131	rq->q->mq_ops->complete(rq);
1132	}
1133
1134	static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1135	{
1136	blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1137	}
1138
1139	static int blk_softirq_cpu_dead(unsigned int cpu)
1140	{
1141	blk_complete_reqs(list: &per_cpu(blk_cpu_done, cpu));
1142	return 0;
1143	}
1144
1145	static void __blk_mq_complete_request_remote(void *data)
1146	{
1147	__raise_softirq_irqoff(nr: BLOCK_SOFTIRQ);
1148	}
1149
1150	static inline bool blk_mq_complete_need_ipi(struct request *rq)
1151	{
1152	int cpu = raw_smp_processor_id();
1153
1154	if (!IS_ENABLED(CONFIG_SMP) ||
1155	!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1156	return false;
1157	/*
1158	* With force threaded interrupts enabled, raising softirq from an SMP
1159	* function call will always result in waking the ksoftirqd thread.
1160	* This is probably worse than completing the request on a different
1161	* cache domain.
1162	*/
1163	if (force_irqthreads())
1164	return false;
1165
1166	/* same CPU or cache domain? Complete locally */
1167	if (cpu == rq->mq_ctx->cpu ||
1168	(!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1169	cpus_share_cache(this_cpu: cpu, that_cpu: rq->mq_ctx->cpu)))
1170	return false;
1171
1172	/* don't try to IPI to an offline CPU */
1173	return cpu_online(cpu: rq->mq_ctx->cpu);
1174	}
1175
1176	static void blk_mq_complete_send_ipi(struct request *rq)
1177	{
1178	unsigned int cpu;
1179
1180	cpu = rq->mq_ctx->cpu;
1181	if (llist_add(new: &rq->ipi_list, head: &per_cpu(blk_cpu_done, cpu)))
1182	smp_call_function_single_async(cpu, csd: &per_cpu(blk_cpu_csd, cpu));
1183	}
1184
1185	static void blk_mq_raise_softirq(struct request *rq)
1186	{
1187	struct llist_head *list;
1188
1189	preempt_disable();
1190	list = this_cpu_ptr(&blk_cpu_done);
1191	if (llist_add(new: &rq->ipi_list, head: list))
1192	raise_softirq(nr: BLOCK_SOFTIRQ);
1193	preempt_enable();
1194	}
1195
1196	bool blk_mq_complete_request_remote(struct request *rq)
1197	{
1198	WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1199
1200	/*
1201	* For request which hctx has only one ctx mapping,
1202	* or a polled request, always complete locally,
1203	* it's pointless to redirect the completion.
1204	*/
1205	if ((rq->mq_hctx->nr_ctx == 1 &&
1206	rq->mq_ctx->cpu == raw_smp_processor_id()) ||
1207	rq->cmd_flags & REQ_POLLED)
1208	return false;
1209
1210	if (blk_mq_complete_need_ipi(rq)) {
1211	blk_mq_complete_send_ipi(rq);
1212	return true;
1213	}
1214
1215	if (rq->q->nr_hw_queues == 1) {
1216	blk_mq_raise_softirq(rq);
1217	return true;
1218	}
1219	return false;
1220	}
1221	EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1222
1223	/**
1224	* blk_mq_complete_request - end I/O on a request
1225	* @rq: the request being processed
1226	*
1227	* Description:
1228	* Complete a request by scheduling the ->complete_rq operation.
1229	**/
1230	void blk_mq_complete_request(struct request *rq)
1231	{
1232	if (!blk_mq_complete_request_remote(rq))
1233	rq->q->mq_ops->complete(rq);
1234	}
1235	EXPORT_SYMBOL(blk_mq_complete_request);
1236
1237	/**
1238	* blk_mq_start_request - Start processing a request
1239	* @rq: Pointer to request to be started
1240	*
1241	* Function used by device drivers to notify the block layer that a request
1242	* is going to be processed now, so blk layer can do proper initializations
1243	* such as starting the timeout timer.
1244	*/
1245	void blk_mq_start_request(struct request *rq)
1246	{
1247	struct request_queue *q = rq->q;
1248
1249	trace_block_rq_issue(rq);
1250
1251	if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1252	rq->io_start_time_ns = ktime_get_ns();
1253	rq->stats_sectors = blk_rq_sectors(rq);
1254	rq->rq_flags |= RQF_STATS;
1255	rq_qos_issue(q, rq);
1256	}
1257
1258	WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1259
1260	blk_add_timer(req: rq);
1261	WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1262	rq->mq_hctx->tags->rqs[rq->tag] = rq;
1263
1264	#ifdef CONFIG_BLK_DEV_INTEGRITY
1265	if (blk_integrity_rq(rq) && req_op(req: rq) == REQ_OP_WRITE)
1266	q->integrity.profile->prepare_fn(rq);
1267	#endif
1268	if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1269	WRITE_ONCE(rq->bio->bi_cookie, rq->mq_hctx->queue_num);
1270	}
1271	EXPORT_SYMBOL(blk_mq_start_request);
1272
1273	/*
1274	* Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1275	* queues. This is important for md arrays to benefit from merging
1276	* requests.
1277	*/
1278	static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1279	{
1280	if (plug->multiple_queues)
1281	return BLK_MAX_REQUEST_COUNT * 2;
1282	return BLK_MAX_REQUEST_COUNT;
1283	}
1284
1285	static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1286	{
1287	struct request *last = rq_list_peek(&plug->mq_list);
1288
1289	if (!plug->rq_count) {
1290	trace_block_plug(q: rq->q);
1291	} else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1292	(!blk_queue_nomerges(rq->q) &&
1293	blk_rq_bytes(rq: last) >= BLK_PLUG_FLUSH_SIZE)) {
1294	blk_mq_flush_plug_list(plug, from_schedule: false);
1295	last = NULL;
1296	trace_block_plug(q: rq->q);
1297	}
1298
1299	if (!plug->multiple_queues && last && last->q != rq->q)
1300	plug->multiple_queues = true;
1301	/*
1302	* Any request allocated from sched tags can't be issued to
1303	* ->queue_rqs() directly
1304	*/
1305	if (!plug->has_elevator && (rq->rq_flags & RQF_SCHED_TAGS))
1306	plug->has_elevator = true;
1307	rq->rq_next = NULL;
1308	rq_list_add(&plug->mq_list, rq);
1309	plug->rq_count++;
1310	}
1311
1312	/**
1313	* blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1314	* @rq: request to insert
1315	* @at_head: insert request at head or tail of queue
1316	*
1317	* Description:
1318	* Insert a fully prepared request at the back of the I/O scheduler queue
1319	* for execution. Don't wait for completion.
1320	*
1321	* Note:
1322	* This function will invoke @done directly if the queue is dead.
1323	*/
1324	void blk_execute_rq_nowait(struct request *rq, bool at_head)
1325	{
1326	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1327
1328	WARN_ON(irqs_disabled());
1329	WARN_ON(!blk_rq_is_passthrough(rq));
1330
1331	blk_account_io_start(req: rq);
1332
1333	/*
1334	* As plugging can be enabled for passthrough requests on a zoned
1335	* device, directly accessing the plug instead of using blk_mq_plug()
1336	* should not have any consequences.
1337	*/
1338	if (current->plug && !at_head) {
1339	blk_add_rq_to_plug(current->plug, rq);
1340	return;
1341	}
1342
1343	blk_mq_insert_request(rq, flags: at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1344	blk_mq_run_hw_queue(hctx, async: hctx->flags & BLK_MQ_F_BLOCKING);
1345	}
1346	EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1347
1348	struct blk_rq_wait {
1349	struct completion done;
1350	blk_status_t ret;
1351	};
1352
1353	static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
1354	{
1355	struct blk_rq_wait *wait = rq->end_io_data;
1356
1357	wait->ret = ret;
1358	complete(&wait->done);
1359	return RQ_END_IO_NONE;
1360	}
1361
1362	bool blk_rq_is_poll(struct request *rq)
1363	{
1364	if (!rq->mq_hctx)
1365	return false;
1366	if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1367	return false;
1368	return true;
1369	}
1370	EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1371
1372	static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1373	{
1374	do {
1375	blk_hctx_poll(q: rq->q, hctx: rq->mq_hctx, NULL, flags: 0);
1376	cond_resched();
1377	} while (!completion_done(x: wait));
1378	}
1379
1380	/**
1381	* blk_execute_rq - insert a request into queue for execution
1382	* @rq: request to insert
1383	* @at_head: insert request at head or tail of queue
1384	*
1385	* Description:
1386	* Insert a fully prepared request at the back of the I/O scheduler queue
1387	* for execution and wait for completion.
1388	* Return: The blk_status_t result provided to blk_mq_end_request().
1389	*/
1390	blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1391	{
1392	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1393	struct blk_rq_wait wait = {
1394	.done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1395	};
1396
1397	WARN_ON(irqs_disabled());
1398	WARN_ON(!blk_rq_is_passthrough(rq));
1399
1400	rq->end_io_data = &wait;
1401	rq->end_io = blk_end_sync_rq;
1402
1403	blk_account_io_start(req: rq);
1404	blk_mq_insert_request(rq, flags: at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1405	blk_mq_run_hw_queue(hctx, async: false);
1406
1407	if (blk_rq_is_poll(rq)) {
1408	blk_rq_poll_completion(rq, wait: &wait.done);
1409	} else {
1410	/*
1411	* Prevent hang_check timer from firing at us during very long
1412	* I/O
1413	*/
1414	unsigned long hang_check = sysctl_hung_task_timeout_secs;
1415
1416	if (hang_check)
1417	while (!wait_for_completion_io_timeout(x: &wait.done,
1418	timeout: hang_check * (HZ/2)))
1419	;
1420	else
1421	wait_for_completion_io(&wait.done);
1422	}
1423
1424	return wait.ret;
1425	}
1426	EXPORT_SYMBOL(blk_execute_rq);
1427
1428	static void __blk_mq_requeue_request(struct request *rq)
1429	{
1430	struct request_queue *q = rq->q;
1431
1432	blk_mq_put_driver_tag(rq);
1433
1434	trace_block_rq_requeue(rq);
1435	rq_qos_requeue(q, rq);
1436
1437	if (blk_mq_request_started(rq)) {
1438	WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1439	rq->rq_flags &= ~RQF_TIMED_OUT;
1440	}
1441	}
1442
1443	void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1444	{
1445	struct request_queue *q = rq->q;
1446	unsigned long flags;
1447
1448	__blk_mq_requeue_request(rq);
1449
1450	/* this request will be re-inserted to io scheduler queue */
1451	blk_mq_sched_requeue_request(rq);
1452
1453	spin_lock_irqsave(&q->requeue_lock, flags);
1454	list_add_tail(new: &rq->queuelist, head: &q->requeue_list);
1455	spin_unlock_irqrestore(lock: &q->requeue_lock, flags);
1456
1457	if (kick_requeue_list)
1458	blk_mq_kick_requeue_list(q);
1459	}
1460	EXPORT_SYMBOL(blk_mq_requeue_request);
1461
1462	static void blk_mq_requeue_work(struct work_struct *work)
1463	{
1464	struct request_queue *q =
1465	container_of(work, struct request_queue, requeue_work.work);
1466	LIST_HEAD(rq_list);
1467	LIST_HEAD(flush_list);
1468	struct request *rq;
1469
1470	spin_lock_irq(lock: &q->requeue_lock);
1471	list_splice_init(list: &q->requeue_list, head: &rq_list);
1472	list_splice_init(list: &q->flush_list, head: &flush_list);
1473	spin_unlock_irq(lock: &q->requeue_lock);
1474
1475	while (!list_empty(head: &rq_list)) {
1476	rq = list_entry(rq_list.next, struct request, queuelist);
1477	/*
1478	* If RQF_DONTPREP ist set, the request has been started by the
1479	* driver already and might have driver-specific data allocated
1480	* already. Insert it into the hctx dispatch list to avoid
1481	* block layer merges for the request.
1482	*/
1483	if (rq->rq_flags & RQF_DONTPREP) {
1484	list_del_init(entry: &rq->queuelist);
1485	blk_mq_request_bypass_insert(rq, flags: 0);
1486	} else {
1487	list_del_init(entry: &rq->queuelist);
1488	blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD);
1489	}
1490	}
1491
1492	while (!list_empty(head: &flush_list)) {
1493	rq = list_entry(flush_list.next, struct request, queuelist);
1494	list_del_init(entry: &rq->queuelist);
1495	blk_mq_insert_request(rq, flags: 0);
1496	}
1497
1498	blk_mq_run_hw_queues(q, async: false);
1499	}
1500
1501	void blk_mq_kick_requeue_list(struct request_queue *q)
1502	{
1503	kblockd_mod_delayed_work_on(cpu: WORK_CPU_UNBOUND, dwork: &q->requeue_work, delay: 0);
1504	}
1505	EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1506
1507	void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1508	unsigned long msecs)
1509	{
1510	kblockd_mod_delayed_work_on(cpu: WORK_CPU_UNBOUND, dwork: &q->requeue_work,
1511	delay: msecs_to_jiffies(m: msecs));
1512	}
1513	EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1514
1515	static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1516	{
1517	/*
1518	* If we find a request that isn't idle we know the queue is busy
1519	* as it's checked in the iter.
1520	* Return false to stop the iteration.
1521	*/
1522	if (blk_mq_request_started(rq)) {
1523	bool *busy = priv;
1524
1525	*busy = true;
1526	return false;
1527	}
1528
1529	return true;
1530	}
1531
1532	bool blk_mq_queue_inflight(struct request_queue *q)
1533	{
1534	bool busy = false;
1535
1536	blk_mq_queue_tag_busy_iter(q, fn: blk_mq_rq_inflight, priv: &busy);
1537	return busy;
1538	}
1539	EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1540
1541	static void blk_mq_rq_timed_out(struct request *req)
1542	{
1543	req->rq_flags |= RQF_TIMED_OUT;
1544	if (req->q->mq_ops->timeout) {
1545	enum blk_eh_timer_return ret;
1546
1547	ret = req->q->mq_ops->timeout(req);
1548	if (ret == BLK_EH_DONE)
1549	return;
1550	WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1551	}
1552
1553	blk_add_timer(req);
1554	}
1555
1556	struct blk_expired_data {
1557	bool has_timedout_rq;
1558	unsigned long next;
1559	unsigned long timeout_start;
1560	};
1561
1562	static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
1563	{
1564	unsigned long deadline;
1565
1566	if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1567	return false;
1568	if (rq->rq_flags & RQF_TIMED_OUT)
1569	return false;
1570
1571	deadline = READ_ONCE(rq->deadline);
1572	if (time_after_eq(expired->timeout_start, deadline))
1573	return true;
1574
1575	if (expired->next == 0)
1576	expired->next = deadline;
1577	else if (time_after(expired->next, deadline))
1578	expired->next = deadline;
1579	return false;
1580	}
1581
1582	void blk_mq_put_rq_ref(struct request *rq)
1583	{
1584	if (is_flush_rq(req: rq)) {
1585	if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1586	blk_mq_free_request(rq);
1587	} else if (req_ref_put_and_test(req: rq)) {
1588	__blk_mq_free_request(rq);
1589	}
1590	}
1591
1592	static bool blk_mq_check_expired(struct request *rq, void *priv)
1593	{
1594	struct blk_expired_data *expired = priv;
1595
1596	/*
1597	* blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1598	* be reallocated underneath the timeout handler's processing, then
1599	* the expire check is reliable. If the request is not expired, then
1600	* it was completed and reallocated as a new request after returning
1601	* from blk_mq_check_expired().
1602	*/
1603	if (blk_mq_req_expired(rq, expired)) {
1604	expired->has_timedout_rq = true;
1605	return false;
1606	}
1607	return true;
1608	}
1609
1610	static bool blk_mq_handle_expired(struct request *rq, void *priv)
1611	{
1612	struct blk_expired_data *expired = priv;
1613
1614	if (blk_mq_req_expired(rq, expired))
1615	blk_mq_rq_timed_out(req: rq);
1616	return true;
1617	}
1618
1619	static void blk_mq_timeout_work(struct work_struct *work)
1620	{
1621	struct request_queue *q =
1622	container_of(work, struct request_queue, timeout_work);
1623	struct blk_expired_data expired = {
1624	.timeout_start = jiffies,
1625	};
1626	struct blk_mq_hw_ctx *hctx;
1627	unsigned long i;
1628
1629	/* A deadlock might occur if a request is stuck requiring a
1630	* timeout at the same time a queue freeze is waiting
1631	* completion, since the timeout code would not be able to
1632	* acquire the queue reference here.
1633	*
1634	* That's why we don't use blk_queue_enter here; instead, we use
1635	* percpu_ref_tryget directly, because we need to be able to
1636	* obtain a reference even in the short window between the queue
1637	* starting to freeze, by dropping the first reference in
1638	* blk_freeze_queue_start, and the moment the last request is
1639	* consumed, marked by the instant q_usage_counter reaches
1640	* zero.
1641	*/
1642	if (!percpu_ref_tryget(ref: &q->q_usage_counter))
1643	return;
1644
1645	/* check if there is any timed-out request */
1646	blk_mq_queue_tag_busy_iter(q, fn: blk_mq_check_expired, priv: &expired);
1647	if (expired.has_timedout_rq) {
1648	/*
1649	* Before walking tags, we must ensure any submit started
1650	* before the current time has finished. Since the submit
1651	* uses srcu or rcu, wait for a synchronization point to
1652	* ensure all running submits have finished
1653	*/
1654	blk_mq_wait_quiesce_done(q->tag_set);
1655
1656	expired.next = 0;
1657	blk_mq_queue_tag_busy_iter(q, fn: blk_mq_handle_expired, priv: &expired);
1658	}
1659
1660	if (expired.next != 0) {
1661	mod_timer(timer: &q->timeout, expires: expired.next);
1662	} else {
1663	/*
1664	* Request timeouts are handled as a forward rolling timer. If
1665	* we end up here it means that no requests are pending and
1666	* also that no request has been pending for a while. Mark
1667	* each hctx as idle.
1668	*/
1669	queue_for_each_hw_ctx(q, hctx, i) {
1670	/* the hctx may be unmapped, so check it here */
1671	if (blk_mq_hw_queue_mapped(hctx))
1672	blk_mq_tag_idle(hctx);
1673	}
1674	}
1675	blk_queue_exit(q);
1676	}
1677
1678	struct flush_busy_ctx_data {
1679	struct blk_mq_hw_ctx *hctx;
1680	struct list_head *list;
1681	};
1682
1683	static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1684	{
1685	struct flush_busy_ctx_data *flush_data = data;
1686	struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1687	struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1688	enum hctx_type type = hctx->type;
1689
1690	spin_lock(lock: &ctx->lock);
1691	list_splice_tail_init(list: &ctx->rq_lists[type], head: flush_data->list);
1692	sbitmap_clear_bit(sb, bitnr);
1693	spin_unlock(lock: &ctx->lock);
1694	return true;
1695	}
1696
1697	/*
1698	* Process software queues that have been marked busy, splicing them
1699	* to the for-dispatch
1700	*/
1701	void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1702	{
1703	struct flush_busy_ctx_data data = {
1704	.hctx = hctx,
1705	.list = list,
1706	};
1707
1708	sbitmap_for_each_set(sb: &hctx->ctx_map, fn: flush_busy_ctx, data: &data);
1709	}
1710	EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1711
1712	struct dispatch_rq_data {
1713	struct blk_mq_hw_ctx *hctx;
1714	struct request *rq;
1715	};
1716
1717	static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1718	void *data)
1719	{
1720	struct dispatch_rq_data *dispatch_data = data;
1721	struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1722	struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1723	enum hctx_type type = hctx->type;
1724
1725	spin_lock(lock: &ctx->lock);
1726	if (!list_empty(head: &ctx->rq_lists[type])) {
1727	dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1728	list_del_init(entry: &dispatch_data->rq->queuelist);
1729	if (list_empty(head: &ctx->rq_lists[type]))
1730	sbitmap_clear_bit(sb, bitnr);
1731	}
1732	spin_unlock(lock: &ctx->lock);
1733
1734	return !dispatch_data->rq;
1735	}
1736
1737	struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1738	struct blk_mq_ctx *start)
1739	{
1740	unsigned off = start ? start->index_hw[hctx->type] : 0;
1741	struct dispatch_rq_data data = {
1742	.hctx = hctx,
1743	.rq = NULL,
1744	};
1745
1746	__sbitmap_for_each_set(sb: &hctx->ctx_map, start: off,
1747	fn: dispatch_rq_from_ctx, data: &data);
1748
1749	return data.rq;
1750	}
1751
1752	bool __blk_mq_alloc_driver_tag(struct request *rq)
1753	{
1754	struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1755	unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1756	int tag;
1757
1758	blk_mq_tag_busy(hctx: rq->mq_hctx);
1759
1760	if (blk_mq_tag_is_reserved(tags: rq->mq_hctx->sched_tags, tag: rq->internal_tag)) {
1761	bt = &rq->mq_hctx->tags->breserved_tags;
1762	tag_offset = 0;
1763	} else {
1764	if (!hctx_may_queue(hctx: rq->mq_hctx, bt))
1765	return false;
1766	}
1767
1768	tag = __sbitmap_queue_get(sbq: bt);
1769	if (tag == BLK_MQ_NO_TAG)
1770	return false;
1771
1772	rq->tag = tag + tag_offset;
1773	blk_mq_inc_active_requests(hctx: rq->mq_hctx);
1774	return true;
1775	}
1776
1777	static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1778	int flags, void *key)
1779	{
1780	struct blk_mq_hw_ctx *hctx;
1781
1782	hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1783
1784	spin_lock(lock: &hctx->dispatch_wait_lock);
1785	if (!list_empty(head: &wait->entry)) {
1786	struct sbitmap_queue *sbq;
1787
1788	list_del_init(entry: &wait->entry);
1789	sbq = &hctx->tags->bitmap_tags;
1790	atomic_dec(v: &sbq->ws_active);
1791	}
1792	spin_unlock(lock: &hctx->dispatch_wait_lock);
1793
1794	blk_mq_run_hw_queue(hctx, async: true);
1795	return 1;
1796	}
1797
1798	/*
1799	* Mark us waiting for a tag. For shared tags, this involves hooking us into
1800	* the tag wakeups. For non-shared tags, we can simply mark us needing a
1801	* restart. For both cases, take care to check the condition again after
1802	* marking us as waiting.
1803	*/
1804	static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1805	struct request *rq)
1806	{
1807	struct sbitmap_queue *sbq;
1808	struct wait_queue_head *wq;
1809	wait_queue_entry_t *wait;
1810	bool ret;
1811
1812	if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1813	!(blk_mq_is_shared_tags(flags: hctx->flags))) {
1814	blk_mq_sched_mark_restart_hctx(hctx);
1815
1816	/*
1817	* It's possible that a tag was freed in the window between the
1818	* allocation failure and adding the hardware queue to the wait
1819	* queue.
1820	*
1821	* Don't clear RESTART here, someone else could have set it.
1822	* At most this will cost an extra queue run.
1823	*/
1824	return blk_mq_get_driver_tag(rq);
1825	}
1826
1827	wait = &hctx->dispatch_wait;
1828	if (!list_empty_careful(head: &wait->entry))
1829	return false;
1830
1831	if (blk_mq_tag_is_reserved(tags: rq->mq_hctx->sched_tags, tag: rq->internal_tag))
1832	sbq = &hctx->tags->breserved_tags;
1833	else
1834	sbq = &hctx->tags->bitmap_tags;
1835	wq = &bt_wait_ptr(bt: sbq, hctx)->wait;
1836
1837	spin_lock_irq(lock: &wq->lock);
1838	spin_lock(lock: &hctx->dispatch_wait_lock);
1839	if (!list_empty(head: &wait->entry)) {
1840	spin_unlock(lock: &hctx->dispatch_wait_lock);
1841	spin_unlock_irq(lock: &wq->lock);
1842	return false;
1843	}
1844
1845	atomic_inc(v: &sbq->ws_active);
1846	wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1847	__add_wait_queue(wq_head: wq, wq_entry: wait);
1848
1849	/*
1850	* It's possible that a tag was freed in the window between the
1851	* allocation failure and adding the hardware queue to the wait
1852	* queue.
1853	*/
1854	ret = blk_mq_get_driver_tag(rq);
1855	if (!ret) {
1856	spin_unlock(lock: &hctx->dispatch_wait_lock);
1857	spin_unlock_irq(lock: &wq->lock);
1858	return false;
1859	}
1860
1861	/*
1862	* We got a tag, remove ourselves from the wait queue to ensure
1863	* someone else gets the wakeup.
1864	*/
1865	list_del_init(entry: &wait->entry);
1866	atomic_dec(v: &sbq->ws_active);
1867	spin_unlock(lock: &hctx->dispatch_wait_lock);
1868	spin_unlock_irq(lock: &wq->lock);
1869
1870	return true;
1871	}
1872
1873	#define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1874	#define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1875	/*
1876	* Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1877	* - EWMA is one simple way to compute running average value
1878	* - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1879	* - take 4 as factor for avoiding to get too small(0) result, and this
1880	* factor doesn't matter because EWMA decreases exponentially
1881	*/
1882	static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1883	{
1884	unsigned int ewma;
1885
1886	ewma = hctx->dispatch_busy;
1887
1888	if (!ewma && !busy)
1889	return;
1890
1891	ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1892	if (busy)
1893	ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1894	ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1895
1896	hctx->dispatch_busy = ewma;
1897	}
1898
1899	#define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1900
1901	static void blk_mq_handle_dev_resource(struct request *rq,
1902	struct list_head *list)
1903	{
1904	list_add(new: &rq->queuelist, head: list);
1905	__blk_mq_requeue_request(rq);
1906	}
1907
1908	static void blk_mq_handle_zone_resource(struct request *rq,
1909	struct list_head *zone_list)
1910	{
1911	/*
1912	* If we end up here it is because we cannot dispatch a request to a
1913	* specific zone due to LLD level zone-write locking or other zone
1914	* related resource not being available. In this case, set the request
1915	* aside in zone_list for retrying it later.
1916	*/
1917	list_add(new: &rq->queuelist, head: zone_list);
1918	__blk_mq_requeue_request(rq);
1919	}
1920
1921	enum prep_dispatch {
1922	PREP_DISPATCH_OK,
1923	PREP_DISPATCH_NO_TAG,
1924	PREP_DISPATCH_NO_BUDGET,
1925	};
1926
1927	static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1928	bool need_budget)
1929	{
1930	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1931	int budget_token = -1;
1932
1933	if (need_budget) {
1934	budget_token = blk_mq_get_dispatch_budget(q: rq->q);
1935	if (budget_token < 0) {
1936	blk_mq_put_driver_tag(rq);
1937	return PREP_DISPATCH_NO_BUDGET;
1938	}
1939	blk_mq_set_rq_budget_token(rq, token: budget_token);
1940	}
1941
1942	if (!blk_mq_get_driver_tag(rq)) {
1943	/*
1944	* The initial allocation attempt failed, so we need to
1945	* rerun the hardware queue when a tag is freed. The
1946	* waitqueue takes care of that. If the queue is run
1947	* before we add this entry back on the dispatch list,
1948	* we'll re-run it below.
1949	*/
1950	if (!blk_mq_mark_tag_wait(hctx, rq)) {
1951	/*
1952	* All budgets not got from this function will be put
1953	* together during handling partial dispatch
1954	*/
1955	if (need_budget)
1956	blk_mq_put_dispatch_budget(q: rq->q, budget_token);
1957	return PREP_DISPATCH_NO_TAG;
1958	}
1959	}
1960
1961	return PREP_DISPATCH_OK;
1962	}
1963
1964	/* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1965	static void blk_mq_release_budgets(struct request_queue *q,
1966	struct list_head *list)
1967	{
1968	struct request *rq;
1969
1970	list_for_each_entry(rq, list, queuelist) {
1971	int budget_token = blk_mq_get_rq_budget_token(rq);
1972
1973	if (budget_token >= 0)
1974	blk_mq_put_dispatch_budget(q, budget_token);
1975	}
1976	}
1977
1978	/*
1979	* blk_mq_commit_rqs will notify driver using bd->last that there is no
1980	* more requests. (See comment in struct blk_mq_ops for commit_rqs for
1981	* details)
1982	* Attention, we should explicitly call this in unusual cases:
1983	* 1) did not queue everything initially scheduled to queue
1984	* 2) the last attempt to queue a request failed
1985	*/
1986	static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued,
1987	bool from_schedule)
1988	{
1989	if (hctx->queue->mq_ops->commit_rqs && queued) {
1990	trace_block_unplug(q: hctx->queue, depth: queued, explicit: !from_schedule);
1991	hctx->queue->mq_ops->commit_rqs(hctx);
1992	}
1993	}
1994
1995	/*
1996	* Returns true if we did some work AND can potentially do more.
1997	*/
1998	bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1999	unsigned int nr_budgets)
2000	{
2001	enum prep_dispatch prep;
2002	struct request_queue *q = hctx->queue;
2003	struct request *rq;
2004	int queued;
2005	blk_status_t ret = BLK_STS_OK;
2006	LIST_HEAD(zone_list);
2007	bool needs_resource = false;
2008
2009	if (list_empty(head: list))
2010	return false;
2011
2012	/*
2013	* Now process all the entries, sending them to the driver.
2014	*/
2015	queued = 0;
2016	do {
2017	struct blk_mq_queue_data bd;
2018
2019	rq = list_first_entry(list, struct request, queuelist);
2020
2021	WARN_ON_ONCE(hctx != rq->mq_hctx);
2022	prep = blk_mq_prep_dispatch_rq(rq, need_budget: !nr_budgets);
2023	if (prep != PREP_DISPATCH_OK)
2024	break;
2025
2026	list_del_init(entry: &rq->queuelist);
2027
2028	bd.rq = rq;
2029	bd.last = list_empty(head: list);
2030
2031	/*
2032	* once the request is queued to lld, no need to cover the
2033	* budget any more
2034	*/
2035	if (nr_budgets)
2036	nr_budgets--;
2037	ret = q->mq_ops->queue_rq(hctx, &bd);
2038	switch (ret) {
2039	case BLK_STS_OK:
2040	queued++;
2041	break;
2042	case BLK_STS_RESOURCE:
2043	needs_resource = true;
2044	fallthrough;
2045	case BLK_STS_DEV_RESOURCE:
2046	blk_mq_handle_dev_resource(rq, list);
2047	goto out;
2048	case BLK_STS_ZONE_RESOURCE:
2049	/*
2050	* Move the request to zone_list and keep going through
2051	* the dispatch list to find more requests the drive can
2052	* accept.
2053	*/
2054	blk_mq_handle_zone_resource(rq, zone_list: &zone_list);
2055	needs_resource = true;
2056	break;
2057	default:
2058	blk_mq_end_request(rq, ret);
2059	}
2060	} while (!list_empty(head: list));
2061	out:
2062	if (!list_empty(head: &zone_list))
2063	list_splice_tail_init(list: &zone_list, head: list);
2064
2065	/* If we didn't flush the entire list, we could have told the driver
2066	* there was more coming, but that turned out to be a lie.
2067	*/
2068	if (!list_empty(head: list) || ret != BLK_STS_OK)
2069	blk_mq_commit_rqs(hctx, queued, from_schedule: false);
2070
2071	/*
2072	* Any items that need requeuing? Stuff them into hctx->dispatch,
2073	* that is where we will continue on next queue run.
2074	*/
2075	if (!list_empty(head: list)) {
2076	bool needs_restart;
2077	/* For non-shared tags, the RESTART check will suffice */
2078	bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2079	((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
2080	blk_mq_is_shared_tags(flags: hctx->flags));
2081
2082	if (nr_budgets)
2083	blk_mq_release_budgets(q, list);
2084
2085	spin_lock(lock: &hctx->lock);
2086	list_splice_tail_init(list, head: &hctx->dispatch);
2087	spin_unlock(lock: &hctx->lock);
2088
2089	/*
2090	* Order adding requests to hctx->dispatch and checking
2091	* SCHED_RESTART flag. The pair of this smp_mb() is the one
2092	* in blk_mq_sched_restart(). Avoid restart code path to
2093	* miss the new added requests to hctx->dispatch, meantime
2094	* SCHED_RESTART is observed here.
2095	*/
2096	smp_mb();
2097
2098	/*
2099	* If SCHED_RESTART was set by the caller of this function and
2100	* it is no longer set that means that it was cleared by another
2101	* thread and hence that a queue rerun is needed.
2102	*
2103	* If 'no_tag' is set, that means that we failed getting
2104	* a driver tag with an I/O scheduler attached. If our dispatch
2105	* waitqueue is no longer active, ensure that we run the queue
2106	* AFTER adding our entries back to the list.
2107	*
2108	* If no I/O scheduler has been configured it is possible that
2109	* the hardware queue got stopped and restarted before requests
2110	* were pushed back onto the dispatch list. Rerun the queue to
2111	* avoid starvation. Notes:
2112	* - blk_mq_run_hw_queue() checks whether or not a queue has
2113	* been stopped before rerunning a queue.
2114	* - Some but not all block drivers stop a queue before
2115	* returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2116	* and dm-rq.
2117	*
2118	* If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2119	* bit is set, run queue after a delay to avoid IO stalls
2120	* that could otherwise occur if the queue is idle. We'll do
2121	* similar if we couldn't get budget or couldn't lock a zone
2122	* and SCHED_RESTART is set.
2123	*/
2124	needs_restart = blk_mq_sched_needs_restart(hctx);
2125	if (prep == PREP_DISPATCH_NO_BUDGET)
2126	needs_resource = true;
2127	if (!needs_restart ||
2128	(no_tag && list_empty_careful(head: &hctx->dispatch_wait.entry)))
2129	blk_mq_run_hw_queue(hctx, async: true);
2130	else if (needs_resource)
2131	blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2132
2133	blk_mq_update_dispatch_busy(hctx, busy: true);
2134	return false;
2135	}
2136
2137	blk_mq_update_dispatch_busy(hctx, busy: false);
2138	return true;
2139	}
2140
2141	static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2142	{
2143	int cpu = cpumask_first_and(srcp1: hctx->cpumask, cpu_online_mask);
2144
2145	if (cpu >= nr_cpu_ids)
2146	cpu = cpumask_first(srcp: hctx->cpumask);
2147	return cpu;
2148	}
2149
2150	/*
2151	* It'd be great if the workqueue API had a way to pass
2152	* in a mask and had some smarts for more clever placement.
2153	* For now we just round-robin here, switching for every
2154	* BLK_MQ_CPU_WORK_BATCH queued items.
2155	*/
2156	static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2157	{
2158	bool tried = false;
2159	int next_cpu = hctx->next_cpu;
2160
2161	if (hctx->queue->nr_hw_queues == 1)
2162	return WORK_CPU_UNBOUND;
2163
2164	if (--hctx->next_cpu_batch <= 0) {
2165	select_cpu:
2166	next_cpu = cpumask_next_and(n: next_cpu, src1p: hctx->cpumask,
2167	cpu_online_mask);
2168	if (next_cpu >= nr_cpu_ids)
2169	next_cpu = blk_mq_first_mapped_cpu(hctx);
2170	hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2171	}
2172
2173	/*
2174	* Do unbound schedule if we can't find a online CPU for this hctx,
2175	* and it should only happen in the path of handling CPU DEAD.
2176	*/
2177	if (!cpu_online(cpu: next_cpu)) {
2178	if (!tried) {
2179	tried = true;
2180	goto select_cpu;
2181	}
2182
2183	/*
2184	* Make sure to re-select CPU next time once after CPUs
2185	* in hctx->cpumask become online again.
2186	*/
2187	hctx->next_cpu = next_cpu;
2188	hctx->next_cpu_batch = 1;
2189	return WORK_CPU_UNBOUND;
2190	}
2191
2192	hctx->next_cpu = next_cpu;
2193	return next_cpu;
2194	}
2195
2196	/**
2197	* blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2198	* @hctx: Pointer to the hardware queue to run.
2199	* @msecs: Milliseconds of delay to wait before running the queue.
2200	*
2201	* Run a hardware queue asynchronously with a delay of @msecs.
2202	*/
2203	void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2204	{
2205	if (unlikely(blk_mq_hctx_stopped(hctx)))
2206	return;
2207	kblockd_mod_delayed_work_on(cpu: blk_mq_hctx_next_cpu(hctx), dwork: &hctx->run_work,
2208	delay: msecs_to_jiffies(m: msecs));
2209	}
2210	EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2211
2212	/**
2213	* blk_mq_run_hw_queue - Start to run a hardware queue.
2214	* @hctx: Pointer to the hardware queue to run.
2215	* @async: If we want to run the queue asynchronously.
2216	*
2217	* Check if the request queue is not in a quiesced state and if there are
2218	* pending requests to be sent. If this is true, run the queue to send requests
2219	* to hardware.
2220	*/
2221	void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2222	{
2223	bool need_run;
2224
2225	/*
2226	* We can't run the queue inline with interrupts disabled.
2227	*/
2228	WARN_ON_ONCE(!async && in_interrupt());
2229
2230	might_sleep_if(!async && hctx->flags & BLK_MQ_F_BLOCKING);
2231
2232	/*
2233	* When queue is quiesced, we may be switching io scheduler, or
2234	* updating nr_hw_queues, or other things, and we can't run queue
2235	* any more, even __blk_mq_hctx_has_pending() can't be called safely.
2236	*
2237	* And queue will be rerun in blk_mq_unquiesce_queue() if it is
2238	* quiesced.
2239	*/
2240	__blk_mq_run_dispatch_ops(hctx->queue, false,
2241	need_run = !blk_queue_quiesced(hctx->queue) &&
2242	blk_mq_hctx_has_pending(hctx));
2243
2244	if (!need_run)
2245	return;
2246
2247	if (async || !cpumask_test_cpu(raw_smp_processor_id(), cpumask: hctx->cpumask)) {
2248	blk_mq_delay_run_hw_queue(hctx, 0);
2249	return;
2250	}
2251
2252	blk_mq_run_dispatch_ops(hctx->queue,
2253	blk_mq_sched_dispatch_requests(hctx));
2254	}
2255	EXPORT_SYMBOL(blk_mq_run_hw_queue);
2256
2257	/*
2258	* Return prefered queue to dispatch from (if any) for non-mq aware IO
2259	* scheduler.
2260	*/
2261	static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2262	{
2263	struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2264	/*
2265	* If the IO scheduler does not respect hardware queues when
2266	* dispatching, we just don't bother with multiple HW queues and
2267	* dispatch from hctx for the current CPU since running multiple queues
2268	* just causes lock contention inside the scheduler and pointless cache
2269	* bouncing.
2270	*/
2271	struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2272
2273	if (!blk_mq_hctx_stopped(hctx))
2274	return hctx;
2275	return NULL;
2276	}
2277
2278	/**
2279	* blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2280	* @q: Pointer to the request queue to run.
2281	* @async: If we want to run the queue asynchronously.
2282	*/
2283	void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2284	{
2285	struct blk_mq_hw_ctx *hctx, *sq_hctx;
2286	unsigned long i;
2287
2288	sq_hctx = NULL;
2289	if (blk_queue_sq_sched(q))
2290	sq_hctx = blk_mq_get_sq_hctx(q);
2291	queue_for_each_hw_ctx(q, hctx, i) {
2292	if (blk_mq_hctx_stopped(hctx))
2293	continue;
2294	/*
2295	* Dispatch from this hctx either if there's no hctx preferred
2296	* by IO scheduler or if it has requests that bypass the
2297	* scheduler.
2298	*/
2299	if (!sq_hctx || sq_hctx == hctx ||
2300	!list_empty_careful(head: &hctx->dispatch))
2301	blk_mq_run_hw_queue(hctx, async);
2302	}
2303	}
2304	EXPORT_SYMBOL(blk_mq_run_hw_queues);
2305
2306	/**
2307	* blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2308	* @q: Pointer to the request queue to run.
2309	* @msecs: Milliseconds of delay to wait before running the queues.
2310	*/
2311	void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2312	{
2313	struct blk_mq_hw_ctx *hctx, *sq_hctx;
2314	unsigned long i;
2315
2316	sq_hctx = NULL;
2317	if (blk_queue_sq_sched(q))
2318	sq_hctx = blk_mq_get_sq_hctx(q);
2319	queue_for_each_hw_ctx(q, hctx, i) {
2320	if (blk_mq_hctx_stopped(hctx))
2321	continue;
2322	/*
2323	* If there is already a run_work pending, leave the
2324	* pending delay untouched. Otherwise, a hctx can stall
2325	* if another hctx is re-delaying the other's work
2326	* before the work executes.
2327	*/
2328	if (delayed_work_pending(&hctx->run_work))
2329	continue;
2330	/*
2331	* Dispatch from this hctx either if there's no hctx preferred
2332	* by IO scheduler or if it has requests that bypass the
2333	* scheduler.
2334	*/
2335	if (!sq_hctx || sq_hctx == hctx ||
2336	!list_empty_careful(head: &hctx->dispatch))
2337	blk_mq_delay_run_hw_queue(hctx, msecs);
2338	}
2339	}
2340	EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2341
2342	/*
2343	* This function is often used for pausing .queue_rq() by driver when
2344	* there isn't enough resource or some conditions aren't satisfied, and
2345	* BLK_STS_RESOURCE is usually returned.
2346	*
2347	* We do not guarantee that dispatch can be drained or blocked
2348	* after blk_mq_stop_hw_queue() returns. Please use
2349	* blk_mq_quiesce_queue() for that requirement.
2350	*/
2351	void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2352	{
2353	cancel_delayed_work(dwork: &hctx->run_work);
2354
2355	set_bit(nr: BLK_MQ_S_STOPPED, addr: &hctx->state);
2356	}
2357	EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2358
2359	/*
2360	* This function is often used for pausing .queue_rq() by driver when
2361	* there isn't enough resource or some conditions aren't satisfied, and
2362	* BLK_STS_RESOURCE is usually returned.
2363	*
2364	* We do not guarantee that dispatch can be drained or blocked
2365	* after blk_mq_stop_hw_queues() returns. Please use
2366	* blk_mq_quiesce_queue() for that requirement.
2367	*/
2368	void blk_mq_stop_hw_queues(struct request_queue *q)
2369	{
2370	struct blk_mq_hw_ctx *hctx;
2371	unsigned long i;
2372
2373	queue_for_each_hw_ctx(q, hctx, i)
2374	blk_mq_stop_hw_queue(hctx);
2375	}
2376	EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2377
2378	void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2379	{
2380	clear_bit(nr: BLK_MQ_S_STOPPED, addr: &hctx->state);
2381
2382	blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);
2383	}
2384	EXPORT_SYMBOL(blk_mq_start_hw_queue);
2385
2386	void blk_mq_start_hw_queues(struct request_queue *q)
2387	{
2388	struct blk_mq_hw_ctx *hctx;
2389	unsigned long i;
2390
2391	queue_for_each_hw_ctx(q, hctx, i)
2392	blk_mq_start_hw_queue(hctx);
2393	}
2394	EXPORT_SYMBOL(blk_mq_start_hw_queues);
2395
2396	void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2397	{
2398	if (!blk_mq_hctx_stopped(hctx))
2399	return;
2400
2401	clear_bit(nr: BLK_MQ_S_STOPPED, addr: &hctx->state);
2402	blk_mq_run_hw_queue(hctx, async);
2403	}
2404	EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2405
2406	void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2407	{
2408	struct blk_mq_hw_ctx *hctx;
2409	unsigned long i;
2410
2411	queue_for_each_hw_ctx(q, hctx, i)
2412	blk_mq_start_stopped_hw_queue(hctx, async ||
2413	(hctx->flags & BLK_MQ_F_BLOCKING));
2414	}
2415	EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2416
2417	static void blk_mq_run_work_fn(struct work_struct *work)
2418	{
2419	struct blk_mq_hw_ctx *hctx =
2420	container_of(work, struct blk_mq_hw_ctx, run_work.work);
2421
2422	blk_mq_run_dispatch_ops(hctx->queue,
2423	blk_mq_sched_dispatch_requests(hctx));
2424	}
2425
2426	/**
2427	* blk_mq_request_bypass_insert - Insert a request at dispatch list.
2428	* @rq: Pointer to request to be inserted.
2429	* @flags: BLK_MQ_INSERT_*
2430	*
2431	* Should only be used carefully, when the caller knows we want to
2432	* bypass a potential IO scheduler on the target device.
2433	*/
2434	static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags)
2435	{
2436	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2437
2438	spin_lock(lock: &hctx->lock);
2439	if (flags & BLK_MQ_INSERT_AT_HEAD)
2440	list_add(new: &rq->queuelist, head: &hctx->dispatch);
2441	else
2442	list_add_tail(new: &rq->queuelist, head: &hctx->dispatch);
2443	spin_unlock(lock: &hctx->lock);
2444	}
2445
2446	static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx,
2447	struct blk_mq_ctx *ctx, struct list_head *list,
2448	bool run_queue_async)
2449	{
2450	struct request *rq;
2451	enum hctx_type type = hctx->type;
2452
2453	/*
2454	* Try to issue requests directly if the hw queue isn't busy to save an
2455	* extra enqueue & dequeue to the sw queue.
2456	*/
2457	if (!hctx->dispatch_busy && !run_queue_async) {
2458	blk_mq_run_dispatch_ops(hctx->queue,
2459	blk_mq_try_issue_list_directly(hctx, list));
2460	if (list_empty(head: list))
2461	goto out;
2462	}
2463
2464	/*
2465	* preemption doesn't flush plug list, so it's possible ctx->cpu is
2466	* offline now
2467	*/
2468	list_for_each_entry(rq, list, queuelist) {
2469	BUG_ON(rq->mq_ctx != ctx);
2470	trace_block_rq_insert(rq);
2471	if (rq->cmd_flags & REQ_NOWAIT)
2472	run_queue_async = true;
2473	}
2474
2475	spin_lock(lock: &ctx->lock);
2476	list_splice_tail_init(list, head: &ctx->rq_lists[type]);
2477	blk_mq_hctx_mark_pending(hctx, ctx);
2478	spin_unlock(lock: &ctx->lock);
2479	out:
2480	blk_mq_run_hw_queue(hctx, run_queue_async);
2481	}
2482
2483	static void blk_mq_insert_request(struct request *rq, blk_insert_t flags)
2484	{
2485	struct request_queue *q = rq->q;
2486	struct blk_mq_ctx *ctx = rq->mq_ctx;
2487	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2488
2489	if (blk_rq_is_passthrough(rq)) {
2490	/*
2491	* Passthrough request have to be added to hctx->dispatch
2492	* directly. The device may be in a situation where it can't
2493	* handle FS request, and always returns BLK_STS_RESOURCE for
2494	* them, which gets them added to hctx->dispatch.
2495	*
2496	* If a passthrough request is required to unblock the queues,
2497	* and it is added to the scheduler queue, there is no chance to
2498	* dispatch it given we prioritize requests in hctx->dispatch.
2499	*/
2500	blk_mq_request_bypass_insert(rq, flags);
2501	} else if (req_op(req: rq) == REQ_OP_FLUSH) {
2502	/*
2503	* Firstly normal IO request is inserted to scheduler queue or
2504	* sw queue, meantime we add flush request to dispatch queue(
2505	* hctx->dispatch) directly and there is at most one in-flight
2506	* flush request for each hw queue, so it doesn't matter to add
2507	* flush request to tail or front of the dispatch queue.
2508	*
2509	* Secondly in case of NCQ, flush request belongs to non-NCQ
2510	* command, and queueing it will fail when there is any
2511	* in-flight normal IO request(NCQ command). When adding flush
2512	* rq to the front of hctx->dispatch, it is easier to introduce
2513	* extra time to flush rq's latency because of S_SCHED_RESTART
2514	* compared with adding to the tail of dispatch queue, then
2515	* chance of flush merge is increased, and less flush requests
2516	* will be issued to controller. It is observed that ~10% time
2517	* is saved in blktests block/004 on disk attached to AHCI/NCQ
2518	* drive when adding flush rq to the front of hctx->dispatch.
2519	*
2520	* Simply queue flush rq to the front of hctx->dispatch so that
2521	* intensive flush workloads can benefit in case of NCQ HW.
2522	*/
2523	blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD);
2524	} else if (q->elevator) {
2525	LIST_HEAD(list);
2526
2527	WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG);
2528
2529	list_add(new: &rq->queuelist, head: &list);
2530	q->elevator->type->ops.insert_requests(hctx, &list, flags);
2531	} else {
2532	trace_block_rq_insert(rq);
2533
2534	spin_lock(lock: &ctx->lock);
2535	if (flags & BLK_MQ_INSERT_AT_HEAD)
2536	list_add(new: &rq->queuelist, head: &ctx->rq_lists[hctx->type]);
2537	else
2538	list_add_tail(new: &rq->queuelist,
2539	head: &ctx->rq_lists[hctx->type]);
2540	blk_mq_hctx_mark_pending(hctx, ctx);
2541	spin_unlock(lock: &ctx->lock);
2542	}
2543	}
2544
2545	static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2546	unsigned int nr_segs)
2547	{
2548	int err;
2549
2550	if (bio->bi_opf & REQ_RAHEAD)
2551	rq->cmd_flags |= REQ_FAILFAST_MASK;
2552
2553	rq->__sector = bio->bi_iter.bi_sector;
2554	blk_rq_bio_prep(rq, bio, nr_segs);
2555
2556	/* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2557	err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2558	WARN_ON_ONCE(err);
2559
2560	blk_account_io_start(req: rq);
2561	}
2562
2563	static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2564	struct request *rq, bool last)
2565	{
2566	struct request_queue *q = rq->q;
2567	struct blk_mq_queue_data bd = {
2568	.rq = rq,
2569	.last = last,
2570	};
2571	blk_status_t ret;
2572
2573	/*
2574	* For OK queue, we are done. For error, caller may kill it.
2575	* Any other error (busy), just add it to our list as we
2576	* previously would have done.
2577	*/
2578	ret = q->mq_ops->queue_rq(hctx, &bd);
2579	switch (ret) {
2580	case BLK_STS_OK:
2581	blk_mq_update_dispatch_busy(hctx, busy: false);
2582	break;
2583	case BLK_STS_RESOURCE:
2584	case BLK_STS_DEV_RESOURCE:
2585	blk_mq_update_dispatch_busy(hctx, busy: true);
2586	__blk_mq_requeue_request(rq);
2587	break;
2588	default:
2589	blk_mq_update_dispatch_busy(hctx, busy: false);
2590	break;
2591	}
2592
2593	return ret;
2594	}
2595
2596	static bool blk_mq_get_budget_and_tag(struct request *rq)
2597	{
2598	int budget_token;
2599
2600	budget_token = blk_mq_get_dispatch_budget(q: rq->q);
2601	if (budget_token < 0)
2602	return false;
2603	blk_mq_set_rq_budget_token(rq, token: budget_token);
2604	if (!blk_mq_get_driver_tag(rq)) {
2605	blk_mq_put_dispatch_budget(q: rq->q, budget_token);
2606	return false;
2607	}
2608	return true;
2609	}
2610
2611	/**
2612	* blk_mq_try_issue_directly - Try to send a request directly to device driver.
2613	* @hctx: Pointer of the associated hardware queue.
2614	* @rq: Pointer to request to be sent.
2615	*
2616	* If the device has enough resources to accept a new request now, send the
2617	* request directly to device driver. Else, insert at hctx->dispatch queue, so
2618	* we can try send it another time in the future. Requests inserted at this
2619	* queue have higher priority.
2620	*/
2621	static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2622	struct request *rq)
2623	{
2624	blk_status_t ret;
2625
2626	if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2627	blk_mq_insert_request(rq, flags: 0);
2628	return;
2629	}
2630
2631	if ((rq->rq_flags & RQF_USE_SCHED) || !blk_mq_get_budget_and_tag(rq)) {
2632	blk_mq_insert_request(rq, flags: 0);
2633	blk_mq_run_hw_queue(hctx, rq->cmd_flags & REQ_NOWAIT);
2634	return;
2635	}
2636
2637	ret = __blk_mq_issue_directly(hctx, rq, last: true);
2638	switch (ret) {
2639	case BLK_STS_OK:
2640	break;
2641	case BLK_STS_RESOURCE:
2642	case BLK_STS_DEV_RESOURCE:
2643	blk_mq_request_bypass_insert(rq, flags: 0);
2644	blk_mq_run_hw_queue(hctx, false);
2645	break;
2646	default:
2647	blk_mq_end_request(rq, ret);
2648	break;
2649	}
2650	}
2651
2652	static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2653	{
2654	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2655
2656	if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2657	blk_mq_insert_request(rq, flags: 0);
2658	return BLK_STS_OK;
2659	}
2660
2661	if (!blk_mq_get_budget_and_tag(rq))
2662	return BLK_STS_RESOURCE;
2663	return __blk_mq_issue_directly(hctx, rq, last);
2664	}
2665
2666	static void blk_mq_plug_issue_direct(struct blk_plug *plug)
2667	{
2668	struct blk_mq_hw_ctx *hctx = NULL;
2669	struct request *rq;
2670	int queued = 0;
2671	blk_status_t ret = BLK_STS_OK;
2672
2673	while ((rq = rq_list_pop(&plug->mq_list))) {
2674	bool last = rq_list_empty(plug->mq_list);
2675
2676	if (hctx != rq->mq_hctx) {
2677	if (hctx) {
2678	blk_mq_commit_rqs(hctx, queued, from_schedule: false);
2679	queued = 0;
2680	}
2681	hctx = rq->mq_hctx;
2682	}
2683
2684	ret = blk_mq_request_issue_directly(rq, last);
2685	switch (ret) {
2686	case BLK_STS_OK:
2687	queued++;
2688	break;
2689	case BLK_STS_RESOURCE:
2690	case BLK_STS_DEV_RESOURCE:
2691	blk_mq_request_bypass_insert(rq, flags: 0);
2692	blk_mq_run_hw_queue(hctx, false);
2693	goto out;
2694	default:
2695	blk_mq_end_request(rq, ret);
2696	break;
2697	}
2698	}
2699
2700	out:
2701	if (ret != BLK_STS_OK)
2702	blk_mq_commit_rqs(hctx, queued, from_schedule: false);
2703	}
2704
2705	static void __blk_mq_flush_plug_list(struct request_queue *q,
2706	struct blk_plug *plug)
2707	{
2708	if (blk_queue_quiesced(q))
2709	return;
2710	q->mq_ops->queue_rqs(&plug->mq_list);
2711	}
2712
2713	static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2714	{
2715	struct blk_mq_hw_ctx *this_hctx = NULL;
2716	struct blk_mq_ctx *this_ctx = NULL;
2717	struct request *requeue_list = NULL;
2718	struct request **requeue_lastp = &requeue_list;
2719	unsigned int depth = 0;
2720	bool is_passthrough = false;
2721	LIST_HEAD(list);
2722
2723	do {
2724	struct request *rq = rq_list_pop(&plug->mq_list);
2725
2726	if (!this_hctx) {
2727	this_hctx = rq->mq_hctx;
2728	this_ctx = rq->mq_ctx;
2729	is_passthrough = blk_rq_is_passthrough(rq);
2730	} else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx ||
2731	is_passthrough != blk_rq_is_passthrough(rq)) {
2732	rq_list_add_tail(&requeue_lastp, rq);
2733	continue;
2734	}
2735	list_add(new: &rq->queuelist, head: &list);
2736	depth++;
2737	} while (!rq_list_empty(plug->mq_list));
2738
2739	plug->mq_list = requeue_list;
2740	trace_block_unplug(q: this_hctx->queue, depth, explicit: !from_sched);
2741
2742	percpu_ref_get(ref: &this_hctx->queue->q_usage_counter);
2743	/* passthrough requests should never be issued to the I/O scheduler */
2744	if (is_passthrough) {
2745	spin_lock(lock: &this_hctx->lock);
2746	list_splice_tail_init(list: &list, head: &this_hctx->dispatch);
2747	spin_unlock(lock: &this_hctx->lock);
2748	blk_mq_run_hw_queue(this_hctx, from_sched);
2749	} else if (this_hctx->queue->elevator) {
2750	this_hctx->queue->elevator->type->ops.insert_requests(this_hctx,
2751	&list, 0);
2752	blk_mq_run_hw_queue(this_hctx, from_sched);
2753	} else {
2754	blk_mq_insert_requests(hctx: this_hctx, ctx: this_ctx, list: &list, run_queue_async: from_sched);
2755	}
2756	percpu_ref_put(ref: &this_hctx->queue->q_usage_counter);
2757	}
2758
2759	void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2760	{
2761	struct request *rq;
2762
2763	/*
2764	* We may have been called recursively midway through handling
2765	* plug->mq_list via a schedule() in the driver's queue_rq() callback.
2766	* To avoid mq_list changing under our feet, clear rq_count early and
2767	* bail out specifically if rq_count is 0 rather than checking
2768	* whether the mq_list is empty.
2769	*/
2770	if (plug->rq_count == 0)
2771	return;
2772	plug->rq_count = 0;
2773
2774	if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2775	struct request_queue *q;
2776
2777	rq = rq_list_peek(&plug->mq_list);
2778	q = rq->q;
2779
2780	/*
2781	* Peek first request and see if we have a ->queue_rqs() hook.
2782	* If we do, we can dispatch the whole plug list in one go. We
2783	* already know at this point that all requests belong to the
2784	* same queue, caller must ensure that's the case.
2785	*/
2786	if (q->mq_ops->queue_rqs) {
2787	blk_mq_run_dispatch_ops(q,
2788	__blk_mq_flush_plug_list(q, plug));
2789	if (rq_list_empty(plug->mq_list))
2790	return;
2791	}
2792
2793	blk_mq_run_dispatch_ops(q,
2794	blk_mq_plug_issue_direct(plug));
2795	if (rq_list_empty(plug->mq_list))
2796	return;
2797	}
2798
2799	do {
2800	blk_mq_dispatch_plug_list(plug, from_sched: from_schedule);
2801	} while (!rq_list_empty(plug->mq_list));
2802	}
2803
2804	static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2805	struct list_head *list)
2806	{
2807	int queued = 0;
2808	blk_status_t ret = BLK_STS_OK;
2809
2810	while (!list_empty(head: list)) {
2811	struct request *rq = list_first_entry(list, struct request,
2812	queuelist);
2813
2814	list_del_init(entry: &rq->queuelist);
2815	ret = blk_mq_request_issue_directly(rq, last: list_empty(head: list));
2816	switch (ret) {
2817	case BLK_STS_OK:
2818	queued++;
2819	break;
2820	case BLK_STS_RESOURCE:
2821	case BLK_STS_DEV_RESOURCE:
2822	blk_mq_request_bypass_insert(rq, flags: 0);
2823	if (list_empty(head: list))
2824	blk_mq_run_hw_queue(hctx, false);
2825	goto out;
2826	default:
2827	blk_mq_end_request(rq, ret);
2828	break;
2829	}
2830	}
2831
2832	out:
2833	if (ret != BLK_STS_OK)
2834	blk_mq_commit_rqs(hctx, queued, from_schedule: false);
2835	}
2836
2837	static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2838	struct bio *bio, unsigned int nr_segs)
2839	{
2840	if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2841	if (blk_attempt_plug_merge(q, bio, nr_segs))
2842	return true;
2843	if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2844	return true;
2845	}
2846	return false;
2847	}
2848
2849	static struct request *blk_mq_get_new_requests(struct request_queue *q,
2850	struct blk_plug *plug,
2851	struct bio *bio,
2852	unsigned int nsegs)
2853	{
2854	struct blk_mq_alloc_data data = {
2855	.q = q,
2856	.nr_tags = 1,
2857	.cmd_flags = bio->bi_opf,
2858	};
2859	struct request *rq;
2860
2861	if (unlikely(bio_queue_enter(bio)))
2862	return NULL;
2863
2864	if (blk_mq_attempt_bio_merge(q, bio, nr_segs: nsegs))
2865	goto queue_exit;
2866
2867	rq_qos_throttle(q, bio);
2868
2869	if (plug) {
2870	data.nr_tags = plug->nr_ios;
2871	plug->nr_ios = 1;
2872	data.cached_rq = &plug->cached_rq;
2873	}
2874
2875	rq = __blk_mq_alloc_requests(data: &data);
2876	if (rq)
2877	return rq;
2878	rq_qos_cleanup(q, bio);
2879	if (bio->bi_opf & REQ_NOWAIT)
2880	bio_wouldblock_error(bio);
2881	queue_exit:
2882	blk_queue_exit(q);
2883	return NULL;
2884	}
2885
2886	static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
2887	struct blk_plug *plug, struct bio **bio, unsigned int nsegs)
2888	{
2889	struct request *rq;
2890	enum hctx_type type, hctx_type;
2891
2892	if (!plug)
2893	return NULL;
2894	rq = rq_list_peek(&plug->cached_rq);
2895	if (!rq || rq->q != q)
2896	return NULL;
2897
2898	if (blk_mq_attempt_bio_merge(q, bio: *bio, nr_segs: nsegs)) {
2899	*bio = NULL;
2900	return NULL;
2901	}
2902
2903	type = blk_mq_get_hctx_type(opf: (*bio)->bi_opf);
2904	hctx_type = rq->mq_hctx->type;
2905	if (type != hctx_type &&
2906	!(type == HCTX_TYPE_READ && hctx_type == HCTX_TYPE_DEFAULT))
2907	return NULL;
2908	if (op_is_flush(op: rq->cmd_flags) != op_is_flush(op: (*bio)->bi_opf))
2909	return NULL;
2910
2911	/*
2912	* If any qos ->throttle() end up blocking, we will have flushed the
2913	* plug and hence killed the cached_rq list as well. Pop this entry
2914	* before we throttle.
2915	*/
2916	plug->cached_rq = rq_list_next(rq);
2917	rq_qos_throttle(q, bio: *bio);
2918
2919	blk_mq_rq_time_init(rq, alloc_time_ns: 0);
2920	rq->cmd_flags = (*bio)->bi_opf;
2921	INIT_LIST_HEAD(list: &rq->queuelist);
2922	return rq;
2923	}
2924
2925	static void bio_set_ioprio(struct bio *bio)
2926	{
2927	/* Nobody set ioprio so far? Initialize it based on task's nice value */
2928	if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE)
2929	bio->bi_ioprio = get_current_ioprio();
2930	blkcg_set_ioprio(bio);
2931	}
2932
2933	/**
2934	* blk_mq_submit_bio - Create and send a request to block device.
2935	* @bio: Bio pointer.
2936	*
2937	* Builds up a request structure from @q and @bio and send to the device. The
2938	* request may not be queued directly to hardware if:
2939	* * This request can be merged with another one
2940	* * We want to place request at plug queue for possible future merging
2941	* * There is an IO scheduler active at this queue
2942	*
2943	* It will not queue the request if there is an error with the bio, or at the
2944	* request creation.
2945	*/
2946	void blk_mq_submit_bio(struct bio *bio)
2947	{
2948	struct request_queue *q = bdev_get_queue(bdev: bio->bi_bdev);
2949	struct blk_plug *plug = blk_mq_plug(bio);
2950	const int is_sync = op_is_sync(op: bio->bi_opf);
2951	struct blk_mq_hw_ctx *hctx;
2952	struct request *rq;
2953	unsigned int nr_segs = 1;
2954	blk_status_t ret;
2955
2956	bio = blk_queue_bounce(bio, q);
2957	if (bio_may_exceed_limits(bio, lim: &q->limits)) {
2958	bio = __bio_split_to_limits(bio, lim: &q->limits, nr_segs: &nr_segs);
2959	if (!bio)
2960	return;
2961	}
2962
2963	if (!bio_integrity_prep(bio))
2964	return;
2965
2966	bio_set_ioprio(bio);
2967
2968	rq = blk_mq_get_cached_request(q, plug, bio: &bio, nsegs: nr_segs);
2969	if (!rq) {
2970	if (!bio)
2971	return;
2972	rq = blk_mq_get_new_requests(q, plug, bio, nsegs: nr_segs);
2973	if (unlikely(!rq))
2974	return;
2975	}
2976
2977	trace_block_getrq(bio);
2978
2979	rq_qos_track(q, rq, bio);
2980
2981	blk_mq_bio_to_request(rq, bio, nr_segs);
2982
2983	ret = blk_crypto_rq_get_keyslot(rq);
2984	if (ret != BLK_STS_OK) {
2985	bio->bi_status = ret;
2986	bio_endio(bio);
2987	blk_mq_free_request(rq);
2988	return;
2989	}
2990
2991	if (op_is_flush(op: bio->bi_opf) && blk_insert_flush(rq))
2992	return;
2993
2994	if (plug) {
2995	blk_add_rq_to_plug(plug, rq);
2996	return;
2997	}
2998
2999	hctx = rq->mq_hctx;
3000	if ((rq->rq_flags & RQF_USE_SCHED) ||
3001	(hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) {
3002	blk_mq_insert_request(rq, flags: 0);
3003	blk_mq_run_hw_queue(hctx, true);
3004	} else {
3005	blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq));
3006	}
3007	}
3008
3009	#ifdef CONFIG_BLK_MQ_STACKING
3010	/**
3011	* blk_insert_cloned_request - Helper for stacking drivers to submit a request
3012	* @rq: the request being queued
3013	*/
3014	blk_status_t blk_insert_cloned_request(struct request *rq)
3015	{
3016	struct request_queue *q = rq->q;
3017	unsigned int max_sectors = blk_queue_get_max_sectors(q, op: req_op(req: rq));
3018	unsigned int max_segments = blk_rq_get_max_segments(rq);
3019	blk_status_t ret;
3020
3021	if (blk_rq_sectors(rq) > max_sectors) {
3022	/*
3023	* SCSI device does not have a good way to return if
3024	* Write Same/Zero is actually supported. If a device rejects
3025	* a non-read/write command (discard, write same,etc.) the
3026	* low-level device driver will set the relevant queue limit to
3027	* 0 to prevent blk-lib from issuing more of the offending
3028	* operations. Commands queued prior to the queue limit being
3029	* reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3030	* errors being propagated to upper layers.
3031	*/
3032	if (max_sectors == 0)
3033	return BLK_STS_NOTSUPP;
3034
3035	printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3036	__func__, blk_rq_sectors(rq), max_sectors);
3037	return BLK_STS_IOERR;
3038	}
3039
3040	/*
3041	* The queue settings related to segment counting may differ from the
3042	* original queue.
3043	*/
3044	rq->nr_phys_segments = blk_recalc_rq_segments(rq);
3045	if (rq->nr_phys_segments > max_segments) {
3046	printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n",
3047	__func__, rq->nr_phys_segments, max_segments);
3048	return BLK_STS_IOERR;
3049	}
3050
3051	if (q->disk && should_fail_request(part: q->disk->part0, bytes: blk_rq_bytes(rq)))
3052	return BLK_STS_IOERR;
3053
3054	ret = blk_crypto_rq_get_keyslot(rq);
3055	if (ret != BLK_STS_OK)
3056	return ret;
3057
3058	blk_account_io_start(req: rq);
3059
3060	/*
3061	* Since we have a scheduler attached on the top device,
3062	* bypass a potential scheduler on the bottom device for
3063	* insert.
3064	*/
3065	blk_mq_run_dispatch_ops(q,
3066	ret = blk_mq_request_issue_directly(rq, true));
3067	if (ret)
3068	blk_account_io_done(req: rq, now: ktime_get_ns());
3069	return ret;
3070	}
3071	EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3072
3073	/**
3074	* blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3075	* @rq: the clone request to be cleaned up
3076	*
3077	* Description:
3078	* Free all bios in @rq for a cloned request.
3079	*/
3080	void blk_rq_unprep_clone(struct request *rq)
3081	{
3082	struct bio *bio;
3083
3084	while ((bio = rq->bio) != NULL) {
3085	rq->bio = bio->bi_next;
3086
3087	bio_put(bio);
3088	}
3089	}
3090	EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3091
3092	/**
3093	* blk_rq_prep_clone - Helper function to setup clone request
3094	* @rq: the request to be setup
3095	* @rq_src: original request to be cloned
3096	* @bs: bio_set that bios for clone are allocated from
3097	* @gfp_mask: memory allocation mask for bio
3098	* @bio_ctr: setup function to be called for each clone bio.
3099	* Returns %0 for success, non %0 for failure.
3100	* @data: private data to be passed to @bio_ctr
3101	*
3102	* Description:
3103	* Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3104	* Also, pages which the original bios are pointing to are not copied
3105	* and the cloned bios just point same pages.
3106	* So cloned bios must be completed before original bios, which means
3107	* the caller must complete @rq before @rq_src.
3108	*/
3109	int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3110	struct bio_set *bs, gfp_t gfp_mask,
3111	int (*bio_ctr)(struct bio *, struct bio *, void *),
3112	void *data)
3113	{
3114	struct bio *bio, *bio_src;
3115
3116	if (!bs)
3117	bs = &fs_bio_set;
3118
3119	__rq_for_each_bio(bio_src, rq_src) {
3120	bio = bio_alloc_clone(bdev: rq->q->disk->part0, bio_src, gfp: gfp_mask,
3121	bs);
3122	if (!bio)
3123	goto free_and_out;
3124
3125	if (bio_ctr && bio_ctr(bio, bio_src, data))
3126	goto free_and_out;
3127
3128	if (rq->bio) {
3129	rq->biotail->bi_next = bio;
3130	rq->biotail = bio;
3131	} else {
3132	rq->bio = rq->biotail = bio;
3133	}
3134	bio = NULL;
3135	}
3136
3137	/* Copy attributes of the original request to the clone request. */
3138	rq->__sector = blk_rq_pos(rq: rq_src);
3139	rq->__data_len = blk_rq_bytes(rq: rq_src);
3140	if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3141	rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3142	rq->special_vec = rq_src->special_vec;
3143	}
3144	rq->nr_phys_segments = rq_src->nr_phys_segments;
3145	rq->ioprio = rq_src->ioprio;
3146
3147	if (rq->bio && blk_crypto_rq_bio_prep(rq, bio: rq->bio, gfp_mask) < 0)
3148	goto free_and_out;
3149
3150	return 0;
3151
3152	free_and_out:
3153	if (bio)
3154	bio_put(bio);
3155	blk_rq_unprep_clone(rq);
3156
3157	return -ENOMEM;
3158	}
3159	EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3160	#endif /* CONFIG_BLK_MQ_STACKING */
3161
3162	/*
3163	* Steal bios from a request and add them to a bio list.
3164	* The request must not have been partially completed before.
3165	*/
3166	void blk_steal_bios(struct bio_list *list, struct request *rq)
3167	{
3168	if (rq->bio) {
3169	if (list->tail)
3170	list->tail->bi_next = rq->bio;
3171	else
3172	list->head = rq->bio;
3173	list->tail = rq->biotail;
3174
3175	rq->bio = NULL;
3176	rq->biotail = NULL;
3177	}
3178
3179	rq->__data_len = 0;
3180	}
3181	EXPORT_SYMBOL_GPL(blk_steal_bios);
3182
3183	static size_t order_to_size(unsigned int order)
3184	{
3185	return (size_t)PAGE_SIZE << order;
3186	}
3187
3188	/* called before freeing request pool in @tags */
3189	static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3190	struct blk_mq_tags *tags)
3191	{
3192	struct page *page;
3193	unsigned long flags;
3194
3195	/*
3196	* There is no need to clear mapping if driver tags is not initialized
3197	* or the mapping belongs to the driver tags.
3198	*/
3199	if (!drv_tags || drv_tags == tags)
3200	return;
3201
3202	list_for_each_entry(page, &tags->page_list, lru) {
3203	unsigned long start = (unsigned long)page_address(page);
3204	unsigned long end = start + order_to_size(order: page->private);
3205	int i;
3206
3207	for (i = 0; i < drv_tags->nr_tags; i++) {
3208	struct request *rq = drv_tags->rqs[i];
3209	unsigned long rq_addr = (unsigned long)rq;
3210
3211	if (rq_addr >= start && rq_addr < end) {
3212	WARN_ON_ONCE(req_ref_read(rq) != 0);
3213	cmpxchg(&drv_tags->rqs[i], rq, NULL);
3214	}
3215	}
3216	}
3217
3218	/*
3219	* Wait until all pending iteration is done.
3220	*
3221	* Request reference is cleared and it is guaranteed to be observed
3222	* after the ->lock is released.
3223	*/
3224	spin_lock_irqsave(&drv_tags->lock, flags);
3225	spin_unlock_irqrestore(lock: &drv_tags->lock, flags);
3226	}
3227
3228	void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3229	unsigned int hctx_idx)
3230	{
3231	struct blk_mq_tags *drv_tags;
3232	struct page *page;
3233
3234	if (list_empty(head: &tags->page_list))
3235	return;
3236
3237	if (blk_mq_is_shared_tags(flags: set->flags))
3238	drv_tags = set->shared_tags;
3239	else
3240	drv_tags = set->tags[hctx_idx];
3241
3242	if (tags->static_rqs && set->ops->exit_request) {
3243	int i;
3244
3245	for (i = 0; i < tags->nr_tags; i++) {
3246	struct request *rq = tags->static_rqs[i];
3247
3248	if (!rq)
3249	continue;
3250	set->ops->exit_request(set, rq, hctx_idx);
3251	tags->static_rqs[i] = NULL;
3252	}
3253	}
3254
3255	blk_mq_clear_rq_mapping(drv_tags, tags);
3256
3257	while (!list_empty(head: &tags->page_list)) {
3258	page = list_first_entry(&tags->page_list, struct page, lru);
3259	list_del_init(entry: &page->lru);
3260	/*
3261	* Remove kmemleak object previously allocated in
3262	* blk_mq_alloc_rqs().
3263	*/
3264	kmemleak_free(page_address(page));
3265	__free_pages(page, order: page->private);
3266	}
3267	}
3268
3269	void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3270	{
3271	kfree(objp: tags->rqs);
3272	tags->rqs = NULL;
3273	kfree(objp: tags->static_rqs);
3274	tags->static_rqs = NULL;
3275
3276	blk_mq_free_tags(tags);
3277	}
3278
3279	static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3280	unsigned int hctx_idx)
3281	{
3282	int i;
3283
3284	for (i = 0; i < set->nr_maps; i++) {
3285	unsigned int start = set->map[i].queue_offset;
3286	unsigned int end = start + set->map[i].nr_queues;
3287
3288	if (hctx_idx >= start && hctx_idx < end)
3289	break;
3290	}
3291
3292	if (i >= set->nr_maps)
3293	i = HCTX_TYPE_DEFAULT;
3294
3295	return i;
3296	}
3297
3298	static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3299	unsigned int hctx_idx)
3300	{
3301	enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3302
3303	return blk_mq_hw_queue_to_node(qmap: &set->map[type], hctx_idx);
3304	}
3305
3306	static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3307	unsigned int hctx_idx,
3308	unsigned int nr_tags,
3309	unsigned int reserved_tags)
3310	{
3311	int node = blk_mq_get_hctx_node(set, hctx_idx);
3312	struct blk_mq_tags *tags;
3313
3314	if (node == NUMA_NO_NODE)
3315	node = set->numa_node;
3316
3317	tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3318	BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3319	if (!tags)
3320	return NULL;
3321
3322	tags->rqs = kcalloc_node(n: nr_tags, size: sizeof(struct request *),
3323	GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3324	node);
3325	if (!tags->rqs)
3326	goto err_free_tags;
3327
3328	tags->static_rqs = kcalloc_node(n: nr_tags, size: sizeof(struct request *),
3329	GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3330	node);
3331	if (!tags->static_rqs)
3332	goto err_free_rqs;
3333
3334	return tags;
3335
3336	err_free_rqs:
3337	kfree(objp: tags->rqs);
3338	err_free_tags:
3339	blk_mq_free_tags(tags);
3340	return NULL;
3341	}
3342
3343	static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3344	unsigned int hctx_idx, int node)
3345	{
3346	int ret;
3347
3348	if (set->ops->init_request) {
3349	ret = set->ops->init_request(set, rq, hctx_idx, node);
3350	if (ret)
3351	return ret;
3352	}
3353
3354	WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3355	return 0;
3356	}
3357
3358	static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3359	struct blk_mq_tags *tags,
3360	unsigned int hctx_idx, unsigned int depth)
3361	{
3362	unsigned int i, j, entries_per_page, max_order = 4;
3363	int node = blk_mq_get_hctx_node(set, hctx_idx);
3364	size_t rq_size, left;
3365
3366	if (node == NUMA_NO_NODE)
3367	node = set->numa_node;
3368
3369	INIT_LIST_HEAD(list: &tags->page_list);
3370
3371	/*
3372	* rq_size is the size of the request plus driver payload, rounded
3373	* to the cacheline size
3374	*/
3375	rq_size = round_up(sizeof(struct request) + set->cmd_size,
3376	cache_line_size());
3377	left = rq_size * depth;
3378
3379	for (i = 0; i < depth; ) {
3380	int this_order = max_order;
3381	struct page *page;
3382	int to_do;
3383	void *p;
3384
3385	while (this_order && left < order_to_size(order: this_order - 1))
3386	this_order--;
3387
3388	do {
3389	page = alloc_pages_node(nid: node,
3390	GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3391	order: this_order);
3392	if (page)
3393	break;
3394	if (!this_order--)
3395	break;
3396	if (order_to_size(order: this_order) < rq_size)
3397	break;
3398	} while (1);
3399
3400	if (!page)
3401	goto fail;
3402
3403	page->private = this_order;
3404	list_add_tail(new: &page->lru, head: &tags->page_list);
3405
3406	p = page_address(page);
3407	/*
3408	* Allow kmemleak to scan these pages as they contain pointers
3409	* to additional allocations like via ops->init_request().
3410	*/
3411	kmemleak_alloc(ptr: p, size: order_to_size(order: this_order), min_count: 1, GFP_NOIO);
3412	entries_per_page = order_to_size(order: this_order) / rq_size;
3413	to_do = min(entries_per_page, depth - i);
3414	left -= to_do * rq_size;
3415	for (j = 0; j < to_do; j++) {
3416	struct request *rq = p;
3417
3418	tags->static_rqs[i] = rq;
3419	if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3420	tags->static_rqs[i] = NULL;
3421	goto fail;
3422	}
3423
3424	p += rq_size;
3425	i++;
3426	}
3427	}
3428	return 0;
3429
3430	fail:
3431	blk_mq_free_rqs(set, tags, hctx_idx);
3432	return -ENOMEM;
3433	}
3434
3435	struct rq_iter_data {
3436	struct blk_mq_hw_ctx *hctx;
3437	bool has_rq;
3438	};
3439
3440	static bool blk_mq_has_request(struct request *rq, void *data)
3441	{
3442	struct rq_iter_data *iter_data = data;
3443
3444	if (rq->mq_hctx != iter_data->hctx)
3445	return true;
3446	iter_data->has_rq = true;
3447	return false;
3448	}
3449
3450	static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3451	{
3452	struct blk_mq_tags *tags = hctx->sched_tags ?
3453	hctx->sched_tags : hctx->tags;
3454	struct rq_iter_data data = {
3455	.hctx = hctx,
3456	};
3457
3458	blk_mq_all_tag_iter(tags, fn: blk_mq_has_request, priv: &data);
3459	return data.has_rq;
3460	}
3461
3462	static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3463	struct blk_mq_hw_ctx *hctx)
3464	{
3465	if (cpumask_first_and(srcp1: hctx->cpumask, cpu_online_mask) != cpu)
3466	return false;
3467	if (cpumask_next_and(n: cpu, src1p: hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3468	return false;
3469	return true;
3470	}
3471
3472	static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3473	{
3474	struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3475	struct blk_mq_hw_ctx, cpuhp_online);
3476
3477	if (!cpumask_test_cpu(cpu, cpumask: hctx->cpumask) ||
3478	!blk_mq_last_cpu_in_hctx(cpu, hctx))
3479	return 0;
3480
3481	/*
3482	* Prevent new request from being allocated on the current hctx.
3483	*
3484	* The smp_mb__after_atomic() Pairs with the implied barrier in
3485	* test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3486	* seen once we return from the tag allocator.
3487	*/
3488	set_bit(nr: BLK_MQ_S_INACTIVE, addr: &hctx->state);
3489	smp_mb__after_atomic();
3490
3491	/*
3492	* Try to grab a reference to the queue and wait for any outstanding
3493	* requests. If we could not grab a reference the queue has been
3494	* frozen and there are no requests.
3495	*/
3496	if (percpu_ref_tryget(ref: &hctx->queue->q_usage_counter)) {
3497	while (blk_mq_hctx_has_requests(hctx))
3498	msleep(msecs: 5);
3499	percpu_ref_put(ref: &hctx->queue->q_usage_counter);
3500	}
3501
3502	return 0;
3503	}
3504
3505	static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3506	{
3507	struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3508	struct blk_mq_hw_ctx, cpuhp_online);
3509
3510	if (cpumask_test_cpu(cpu, cpumask: hctx->cpumask))
3511	clear_bit(nr: BLK_MQ_S_INACTIVE, addr: &hctx->state);
3512	return 0;
3513	}
3514
3515	/*
3516	* 'cpu' is going away. splice any existing rq_list entries from this
3517	* software queue to the hw queue dispatch list, and ensure that it
3518	* gets run.
3519	*/
3520	static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3521	{
3522	struct blk_mq_hw_ctx *hctx;
3523	struct blk_mq_ctx *ctx;
3524	LIST_HEAD(tmp);
3525	enum hctx_type type;
3526
3527	hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3528	if (!cpumask_test_cpu(cpu, cpumask: hctx->cpumask))
3529	return 0;
3530
3531	ctx = __blk_mq_get_ctx(q: hctx->queue, cpu);
3532	type = hctx->type;
3533
3534	spin_lock(lock: &ctx->lock);
3535	if (!list_empty(head: &ctx->rq_lists[type])) {
3536	list_splice_init(list: &ctx->rq_lists[type], head: &tmp);
3537	blk_mq_hctx_clear_pending(hctx, ctx);
3538	}
3539	spin_unlock(lock: &ctx->lock);
3540
3541	if (list_empty(head: &tmp))
3542	return 0;
3543
3544	spin_lock(lock: &hctx->lock);
3545	list_splice_tail_init(list: &tmp, head: &hctx->dispatch);
3546	spin_unlock(lock: &hctx->lock);
3547
3548	blk_mq_run_hw_queue(hctx, true);
3549	return 0;
3550	}
3551
3552	static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3553	{
3554	if (!(hctx->flags & BLK_MQ_F_STACKING))
3555	cpuhp_state_remove_instance_nocalls(state: CPUHP_AP_BLK_MQ_ONLINE,
3556	node: &hctx->cpuhp_online);
3557	cpuhp_state_remove_instance_nocalls(state: CPUHP_BLK_MQ_DEAD,
3558	node: &hctx->cpuhp_dead);
3559	}
3560
3561	/*
3562	* Before freeing hw queue, clearing the flush request reference in
3563	* tags->rqs[] for avoiding potential UAF.
3564	*/
3565	static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3566	unsigned int queue_depth, struct request *flush_rq)
3567	{
3568	int i;
3569	unsigned long flags;
3570
3571	/* The hw queue may not be mapped yet */
3572	if (!tags)
3573	return;
3574
3575	WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3576
3577	for (i = 0; i < queue_depth; i++)
3578	cmpxchg(&tags->rqs[i], flush_rq, NULL);
3579
3580	/*
3581	* Wait until all pending iteration is done.
3582	*
3583	* Request reference is cleared and it is guaranteed to be observed
3584	* after the ->lock is released.
3585	*/
3586	spin_lock_irqsave(&tags->lock, flags);
3587	spin_unlock_irqrestore(lock: &tags->lock, flags);
3588	}
3589
3590	/* hctx->ctxs will be freed in queue's release handler */
3591	static void blk_mq_exit_hctx(struct request_queue *q,
3592	struct blk_mq_tag_set *set,
3593	struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3594	{
3595	struct request *flush_rq = hctx->fq->flush_rq;
3596
3597	if (blk_mq_hw_queue_mapped(hctx))
3598	blk_mq_tag_idle(hctx);
3599
3600	if (blk_queue_init_done(q))
3601	blk_mq_clear_flush_rq_mapping(tags: set->tags[hctx_idx],
3602	queue_depth: set->queue_depth, flush_rq);
3603	if (set->ops->exit_request)
3604	set->ops->exit_request(set, flush_rq, hctx_idx);
3605
3606	if (set->ops->exit_hctx)
3607	set->ops->exit_hctx(hctx, hctx_idx);
3608
3609	blk_mq_remove_cpuhp(hctx);
3610
3611	xa_erase(&q->hctx_table, index: hctx_idx);
3612
3613	spin_lock(lock: &q->unused_hctx_lock);
3614	list_add(new: &hctx->hctx_list, head: &q->unused_hctx_list);
3615	spin_unlock(lock: &q->unused_hctx_lock);
3616	}
3617
3618	static void blk_mq_exit_hw_queues(struct request_queue *q,
3619	struct blk_mq_tag_set *set, int nr_queue)
3620	{
3621	struct blk_mq_hw_ctx *hctx;
3622	unsigned long i;
3623
3624	queue_for_each_hw_ctx(q, hctx, i) {
3625	if (i == nr_queue)
3626	break;
3627	blk_mq_exit_hctx(q, set, hctx, hctx_idx: i);
3628	}
3629	}
3630
3631	static int blk_mq_init_hctx(struct request_queue *q,
3632	struct blk_mq_tag_set *set,
3633	struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3634	{
3635	hctx->queue_num = hctx_idx;
3636
3637	if (!(hctx->flags & BLK_MQ_F_STACKING))
3638	cpuhp_state_add_instance_nocalls(state: CPUHP_AP_BLK_MQ_ONLINE,
3639	node: &hctx->cpuhp_online);
3640	cpuhp_state_add_instance_nocalls(state: CPUHP_BLK_MQ_DEAD, node: &hctx->cpuhp_dead);
3641
3642	hctx->tags = set->tags[hctx_idx];
3643
3644	if (set->ops->init_hctx &&
3645	set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3646	goto unregister_cpu_notifier;
3647
3648	if (blk_mq_init_request(set, rq: hctx->fq->flush_rq, hctx_idx,
3649	node: hctx->numa_node))
3650	goto exit_hctx;
3651
3652	if (xa_insert(xa: &q->hctx_table, index: hctx_idx, entry: hctx, GFP_KERNEL))
3653	goto exit_flush_rq;
3654
3655	return 0;
3656
3657	exit_flush_rq:
3658	if (set->ops->exit_request)
3659	set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3660	exit_hctx:
3661	if (set->ops->exit_hctx)
3662	set->ops->exit_hctx(hctx, hctx_idx);
3663	unregister_cpu_notifier:
3664	blk_mq_remove_cpuhp(hctx);
3665	return -1;
3666	}
3667
3668	static struct blk_mq_hw_ctx *
3669	blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3670	int node)
3671	{
3672	struct blk_mq_hw_ctx *hctx;
3673	gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3674
3675	hctx = kzalloc_node(size: sizeof(struct blk_mq_hw_ctx), flags: gfp, node);
3676	if (!hctx)
3677	goto fail_alloc_hctx;
3678
3679	if (!zalloc_cpumask_var_node(mask: &hctx->cpumask, flags: gfp, node))
3680	goto free_hctx;
3681
3682	atomic_set(v: &hctx->nr_active, i: 0);
3683	if (node == NUMA_NO_NODE)
3684	node = set->numa_node;
3685	hctx->numa_node = node;
3686
3687	INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3688	spin_lock_init(&hctx->lock);
3689	INIT_LIST_HEAD(list: &hctx->dispatch);
3690	hctx->queue = q;
3691	hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3692
3693	INIT_LIST_HEAD(list: &hctx->hctx_list);
3694
3695	/*
3696	* Allocate space for all possible cpus to avoid allocation at
3697	* runtime
3698	*/
3699	hctx->ctxs = kmalloc_array_node(n: nr_cpu_ids, size: sizeof(void *),
3700	flags: gfp, node);
3701	if (!hctx->ctxs)
3702	goto free_cpumask;
3703
3704	if (sbitmap_init_node(sb: &hctx->ctx_map, depth: nr_cpu_ids, ilog2(8),
3705	flags: gfp, node, round_robin: false, alloc_hint: false))
3706	goto free_ctxs;
3707	hctx->nr_ctx = 0;
3708
3709	spin_lock_init(&hctx->dispatch_wait_lock);
3710	init_waitqueue_func_entry(wq_entry: &hctx->dispatch_wait, func: blk_mq_dispatch_wake);
3711	INIT_LIST_HEAD(list: &hctx->dispatch_wait.entry);
3712
3713	hctx->fq = blk_alloc_flush_queue(node: hctx->numa_node, cmd_size: set->cmd_size, flags: gfp);
3714	if (!hctx->fq)
3715	goto free_bitmap;
3716
3717	blk_mq_hctx_kobj_init(hctx);
3718
3719	return hctx;
3720
3721	free_bitmap:
3722	sbitmap_free(sb: &hctx->ctx_map);
3723	free_ctxs:
3724	kfree(objp: hctx->ctxs);
3725	free_cpumask:
3726	free_cpumask_var(mask: hctx->cpumask);
3727	free_hctx:
3728	kfree(objp: hctx);
3729	fail_alloc_hctx:
3730	return NULL;
3731	}
3732
3733	static void blk_mq_init_cpu_queues(struct request_queue *q,
3734	unsigned int nr_hw_queues)
3735	{
3736	struct blk_mq_tag_set *set = q->tag_set;
3737	unsigned int i, j;
3738
3739	for_each_possible_cpu(i) {
3740	struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3741	struct blk_mq_hw_ctx *hctx;
3742	int k;
3743
3744	__ctx->cpu = i;
3745	spin_lock_init(&__ctx->lock);
3746	for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3747	INIT_LIST_HEAD(list: &__ctx->rq_lists[k]);
3748
3749	__ctx->queue = q;
3750
3751	/*
3752	* Set local node, IFF we have more than one hw queue. If
3753	* not, we remain on the home node of the device
3754	*/
3755	for (j = 0; j < set->nr_maps; j++) {
3756	hctx = blk_mq_map_queue_type(q, type: j, cpu: i);
3757	if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3758	hctx->numa_node = cpu_to_node(cpu: i);
3759	}
3760	}
3761	}
3762
3763	struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3764	unsigned int hctx_idx,
3765	unsigned int depth)
3766	{
3767	struct blk_mq_tags *tags;
3768	int ret;
3769
3770	tags = blk_mq_alloc_rq_map(set, hctx_idx, nr_tags: depth, reserved_tags: set->reserved_tags);
3771	if (!tags)
3772	return NULL;
3773
3774	ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3775	if (ret) {
3776	blk_mq_free_rq_map(tags);
3777	return NULL;
3778	}
3779
3780	return tags;
3781	}
3782
3783	static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3784	int hctx_idx)
3785	{
3786	if (blk_mq_is_shared_tags(flags: set->flags)) {
3787	set->tags[hctx_idx] = set->shared_tags;
3788
3789	return true;
3790	}
3791
3792	set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3793	depth: set->queue_depth);
3794
3795	return set->tags[hctx_idx];
3796	}
3797
3798	void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3799	struct blk_mq_tags *tags,
3800	unsigned int hctx_idx)
3801	{
3802	if (tags) {
3803	blk_mq_free_rqs(set, tags, hctx_idx);
3804	blk_mq_free_rq_map(tags);
3805	}
3806	}
3807
3808	static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3809	unsigned int hctx_idx)
3810	{
3811	if (!blk_mq_is_shared_tags(flags: set->flags))
3812	blk_mq_free_map_and_rqs(set, tags: set->tags[hctx_idx], hctx_idx);
3813
3814	set->tags[hctx_idx] = NULL;
3815	}
3816
3817	static void blk_mq_map_swqueue(struct request_queue *q)
3818	{
3819	unsigned int j, hctx_idx;
3820	unsigned long i;
3821	struct blk_mq_hw_ctx *hctx;
3822	struct blk_mq_ctx *ctx;
3823	struct blk_mq_tag_set *set = q->tag_set;
3824
3825	queue_for_each_hw_ctx(q, hctx, i) {
3826	cpumask_clear(dstp: hctx->cpumask);
3827	hctx->nr_ctx = 0;
3828	hctx->dispatch_from = NULL;
3829	}
3830
3831	/*
3832	* Map software to hardware queues.
3833	*
3834	* If the cpu isn't present, the cpu is mapped to first hctx.
3835	*/
3836	for_each_possible_cpu(i) {
3837
3838	ctx = per_cpu_ptr(q->queue_ctx, i);
3839	for (j = 0; j < set->nr_maps; j++) {
3840	if (!set->map[j].nr_queues) {
3841	ctx->hctxs[j] = blk_mq_map_queue_type(q,
3842	type: HCTX_TYPE_DEFAULT, cpu: i);
3843	continue;
3844	}
3845	hctx_idx = set->map[j].mq_map[i];
3846	/* unmapped hw queue can be remapped after CPU topo changed */
3847	if (!set->tags[hctx_idx] &&
3848	!__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3849	/*
3850	* If tags initialization fail for some hctx,
3851	* that hctx won't be brought online. In this
3852	* case, remap the current ctx to hctx[0] which
3853	* is guaranteed to always have tags allocated
3854	*/
3855	set->map[j].mq_map[i] = 0;
3856	}
3857
3858	hctx = blk_mq_map_queue_type(q, type: j, cpu: i);
3859	ctx->hctxs[j] = hctx;
3860	/*
3861	* If the CPU is already set in the mask, then we've
3862	* mapped this one already. This can happen if
3863	* devices share queues across queue maps.
3864	*/
3865	if (cpumask_test_cpu(cpu: i, cpumask: hctx->cpumask))
3866	continue;
3867
3868	cpumask_set_cpu(cpu: i, dstp: hctx->cpumask);
3869	hctx->type = j;
3870	ctx->index_hw[hctx->type] = hctx->nr_ctx;
3871	hctx->ctxs[hctx->nr_ctx++] = ctx;
3872
3873	/*
3874	* If the nr_ctx type overflows, we have exceeded the
3875	* amount of sw queues we can support.
3876	*/
3877	BUG_ON(!hctx->nr_ctx);
3878	}
3879
3880	for (; j < HCTX_MAX_TYPES; j++)
3881	ctx->hctxs[j] = blk_mq_map_queue_type(q,
3882	type: HCTX_TYPE_DEFAULT, cpu: i);
3883	}
3884
3885	queue_for_each_hw_ctx(q, hctx, i) {
3886	/*
3887	* If no software queues are mapped to this hardware queue,
3888	* disable it and free the request entries.
3889	*/
3890	if (!hctx->nr_ctx) {
3891	/* Never unmap queue 0. We need it as a
3892	* fallback in case of a new remap fails
3893	* allocation
3894	*/
3895	if (i)
3896	__blk_mq_free_map_and_rqs(set, hctx_idx: i);
3897
3898	hctx->tags = NULL;
3899	continue;
3900	}
3901
3902	hctx->tags = set->tags[i];
3903	WARN_ON(!hctx->tags);
3904
3905	/*
3906	* Set the map size to the number of mapped software queues.
3907	* This is more accurate and more efficient than looping
3908	* over all possibly mapped software queues.
3909	*/
3910	sbitmap_resize(sb: &hctx->ctx_map, depth: hctx->nr_ctx);
3911
3912	/*
3913	* Initialize batch roundrobin counts
3914	*/
3915	hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3916	hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3917	}
3918	}
3919
3920	/*
3921	* Caller needs to ensure that we're either frozen/quiesced, or that
3922	* the queue isn't live yet.
3923	*/
3924	static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3925	{
3926	struct blk_mq_hw_ctx *hctx;
3927	unsigned long i;
3928
3929	queue_for_each_hw_ctx(q, hctx, i) {
3930	if (shared) {
3931	hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3932	} else {
3933	blk_mq_tag_idle(hctx);
3934	hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3935	}
3936	}
3937	}
3938
3939	static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3940	bool shared)
3941	{
3942	struct request_queue *q;
3943
3944	lockdep_assert_held(&set->tag_list_lock);
3945
3946	list_for_each_entry(q, &set->tag_list, tag_set_list) {
3947	blk_mq_freeze_queue(q);
3948	queue_set_hctx_shared(q, shared);
3949	blk_mq_unfreeze_queue(q);
3950	}
3951	}
3952
3953	static void blk_mq_del_queue_tag_set(struct request_queue *q)
3954	{
3955	struct blk_mq_tag_set *set = q->tag_set;
3956
3957	mutex_lock(&set->tag_list_lock);
3958	list_del(entry: &q->tag_set_list);
3959	if (list_is_singular(head: &set->tag_list)) {
3960	/* just transitioned to unshared */
3961	set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3962	/* update existing queue */
3963	blk_mq_update_tag_set_shared(set, shared: false);
3964	}
3965	mutex_unlock(lock: &set->tag_list_lock);
3966	INIT_LIST_HEAD(list: &q->tag_set_list);
3967	}
3968
3969	static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3970	struct request_queue *q)
3971	{
3972	mutex_lock(&set->tag_list_lock);
3973
3974	/*
3975	* Check to see if we're transitioning to shared (from 1 to 2 queues).
3976	*/
3977	if (!list_empty(head: &set->tag_list) &&
3978	!(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3979	set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3980	/* update existing queue */
3981	blk_mq_update_tag_set_shared(set, shared: true);
3982	}
3983	if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3984	queue_set_hctx_shared(q, shared: true);
3985	list_add_tail(new: &q->tag_set_list, head: &set->tag_list);
3986
3987	mutex_unlock(lock: &set->tag_list_lock);
3988	}
3989
3990	/* All allocations will be freed in release handler of q->mq_kobj */
3991	static int blk_mq_alloc_ctxs(struct request_queue *q)
3992	{
3993	struct blk_mq_ctxs *ctxs;
3994	int cpu;
3995
3996	ctxs = kzalloc(size: sizeof(*ctxs), GFP_KERNEL);
3997	if (!ctxs)
3998	return -ENOMEM;
3999
4000	ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
4001	if (!ctxs->queue_ctx)
4002	goto fail;
4003
4004	for_each_possible_cpu(cpu) {
4005	struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
4006	ctx->ctxs = ctxs;
4007	}
4008
4009	q->mq_kobj = &ctxs->kobj;
4010	q->queue_ctx = ctxs->queue_ctx;
4011
4012	return 0;
4013	fail:
4014	kfree(objp: ctxs);
4015	return -ENOMEM;
4016	}
4017
4018	/*
4019	* It is the actual release handler for mq, but we do it from
4020	* request queue's release handler for avoiding use-after-free
4021	* and headache because q->mq_kobj shouldn't have been introduced,
4022	* but we can't group ctx/kctx kobj without it.
4023	*/
4024	void blk_mq_release(struct request_queue *q)
4025	{
4026	struct blk_mq_hw_ctx *hctx, *next;
4027	unsigned long i;
4028
4029	queue_for_each_hw_ctx(q, hctx, i)
4030	WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4031
4032	/* all hctx are in .unused_hctx_list now */
4033	list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4034	list_del_init(entry: &hctx->hctx_list);
4035	kobject_put(kobj: &hctx->kobj);
4036	}
4037
4038	xa_destroy(&q->hctx_table);
4039
4040	/*
4041	* release .mq_kobj and sw queue's kobject now because
4042	* both share lifetime with request queue.
4043	*/
4044	blk_mq_sysfs_deinit(q);
4045	}
4046
4047	static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
4048	void *queuedata)
4049	{
4050	struct request_queue *q;
4051	int ret;
4052
4053	q = blk_alloc_queue(node_id: set->numa_node);
4054	if (!q)
4055	return ERR_PTR(error: -ENOMEM);
4056	q->queuedata = queuedata;
4057	ret = blk_mq_init_allocated_queue(set, q);
4058	if (ret) {
4059	blk_put_queue(q);
4060	return ERR_PTR(error: ret);
4061	}
4062	return q;
4063	}
4064
4065	struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
4066	{
4067	return blk_mq_init_queue_data(set, NULL);
4068	}
4069	EXPORT_SYMBOL(blk_mq_init_queue);
4070
4071	/**
4072	* blk_mq_destroy_queue - shutdown a request queue
4073	* @q: request queue to shutdown
4074	*
4075	* This shuts down a request queue allocated by blk_mq_init_queue(). All future
4076	* requests will be failed with -ENODEV. The caller is responsible for dropping
4077	* the reference from blk_mq_init_queue() by calling blk_put_queue().
4078	*
4079	* Context: can sleep
4080	*/
4081	void blk_mq_destroy_queue(struct request_queue *q)
4082	{
4083	WARN_ON_ONCE(!queue_is_mq(q));
4084	WARN_ON_ONCE(blk_queue_registered(q));
4085
4086	might_sleep();
4087
4088	blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4089	blk_queue_start_drain(q);
4090	blk_mq_freeze_queue_wait(q);
4091
4092	blk_sync_queue(q);
4093	blk_mq_cancel_work_sync(q);
4094	blk_mq_exit_queue(q);
4095	}
4096	EXPORT_SYMBOL(blk_mq_destroy_queue);
4097
4098	struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
4099	struct lock_class_key *lkclass)
4100	{
4101	struct request_queue *q;
4102	struct gendisk *disk;
4103
4104	q = blk_mq_init_queue_data(set, queuedata);
4105	if (IS_ERR(ptr: q))
4106	return ERR_CAST(ptr: q);
4107
4108	disk = __alloc_disk_node(q, node_id: set->numa_node, lkclass);
4109	if (!disk) {
4110	blk_mq_destroy_queue(q);
4111	blk_put_queue(q);
4112	return ERR_PTR(error: -ENOMEM);
4113	}
4114	set_bit(GD_OWNS_QUEUE, addr: &disk->state);
4115	return disk;
4116	}
4117	EXPORT_SYMBOL(__blk_mq_alloc_disk);
4118
4119	struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4120	struct lock_class_key *lkclass)
4121	{
4122	struct gendisk *disk;
4123
4124	if (!blk_get_queue(q))
4125	return NULL;
4126	disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4127	if (!disk)
4128	blk_put_queue(q);
4129	return disk;
4130	}
4131	EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4132
4133	static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4134	struct blk_mq_tag_set *set, struct request_queue *q,
4135	int hctx_idx, int node)
4136	{
4137	struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4138
4139	/* reuse dead hctx first */
4140	spin_lock(lock: &q->unused_hctx_lock);
4141	list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4142	if (tmp->numa_node == node) {
4143	hctx = tmp;
4144	break;
4145	}
4146	}
4147	if (hctx)
4148	list_del_init(entry: &hctx->hctx_list);
4149	spin_unlock(lock: &q->unused_hctx_lock);
4150
4151	if (!hctx)
4152	hctx = blk_mq_alloc_hctx(q, set, node);
4153	if (!hctx)
4154	goto fail;
4155
4156	if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4157	goto free_hctx;
4158
4159	return hctx;
4160
4161	free_hctx:
4162	kobject_put(kobj: &hctx->kobj);
4163	fail:
4164	return NULL;
4165	}
4166
4167	static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4168	struct request_queue *q)
4169	{
4170	struct blk_mq_hw_ctx *hctx;
4171	unsigned long i, j;
4172
4173	/* protect against switching io scheduler */
4174	mutex_lock(&q->sysfs_lock);
4175	for (i = 0; i < set->nr_hw_queues; i++) {
4176	int old_node;
4177	int node = blk_mq_get_hctx_node(set, hctx_idx: i);
4178	struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, index: i);
4179
4180	if (old_hctx) {
4181	old_node = old_hctx->numa_node;
4182	blk_mq_exit_hctx(q, set, hctx: old_hctx, hctx_idx: i);
4183	}
4184
4185	if (!blk_mq_alloc_and_init_hctx(set, q, hctx_idx: i, node)) {
4186	if (!old_hctx)
4187	break;
4188	pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4189	node, old_node);
4190	hctx = blk_mq_alloc_and_init_hctx(set, q, hctx_idx: i, node: old_node);
4191	WARN_ON_ONCE(!hctx);
4192	}
4193	}
4194	/*
4195	* Increasing nr_hw_queues fails. Free the newly allocated
4196	* hctxs and keep the previous q->nr_hw_queues.
4197	*/
4198	if (i != set->nr_hw_queues) {
4199	j = q->nr_hw_queues;
4200	} else {
4201	j = i;
4202	q->nr_hw_queues = set->nr_hw_queues;
4203	}
4204
4205	xa_for_each_start(&q->hctx_table, j, hctx, j)
4206	blk_mq_exit_hctx(q, set, hctx, hctx_idx: j);
4207	mutex_unlock(lock: &q->sysfs_lock);
4208	}
4209
4210	static void blk_mq_update_poll_flag(struct request_queue *q)
4211	{
4212	struct blk_mq_tag_set *set = q->tag_set;
4213
4214	if (set->nr_maps > HCTX_TYPE_POLL &&
4215	set->map[HCTX_TYPE_POLL].nr_queues)
4216	blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4217	else
4218	blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4219	}
4220
4221	int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4222	struct request_queue *q)
4223	{
4224	/* mark the queue as mq asap */
4225	q->mq_ops = set->ops;
4226
4227	if (blk_mq_alloc_ctxs(q))
4228	goto err_exit;
4229
4230	/* init q->mq_kobj and sw queues' kobjects */
4231	blk_mq_sysfs_init(q);
4232
4233	INIT_LIST_HEAD(list: &q->unused_hctx_list);
4234	spin_lock_init(&q->unused_hctx_lock);
4235
4236	xa_init(xa: &q->hctx_table);
4237
4238	blk_mq_realloc_hw_ctxs(set, q);
4239	if (!q->nr_hw_queues)
4240	goto err_hctxs;
4241
4242	INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4243	blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4244
4245	q->tag_set = set;
4246
4247	q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4248	blk_mq_update_poll_flag(q);
4249
4250	INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4251	INIT_LIST_HEAD(list: &q->flush_list);
4252	INIT_LIST_HEAD(list: &q->requeue_list);
4253	spin_lock_init(&q->requeue_lock);
4254
4255	q->nr_requests = set->queue_depth;
4256
4257	blk_mq_init_cpu_queues(q, nr_hw_queues: set->nr_hw_queues);
4258	blk_mq_add_queue_tag_set(set, q);
4259	blk_mq_map_swqueue(q);
4260	return 0;
4261
4262	err_hctxs:
4263	blk_mq_release(q);
4264	err_exit:
4265	q->mq_ops = NULL;
4266	return -ENOMEM;
4267	}
4268	EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4269
4270	/* tags can _not_ be used after returning from blk_mq_exit_queue */
4271	void blk_mq_exit_queue(struct request_queue *q)
4272	{
4273	struct blk_mq_tag_set *set = q->tag_set;
4274
4275	/* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4276	blk_mq_exit_hw_queues(q, set, nr_queue: set->nr_hw_queues);
4277	/* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4278	blk_mq_del_queue_tag_set(q);
4279	}
4280
4281	static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4282	{
4283	int i;
4284
4285	if (blk_mq_is_shared_tags(flags: set->flags)) {
4286	set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4287	BLK_MQ_NO_HCTX_IDX,
4288	depth: set->queue_depth);
4289	if (!set->shared_tags)
4290	return -ENOMEM;
4291	}
4292
4293	for (i = 0; i < set->nr_hw_queues; i++) {
4294	if (!__blk_mq_alloc_map_and_rqs(set, hctx_idx: i))
4295	goto out_unwind;
4296	cond_resched();
4297	}
4298
4299	return 0;
4300
4301	out_unwind:
4302	while (--i >= 0)
4303	__blk_mq_free_map_and_rqs(set, hctx_idx: i);
4304
4305	if (blk_mq_is_shared_tags(flags: set->flags)) {
4306	blk_mq_free_map_and_rqs(set, tags: set->shared_tags,
4307	BLK_MQ_NO_HCTX_IDX);
4308	}
4309
4310	return -ENOMEM;
4311	}
4312
4313	/*
4314	* Allocate the request maps associated with this tag_set. Note that this
4315	* may reduce the depth asked for, if memory is tight. set->queue_depth
4316	* will be updated to reflect the allocated depth.
4317	*/
4318	static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4319	{
4320	unsigned int depth;
4321	int err;
4322
4323	depth = set->queue_depth;
4324	do {
4325	err = __blk_mq_alloc_rq_maps(set);
4326	if (!err)
4327	break;
4328
4329	set->queue_depth >>= 1;
4330	if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4331	err = -ENOMEM;
4332	break;
4333	}
4334	} while (set->queue_depth);
4335
4336	if (!set->queue_depth || err) {
4337	pr_err("blk-mq: failed to allocate request map\n");
4338	return -ENOMEM;
4339	}
4340
4341	if (depth != set->queue_depth)
4342	pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4343	depth, set->queue_depth);
4344
4345	return 0;
4346	}
4347
4348	static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4349	{
4350	/*
4351	* blk_mq_map_queues() and multiple .map_queues() implementations
4352	* expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4353	* number of hardware queues.
4354	*/
4355	if (set->nr_maps == 1)
4356	set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4357
4358	if (set->ops->map_queues && !is_kdump_kernel()) {
4359	int i;
4360
4361	/*
4362	* transport .map_queues is usually done in the following
4363	* way:
4364	*
4365	* for (queue = 0; queue < set->nr_hw_queues; queue++) {
4366	* mask = get_cpu_mask(queue)
4367	* for_each_cpu(cpu, mask)
4368	* set->map[x].mq_map[cpu] = queue;
4369	* }
4370	*
4371	* When we need to remap, the table has to be cleared for
4372	* killing stale mapping since one CPU may not be mapped
4373	* to any hw queue.
4374	*/
4375	for (i = 0; i < set->nr_maps; i++)
4376	blk_mq_clear_mq_map(qmap: &set->map[i]);
4377
4378	set->ops->map_queues(set);
4379	} else {
4380	BUG_ON(set->nr_maps > 1);
4381	blk_mq_map_queues(qmap: &set->map[HCTX_TYPE_DEFAULT]);
4382	}
4383	}
4384
4385	static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4386	int new_nr_hw_queues)
4387	{
4388	struct blk_mq_tags **new_tags;
4389	int i;
4390
4391	if (set->nr_hw_queues >= new_nr_hw_queues)
4392	goto done;
4393
4394	new_tags = kcalloc_node(n: new_nr_hw_queues, size: sizeof(struct blk_mq_tags *),
4395	GFP_KERNEL, node: set->numa_node);
4396	if (!new_tags)
4397	return -ENOMEM;
4398
4399	if (set->tags)
4400	memcpy(new_tags, set->tags, set->nr_hw_queues *
4401	sizeof(*set->tags));
4402	kfree(objp: set->tags);
4403	set->tags = new_tags;
4404
4405	for (i = set->nr_hw_queues; i < new_nr_hw_queues; i++) {
4406	if (!__blk_mq_alloc_map_and_rqs(set, hctx_idx: i)) {
4407	while (--i >= set->nr_hw_queues)
4408	__blk_mq_free_map_and_rqs(set, hctx_idx: i);
4409	return -ENOMEM;
4410	}
4411	cond_resched();
4412	}
4413
4414	done:
4415	set->nr_hw_queues = new_nr_hw_queues;
4416	return 0;
4417	}
4418
4419	/*
4420	* Alloc a tag set to be associated with one or more request queues.
4421	* May fail with EINVAL for various error conditions. May adjust the
4422	* requested depth down, if it's too large. In that case, the set
4423	* value will be stored in set->queue_depth.
4424	*/
4425	int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4426	{
4427	int i, ret;
4428
4429	BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4430
4431	if (!set->nr_hw_queues)
4432	return -EINVAL;
4433	if (!set->queue_depth)
4434	return -EINVAL;
4435	if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4436	return -EINVAL;
4437
4438	if (!set->ops->queue_rq)
4439	return -EINVAL;
4440
4441	if (!set->ops->get_budget ^ !set->ops->put_budget)
4442	return -EINVAL;
4443
4444	if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4445	pr_info("blk-mq: reduced tag depth to %u\n",
4446	BLK_MQ_MAX_DEPTH);
4447	set->queue_depth = BLK_MQ_MAX_DEPTH;
4448	}
4449
4450	if (!set->nr_maps)
4451	set->nr_maps = 1;
4452	else if (set->nr_maps > HCTX_MAX_TYPES)
4453	return -EINVAL;
4454
4455	/*
4456	* If a crashdump is active, then we are potentially in a very
4457	* memory constrained environment. Limit us to 1 queue and
4458	* 64 tags to prevent using too much memory.
4459	*/
4460	if (is_kdump_kernel()) {
4461	set->nr_hw_queues = 1;
4462	set->nr_maps = 1;
4463	set->queue_depth = min(64U, set->queue_depth);
4464	}
4465	/*
4466	* There is no use for more h/w queues than cpus if we just have
4467	* a single map
4468	*/
4469	if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4470	set->nr_hw_queues = nr_cpu_ids;
4471
4472	if (set->flags & BLK_MQ_F_BLOCKING) {
4473	set->srcu = kmalloc(size: sizeof(*set->srcu), GFP_KERNEL);
4474	if (!set->srcu)
4475	return -ENOMEM;
4476	ret = init_srcu_struct(set->srcu);
4477	if (ret)
4478	goto out_free_srcu;
4479	}
4480
4481	ret = -ENOMEM;
4482	set->tags = kcalloc_node(n: set->nr_hw_queues,
4483	size: sizeof(struct blk_mq_tags *), GFP_KERNEL,
4484	node: set->numa_node);
4485	if (!set->tags)
4486	goto out_cleanup_srcu;
4487
4488	for (i = 0; i < set->nr_maps; i++) {
4489	set->map[i].mq_map = kcalloc_node(n: nr_cpu_ids,
4490	size: sizeof(set->map[i].mq_map[0]),
4491	GFP_KERNEL, node: set->numa_node);
4492	if (!set->map[i].mq_map)
4493	goto out_free_mq_map;
4494	set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4495	}
4496
4497	blk_mq_update_queue_map(set);
4498
4499	ret = blk_mq_alloc_set_map_and_rqs(set);
4500	if (ret)
4501	goto out_free_mq_map;
4502
4503	mutex_init(&set->tag_list_lock);
4504	INIT_LIST_HEAD(list: &set->tag_list);
4505
4506	return 0;
4507
4508	out_free_mq_map:
4509	for (i = 0; i < set->nr_maps; i++) {
4510	kfree(objp: set->map[i].mq_map);
4511	set->map[i].mq_map = NULL;
4512	}
4513	kfree(objp: set->tags);
4514	set->tags = NULL;
4515	out_cleanup_srcu:
4516	if (set->flags & BLK_MQ_F_BLOCKING)
4517	cleanup_srcu_struct(ssp: set->srcu);
4518	out_free_srcu:
4519	if (set->flags & BLK_MQ_F_BLOCKING)
4520	kfree(objp: set->srcu);
4521	return ret;
4522	}
4523	EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4524
4525	/* allocate and initialize a tagset for a simple single-queue device */
4526	int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4527	const struct blk_mq_ops *ops, unsigned int queue_depth,
4528	unsigned int set_flags)
4529	{
4530	memset(set, 0, sizeof(*set));
4531	set->ops = ops;
4532	set->nr_hw_queues = 1;
4533	set->nr_maps = 1;
4534	set->queue_depth = queue_depth;
4535	set->numa_node = NUMA_NO_NODE;
4536	set->flags = set_flags;
4537	return blk_mq_alloc_tag_set(set);
4538	}
4539	EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4540
4541	void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4542	{
4543	int i, j;
4544
4545	for (i = 0; i < set->nr_hw_queues; i++)
4546	__blk_mq_free_map_and_rqs(set, hctx_idx: i);
4547
4548	if (blk_mq_is_shared_tags(flags: set->flags)) {
4549	blk_mq_free_map_and_rqs(set, tags: set->shared_tags,
4550	BLK_MQ_NO_HCTX_IDX);
4551	}
4552
4553	for (j = 0; j < set->nr_maps; j++) {
4554	kfree(objp: set->map[j].mq_map);
4555	set->map[j].mq_map = NULL;
4556	}
4557
4558	kfree(objp: set->tags);
4559	set->tags = NULL;
4560	if (set->flags & BLK_MQ_F_BLOCKING) {
4561	cleanup_srcu_struct(ssp: set->srcu);
4562	kfree(objp: set->srcu);
4563	}
4564	}
4565	EXPORT_SYMBOL(blk_mq_free_tag_set);
4566
4567	int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4568	{
4569	struct blk_mq_tag_set *set = q->tag_set;
4570	struct blk_mq_hw_ctx *hctx;
4571	int ret;
4572	unsigned long i;
4573
4574	if (!set)
4575	return -EINVAL;
4576
4577	if (q->nr_requests == nr)
4578	return 0;
4579
4580	blk_mq_freeze_queue(q);
4581	blk_mq_quiesce_queue(q);
4582
4583	ret = 0;
4584	queue_for_each_hw_ctx(q, hctx, i) {
4585	if (!hctx->tags)
4586	continue;
4587	/*
4588	* If we're using an MQ scheduler, just update the scheduler
4589	* queue depth. This is similar to what the old code would do.
4590	*/
4591	if (hctx->sched_tags) {
4592	ret = blk_mq_tag_update_depth(hctx, tags: &hctx->sched_tags,
4593	depth: nr, can_grow: true);
4594	} else {
4595	ret = blk_mq_tag_update_depth(hctx, tags: &hctx->tags, depth: nr,
4596	can_grow: false);
4597	}
4598	if (ret)
4599	break;
4600	if (q->elevator && q->elevator->type->ops.depth_updated)
4601	q->elevator->type->ops.depth_updated(hctx);
4602	}
4603	if (!ret) {
4604	q->nr_requests = nr;
4605	if (blk_mq_is_shared_tags(flags: set->flags)) {
4606	if (q->elevator)
4607	blk_mq_tag_update_sched_shared_tags(q);
4608	else
4609	blk_mq_tag_resize_shared_tags(set, size: nr);
4610	}
4611	}
4612
4613	blk_mq_unquiesce_queue(q);
4614	blk_mq_unfreeze_queue(q);
4615
4616	return ret;
4617	}
4618
4619	/*
4620	* request_queue and elevator_type pair.
4621	* It is just used by __blk_mq_update_nr_hw_queues to cache
4622	* the elevator_type associated with a request_queue.
4623	*/
4624	struct blk_mq_qe_pair {
4625	struct list_head node;
4626	struct request_queue *q;
4627	struct elevator_type *type;
4628	};
4629
4630	/*
4631	* Cache the elevator_type in qe pair list and switch the
4632	* io scheduler to 'none'
4633	*/
4634	static bool blk_mq_elv_switch_none(struct list_head *head,
4635	struct request_queue *q)
4636	{
4637	struct blk_mq_qe_pair *qe;
4638
4639	qe = kmalloc(size: sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4640	if (!qe)
4641	return false;
4642
4643	/* q->elevator needs protection from ->sysfs_lock */
4644	mutex_lock(&q->sysfs_lock);
4645
4646	/* the check has to be done with holding sysfs_lock */
4647	if (!q->elevator) {
4648	kfree(objp: qe);
4649	goto unlock;
4650	}
4651
4652	INIT_LIST_HEAD(list: &qe->node);
4653	qe->q = q;
4654	qe->type = q->elevator->type;
4655	/* keep a reference to the elevator module as we'll switch back */
4656	__elevator_get(e: qe->type);
4657	list_add(new: &qe->node, head);
4658	elevator_disable(q);
4659	unlock:
4660	mutex_unlock(lock: &q->sysfs_lock);
4661
4662	return true;
4663	}
4664
4665	static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4666	struct request_queue *q)
4667	{
4668	struct blk_mq_qe_pair *qe;
4669
4670	list_for_each_entry(qe, head, node)
4671	if (qe->q == q)
4672	return qe;
4673
4674	return NULL;
4675	}
4676
4677	static void blk_mq_elv_switch_back(struct list_head *head,
4678	struct request_queue *q)
4679	{
4680	struct blk_mq_qe_pair *qe;
4681	struct elevator_type *t;
4682
4683	qe = blk_lookup_qe_pair(head, q);
4684	if (!qe)
4685	return;
4686	t = qe->type;
4687	list_del(entry: &qe->node);
4688	kfree(objp: qe);
4689
4690	mutex_lock(&q->sysfs_lock);
4691	elevator_switch(q, new_e: t);
4692	/* drop the reference acquired in blk_mq_elv_switch_none */
4693	elevator_put(e: t);
4694	mutex_unlock(lock: &q->sysfs_lock);
4695	}
4696
4697	static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4698	int nr_hw_queues)
4699	{
4700	struct request_queue *q;
4701	LIST_HEAD(head);
4702	int prev_nr_hw_queues = set->nr_hw_queues;
4703	int i;
4704
4705	lockdep_assert_held(&set->tag_list_lock);
4706
4707	if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4708	nr_hw_queues = nr_cpu_ids;
4709	if (nr_hw_queues < 1)
4710	return;
4711	if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4712	return;
4713
4714	list_for_each_entry(q, &set->tag_list, tag_set_list)
4715	blk_mq_freeze_queue(q);
4716	/*
4717	* Switch IO scheduler to 'none', cleaning up the data associated
4718	* with the previous scheduler. We will switch back once we are done
4719	* updating the new sw to hw queue mappings.
4720	*/
4721	list_for_each_entry(q, &set->tag_list, tag_set_list)
4722	if (!blk_mq_elv_switch_none(head: &head, q))
4723	goto switch_back;
4724
4725	list_for_each_entry(q, &set->tag_list, tag_set_list) {
4726	blk_mq_debugfs_unregister_hctxs(q);
4727	blk_mq_sysfs_unregister_hctxs(q);
4728	}
4729
4730	if (blk_mq_realloc_tag_set_tags(set, new_nr_hw_queues: nr_hw_queues) < 0)
4731	goto reregister;
4732
4733	fallback:
4734	blk_mq_update_queue_map(set);
4735	list_for_each_entry(q, &set->tag_list, tag_set_list) {
4736	blk_mq_realloc_hw_ctxs(set, q);
4737	blk_mq_update_poll_flag(q);
4738	if (q->nr_hw_queues != set->nr_hw_queues) {
4739	int i = prev_nr_hw_queues;
4740
4741	pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4742	nr_hw_queues, prev_nr_hw_queues);
4743	for (; i < set->nr_hw_queues; i++)
4744	__blk_mq_free_map_and_rqs(set, hctx_idx: i);
4745
4746	set->nr_hw_queues = prev_nr_hw_queues;
4747	goto fallback;
4748	}
4749	blk_mq_map_swqueue(q);
4750	}
4751
4752	reregister:
4753	list_for_each_entry(q, &set->tag_list, tag_set_list) {
4754	blk_mq_sysfs_register_hctxs(q);
4755	blk_mq_debugfs_register_hctxs(q);
4756	}
4757
4758	switch_back:
4759	list_for_each_entry(q, &set->tag_list, tag_set_list)
4760	blk_mq_elv_switch_back(head: &head, q);
4761
4762	list_for_each_entry(q, &set->tag_list, tag_set_list)
4763	blk_mq_unfreeze_queue(q);
4764
4765	/* Free the excess tags when nr_hw_queues shrink. */
4766	for (i = set->nr_hw_queues; i < prev_nr_hw_queues; i++)
4767	__blk_mq_free_map_and_rqs(set, hctx_idx: i);
4768	}
4769
4770	void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4771	{
4772	mutex_lock(&set->tag_list_lock);
4773	__blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4774	mutex_unlock(lock: &set->tag_list_lock);
4775	}
4776	EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4777
4778	static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
4779	struct io_comp_batch *iob, unsigned int flags)
4780	{
4781	long state = get_current_state();
4782	int ret;
4783
4784	do {
4785	ret = q->mq_ops->poll(hctx, iob);
4786	if (ret > 0) {
4787	__set_current_state(TASK_RUNNING);
4788	return ret;
4789	}
4790
4791	if (signal_pending_state(state, current))
4792	__set_current_state(TASK_RUNNING);
4793	if (task_is_running(current))
4794	return 1;
4795
4796	if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4797	break;
4798	cpu_relax();
4799	} while (!need_resched());
4800
4801	__set_current_state(TASK_RUNNING);
4802	return 0;
4803	}
4804
4805	int blk_mq_poll(struct request_queue *q, blk_qc_t cookie,
4806	struct io_comp_batch *iob, unsigned int flags)
4807	{
4808	struct blk_mq_hw_ctx *hctx = xa_load(&q->hctx_table, index: cookie);
4809
4810	return blk_hctx_poll(q, hctx, iob, flags);
4811	}
4812
4813	int blk_rq_poll(struct request *rq, struct io_comp_batch *iob,
4814	unsigned int poll_flags)
4815	{
4816	struct request_queue *q = rq->q;
4817	int ret;
4818
4819	if (!blk_rq_is_poll(rq))
4820	return 0;
4821	if (!percpu_ref_tryget(ref: &q->q_usage_counter))
4822	return 0;
4823
4824	ret = blk_hctx_poll(q, hctx: rq->mq_hctx, iob, flags: poll_flags);
4825	blk_queue_exit(q);
4826
4827	return ret;
4828	}
4829	EXPORT_SYMBOL_GPL(blk_rq_poll);
4830
4831	unsigned int blk_mq_rq_cpu(struct request *rq)
4832	{
4833	return rq->mq_ctx->cpu;
4834	}
4835	EXPORT_SYMBOL(blk_mq_rq_cpu);
4836
4837	void blk_mq_cancel_work_sync(struct request_queue *q)
4838	{
4839	struct blk_mq_hw_ctx *hctx;
4840	unsigned long i;
4841
4842	cancel_delayed_work_sync(dwork: &q->requeue_work);
4843
4844	queue_for_each_hw_ctx(q, hctx, i)
4845	cancel_delayed_work_sync(dwork: &hctx->run_work);
4846	}
4847
4848	static int __init blk_mq_init(void)
4849	{
4850	int i;
4851
4852	for_each_possible_cpu(i)
4853	init_llist_head(list: &per_cpu(blk_cpu_done, i));
4854	for_each_possible_cpu(i)
4855	INIT_CSD(&per_cpu(blk_cpu_csd, i),
4856	__blk_mq_complete_request_remote, NULL);
4857	open_softirq(nr: BLOCK_SOFTIRQ, action: blk_done_softirq);
4858
4859	cpuhp_setup_state_nocalls(state: CPUHP_BLOCK_SOFTIRQ_DEAD,
4860	name: "block/softirq:dead", NULL,
4861	teardown: blk_softirq_cpu_dead);
4862	cpuhp_setup_state_multi(state: CPUHP_BLK_MQ_DEAD, name: "block/mq:dead", NULL,
4863	teardown: blk_mq_hctx_notify_dead);
4864	cpuhp_setup_state_multi(state: CPUHP_AP_BLK_MQ_ONLINE, name: "block/mq:online",
4865	startup: blk_mq_hctx_notify_online,
4866	teardown: blk_mq_hctx_notify_offline);
4867	return 0;
4868	}
4869	subsys_initcall(blk_mq_init);
4870