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/topology.h>
25	#include <linux/sched/signal.h>
26	#include <linux/delay.h>
27	#include <linux/crash_dump.h>
28	#include <linux/prefetch.h>
29	#include <linux/blk-crypto.h>
30	#include <linux/part_stat.h>
31
32	#include <trace/events/block.h>
33
34	#include <linux/t10-pi.h>
35	#include "blk.h"
36	#include "blk-mq.h"
37	#include "blk-mq-debugfs.h"
38	#include "blk-pm.h"
39	#include "blk-stat.h"
40	#include "blk-mq-sched.h"
41	#include "blk-rq-qos.h"
42
43	static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
44	static DEFINE_PER_CPU(call_single_data_t, blk_cpu_csd);
45
46	static void blk_mq_insert_request(struct request *rq, blk_insert_t flags);
47	static void blk_mq_request_bypass_insert(struct request *rq,
48	blk_insert_t flags);
49	static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
50	struct list_head *list);
51	static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
52	struct io_comp_batch *iob, unsigned int flags);
53
54	/*
55	* Check if any of the ctx, dispatch list or elevator
56	* have pending work in this hardware queue.
57	*/
58	static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
59	{
60	return !list_empty_careful(head: &hctx->dispatch) ||
61	sbitmap_any_bit_set(sb: &hctx->ctx_map) ||
62	blk_mq_sched_has_work(hctx);
63	}
64
65	/*
66	* Mark this ctx as having pending work in this hardware queue
67	*/
68	static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
69	struct blk_mq_ctx *ctx)
70	{
71	const int bit = ctx->index_hw[hctx->type];
72
73	if (!sbitmap_test_bit(sb: &hctx->ctx_map, bitnr: bit))
74	sbitmap_set_bit(sb: &hctx->ctx_map, bitnr: bit);
75	}
76
77	static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
78	struct blk_mq_ctx *ctx)
79	{
80	const int bit = ctx->index_hw[hctx->type];
81
82	sbitmap_clear_bit(sb: &hctx->ctx_map, bitnr: bit);
83	}
84
85	struct mq_inflight {
86	struct block_device *part;
87	unsigned int inflight[2];
88	};
89
90	static bool blk_mq_check_inflight(struct request *rq, void *priv)
91	{
92	struct mq_inflight *mi = priv;
93
94	if (rq->part && blk_do_io_stat(rq) &&
95	(!mi->part->bd_partno || rq->part == mi->part) &&
96	blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
97	mi->inflight[rq_data_dir(rq)]++;
98
99	return true;
100	}
101
102	unsigned int blk_mq_in_flight(struct request_queue *q,
103	struct block_device *part)
104	{
105	struct mq_inflight mi = { .part = part };
106
107	blk_mq_queue_tag_busy_iter(q, fn: blk_mq_check_inflight, priv: &mi);
108
109	return mi.inflight[0] + mi.inflight[1];
110	}
111
112	void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
113	unsigned int inflight[2])
114	{
115	struct mq_inflight mi = { .part = part };
116
117	blk_mq_queue_tag_busy_iter(q, fn: blk_mq_check_inflight, priv: &mi);
118	inflight[0] = mi.inflight[0];
119	inflight[1] = mi.inflight[1];
120	}
121
122	void blk_freeze_queue_start(struct request_queue *q)
123	{
124	mutex_lock(&q->mq_freeze_lock);
125	if (++q->mq_freeze_depth == 1) {
126	percpu_ref_kill(ref: &q->q_usage_counter);
127	mutex_unlock(lock: &q->mq_freeze_lock);
128	if (queue_is_mq(q))
129	blk_mq_run_hw_queues(q, async: false);
130	} else {
131	mutex_unlock(lock: &q->mq_freeze_lock);
132	}
133	}
134	EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
135
136	void blk_mq_freeze_queue_wait(struct request_queue *q)
137	{
138	wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
139	}
140	EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
141
142	int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
143	unsigned long timeout)
144	{
145	return wait_event_timeout(q->mq_freeze_wq,
146	percpu_ref_is_zero(&q->q_usage_counter),
147	timeout);
148	}
149	EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
150
151	/*
152	* Guarantee no request is in use, so we can change any data structure of
153	* the queue afterward.
154	*/
155	void blk_freeze_queue(struct request_queue *q)
156	{
157	/*
158	* In the !blk_mq case we are only calling this to kill the
159	* q_usage_counter, otherwise this increases the freeze depth
160	* and waits for it to return to zero. For this reason there is
161	* no blk_unfreeze_queue(), and blk_freeze_queue() is not
162	* exported to drivers as the only user for unfreeze is blk_mq.
163	*/
164	blk_freeze_queue_start(q);
165	blk_mq_freeze_queue_wait(q);
166	}
167
168	void blk_mq_freeze_queue(struct request_queue *q)
169	{
170	/*
171	* ...just an alias to keep freeze and unfreeze actions balanced
172	* in the blk_mq_* namespace
173	*/
174	blk_freeze_queue(q);
175	}
176	EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
177
178	void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
179	{
180	mutex_lock(&q->mq_freeze_lock);
181	if (force_atomic)
182	q->q_usage_counter.data->force_atomic = true;
183	q->mq_freeze_depth--;
184	WARN_ON_ONCE(q->mq_freeze_depth < 0);
185	if (!q->mq_freeze_depth) {
186	percpu_ref_resurrect(ref: &q->q_usage_counter);
187	wake_up_all(&q->mq_freeze_wq);
188	}
189	mutex_unlock(lock: &q->mq_freeze_lock);
190	}
191
192	void blk_mq_unfreeze_queue(struct request_queue *q)
193	{
194	__blk_mq_unfreeze_queue(q, force_atomic: false);
195	}
196	EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
197
198	/*
199	* FIXME: replace the scsi_internal_device_*block_nowait() calls in the
200	* mpt3sas driver such that this function can be removed.
201	*/
202	void blk_mq_quiesce_queue_nowait(struct request_queue *q)
203	{
204	unsigned long flags;
205
206	spin_lock_irqsave(&q->queue_lock, flags);
207	if (!q->quiesce_depth++)
208	blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
209	spin_unlock_irqrestore(lock: &q->queue_lock, flags);
210	}
211	EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
212
213	/**
214	* blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
215	* @set: tag_set to wait on
216	*
217	* Note: it is driver's responsibility for making sure that quiesce has
218	* been started on or more of the request_queues of the tag_set. This
219	* function only waits for the quiesce on those request_queues that had
220	* the quiesce flag set using blk_mq_quiesce_queue_nowait.
221	*/
222	void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set)
223	{
224	if (set->flags & BLK_MQ_F_BLOCKING)
225	synchronize_srcu(ssp: set->srcu);
226	else
227	synchronize_rcu();
228	}
229	EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
230
231	/**
232	* blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
233	* @q: request queue.
234	*
235	* Note: this function does not prevent that the struct request end_io()
236	* callback function is invoked. Once this function is returned, we make
237	* sure no dispatch can happen until the queue is unquiesced via
238	* blk_mq_unquiesce_queue().
239	*/
240	void blk_mq_quiesce_queue(struct request_queue *q)
241	{
242	blk_mq_quiesce_queue_nowait(q);
243	/* nothing to wait for non-mq queues */
244	if (queue_is_mq(q))
245	blk_mq_wait_quiesce_done(q->tag_set);
246	}
247	EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
248
249	/*
250	* blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
251	* @q: request queue.
252	*
253	* This function recovers queue into the state before quiescing
254	* which is done by blk_mq_quiesce_queue.
255	*/
256	void blk_mq_unquiesce_queue(struct request_queue *q)
257	{
258	unsigned long flags;
259	bool run_queue = false;
260
261	spin_lock_irqsave(&q->queue_lock, flags);
262	if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
263	;
264	} else if (!--q->quiesce_depth) {
265	blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
266	run_queue = true;
267	}
268	spin_unlock_irqrestore(lock: &q->queue_lock, flags);
269
270	/* dispatch requests which are inserted during quiescing */
271	if (run_queue)
272	blk_mq_run_hw_queues(q, async: true);
273	}
274	EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
275
276	void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set)
277	{
278	struct request_queue *q;
279
280	mutex_lock(&set->tag_list_lock);
281	list_for_each_entry(q, &set->tag_list, tag_set_list) {
282	if (!blk_queue_skip_tagset_quiesce(q))
283	blk_mq_quiesce_queue_nowait(q);
284	}
285	blk_mq_wait_quiesce_done(set);
286	mutex_unlock(lock: &set->tag_list_lock);
287	}
288	EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset);
289
290	void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set)
291	{
292	struct request_queue *q;
293
294	mutex_lock(&set->tag_list_lock);
295	list_for_each_entry(q, &set->tag_list, tag_set_list) {
296	if (!blk_queue_skip_tagset_quiesce(q))
297	blk_mq_unquiesce_queue(q);
298	}
299	mutex_unlock(lock: &set->tag_list_lock);
300	}
301	EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset);
302
303	void blk_mq_wake_waiters(struct request_queue *q)
304	{
305	struct blk_mq_hw_ctx *hctx;
306	unsigned long i;
307
308	queue_for_each_hw_ctx(q, hctx, i)
309	if (blk_mq_hw_queue_mapped(hctx))
310	blk_mq_tag_wakeup_all(tags: hctx->tags, true);
311	}
312
313	void blk_rq_init(struct request_queue *q, struct request *rq)
314	{
315	memset(rq, 0, sizeof(*rq));
316
317	INIT_LIST_HEAD(list: &rq->queuelist);
318	rq->q = q;
319	rq->__sector = (sector_t) -1;
320	INIT_HLIST_NODE(h: &rq->hash);
321	RB_CLEAR_NODE(&rq->rb_node);
322	rq->tag = BLK_MQ_NO_TAG;
323	rq->internal_tag = BLK_MQ_NO_TAG;
324	rq->start_time_ns = blk_time_get_ns();
325	rq->part = NULL;
326	blk_crypto_rq_set_defaults(rq);
327	}
328	EXPORT_SYMBOL(blk_rq_init);
329
330	/* Set start and alloc time when the allocated request is actually used */
331	static inline void blk_mq_rq_time_init(struct request *rq, u64 alloc_time_ns)
332	{
333	if (blk_mq_need_time_stamp(rq))
334	rq->start_time_ns = blk_time_get_ns();
335	else
336	rq->start_time_ns = 0;
337
338	#ifdef CONFIG_BLK_RQ_ALLOC_TIME
339	if (blk_queue_rq_alloc_time(rq->q))
340	rq->alloc_time_ns = alloc_time_ns ?: rq->start_time_ns;
341	else
342	rq->alloc_time_ns = 0;
343	#endif
344	}
345
346	static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
347	struct blk_mq_tags *tags, unsigned int tag)
348	{
349	struct blk_mq_ctx *ctx = data->ctx;
350	struct blk_mq_hw_ctx *hctx = data->hctx;
351	struct request_queue *q = data->q;
352	struct request *rq = tags->static_rqs[tag];
353
354	rq->q = q;
355	rq->mq_ctx = ctx;
356	rq->mq_hctx = hctx;
357	rq->cmd_flags = data->cmd_flags;
358
359	if (data->flags & BLK_MQ_REQ_PM)
360	data->rq_flags |= RQF_PM;
361	if (blk_queue_io_stat(q))
362	data->rq_flags |= RQF_IO_STAT;
363	rq->rq_flags = data->rq_flags;
364
365	if (data->rq_flags & RQF_SCHED_TAGS) {
366	rq->tag = BLK_MQ_NO_TAG;
367	rq->internal_tag = tag;
368	} else {
369	rq->tag = tag;
370	rq->internal_tag = BLK_MQ_NO_TAG;
371	}
372	rq->timeout = 0;
373
374	rq->part = NULL;
375	rq->io_start_time_ns = 0;
376	rq->stats_sectors = 0;
377	rq->nr_phys_segments = 0;
378	#if defined(CONFIG_BLK_DEV_INTEGRITY)
379	rq->nr_integrity_segments = 0;
380	#endif
381	rq->end_io = NULL;
382	rq->end_io_data = NULL;
383
384	blk_crypto_rq_set_defaults(rq);
385	INIT_LIST_HEAD(list: &rq->queuelist);
386	/* tag was already set */
387	WRITE_ONCE(rq->deadline, 0);
388	req_ref_set(req: rq, value: 1);
389
390	if (rq->rq_flags & RQF_USE_SCHED) {
391	struct elevator_queue *e = data->q->elevator;
392
393	INIT_HLIST_NODE(h: &rq->hash);
394	RB_CLEAR_NODE(&rq->rb_node);
395
396	if (e->type->ops.prepare_request)
397	e->type->ops.prepare_request(rq);
398	}
399
400	return rq;
401	}
402
403	static inline struct request *
404	__blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data)
405	{
406	unsigned int tag, tag_offset;
407	struct blk_mq_tags *tags;
408	struct request *rq;
409	unsigned long tag_mask;
410	int i, nr = 0;
411
412	tag_mask = blk_mq_get_tags(data, nr_tags: data->nr_tags, offset: &tag_offset);
413	if (unlikely(!tag_mask))
414	return NULL;
415
416	tags = blk_mq_tags_from_data(data);
417	for (i = 0; tag_mask; i++) {
418	if (!(tag_mask & (1UL << i)))
419	continue;
420	tag = tag_offset + i;
421	prefetch(tags->static_rqs[tag]);
422	tag_mask &= ~(1UL << i);
423	rq = blk_mq_rq_ctx_init(data, tags, tag);
424	rq_list_add(data->cached_rq, rq);
425	nr++;
426	}
427	if (!(data->rq_flags & RQF_SCHED_TAGS))
428	blk_mq_add_active_requests(hctx: data->hctx, val: nr);
429	/* caller already holds a reference, add for remainder */
430	percpu_ref_get_many(ref: &data->q->q_usage_counter, nr: nr - 1);
431	data->nr_tags -= nr;
432
433	return rq_list_pop(data->cached_rq);
434	}
435
436	static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
437	{
438	struct request_queue *q = data->q;
439	u64 alloc_time_ns = 0;
440	struct request *rq;
441	unsigned int tag;
442
443	/* alloc_time includes depth and tag waits */
444	if (blk_queue_rq_alloc_time(q))
445	alloc_time_ns = blk_time_get_ns();
446
447	if (data->cmd_flags & REQ_NOWAIT)
448	data->flags |= BLK_MQ_REQ_NOWAIT;
449
450	if (q->elevator) {
451	/*
452	* All requests use scheduler tags when an I/O scheduler is
453	* enabled for the queue.
454	*/
455	data->rq_flags |= RQF_SCHED_TAGS;
456
457	/*
458	* Flush/passthrough requests are special and go directly to the
459	* dispatch list.
460	*/
461	if ((data->cmd_flags & REQ_OP_MASK) != REQ_OP_FLUSH &&
462	!blk_op_is_passthrough(op: data->cmd_flags)) {
463	struct elevator_mq_ops *ops = &q->elevator->type->ops;
464
465	WARN_ON_ONCE(data->flags & BLK_MQ_REQ_RESERVED);
466
467	data->rq_flags |= RQF_USE_SCHED;
468	if (ops->limit_depth)
469	ops->limit_depth(data->cmd_flags, data);
470	}
471	}
472
473	retry:
474	data->ctx = blk_mq_get_ctx(q);
475	data->hctx = blk_mq_map_queue(q, opf: data->cmd_flags, ctx: data->ctx);
476	if (!(data->rq_flags & RQF_SCHED_TAGS))
477	blk_mq_tag_busy(hctx: data->hctx);
478
479	if (data->flags & BLK_MQ_REQ_RESERVED)
480	data->rq_flags |= RQF_RESV;
481
482	/*
483	* Try batched alloc if we want more than 1 tag.
484	*/
485	if (data->nr_tags > 1) {
486	rq = __blk_mq_alloc_requests_batch(data);
487	if (rq) {
488	blk_mq_rq_time_init(rq, alloc_time_ns);
489	return rq;
490	}
491	data->nr_tags = 1;
492	}
493
494	/*
495	* Waiting allocations only fail because of an inactive hctx. In that
496	* case just retry the hctx assignment and tag allocation as CPU hotplug
497	* should have migrated us to an online CPU by now.
498	*/
499	tag = blk_mq_get_tag(data);
500	if (tag == BLK_MQ_NO_TAG) {
501	if (data->flags & BLK_MQ_REQ_NOWAIT)
502	return NULL;
503	/*
504	* Give up the CPU and sleep for a random short time to
505	* ensure that thread using a realtime scheduling class
506	* are migrated off the CPU, and thus off the hctx that
507	* is going away.
508	*/
509	msleep(msecs: 3);
510	goto retry;
511	}
512
513	if (!(data->rq_flags & RQF_SCHED_TAGS))
514	blk_mq_inc_active_requests(hctx: data->hctx);
515	rq = blk_mq_rq_ctx_init(data, tags: blk_mq_tags_from_data(data), tag);
516	blk_mq_rq_time_init(rq, alloc_time_ns);
517	return rq;
518	}
519
520	static struct request *blk_mq_rq_cache_fill(struct request_queue *q,
521	struct blk_plug *plug,
522	blk_opf_t opf,
523	blk_mq_req_flags_t flags)
524	{
525	struct blk_mq_alloc_data data = {
526	.q = q,
527	.flags = flags,
528	.cmd_flags = opf,
529	.nr_tags = plug->nr_ios,
530	.cached_rq = &plug->cached_rq,
531	};
532	struct request *rq;
533
534	if (blk_queue_enter(q, flags))
535	return NULL;
536
537	plug->nr_ios = 1;
538
539	rq = __blk_mq_alloc_requests(data: &data);
540	if (unlikely(!rq))
541	blk_queue_exit(q);
542	return rq;
543	}
544
545	static struct request *blk_mq_alloc_cached_request(struct request_queue *q,
546	blk_opf_t opf,
547	blk_mq_req_flags_t flags)
548	{
549	struct blk_plug *plug = current->plug;
550	struct request *rq;
551
552	if (!plug)
553	return NULL;
554
555	if (rq_list_empty(plug->cached_rq)) {
556	if (plug->nr_ios == 1)
557	return NULL;
558	rq = blk_mq_rq_cache_fill(q, plug, opf, flags);
559	if (!rq)
560	return NULL;
561	} else {
562	rq = rq_list_peek(&plug->cached_rq);
563	if (!rq || rq->q != q)
564	return NULL;
565
566	if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type)
567	return NULL;
568	if (op_is_flush(op: rq->cmd_flags) != op_is_flush(op: opf))
569	return NULL;
570
571	plug->cached_rq = rq_list_next(rq);
572	blk_mq_rq_time_init(rq, alloc_time_ns: 0);
573	}
574
575	rq->cmd_flags = opf;
576	INIT_LIST_HEAD(list: &rq->queuelist);
577	return rq;
578	}
579
580	struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf,
581	blk_mq_req_flags_t flags)
582	{
583	struct request *rq;
584
585	rq = blk_mq_alloc_cached_request(q, opf, flags);
586	if (!rq) {
587	struct blk_mq_alloc_data data = {
588	.q = q,
589	.flags = flags,
590	.cmd_flags = opf,
591	.nr_tags = 1,
592	};
593	int ret;
594
595	ret = blk_queue_enter(q, flags);
596	if (ret)
597	return ERR_PTR(error: ret);
598
599	rq = __blk_mq_alloc_requests(data: &data);
600	if (!rq)
601	goto out_queue_exit;
602	}
603	rq->__data_len = 0;
604	rq->__sector = (sector_t) -1;
605	rq->bio = rq->biotail = NULL;
606	return rq;
607	out_queue_exit:
608	blk_queue_exit(q);
609	return ERR_PTR(error: -EWOULDBLOCK);
610	}
611	EXPORT_SYMBOL(blk_mq_alloc_request);
612
613	struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
614	blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx)
615	{
616	struct blk_mq_alloc_data data = {
617	.q = q,
618	.flags = flags,
619	.cmd_flags = opf,
620	.nr_tags = 1,
621	};
622	u64 alloc_time_ns = 0;
623	struct request *rq;
624	unsigned int cpu;
625	unsigned int tag;
626	int ret;
627
628	/* alloc_time includes depth and tag waits */
629	if (blk_queue_rq_alloc_time(q))
630	alloc_time_ns = blk_time_get_ns();
631
632	/*
633	* If the tag allocator sleeps we could get an allocation for a
634	* different hardware context. No need to complicate the low level
635	* allocator for this for the rare use case of a command tied to
636	* a specific queue.
637	*/
638	if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) ||
639	WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED)))
640	return ERR_PTR(error: -EINVAL);
641
642	if (hctx_idx >= q->nr_hw_queues)
643	return ERR_PTR(error: -EIO);
644
645	ret = blk_queue_enter(q, flags);
646	if (ret)
647	return ERR_PTR(error: ret);
648
649	/*
650	* Check if the hardware context is actually mapped to anything.
651	* If not tell the caller that it should skip this queue.
652	*/
653	ret = -EXDEV;
654	data.hctx = xa_load(&q->hctx_table, index: hctx_idx);
655	if (!blk_mq_hw_queue_mapped(hctx: data.hctx))
656	goto out_queue_exit;
657	cpu = cpumask_first_and(srcp1: data.hctx->cpumask, cpu_online_mask);
658	if (cpu >= nr_cpu_ids)
659	goto out_queue_exit;
660	data.ctx = __blk_mq_get_ctx(q, cpu);
661
662	if (q->elevator)
663	data.rq_flags |= RQF_SCHED_TAGS;
664	else
665	blk_mq_tag_busy(hctx: data.hctx);
666
667	if (flags & BLK_MQ_REQ_RESERVED)
668	data.rq_flags |= RQF_RESV;
669
670	ret = -EWOULDBLOCK;
671	tag = blk_mq_get_tag(data: &data);
672	if (tag == BLK_MQ_NO_TAG)
673	goto out_queue_exit;
674	if (!(data.rq_flags & RQF_SCHED_TAGS))
675	blk_mq_inc_active_requests(hctx: data.hctx);
676	rq = blk_mq_rq_ctx_init(data: &data, tags: blk_mq_tags_from_data(data: &data), tag);
677	blk_mq_rq_time_init(rq, alloc_time_ns);
678	rq->__data_len = 0;
679	rq->__sector = (sector_t) -1;
680	rq->bio = rq->biotail = NULL;
681	return rq;
682
683	out_queue_exit:
684	blk_queue_exit(q);
685	return ERR_PTR(error: ret);
686	}
687	EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
688
689	static void blk_mq_finish_request(struct request *rq)
690	{
691	struct request_queue *q = rq->q;
692
693	if (rq->rq_flags & RQF_USE_SCHED) {
694	q->elevator->type->ops.finish_request(rq);
695	/*
696	* For postflush request that may need to be
697	* completed twice, we should clear this flag
698	* to avoid double finish_request() on the rq.
699	*/
700	rq->rq_flags &= ~RQF_USE_SCHED;
701	}
702	}
703
704	static void __blk_mq_free_request(struct request *rq)
705	{
706	struct request_queue *q = rq->q;
707	struct blk_mq_ctx *ctx = rq->mq_ctx;
708	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
709	const int sched_tag = rq->internal_tag;
710
711	blk_crypto_free_request(rq);
712	blk_pm_mark_last_busy(rq);
713	rq->mq_hctx = NULL;
714
715	if (rq->tag != BLK_MQ_NO_TAG) {
716	blk_mq_dec_active_requests(hctx);
717	blk_mq_put_tag(tags: hctx->tags, ctx, tag: rq->tag);
718	}
719	if (sched_tag != BLK_MQ_NO_TAG)
720	blk_mq_put_tag(tags: hctx->sched_tags, ctx, tag: sched_tag);
721	blk_mq_sched_restart(hctx);
722	blk_queue_exit(q);
723	}
724
725	void blk_mq_free_request(struct request *rq)
726	{
727	struct request_queue *q = rq->q;
728
729	blk_mq_finish_request(rq);
730
731	if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
732	laptop_io_completion(info: q->disk->bdi);
733
734	rq_qos_done(q, rq);
735
736	WRITE_ONCE(rq->state, MQ_RQ_IDLE);
737	if (req_ref_put_and_test(req: rq))
738	__blk_mq_free_request(rq);
739	}
740	EXPORT_SYMBOL_GPL(blk_mq_free_request);
741
742	void blk_mq_free_plug_rqs(struct blk_plug *plug)
743	{
744	struct request *rq;
745
746	while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
747	blk_mq_free_request(rq);
748	}
749
750	void blk_dump_rq_flags(struct request *rq, char *msg)
751	{
752	printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
753	rq->q->disk ? rq->q->disk->disk_name : "?",
754	(__force unsigned long long) rq->cmd_flags);
755
756	printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
757	(unsigned long long)blk_rq_pos(rq),
758	blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
759	printk(KERN_INFO " bio %p, biotail %p, len %u\n",
760	rq->bio, rq->biotail, blk_rq_bytes(rq));
761	}
762	EXPORT_SYMBOL(blk_dump_rq_flags);
763
764	static void req_bio_endio(struct request *rq, struct bio *bio,
765	unsigned int nbytes, blk_status_t error)
766	{
767	if (unlikely(error)) {
768	bio->bi_status = error;
769	} else if (req_op(req: rq) == REQ_OP_ZONE_APPEND) {
770	/*
771	* Partial zone append completions cannot be supported as the
772	* BIO fragments may end up not being written sequentially.
773	*/
774	if (bio->bi_iter.bi_size != nbytes)
775	bio->bi_status = BLK_STS_IOERR;
776	else
777	bio->bi_iter.bi_sector = rq->__sector;
778	}
779
780	bio_advance(bio, nbytes);
781
782	if (unlikely(rq->rq_flags & RQF_QUIET))
783	bio_set_flag(bio, bit: BIO_QUIET);
784	/* don't actually finish bio if it's part of flush sequence */
785	if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
786	bio_endio(bio);
787	}
788
789	static void blk_account_io_completion(struct request *req, unsigned int bytes)
790	{
791	if (req->part && blk_do_io_stat(rq: req)) {
792	const int sgrp = op_stat_group(op: req_op(req));
793
794	part_stat_lock();
795	part_stat_add(req->part, sectors[sgrp], bytes >> 9);
796	part_stat_unlock();
797	}
798	}
799
800	static void blk_print_req_error(struct request *req, blk_status_t status)
801	{
802	printk_ratelimited(KERN_ERR
803	"%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
804	"phys_seg %u prio class %u\n",
805	blk_status_to_str(status),
806	req->q->disk ? req->q->disk->disk_name : "?",
807	blk_rq_pos(req), (__force u32)req_op(req),
808	blk_op_str(req_op(req)),
809	(__force u32)(req->cmd_flags & ~REQ_OP_MASK),
810	req->nr_phys_segments,
811	IOPRIO_PRIO_CLASS(req->ioprio));
812	}
813
814	/*
815	* Fully end IO on a request. Does not support partial completions, or
816	* errors.
817	*/
818	static void blk_complete_request(struct request *req)
819	{
820	const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
821	int total_bytes = blk_rq_bytes(rq: req);
822	struct bio *bio = req->bio;
823
824	trace_block_rq_complete(rq: req, BLK_STS_OK, nr_bytes: total_bytes);
825
826	if (!bio)
827	return;
828
829	#ifdef CONFIG_BLK_DEV_INTEGRITY
830	if (blk_integrity_rq(rq: req) && req_op(req) == REQ_OP_READ)
831	req->q->integrity.profile->complete_fn(req, total_bytes);
832	#endif
833
834	/*
835	* Upper layers may call blk_crypto_evict_key() anytime after the last
836	* bio_endio(). Therefore, the keyslot must be released before that.
837	*/
838	blk_crypto_rq_put_keyslot(rq: req);
839
840	blk_account_io_completion(req, bytes: total_bytes);
841
842	do {
843	struct bio *next = bio->bi_next;
844
845	/* Completion has already been traced */
846	bio_clear_flag(bio, bit: BIO_TRACE_COMPLETION);
847
848	if (req_op(req) == REQ_OP_ZONE_APPEND)
849	bio->bi_iter.bi_sector = req->__sector;
850
851	if (!is_flush)
852	bio_endio(bio);
853	bio = next;
854	} while (bio);
855
856	/*
857	* Reset counters so that the request stacking driver
858	* can find how many bytes remain in the request
859	* later.
860	*/
861	if (!req->end_io) {
862	req->bio = NULL;
863	req->__data_len = 0;
864	}
865	}
866
867	/**
868	* blk_update_request - Complete multiple bytes without completing the request
869	* @req: the request being processed
870	* @error: block status code
871	* @nr_bytes: number of bytes to complete for @req
872	*
873	* Description:
874	* Ends I/O on a number of bytes attached to @req, but doesn't complete
875	* the request structure even if @req doesn't have leftover.
876	* If @req has leftover, sets it up for the next range of segments.
877	*
878	* Passing the result of blk_rq_bytes() as @nr_bytes guarantees
879	* %false return from this function.
880	*
881	* Note:
882	* The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
883	* except in the consistency check at the end of this function.
884	*
885	* Return:
886	* %false - this request doesn't have any more data
887	* %true - this request has more data
888	**/
889	bool blk_update_request(struct request *req, blk_status_t error,
890	unsigned int nr_bytes)
891	{
892	int total_bytes;
893
894	trace_block_rq_complete(rq: req, error, nr_bytes);
895
896	if (!req->bio)
897	return false;
898
899	#ifdef CONFIG_BLK_DEV_INTEGRITY
900	if (blk_integrity_rq(rq: req) && req_op(req) == REQ_OP_READ &&
901	error == BLK_STS_OK)
902	req->q->integrity.profile->complete_fn(req, nr_bytes);
903	#endif
904
905	/*
906	* Upper layers may call blk_crypto_evict_key() anytime after the last
907	* bio_endio(). Therefore, the keyslot must be released before that.
908	*/
909	if (blk_crypto_rq_has_keyslot(rq: req) && nr_bytes >= blk_rq_bytes(rq: req))
910	__blk_crypto_rq_put_keyslot(rq: req);
911
912	if (unlikely(error && !blk_rq_is_passthrough(req) &&
913	!(req->rq_flags & RQF_QUIET)) &&
914	!test_bit(GD_DEAD, &req->q->disk->state)) {
915	blk_print_req_error(req, status: error);
916	trace_block_rq_error(rq: req, error, nr_bytes);
917	}
918
919	blk_account_io_completion(req, bytes: nr_bytes);
920
921	total_bytes = 0;
922	while (req->bio) {
923	struct bio *bio = req->bio;
924	unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
925
926	if (bio_bytes == bio->bi_iter.bi_size)
927	req->bio = bio->bi_next;
928
929	/* Completion has already been traced */
930	bio_clear_flag(bio, bit: BIO_TRACE_COMPLETION);
931	req_bio_endio(rq: req, bio, nbytes: bio_bytes, error);
932
933	total_bytes += bio_bytes;
934	nr_bytes -= bio_bytes;
935
936	if (!nr_bytes)
937	break;
938	}
939
940	/*
941	* completely done
942	*/
943	if (!req->bio) {
944	/*
945	* Reset counters so that the request stacking driver
946	* can find how many bytes remain in the request
947	* later.
948	*/
949	req->__data_len = 0;
950	return false;
951	}
952
953	req->__data_len -= total_bytes;
954
955	/* update sector only for requests with clear definition of sector */
956	if (!blk_rq_is_passthrough(rq: req))
957	req->__sector += total_bytes >> 9;
958
959	/* mixed attributes always follow the first bio */
960	if (req->rq_flags & RQF_MIXED_MERGE) {
961	req->cmd_flags &= ~REQ_FAILFAST_MASK;
962	req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
963	}
964
965	if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
966	/*
967	* If total number of sectors is less than the first segment
968	* size, something has gone terribly wrong.
969	*/
970	if (blk_rq_bytes(rq: req) < blk_rq_cur_bytes(rq: req)) {
971	blk_dump_rq_flags(req, "request botched");
972	req->__data_len = blk_rq_cur_bytes(rq: req);
973	}
974
975	/* recalculate the number of segments */
976	req->nr_phys_segments = blk_recalc_rq_segments(rq: req);
977	}
978
979	return true;
980	}
981	EXPORT_SYMBOL_GPL(blk_update_request);
982
983	static inline void blk_account_io_done(struct request *req, u64 now)
984	{
985	trace_block_io_done(rq: req);
986
987	/*
988	* Account IO completion. flush_rq isn't accounted as a
989	* normal IO on queueing nor completion. Accounting the
990	* containing request is enough.
991	*/
992	if (blk_do_io_stat(rq: req) && req->part &&
993	!(req->rq_flags & RQF_FLUSH_SEQ)) {
994	const int sgrp = op_stat_group(op: req_op(req));
995
996	part_stat_lock();
997	update_io_ticks(part: req->part, now: jiffies, end: true);
998	part_stat_inc(req->part, ios[sgrp]);
999	part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
1000	part_stat_unlock();
1001	}
1002	}
1003
1004	static inline void blk_account_io_start(struct request *req)
1005	{
1006	trace_block_io_start(rq: req);
1007
1008	if (blk_do_io_stat(rq: req)) {
1009	/*
1010	* All non-passthrough requests are created from a bio with one
1011	* exception: when a flush command that is part of a flush sequence
1012	* generated by the state machine in blk-flush.c is cloned onto the
1013	* lower device by dm-multipath we can get here without a bio.
1014	*/
1015	if (req->bio)
1016	req->part = req->bio->bi_bdev;
1017	else
1018	req->part = req->q->disk->part0;
1019
1020	part_stat_lock();
1021	update_io_ticks(part: req->part, now: jiffies, end: false);
1022	part_stat_unlock();
1023	}
1024	}
1025
1026	static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
1027	{
1028	if (rq->rq_flags & RQF_STATS)
1029	blk_stat_add(rq, now);
1030
1031	blk_mq_sched_completed_request(rq, now);
1032	blk_account_io_done(req: rq, now);
1033	}
1034
1035	inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
1036	{
1037	if (blk_mq_need_time_stamp(rq))
1038	__blk_mq_end_request_acct(rq, now: blk_time_get_ns());
1039
1040	blk_mq_finish_request(rq);
1041
1042	if (rq->end_io) {
1043	rq_qos_done(q: rq->q, rq);
1044	if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1045	blk_mq_free_request(rq);
1046	} else {
1047	blk_mq_free_request(rq);
1048	}
1049	}
1050	EXPORT_SYMBOL(__blk_mq_end_request);
1051
1052	void blk_mq_end_request(struct request *rq, blk_status_t error)
1053	{
1054	if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1055	BUG();
1056	__blk_mq_end_request(rq, error);
1057	}
1058	EXPORT_SYMBOL(blk_mq_end_request);
1059
1060	#define TAG_COMP_BATCH 32
1061
1062	static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1063	int *tag_array, int nr_tags)
1064	{
1065	struct request_queue *q = hctx->queue;
1066
1067	blk_mq_sub_active_requests(hctx, val: nr_tags);
1068
1069	blk_mq_put_tags(tags: hctx->tags, tag_array, nr_tags);
1070	percpu_ref_put_many(ref: &q->q_usage_counter, nr: nr_tags);
1071	}
1072
1073	void blk_mq_end_request_batch(struct io_comp_batch *iob)
1074	{
1075	int tags[TAG_COMP_BATCH], nr_tags = 0;
1076	struct blk_mq_hw_ctx *cur_hctx = NULL;
1077	struct request *rq;
1078	u64 now = 0;
1079
1080	if (iob->need_ts)
1081	now = blk_time_get_ns();
1082
1083	while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1084	prefetch(rq->bio);
1085	prefetch(rq->rq_next);
1086
1087	blk_complete_request(req: rq);
1088	if (iob->need_ts)
1089	__blk_mq_end_request_acct(rq, now);
1090
1091	blk_mq_finish_request(rq);
1092
1093	rq_qos_done(q: rq->q, rq);
1094
1095	/*
1096	* If end_io handler returns NONE, then it still has
1097	* ownership of the request.
1098	*/
1099	if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
1100	continue;
1101
1102	WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1103	if (!req_ref_put_and_test(req: rq))
1104	continue;
1105
1106	blk_crypto_free_request(rq);
1107	blk_pm_mark_last_busy(rq);
1108
1109	if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1110	if (cur_hctx)
1111	blk_mq_flush_tag_batch(hctx: cur_hctx, tag_array: tags, nr_tags);
1112	nr_tags = 0;
1113	cur_hctx = rq->mq_hctx;
1114	}
1115	tags[nr_tags++] = rq->tag;
1116	}
1117
1118	if (nr_tags)
1119	blk_mq_flush_tag_batch(hctx: cur_hctx, tag_array: tags, nr_tags);
1120	}
1121	EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1122
1123	static void blk_complete_reqs(struct llist_head *list)
1124	{
1125	struct llist_node *entry = llist_reverse_order(head: llist_del_all(head: list));
1126	struct request *rq, *next;
1127
1128	llist_for_each_entry_safe(rq, next, entry, ipi_list)
1129	rq->q->mq_ops->complete(rq);
1130	}
1131
1132	static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1133	{
1134	blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1135	}
1136
1137	static int blk_softirq_cpu_dead(unsigned int cpu)
1138	{
1139	blk_complete_reqs(list: &per_cpu(blk_cpu_done, cpu));
1140	return 0;
1141	}
1142
1143	static void __blk_mq_complete_request_remote(void *data)
1144	{
1145	__raise_softirq_irqoff(nr: BLOCK_SOFTIRQ);
1146	}
1147
1148	static inline bool blk_mq_complete_need_ipi(struct request *rq)
1149	{
1150	int cpu = raw_smp_processor_id();
1151
1152	if (!IS_ENABLED(CONFIG_SMP) ||
1153	!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1154	return false;
1155	/*
1156	* With force threaded interrupts enabled, raising softirq from an SMP
1157	* function call will always result in waking the ksoftirqd thread.
1158	* This is probably worse than completing the request on a different
1159	* cache domain.
1160	*/
1161	if (force_irqthreads())
1162	return false;
1163
1164	/* same CPU or cache domain and capacity? Complete locally */
1165	if (cpu == rq->mq_ctx->cpu ||
1166	(!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1167	cpus_share_cache(this_cpu: cpu, that_cpu: rq->mq_ctx->cpu) &&
1168	cpus_equal_capacity(this_cpu: cpu, that_cpu: rq->mq_ctx->cpu)))
1169	return false;
1170
1171	/* don't try to IPI to an offline CPU */
1172	return cpu_online(cpu: rq->mq_ctx->cpu);
1173	}
1174
1175	static void blk_mq_complete_send_ipi(struct request *rq)
1176	{
1177	unsigned int cpu;
1178
1179	cpu = rq->mq_ctx->cpu;
1180	if (llist_add(new: &rq->ipi_list, head: &per_cpu(blk_cpu_done, cpu)))
1181	smp_call_function_single_async(cpu, csd: &per_cpu(blk_cpu_csd, cpu));
1182	}
1183
1184	static void blk_mq_raise_softirq(struct request *rq)
1185	{
1186	struct llist_head *list;
1187
1188	preempt_disable();
1189	list = this_cpu_ptr(&blk_cpu_done);
1190	if (llist_add(new: &rq->ipi_list, head: list))
1191	raise_softirq(nr: BLOCK_SOFTIRQ);
1192	preempt_enable();
1193	}
1194
1195	bool blk_mq_complete_request_remote(struct request *rq)
1196	{
1197	WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1198
1199	/*
1200	* For request which hctx has only one ctx mapping,
1201	* or a polled request, always complete locally,
1202	* it's pointless to redirect the completion.
1203	*/
1204	if ((rq->mq_hctx->nr_ctx == 1 &&
1205	rq->mq_ctx->cpu == raw_smp_processor_id()) ||
1206	rq->cmd_flags & REQ_POLLED)
1207	return false;
1208
1209	if (blk_mq_complete_need_ipi(rq)) {
1210	blk_mq_complete_send_ipi(rq);
1211	return true;
1212	}
1213
1214	if (rq->q->nr_hw_queues == 1) {
1215	blk_mq_raise_softirq(rq);
1216	return true;
1217	}
1218	return false;
1219	}
1220	EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1221
1222	/**
1223	* blk_mq_complete_request - end I/O on a request
1224	* @rq: the request being processed
1225	*
1226	* Description:
1227	* Complete a request by scheduling the ->complete_rq operation.
1228	**/
1229	void blk_mq_complete_request(struct request *rq)
1230	{
1231	if (!blk_mq_complete_request_remote(rq))
1232	rq->q->mq_ops->complete(rq);
1233	}
1234	EXPORT_SYMBOL(blk_mq_complete_request);
1235
1236	/**
1237	* blk_mq_start_request - Start processing a request
1238	* @rq: Pointer to request to be started
1239	*
1240	* Function used by device drivers to notify the block layer that a request
1241	* is going to be processed now, so blk layer can do proper initializations
1242	* such as starting the timeout timer.
1243	*/
1244	void blk_mq_start_request(struct request *rq)
1245	{
1246	struct request_queue *q = rq->q;
1247
1248	trace_block_rq_issue(rq);
1249
1250	if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags) &&
1251	!blk_rq_is_passthrough(rq)) {
1252	rq->io_start_time_ns = blk_time_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	blk_wait_io(done: &wait.done);
1411
1412	return wait.ret;
1413	}
1414	EXPORT_SYMBOL(blk_execute_rq);
1415
1416	static void __blk_mq_requeue_request(struct request *rq)
1417	{
1418	struct request_queue *q = rq->q;
1419
1420	blk_mq_put_driver_tag(rq);
1421
1422	trace_block_rq_requeue(rq);
1423	rq_qos_requeue(q, rq);
1424
1425	if (blk_mq_request_started(rq)) {
1426	WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1427	rq->rq_flags &= ~RQF_TIMED_OUT;
1428	}
1429	}
1430
1431	void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1432	{
1433	struct request_queue *q = rq->q;
1434	unsigned long flags;
1435
1436	__blk_mq_requeue_request(rq);
1437
1438	/* this request will be re-inserted to io scheduler queue */
1439	blk_mq_sched_requeue_request(rq);
1440
1441	spin_lock_irqsave(&q->requeue_lock, flags);
1442	list_add_tail(new: &rq->queuelist, head: &q->requeue_list);
1443	spin_unlock_irqrestore(lock: &q->requeue_lock, flags);
1444
1445	if (kick_requeue_list)
1446	blk_mq_kick_requeue_list(q);
1447	}
1448	EXPORT_SYMBOL(blk_mq_requeue_request);
1449
1450	static void blk_mq_requeue_work(struct work_struct *work)
1451	{
1452	struct request_queue *q =
1453	container_of(work, struct request_queue, requeue_work.work);
1454	LIST_HEAD(rq_list);
1455	LIST_HEAD(flush_list);
1456	struct request *rq;
1457
1458	spin_lock_irq(lock: &q->requeue_lock);
1459	list_splice_init(list: &q->requeue_list, head: &rq_list);
1460	list_splice_init(list: &q->flush_list, head: &flush_list);
1461	spin_unlock_irq(lock: &q->requeue_lock);
1462
1463	while (!list_empty(head: &rq_list)) {
1464	rq = list_entry(rq_list.next, struct request, queuelist);
1465	/*
1466	* If RQF_DONTPREP ist set, the request has been started by the
1467	* driver already and might have driver-specific data allocated
1468	* already. Insert it into the hctx dispatch list to avoid
1469	* block layer merges for the request.
1470	*/
1471	if (rq->rq_flags & RQF_DONTPREP) {
1472	list_del_init(entry: &rq->queuelist);
1473	blk_mq_request_bypass_insert(rq, flags: 0);
1474	} else {
1475	list_del_init(entry: &rq->queuelist);
1476	blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD);
1477	}
1478	}
1479
1480	while (!list_empty(head: &flush_list)) {
1481	rq = list_entry(flush_list.next, struct request, queuelist);
1482	list_del_init(entry: &rq->queuelist);
1483	blk_mq_insert_request(rq, flags: 0);
1484	}
1485
1486	blk_mq_run_hw_queues(q, async: false);
1487	}
1488
1489	void blk_mq_kick_requeue_list(struct request_queue *q)
1490	{
1491	kblockd_mod_delayed_work_on(cpu: WORK_CPU_UNBOUND, dwork: &q->requeue_work, delay: 0);
1492	}
1493	EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1494
1495	void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1496	unsigned long msecs)
1497	{
1498	kblockd_mod_delayed_work_on(cpu: WORK_CPU_UNBOUND, dwork: &q->requeue_work,
1499	delay: msecs_to_jiffies(m: msecs));
1500	}
1501	EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1502
1503	static bool blk_is_flush_data_rq(struct request *rq)
1504	{
1505	return (rq->rq_flags & RQF_FLUSH_SEQ) && !is_flush_rq(req: rq);
1506	}
1507
1508	static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1509	{
1510	/*
1511	* If we find a request that isn't idle we know the queue is busy
1512	* as it's checked in the iter.
1513	* Return false to stop the iteration.
1514	*
1515	* In case of queue quiesce, if one flush data request is completed,
1516	* don't count it as inflight given the flush sequence is suspended,
1517	* and the original flush data request is invisible to driver, just
1518	* like other pending requests because of quiesce
1519	*/
1520	if (blk_mq_request_started(rq) && !(blk_queue_quiesced(rq->q) &&
1521	blk_is_flush_data_rq(rq) &&
1522	blk_mq_request_completed(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	* Add one explicit barrier since blk_mq_get_driver_tag() may
1851	* not imply barrier in case of failure.
1852	*
1853	* Order adding us to wait queue and allocating driver tag.
1854	*
1855	* The pair is the one implied in sbitmap_queue_wake_up() which
1856	* orders clearing sbitmap tag bits and waitqueue_active() in
1857	* __sbitmap_queue_wake_up(), since waitqueue_active() is lockless
1858	*
1859	* Otherwise, re-order of adding wait queue and getting driver tag
1860	* may cause __sbitmap_queue_wake_up() to wake up nothing because
1861	* the waitqueue_active() may not observe us in wait queue.
1862	*/
1863	smp_mb();
1864
1865	/*
1866	* It's possible that a tag was freed in the window between the
1867	* allocation failure and adding the hardware queue to the wait
1868	* queue.
1869	*/
1870	ret = blk_mq_get_driver_tag(rq);
1871	if (!ret) {
1872	spin_unlock(lock: &hctx->dispatch_wait_lock);
1873	spin_unlock_irq(lock: &wq->lock);
1874	return false;
1875	}
1876
1877	/*
1878	* We got a tag, remove ourselves from the wait queue to ensure
1879	* someone else gets the wakeup.
1880	*/
1881	list_del_init(entry: &wait->entry);
1882	atomic_dec(v: &sbq->ws_active);
1883	spin_unlock(lock: &hctx->dispatch_wait_lock);
1884	spin_unlock_irq(lock: &wq->lock);
1885
1886	return true;
1887	}
1888
1889	#define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1890	#define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1891	/*
1892	* Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1893	* - EWMA is one simple way to compute running average value
1894	* - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1895	* - take 4 as factor for avoiding to get too small(0) result, and this
1896	* factor doesn't matter because EWMA decreases exponentially
1897	*/
1898	static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1899	{
1900	unsigned int ewma;
1901
1902	ewma = hctx->dispatch_busy;
1903
1904	if (!ewma && !busy)
1905	return;
1906
1907	ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1908	if (busy)
1909	ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1910	ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1911
1912	hctx->dispatch_busy = ewma;
1913	}
1914
1915	#define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1916
1917	static void blk_mq_handle_dev_resource(struct request *rq,
1918	struct list_head *list)
1919	{
1920	list_add(new: &rq->queuelist, head: list);
1921	__blk_mq_requeue_request(rq);
1922	}
1923
1924	static void blk_mq_handle_zone_resource(struct request *rq,
1925	struct list_head *zone_list)
1926	{
1927	/*
1928	* If we end up here it is because we cannot dispatch a request to a
1929	* specific zone due to LLD level zone-write locking or other zone
1930	* related resource not being available. In this case, set the request
1931	* aside in zone_list for retrying it later.
1932	*/
1933	list_add(new: &rq->queuelist, head: zone_list);
1934	__blk_mq_requeue_request(rq);
1935	}
1936
1937	enum prep_dispatch {
1938	PREP_DISPATCH_OK,
1939	PREP_DISPATCH_NO_TAG,
1940	PREP_DISPATCH_NO_BUDGET,
1941	};
1942
1943	static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1944	bool need_budget)
1945	{
1946	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1947	int budget_token = -1;
1948
1949	if (need_budget) {
1950	budget_token = blk_mq_get_dispatch_budget(q: rq->q);
1951	if (budget_token < 0) {
1952	blk_mq_put_driver_tag(rq);
1953	return PREP_DISPATCH_NO_BUDGET;
1954	}
1955	blk_mq_set_rq_budget_token(rq, token: budget_token);
1956	}
1957
1958	if (!blk_mq_get_driver_tag(rq)) {
1959	/*
1960	* The initial allocation attempt failed, so we need to
1961	* rerun the hardware queue when a tag is freed. The
1962	* waitqueue takes care of that. If the queue is run
1963	* before we add this entry back on the dispatch list,
1964	* we'll re-run it below.
1965	*/
1966	if (!blk_mq_mark_tag_wait(hctx, rq)) {
1967	/*
1968	* All budgets not got from this function will be put
1969	* together during handling partial dispatch
1970	*/
1971	if (need_budget)
1972	blk_mq_put_dispatch_budget(q: rq->q, budget_token);
1973	return PREP_DISPATCH_NO_TAG;
1974	}
1975	}
1976
1977	return PREP_DISPATCH_OK;
1978	}
1979
1980	/* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1981	static void blk_mq_release_budgets(struct request_queue *q,
1982	struct list_head *list)
1983	{
1984	struct request *rq;
1985
1986	list_for_each_entry(rq, list, queuelist) {
1987	int budget_token = blk_mq_get_rq_budget_token(rq);
1988
1989	if (budget_token >= 0)
1990	blk_mq_put_dispatch_budget(q, budget_token);
1991	}
1992	}
1993
1994	/*
1995	* blk_mq_commit_rqs will notify driver using bd->last that there is no
1996	* more requests. (See comment in struct blk_mq_ops for commit_rqs for
1997	* details)
1998	* Attention, we should explicitly call this in unusual cases:
1999	* 1) did not queue everything initially scheduled to queue
2000	* 2) the last attempt to queue a request failed
2001	*/
2002	static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued,
2003	bool from_schedule)
2004	{
2005	if (hctx->queue->mq_ops->commit_rqs && queued) {
2006	trace_block_unplug(q: hctx->queue, depth: queued, explicit: !from_schedule);
2007	hctx->queue->mq_ops->commit_rqs(hctx);
2008	}
2009	}
2010
2011	/*
2012	* Returns true if we did some work AND can potentially do more.
2013	*/
2014	bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
2015	unsigned int nr_budgets)
2016	{
2017	enum prep_dispatch prep;
2018	struct request_queue *q = hctx->queue;
2019	struct request *rq;
2020	int queued;
2021	blk_status_t ret = BLK_STS_OK;
2022	LIST_HEAD(zone_list);
2023	bool needs_resource = false;
2024
2025	if (list_empty(head: list))
2026	return false;
2027
2028	/*
2029	* Now process all the entries, sending them to the driver.
2030	*/
2031	queued = 0;
2032	do {
2033	struct blk_mq_queue_data bd;
2034
2035	rq = list_first_entry(list, struct request, queuelist);
2036
2037	WARN_ON_ONCE(hctx != rq->mq_hctx);
2038	prep = blk_mq_prep_dispatch_rq(rq, need_budget: !nr_budgets);
2039	if (prep != PREP_DISPATCH_OK)
2040	break;
2041
2042	list_del_init(entry: &rq->queuelist);
2043
2044	bd.rq = rq;
2045	bd.last = list_empty(head: list);
2046
2047	/*
2048	* once the request is queued to lld, no need to cover the
2049	* budget any more
2050	*/
2051	if (nr_budgets)
2052	nr_budgets--;
2053	ret = q->mq_ops->queue_rq(hctx, &bd);
2054	switch (ret) {
2055	case BLK_STS_OK:
2056	queued++;
2057	break;
2058	case BLK_STS_RESOURCE:
2059	needs_resource = true;
2060	fallthrough;
2061	case BLK_STS_DEV_RESOURCE:
2062	blk_mq_handle_dev_resource(rq, list);
2063	goto out;
2064	case BLK_STS_ZONE_RESOURCE:
2065	/*
2066	* Move the request to zone_list and keep going through
2067	* the dispatch list to find more requests the drive can
2068	* accept.
2069	*/
2070	blk_mq_handle_zone_resource(rq, zone_list: &zone_list);
2071	needs_resource = true;
2072	break;
2073	default:
2074	blk_mq_end_request(rq, ret);
2075	}
2076	} while (!list_empty(head: list));
2077	out:
2078	if (!list_empty(head: &zone_list))
2079	list_splice_tail_init(list: &zone_list, head: list);
2080
2081	/* If we didn't flush the entire list, we could have told the driver
2082	* there was more coming, but that turned out to be a lie.
2083	*/
2084	if (!list_empty(head: list) || ret != BLK_STS_OK)
2085	blk_mq_commit_rqs(hctx, queued, from_schedule: false);
2086
2087	/*
2088	* Any items that need requeuing? Stuff them into hctx->dispatch,
2089	* that is where we will continue on next queue run.
2090	*/
2091	if (!list_empty(head: list)) {
2092	bool needs_restart;
2093	/* For non-shared tags, the RESTART check will suffice */
2094	bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2095	((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
2096	blk_mq_is_shared_tags(flags: hctx->flags));
2097
2098	if (nr_budgets)
2099	blk_mq_release_budgets(q, list);
2100
2101	spin_lock(lock: &hctx->lock);
2102	list_splice_tail_init(list, head: &hctx->dispatch);
2103	spin_unlock(lock: &hctx->lock);
2104
2105	/*
2106	* Order adding requests to hctx->dispatch and checking
2107	* SCHED_RESTART flag. The pair of this smp_mb() is the one
2108	* in blk_mq_sched_restart(). Avoid restart code path to
2109	* miss the new added requests to hctx->dispatch, meantime
2110	* SCHED_RESTART is observed here.
2111	*/
2112	smp_mb();
2113
2114	/*
2115	* If SCHED_RESTART was set by the caller of this function and
2116	* it is no longer set that means that it was cleared by another
2117	* thread and hence that a queue rerun is needed.
2118	*
2119	* If 'no_tag' is set, that means that we failed getting
2120	* a driver tag with an I/O scheduler attached. If our dispatch
2121	* waitqueue is no longer active, ensure that we run the queue
2122	* AFTER adding our entries back to the list.
2123	*
2124	* If no I/O scheduler has been configured it is possible that
2125	* the hardware queue got stopped and restarted before requests
2126	* were pushed back onto the dispatch list. Rerun the queue to
2127	* avoid starvation. Notes:
2128	* - blk_mq_run_hw_queue() checks whether or not a queue has
2129	* been stopped before rerunning a queue.
2130	* - Some but not all block drivers stop a queue before
2131	* returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2132	* and dm-rq.
2133	*
2134	* If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2135	* bit is set, run queue after a delay to avoid IO stalls
2136	* that could otherwise occur if the queue is idle. We'll do
2137	* similar if we couldn't get budget or couldn't lock a zone
2138	* and SCHED_RESTART is set.
2139	*/
2140	needs_restart = blk_mq_sched_needs_restart(hctx);
2141	if (prep == PREP_DISPATCH_NO_BUDGET)
2142	needs_resource = true;
2143	if (!needs_restart ||
2144	(no_tag && list_empty_careful(head: &hctx->dispatch_wait.entry)))
2145	blk_mq_run_hw_queue(hctx, async: true);
2146	else if (needs_resource)
2147	blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2148
2149	blk_mq_update_dispatch_busy(hctx, busy: true);
2150	return false;
2151	}
2152
2153	blk_mq_update_dispatch_busy(hctx, busy: false);
2154	return true;
2155	}
2156
2157	static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2158	{
2159	int cpu = cpumask_first_and(srcp1: hctx->cpumask, cpu_online_mask);
2160
2161	if (cpu >= nr_cpu_ids)
2162	cpu = cpumask_first(srcp: hctx->cpumask);
2163	return cpu;
2164	}
2165
2166	/*
2167	* It'd be great if the workqueue API had a way to pass
2168	* in a mask and had some smarts for more clever placement.
2169	* For now we just round-robin here, switching for every
2170	* BLK_MQ_CPU_WORK_BATCH queued items.
2171	*/
2172	static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2173	{
2174	bool tried = false;
2175	int next_cpu = hctx->next_cpu;
2176
2177	if (hctx->queue->nr_hw_queues == 1)
2178	return WORK_CPU_UNBOUND;
2179
2180	if (--hctx->next_cpu_batch <= 0) {
2181	select_cpu:
2182	next_cpu = cpumask_next_and(n: next_cpu, src1p: hctx->cpumask,
2183	cpu_online_mask);
2184	if (next_cpu >= nr_cpu_ids)
2185	next_cpu = blk_mq_first_mapped_cpu(hctx);
2186	hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2187	}
2188
2189	/*
2190	* Do unbound schedule if we can't find a online CPU for this hctx,
2191	* and it should only happen in the path of handling CPU DEAD.
2192	*/
2193	if (!cpu_online(cpu: next_cpu)) {
2194	if (!tried) {
2195	tried = true;
2196	goto select_cpu;
2197	}
2198
2199	/*
2200	* Make sure to re-select CPU next time once after CPUs
2201	* in hctx->cpumask become online again.
2202	*/
2203	hctx->next_cpu = next_cpu;
2204	hctx->next_cpu_batch = 1;
2205	return WORK_CPU_UNBOUND;
2206	}
2207
2208	hctx->next_cpu = next_cpu;
2209	return next_cpu;
2210	}
2211
2212	/**
2213	* blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2214	* @hctx: Pointer to the hardware queue to run.
2215	* @msecs: Milliseconds of delay to wait before running the queue.
2216	*
2217	* Run a hardware queue asynchronously with a delay of @msecs.
2218	*/
2219	void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2220	{
2221	if (unlikely(blk_mq_hctx_stopped(hctx)))
2222	return;
2223	kblockd_mod_delayed_work_on(cpu: blk_mq_hctx_next_cpu(hctx), dwork: &hctx->run_work,
2224	delay: msecs_to_jiffies(m: msecs));
2225	}
2226	EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2227
2228	/**
2229	* blk_mq_run_hw_queue - Start to run a hardware queue.
2230	* @hctx: Pointer to the hardware queue to run.
2231	* @async: If we want to run the queue asynchronously.
2232	*
2233	* Check if the request queue is not in a quiesced state and if there are
2234	* pending requests to be sent. If this is true, run the queue to send requests
2235	* to hardware.
2236	*/
2237	void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2238	{
2239	bool need_run;
2240
2241	/*
2242	* We can't run the queue inline with interrupts disabled.
2243	*/
2244	WARN_ON_ONCE(!async && in_interrupt());
2245
2246	might_sleep_if(!async && hctx->flags & BLK_MQ_F_BLOCKING);
2247
2248	/*
2249	* When queue is quiesced, we may be switching io scheduler, or
2250	* updating nr_hw_queues, or other things, and we can't run queue
2251	* any more, even __blk_mq_hctx_has_pending() can't be called safely.
2252	*
2253	* And queue will be rerun in blk_mq_unquiesce_queue() if it is
2254	* quiesced.
2255	*/
2256	__blk_mq_run_dispatch_ops(hctx->queue, false,
2257	need_run = !blk_queue_quiesced(hctx->queue) &&
2258	blk_mq_hctx_has_pending(hctx));
2259
2260	if (!need_run)
2261	return;
2262
2263	if (async || !cpumask_test_cpu(raw_smp_processor_id(), cpumask: hctx->cpumask)) {
2264	blk_mq_delay_run_hw_queue(hctx, 0);
2265	return;
2266	}
2267
2268	blk_mq_run_dispatch_ops(hctx->queue,
2269	blk_mq_sched_dispatch_requests(hctx));
2270	}
2271	EXPORT_SYMBOL(blk_mq_run_hw_queue);
2272
2273	/*
2274	* Return prefered queue to dispatch from (if any) for non-mq aware IO
2275	* scheduler.
2276	*/
2277	static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2278	{
2279	struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2280	/*
2281	* If the IO scheduler does not respect hardware queues when
2282	* dispatching, we just don't bother with multiple HW queues and
2283	* dispatch from hctx for the current CPU since running multiple queues
2284	* just causes lock contention inside the scheduler and pointless cache
2285	* bouncing.
2286	*/
2287	struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2288
2289	if (!blk_mq_hctx_stopped(hctx))
2290	return hctx;
2291	return NULL;
2292	}
2293
2294	/**
2295	* blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2296	* @q: Pointer to the request queue to run.
2297	* @async: If we want to run the queue asynchronously.
2298	*/
2299	void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2300	{
2301	struct blk_mq_hw_ctx *hctx, *sq_hctx;
2302	unsigned long i;
2303
2304	sq_hctx = NULL;
2305	if (blk_queue_sq_sched(q))
2306	sq_hctx = blk_mq_get_sq_hctx(q);
2307	queue_for_each_hw_ctx(q, hctx, i) {
2308	if (blk_mq_hctx_stopped(hctx))
2309	continue;
2310	/*
2311	* Dispatch from this hctx either if there's no hctx preferred
2312	* by IO scheduler or if it has requests that bypass the
2313	* scheduler.
2314	*/
2315	if (!sq_hctx || sq_hctx == hctx ||
2316	!list_empty_careful(head: &hctx->dispatch))
2317	blk_mq_run_hw_queue(hctx, async);
2318	}
2319	}
2320	EXPORT_SYMBOL(blk_mq_run_hw_queues);
2321
2322	/**
2323	* blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2324	* @q: Pointer to the request queue to run.
2325	* @msecs: Milliseconds of delay to wait before running the queues.
2326	*/
2327	void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2328	{
2329	struct blk_mq_hw_ctx *hctx, *sq_hctx;
2330	unsigned long i;
2331
2332	sq_hctx = NULL;
2333	if (blk_queue_sq_sched(q))
2334	sq_hctx = blk_mq_get_sq_hctx(q);
2335	queue_for_each_hw_ctx(q, hctx, i) {
2336	if (blk_mq_hctx_stopped(hctx))
2337	continue;
2338	/*
2339	* If there is already a run_work pending, leave the
2340	* pending delay untouched. Otherwise, a hctx can stall
2341	* if another hctx is re-delaying the other's work
2342	* before the work executes.
2343	*/
2344	if (delayed_work_pending(&hctx->run_work))
2345	continue;
2346	/*
2347	* Dispatch from this hctx either if there's no hctx preferred
2348	* by IO scheduler or if it has requests that bypass the
2349	* scheduler.
2350	*/
2351	if (!sq_hctx || sq_hctx == hctx ||
2352	!list_empty_careful(head: &hctx->dispatch))
2353	blk_mq_delay_run_hw_queue(hctx, msecs);
2354	}
2355	}
2356	EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2357
2358	/*
2359	* This function is often used for pausing .queue_rq() by driver when
2360	* there isn't enough resource or some conditions aren't satisfied, and
2361	* BLK_STS_RESOURCE is usually returned.
2362	*
2363	* We do not guarantee that dispatch can be drained or blocked
2364	* after blk_mq_stop_hw_queue() returns. Please use
2365	* blk_mq_quiesce_queue() for that requirement.
2366	*/
2367	void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2368	{
2369	cancel_delayed_work(dwork: &hctx->run_work);
2370
2371	set_bit(nr: BLK_MQ_S_STOPPED, addr: &hctx->state);
2372	}
2373	EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2374
2375	/*
2376	* This function is often used for pausing .queue_rq() by driver when
2377	* there isn't enough resource or some conditions aren't satisfied, and
2378	* BLK_STS_RESOURCE is usually returned.
2379	*
2380	* We do not guarantee that dispatch can be drained or blocked
2381	* after blk_mq_stop_hw_queues() returns. Please use
2382	* blk_mq_quiesce_queue() for that requirement.
2383	*/
2384	void blk_mq_stop_hw_queues(struct request_queue *q)
2385	{
2386	struct blk_mq_hw_ctx *hctx;
2387	unsigned long i;
2388
2389	queue_for_each_hw_ctx(q, hctx, i)
2390	blk_mq_stop_hw_queue(hctx);
2391	}
2392	EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2393
2394	void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2395	{
2396	clear_bit(nr: BLK_MQ_S_STOPPED, addr: &hctx->state);
2397
2398	blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);
2399	}
2400	EXPORT_SYMBOL(blk_mq_start_hw_queue);
2401
2402	void blk_mq_start_hw_queues(struct request_queue *q)
2403	{
2404	struct blk_mq_hw_ctx *hctx;
2405	unsigned long i;
2406
2407	queue_for_each_hw_ctx(q, hctx, i)
2408	blk_mq_start_hw_queue(hctx);
2409	}
2410	EXPORT_SYMBOL(blk_mq_start_hw_queues);
2411
2412	void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2413	{
2414	if (!blk_mq_hctx_stopped(hctx))
2415	return;
2416
2417	clear_bit(nr: BLK_MQ_S_STOPPED, addr: &hctx->state);
2418	blk_mq_run_hw_queue(hctx, async);
2419	}
2420	EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2421
2422	void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2423	{
2424	struct blk_mq_hw_ctx *hctx;
2425	unsigned long i;
2426
2427	queue_for_each_hw_ctx(q, hctx, i)
2428	blk_mq_start_stopped_hw_queue(hctx, async ||
2429	(hctx->flags & BLK_MQ_F_BLOCKING));
2430	}
2431	EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2432
2433	static void blk_mq_run_work_fn(struct work_struct *work)
2434	{
2435	struct blk_mq_hw_ctx *hctx =
2436	container_of(work, struct blk_mq_hw_ctx, run_work.work);
2437
2438	blk_mq_run_dispatch_ops(hctx->queue,
2439	blk_mq_sched_dispatch_requests(hctx));
2440	}
2441
2442	/**
2443	* blk_mq_request_bypass_insert - Insert a request at dispatch list.
2444	* @rq: Pointer to request to be inserted.
2445	* @flags: BLK_MQ_INSERT_*
2446	*
2447	* Should only be used carefully, when the caller knows we want to
2448	* bypass a potential IO scheduler on the target device.
2449	*/
2450	static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags)
2451	{
2452	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2453
2454	spin_lock(lock: &hctx->lock);
2455	if (flags & BLK_MQ_INSERT_AT_HEAD)
2456	list_add(new: &rq->queuelist, head: &hctx->dispatch);
2457	else
2458	list_add_tail(new: &rq->queuelist, head: &hctx->dispatch);
2459	spin_unlock(lock: &hctx->lock);
2460	}
2461
2462	static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx,
2463	struct blk_mq_ctx *ctx, struct list_head *list,
2464	bool run_queue_async)
2465	{
2466	struct request *rq;
2467	enum hctx_type type = hctx->type;
2468
2469	/*
2470	* Try to issue requests directly if the hw queue isn't busy to save an
2471	* extra enqueue & dequeue to the sw queue.
2472	*/
2473	if (!hctx->dispatch_busy && !run_queue_async) {
2474	blk_mq_run_dispatch_ops(hctx->queue,
2475	blk_mq_try_issue_list_directly(hctx, list));
2476	if (list_empty(head: list))
2477	goto out;
2478	}
2479
2480	/*
2481	* preemption doesn't flush plug list, so it's possible ctx->cpu is
2482	* offline now
2483	*/
2484	list_for_each_entry(rq, list, queuelist) {
2485	BUG_ON(rq->mq_ctx != ctx);
2486	trace_block_rq_insert(rq);
2487	if (rq->cmd_flags & REQ_NOWAIT)
2488	run_queue_async = true;
2489	}
2490
2491	spin_lock(lock: &ctx->lock);
2492	list_splice_tail_init(list, head: &ctx->rq_lists[type]);
2493	blk_mq_hctx_mark_pending(hctx, ctx);
2494	spin_unlock(lock: &ctx->lock);
2495	out:
2496	blk_mq_run_hw_queue(hctx, run_queue_async);
2497	}
2498
2499	static void blk_mq_insert_request(struct request *rq, blk_insert_t flags)
2500	{
2501	struct request_queue *q = rq->q;
2502	struct blk_mq_ctx *ctx = rq->mq_ctx;
2503	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2504
2505	if (blk_rq_is_passthrough(rq)) {
2506	/*
2507	* Passthrough request have to be added to hctx->dispatch
2508	* directly. The device may be in a situation where it can't
2509	* handle FS request, and always returns BLK_STS_RESOURCE for
2510	* them, which gets them added to hctx->dispatch.
2511	*
2512	* If a passthrough request is required to unblock the queues,
2513	* and it is added to the scheduler queue, there is no chance to
2514	* dispatch it given we prioritize requests in hctx->dispatch.
2515	*/
2516	blk_mq_request_bypass_insert(rq, flags);
2517	} else if (req_op(req: rq) == REQ_OP_FLUSH) {
2518	/*
2519	* Firstly normal IO request is inserted to scheduler queue or
2520	* sw queue, meantime we add flush request to dispatch queue(
2521	* hctx->dispatch) directly and there is at most one in-flight
2522	* flush request for each hw queue, so it doesn't matter to add
2523	* flush request to tail or front of the dispatch queue.
2524	*
2525	* Secondly in case of NCQ, flush request belongs to non-NCQ
2526	* command, and queueing it will fail when there is any
2527	* in-flight normal IO request(NCQ command). When adding flush
2528	* rq to the front of hctx->dispatch, it is easier to introduce
2529	* extra time to flush rq's latency because of S_SCHED_RESTART
2530	* compared with adding to the tail of dispatch queue, then
2531	* chance of flush merge is increased, and less flush requests
2532	* will be issued to controller. It is observed that ~10% time
2533	* is saved in blktests block/004 on disk attached to AHCI/NCQ
2534	* drive when adding flush rq to the front of hctx->dispatch.
2535	*
2536	* Simply queue flush rq to the front of hctx->dispatch so that
2537	* intensive flush workloads can benefit in case of NCQ HW.
2538	*/
2539	blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD);
2540	} else if (q->elevator) {
2541	LIST_HEAD(list);
2542
2543	WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG);
2544
2545	list_add(new: &rq->queuelist, head: &list);
2546	q->elevator->type->ops.insert_requests(hctx, &list, flags);
2547	} else {
2548	trace_block_rq_insert(rq);
2549
2550	spin_lock(lock: &ctx->lock);
2551	if (flags & BLK_MQ_INSERT_AT_HEAD)
2552	list_add(new: &rq->queuelist, head: &ctx->rq_lists[hctx->type]);
2553	else
2554	list_add_tail(new: &rq->queuelist,
2555	head: &ctx->rq_lists[hctx->type]);
2556	blk_mq_hctx_mark_pending(hctx, ctx);
2557	spin_unlock(lock: &ctx->lock);
2558	}
2559	}
2560
2561	static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2562	unsigned int nr_segs)
2563	{
2564	int err;
2565
2566	if (bio->bi_opf & REQ_RAHEAD)
2567	rq->cmd_flags |= REQ_FAILFAST_MASK;
2568
2569	rq->__sector = bio->bi_iter.bi_sector;
2570	rq->write_hint = bio->bi_write_hint;
2571	blk_rq_bio_prep(rq, bio, nr_segs);
2572
2573	/* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2574	err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2575	WARN_ON_ONCE(err);
2576
2577	blk_account_io_start(req: rq);
2578	}
2579
2580	static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2581	struct request *rq, bool last)
2582	{
2583	struct request_queue *q = rq->q;
2584	struct blk_mq_queue_data bd = {
2585	.rq = rq,
2586	.last = last,
2587	};
2588	blk_status_t ret;
2589
2590	/*
2591	* For OK queue, we are done. For error, caller may kill it.
2592	* Any other error (busy), just add it to our list as we
2593	* previously would have done.
2594	*/
2595	ret = q->mq_ops->queue_rq(hctx, &bd);
2596	switch (ret) {
2597	case BLK_STS_OK:
2598	blk_mq_update_dispatch_busy(hctx, busy: false);
2599	break;
2600	case BLK_STS_RESOURCE:
2601	case BLK_STS_DEV_RESOURCE:
2602	blk_mq_update_dispatch_busy(hctx, busy: true);
2603	__blk_mq_requeue_request(rq);
2604	break;
2605	default:
2606	blk_mq_update_dispatch_busy(hctx, busy: false);
2607	break;
2608	}
2609
2610	return ret;
2611	}
2612
2613	static bool blk_mq_get_budget_and_tag(struct request *rq)
2614	{
2615	int budget_token;
2616
2617	budget_token = blk_mq_get_dispatch_budget(q: rq->q);
2618	if (budget_token < 0)
2619	return false;
2620	blk_mq_set_rq_budget_token(rq, token: budget_token);
2621	if (!blk_mq_get_driver_tag(rq)) {
2622	blk_mq_put_dispatch_budget(q: rq->q, budget_token);
2623	return false;
2624	}
2625	return true;
2626	}
2627
2628	/**
2629	* blk_mq_try_issue_directly - Try to send a request directly to device driver.
2630	* @hctx: Pointer of the associated hardware queue.
2631	* @rq: Pointer to request to be sent.
2632	*
2633	* If the device has enough resources to accept a new request now, send the
2634	* request directly to device driver. Else, insert at hctx->dispatch queue, so
2635	* we can try send it another time in the future. Requests inserted at this
2636	* queue have higher priority.
2637	*/
2638	static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2639	struct request *rq)
2640	{
2641	blk_status_t ret;
2642
2643	if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2644	blk_mq_insert_request(rq, flags: 0);
2645	return;
2646	}
2647
2648	if ((rq->rq_flags & RQF_USE_SCHED) || !blk_mq_get_budget_and_tag(rq)) {
2649	blk_mq_insert_request(rq, flags: 0);
2650	blk_mq_run_hw_queue(hctx, rq->cmd_flags & REQ_NOWAIT);
2651	return;
2652	}
2653
2654	ret = __blk_mq_issue_directly(hctx, rq, last: true);
2655	switch (ret) {
2656	case BLK_STS_OK:
2657	break;
2658	case BLK_STS_RESOURCE:
2659	case BLK_STS_DEV_RESOURCE:
2660	blk_mq_request_bypass_insert(rq, flags: 0);
2661	blk_mq_run_hw_queue(hctx, false);
2662	break;
2663	default:
2664	blk_mq_end_request(rq, ret);
2665	break;
2666	}
2667	}
2668
2669	static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2670	{
2671	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2672
2673	if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2674	blk_mq_insert_request(rq, flags: 0);
2675	return BLK_STS_OK;
2676	}
2677
2678	if (!blk_mq_get_budget_and_tag(rq))
2679	return BLK_STS_RESOURCE;
2680	return __blk_mq_issue_directly(hctx, rq, last);
2681	}
2682
2683	static void blk_mq_plug_issue_direct(struct blk_plug *plug)
2684	{
2685	struct blk_mq_hw_ctx *hctx = NULL;
2686	struct request *rq;
2687	int queued = 0;
2688	blk_status_t ret = BLK_STS_OK;
2689
2690	while ((rq = rq_list_pop(&plug->mq_list))) {
2691	bool last = rq_list_empty(plug->mq_list);
2692
2693	if (hctx != rq->mq_hctx) {
2694	if (hctx) {
2695	blk_mq_commit_rqs(hctx, queued, from_schedule: false);
2696	queued = 0;
2697	}
2698	hctx = rq->mq_hctx;
2699	}
2700
2701	ret = blk_mq_request_issue_directly(rq, last);
2702	switch (ret) {
2703	case BLK_STS_OK:
2704	queued++;
2705	break;
2706	case BLK_STS_RESOURCE:
2707	case BLK_STS_DEV_RESOURCE:
2708	blk_mq_request_bypass_insert(rq, flags: 0);
2709	blk_mq_run_hw_queue(hctx, false);
2710	goto out;
2711	default:
2712	blk_mq_end_request(rq, ret);
2713	break;
2714	}
2715	}
2716
2717	out:
2718	if (ret != BLK_STS_OK)
2719	blk_mq_commit_rqs(hctx, queued, from_schedule: false);
2720	}
2721
2722	static void __blk_mq_flush_plug_list(struct request_queue *q,
2723	struct blk_plug *plug)
2724	{
2725	if (blk_queue_quiesced(q))
2726	return;
2727	q->mq_ops->queue_rqs(&plug->mq_list);
2728	}
2729
2730	static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2731	{
2732	struct blk_mq_hw_ctx *this_hctx = NULL;
2733	struct blk_mq_ctx *this_ctx = NULL;
2734	struct request *requeue_list = NULL;
2735	struct request **requeue_lastp = &requeue_list;
2736	unsigned int depth = 0;
2737	bool is_passthrough = false;
2738	LIST_HEAD(list);
2739
2740	do {
2741	struct request *rq = rq_list_pop(&plug->mq_list);
2742
2743	if (!this_hctx) {
2744	this_hctx = rq->mq_hctx;
2745	this_ctx = rq->mq_ctx;
2746	is_passthrough = blk_rq_is_passthrough(rq);
2747	} else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx ||
2748	is_passthrough != blk_rq_is_passthrough(rq)) {
2749	rq_list_add_tail(&requeue_lastp, rq);
2750	continue;
2751	}
2752	list_add(new: &rq->queuelist, head: &list);
2753	depth++;
2754	} while (!rq_list_empty(plug->mq_list));
2755
2756	plug->mq_list = requeue_list;
2757	trace_block_unplug(q: this_hctx->queue, depth, explicit: !from_sched);
2758
2759	percpu_ref_get(ref: &this_hctx->queue->q_usage_counter);
2760	/* passthrough requests should never be issued to the I/O scheduler */
2761	if (is_passthrough) {
2762	spin_lock(lock: &this_hctx->lock);
2763	list_splice_tail_init(list: &list, head: &this_hctx->dispatch);
2764	spin_unlock(lock: &this_hctx->lock);
2765	blk_mq_run_hw_queue(this_hctx, from_sched);
2766	} else if (this_hctx->queue->elevator) {
2767	this_hctx->queue->elevator->type->ops.insert_requests(this_hctx,
2768	&list, 0);
2769	blk_mq_run_hw_queue(this_hctx, from_sched);
2770	} else {
2771	blk_mq_insert_requests(hctx: this_hctx, ctx: this_ctx, list: &list, run_queue_async: from_sched);
2772	}
2773	percpu_ref_put(ref: &this_hctx->queue->q_usage_counter);
2774	}
2775
2776	void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2777	{
2778	struct request *rq;
2779
2780	/*
2781	* We may have been called recursively midway through handling
2782	* plug->mq_list via a schedule() in the driver's queue_rq() callback.
2783	* To avoid mq_list changing under our feet, clear rq_count early and
2784	* bail out specifically if rq_count is 0 rather than checking
2785	* whether the mq_list is empty.
2786	*/
2787	if (plug->rq_count == 0)
2788	return;
2789	plug->rq_count = 0;
2790
2791	if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2792	struct request_queue *q;
2793
2794	rq = rq_list_peek(&plug->mq_list);
2795	q = rq->q;
2796
2797	/*
2798	* Peek first request and see if we have a ->queue_rqs() hook.
2799	* If we do, we can dispatch the whole plug list in one go. We
2800	* already know at this point that all requests belong to the
2801	* same queue, caller must ensure that's the case.
2802	*/
2803	if (q->mq_ops->queue_rqs) {
2804	blk_mq_run_dispatch_ops(q,
2805	__blk_mq_flush_plug_list(q, plug));
2806	if (rq_list_empty(plug->mq_list))
2807	return;
2808	}
2809
2810	blk_mq_run_dispatch_ops(q,
2811	blk_mq_plug_issue_direct(plug));
2812	if (rq_list_empty(plug->mq_list))
2813	return;
2814	}
2815
2816	do {
2817	blk_mq_dispatch_plug_list(plug, from_sched: from_schedule);
2818	} while (!rq_list_empty(plug->mq_list));
2819	}
2820
2821	static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2822	struct list_head *list)
2823	{
2824	int queued = 0;
2825	blk_status_t ret = BLK_STS_OK;
2826
2827	while (!list_empty(head: list)) {
2828	struct request *rq = list_first_entry(list, struct request,
2829	queuelist);
2830
2831	list_del_init(entry: &rq->queuelist);
2832	ret = blk_mq_request_issue_directly(rq, last: list_empty(head: list));
2833	switch (ret) {
2834	case BLK_STS_OK:
2835	queued++;
2836	break;
2837	case BLK_STS_RESOURCE:
2838	case BLK_STS_DEV_RESOURCE:
2839	blk_mq_request_bypass_insert(rq, flags: 0);
2840	if (list_empty(head: list))
2841	blk_mq_run_hw_queue(hctx, false);
2842	goto out;
2843	default:
2844	blk_mq_end_request(rq, ret);
2845	break;
2846	}
2847	}
2848
2849	out:
2850	if (ret != BLK_STS_OK)
2851	blk_mq_commit_rqs(hctx, queued, from_schedule: false);
2852	}
2853
2854	static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2855	struct bio *bio, unsigned int nr_segs)
2856	{
2857	if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2858	if (blk_attempt_plug_merge(q, bio, nr_segs))
2859	return true;
2860	if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2861	return true;
2862	}
2863	return false;
2864	}
2865
2866	static struct request *blk_mq_get_new_requests(struct request_queue *q,
2867	struct blk_plug *plug,
2868	struct bio *bio,
2869	unsigned int nsegs)
2870	{
2871	struct blk_mq_alloc_data data = {
2872	.q = q,
2873	.nr_tags = 1,
2874	.cmd_flags = bio->bi_opf,
2875	};
2876	struct request *rq;
2877
2878	rq_qos_throttle(q, bio);
2879
2880	if (plug) {
2881	data.nr_tags = plug->nr_ios;
2882	plug->nr_ios = 1;
2883	data.cached_rq = &plug->cached_rq;
2884	}
2885
2886	rq = __blk_mq_alloc_requests(data: &data);
2887	if (rq)
2888	return rq;
2889	rq_qos_cleanup(q, bio);
2890	if (bio->bi_opf & REQ_NOWAIT)
2891	bio_wouldblock_error(bio);
2892	return NULL;
2893	}
2894
2895	/*
2896	* Check if there is a suitable cached request and return it.
2897	*/
2898	static struct request *blk_mq_peek_cached_request(struct blk_plug *plug,
2899	struct request_queue *q, blk_opf_t opf)
2900	{
2901	enum hctx_type type = blk_mq_get_hctx_type(opf);
2902	struct request *rq;
2903
2904	if (!plug)
2905	return NULL;
2906	rq = rq_list_peek(&plug->cached_rq);
2907	if (!rq || rq->q != q)
2908	return NULL;
2909	if (type != rq->mq_hctx->type &&
2910	(type != HCTX_TYPE_READ || rq->mq_hctx->type != HCTX_TYPE_DEFAULT))
2911	return NULL;
2912	if (op_is_flush(op: rq->cmd_flags) != op_is_flush(op: opf))
2913	return NULL;
2914	return rq;
2915	}
2916
2917	static void blk_mq_use_cached_rq(struct request *rq, struct blk_plug *plug,
2918	struct bio *bio)
2919	{
2920	WARN_ON_ONCE(rq_list_peek(&plug->cached_rq) != rq);
2921
2922	/*
2923	* If any qos ->throttle() end up blocking, we will have flushed the
2924	* plug and hence killed the cached_rq list as well. Pop this entry
2925	* before we throttle.
2926	*/
2927	plug->cached_rq = rq_list_next(rq);
2928	rq_qos_throttle(q: rq->q, bio);
2929
2930	blk_mq_rq_time_init(rq, alloc_time_ns: 0);
2931	rq->cmd_flags = bio->bi_opf;
2932	INIT_LIST_HEAD(list: &rq->queuelist);
2933	}
2934
2935	/**
2936	* blk_mq_submit_bio - Create and send a request to block device.
2937	* @bio: Bio pointer.
2938	*
2939	* Builds up a request structure from @q and @bio and send to the device. The
2940	* request may not be queued directly to hardware if:
2941	* * This request can be merged with another one
2942	* * We want to place request at plug queue for possible future merging
2943	* * There is an IO scheduler active at this queue
2944	*
2945	* It will not queue the request if there is an error with the bio, or at the
2946	* request creation.
2947	*/
2948	void blk_mq_submit_bio(struct bio *bio)
2949	{
2950	struct request_queue *q = bdev_get_queue(bdev: bio->bi_bdev);
2951	struct blk_plug *plug = blk_mq_plug(bio);
2952	const int is_sync = op_is_sync(op: bio->bi_opf);
2953	struct blk_mq_hw_ctx *hctx;
2954	unsigned int nr_segs = 1;
2955	struct request *rq;
2956	blk_status_t ret;
2957
2958	bio = blk_queue_bounce(bio, q);
2959
2960	/*
2961	* If the plug has a cached request for this queue, try use it.
2962	*
2963	* The cached request already holds a q_usage_counter reference and we
2964	* don't have to acquire a new one if we use it.
2965	*/
2966	rq = blk_mq_peek_cached_request(plug, q, opf: bio->bi_opf);
2967	if (!rq) {
2968	if (unlikely(bio_queue_enter(bio)))
2969	return;
2970	}
2971
2972	if (unlikely(bio_may_exceed_limits(bio, &q->limits))) {
2973	bio = __bio_split_to_limits(bio, lim: &q->limits, nr_segs: &nr_segs);
2974	if (!bio)
2975	goto queue_exit;
2976	}
2977	if (!bio_integrity_prep(bio))
2978	goto queue_exit;
2979
2980	if (blk_mq_attempt_bio_merge(q, bio, nr_segs))
2981	goto queue_exit;
2982
2983	if (!rq) {
2984	rq = blk_mq_get_new_requests(q, plug, bio, nsegs: nr_segs);
2985	if (unlikely(!rq))
2986	goto queue_exit;
2987	} else {
2988	blk_mq_use_cached_rq(rq, plug, bio);
2989	}
2990
2991	trace_block_getrq(bio);
2992
2993	rq_qos_track(q, rq, bio);
2994
2995	blk_mq_bio_to_request(rq, bio, nr_segs);
2996
2997	ret = blk_crypto_rq_get_keyslot(rq);
2998	if (ret != BLK_STS_OK) {
2999	bio->bi_status = ret;
3000	bio_endio(bio);
3001	blk_mq_free_request(rq);
3002	return;
3003	}
3004
3005	if (op_is_flush(op: bio->bi_opf) && blk_insert_flush(rq))
3006	return;
3007
3008	if (plug) {
3009	blk_add_rq_to_plug(plug, rq);
3010	return;
3011	}
3012
3013	hctx = rq->mq_hctx;
3014	if ((rq->rq_flags & RQF_USE_SCHED) ||
3015	(hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) {
3016	blk_mq_insert_request(rq, flags: 0);
3017	blk_mq_run_hw_queue(hctx, true);
3018	} else {
3019	blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq));
3020	}
3021	return;
3022
3023	queue_exit:
3024	/*
3025	* Don't drop the queue reference if we were trying to use a cached
3026	* request and thus didn't acquire one.
3027	*/
3028	if (!rq)
3029	blk_queue_exit(q);
3030	}
3031
3032	#ifdef CONFIG_BLK_MQ_STACKING
3033	/**
3034	* blk_insert_cloned_request - Helper for stacking drivers to submit a request
3035	* @rq: the request being queued
3036	*/
3037	blk_status_t blk_insert_cloned_request(struct request *rq)
3038	{
3039	struct request_queue *q = rq->q;
3040	unsigned int max_sectors = blk_queue_get_max_sectors(q, op: req_op(req: rq));
3041	unsigned int max_segments = blk_rq_get_max_segments(rq);
3042	blk_status_t ret;
3043
3044	if (blk_rq_sectors(rq) > max_sectors) {
3045	/*
3046	* SCSI device does not have a good way to return if
3047	* Write Same/Zero is actually supported. If a device rejects
3048	* a non-read/write command (discard, write same,etc.) the
3049	* low-level device driver will set the relevant queue limit to
3050	* 0 to prevent blk-lib from issuing more of the offending
3051	* operations. Commands queued prior to the queue limit being
3052	* reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3053	* errors being propagated to upper layers.
3054	*/
3055	if (max_sectors == 0)
3056	return BLK_STS_NOTSUPP;
3057
3058	printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3059	__func__, blk_rq_sectors(rq), max_sectors);
3060	return BLK_STS_IOERR;
3061	}
3062
3063	/*
3064	* The queue settings related to segment counting may differ from the
3065	* original queue.
3066	*/
3067	rq->nr_phys_segments = blk_recalc_rq_segments(rq);
3068	if (rq->nr_phys_segments > max_segments) {
3069	printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n",
3070	__func__, rq->nr_phys_segments, max_segments);
3071	return BLK_STS_IOERR;
3072	}
3073
3074	if (q->disk && should_fail_request(part: q->disk->part0, bytes: blk_rq_bytes(rq)))
3075	return BLK_STS_IOERR;
3076
3077	ret = blk_crypto_rq_get_keyslot(rq);
3078	if (ret != BLK_STS_OK)
3079	return ret;
3080
3081	blk_account_io_start(req: rq);
3082
3083	/*
3084	* Since we have a scheduler attached on the top device,
3085	* bypass a potential scheduler on the bottom device for
3086	* insert.
3087	*/
3088	blk_mq_run_dispatch_ops(q,
3089	ret = blk_mq_request_issue_directly(rq, true));
3090	if (ret)
3091	blk_account_io_done(req: rq, now: blk_time_get_ns());
3092	return ret;
3093	}
3094	EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3095
3096	/**
3097	* blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3098	* @rq: the clone request to be cleaned up
3099	*
3100	* Description:
3101	* Free all bios in @rq for a cloned request.
3102	*/
3103	void blk_rq_unprep_clone(struct request *rq)
3104	{
3105	struct bio *bio;
3106
3107	while ((bio = rq->bio) != NULL) {
3108	rq->bio = bio->bi_next;
3109
3110	bio_put(bio);
3111	}
3112	}
3113	EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3114
3115	/**
3116	* blk_rq_prep_clone - Helper function to setup clone request
3117	* @rq: the request to be setup
3118	* @rq_src: original request to be cloned
3119	* @bs: bio_set that bios for clone are allocated from
3120	* @gfp_mask: memory allocation mask for bio
3121	* @bio_ctr: setup function to be called for each clone bio.
3122	* Returns %0 for success, non %0 for failure.
3123	* @data: private data to be passed to @bio_ctr
3124	*
3125	* Description:
3126	* Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3127	* Also, pages which the original bios are pointing to are not copied
3128	* and the cloned bios just point same pages.
3129	* So cloned bios must be completed before original bios, which means
3130	* the caller must complete @rq before @rq_src.
3131	*/
3132	int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3133	struct bio_set *bs, gfp_t gfp_mask,
3134	int (*bio_ctr)(struct bio *, struct bio *, void *),
3135	void *data)
3136	{
3137	struct bio *bio, *bio_src;
3138
3139	if (!bs)
3140	bs = &fs_bio_set;
3141
3142	__rq_for_each_bio(bio_src, rq_src) {
3143	bio = bio_alloc_clone(bdev: rq->q->disk->part0, bio_src, gfp: gfp_mask,
3144	bs);
3145	if (!bio)
3146	goto free_and_out;
3147
3148	if (bio_ctr && bio_ctr(bio, bio_src, data))
3149	goto free_and_out;
3150
3151	if (rq->bio) {
3152	rq->biotail->bi_next = bio;
3153	rq->biotail = bio;
3154	} else {
3155	rq->bio = rq->biotail = bio;
3156	}
3157	bio = NULL;
3158	}
3159
3160	/* Copy attributes of the original request to the clone request. */
3161	rq->__sector = blk_rq_pos(rq: rq_src);
3162	rq->__data_len = blk_rq_bytes(rq: rq_src);
3163	if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3164	rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3165	rq->special_vec = rq_src->special_vec;
3166	}
3167	rq->nr_phys_segments = rq_src->nr_phys_segments;
3168	rq->ioprio = rq_src->ioprio;
3169	rq->write_hint = rq_src->write_hint;
3170
3171	if (rq->bio && blk_crypto_rq_bio_prep(rq, bio: rq->bio, gfp_mask) < 0)
3172	goto free_and_out;
3173
3174	return 0;
3175
3176	free_and_out:
3177	if (bio)
3178	bio_put(bio);
3179	blk_rq_unprep_clone(rq);
3180
3181	return -ENOMEM;
3182	}
3183	EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3184	#endif /* CONFIG_BLK_MQ_STACKING */
3185
3186	/*
3187	* Steal bios from a request and add them to a bio list.
3188	* The request must not have been partially completed before.
3189	*/
3190	void blk_steal_bios(struct bio_list *list, struct request *rq)
3191	{
3192	if (rq->bio) {
3193	if (list->tail)
3194	list->tail->bi_next = rq->bio;
3195	else
3196	list->head = rq->bio;
3197	list->tail = rq->biotail;
3198
3199	rq->bio = NULL;
3200	rq->biotail = NULL;
3201	}
3202
3203	rq->__data_len = 0;
3204	}
3205	EXPORT_SYMBOL_GPL(blk_steal_bios);
3206
3207	static size_t order_to_size(unsigned int order)
3208	{
3209	return (size_t)PAGE_SIZE << order;
3210	}
3211
3212	/* called before freeing request pool in @tags */
3213	static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3214	struct blk_mq_tags *tags)
3215	{
3216	struct page *page;
3217	unsigned long flags;
3218
3219	/*
3220	* There is no need to clear mapping if driver tags is not initialized
3221	* or the mapping belongs to the driver tags.
3222	*/
3223	if (!drv_tags || drv_tags == tags)
3224	return;
3225
3226	list_for_each_entry(page, &tags->page_list, lru) {
3227	unsigned long start = (unsigned long)page_address(page);
3228	unsigned long end = start + order_to_size(order: page->private);
3229	int i;
3230
3231	for (i = 0; i < drv_tags->nr_tags; i++) {
3232	struct request *rq = drv_tags->rqs[i];
3233	unsigned long rq_addr = (unsigned long)rq;
3234
3235	if (rq_addr >= start && rq_addr < end) {
3236	WARN_ON_ONCE(req_ref_read(rq) != 0);
3237	cmpxchg(&drv_tags->rqs[i], rq, NULL);
3238	}
3239	}
3240	}
3241
3242	/*
3243	* Wait until all pending iteration is done.
3244	*
3245	* Request reference is cleared and it is guaranteed to be observed
3246	* after the ->lock is released.
3247	*/
3248	spin_lock_irqsave(&drv_tags->lock, flags);
3249	spin_unlock_irqrestore(lock: &drv_tags->lock, flags);
3250	}
3251
3252	void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3253	unsigned int hctx_idx)
3254	{
3255	struct blk_mq_tags *drv_tags;
3256	struct page *page;
3257
3258	if (list_empty(head: &tags->page_list))
3259	return;
3260
3261	if (blk_mq_is_shared_tags(flags: set->flags))
3262	drv_tags = set->shared_tags;
3263	else
3264	drv_tags = set->tags[hctx_idx];
3265
3266	if (tags->static_rqs && set->ops->exit_request) {
3267	int i;
3268
3269	for (i = 0; i < tags->nr_tags; i++) {
3270	struct request *rq = tags->static_rqs[i];
3271
3272	if (!rq)
3273	continue;
3274	set->ops->exit_request(set, rq, hctx_idx);
3275	tags->static_rqs[i] = NULL;
3276	}
3277	}
3278
3279	blk_mq_clear_rq_mapping(drv_tags, tags);
3280
3281	while (!list_empty(head: &tags->page_list)) {
3282	page = list_first_entry(&tags->page_list, struct page, lru);
3283	list_del_init(entry: &page->lru);
3284	/*
3285	* Remove kmemleak object previously allocated in
3286	* blk_mq_alloc_rqs().
3287	*/
3288	kmemleak_free(page_address(page));
3289	__free_pages(page, order: page->private);
3290	}
3291	}
3292
3293	void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3294	{
3295	kfree(objp: tags->rqs);
3296	tags->rqs = NULL;
3297	kfree(objp: tags->static_rqs);
3298	tags->static_rqs = NULL;
3299
3300	blk_mq_free_tags(tags);
3301	}
3302
3303	static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3304	unsigned int hctx_idx)
3305	{
3306	int i;
3307
3308	for (i = 0; i < set->nr_maps; i++) {
3309	unsigned int start = set->map[i].queue_offset;
3310	unsigned int end = start + set->map[i].nr_queues;
3311
3312	if (hctx_idx >= start && hctx_idx < end)
3313	break;
3314	}
3315
3316	if (i >= set->nr_maps)
3317	i = HCTX_TYPE_DEFAULT;
3318
3319	return i;
3320	}
3321
3322	static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3323	unsigned int hctx_idx)
3324	{
3325	enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3326
3327	return blk_mq_hw_queue_to_node(qmap: &set->map[type], hctx_idx);
3328	}
3329
3330	static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3331	unsigned int hctx_idx,
3332	unsigned int nr_tags,
3333	unsigned int reserved_tags)
3334	{
3335	int node = blk_mq_get_hctx_node(set, hctx_idx);
3336	struct blk_mq_tags *tags;
3337
3338	if (node == NUMA_NO_NODE)
3339	node = set->numa_node;
3340
3341	tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3342	BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3343	if (!tags)
3344	return NULL;
3345
3346	tags->rqs = kcalloc_node(n: nr_tags, size: sizeof(struct request *),
3347	GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3348	node);
3349	if (!tags->rqs)
3350	goto err_free_tags;
3351
3352	tags->static_rqs = kcalloc_node(n: nr_tags, size: sizeof(struct request *),
3353	GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3354	node);
3355	if (!tags->static_rqs)
3356	goto err_free_rqs;
3357
3358	return tags;
3359
3360	err_free_rqs:
3361	kfree(objp: tags->rqs);
3362	err_free_tags:
3363	blk_mq_free_tags(tags);
3364	return NULL;
3365	}
3366
3367	static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3368	unsigned int hctx_idx, int node)
3369	{
3370	int ret;
3371
3372	if (set->ops->init_request) {
3373	ret = set->ops->init_request(set, rq, hctx_idx, node);
3374	if (ret)
3375	return ret;
3376	}
3377
3378	WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3379	return 0;
3380	}
3381
3382	static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3383	struct blk_mq_tags *tags,
3384	unsigned int hctx_idx, unsigned int depth)
3385	{
3386	unsigned int i, j, entries_per_page, max_order = 4;
3387	int node = blk_mq_get_hctx_node(set, hctx_idx);
3388	size_t rq_size, left;
3389
3390	if (node == NUMA_NO_NODE)
3391	node = set->numa_node;
3392
3393	INIT_LIST_HEAD(list: &tags->page_list);
3394
3395	/*
3396	* rq_size is the size of the request plus driver payload, rounded
3397	* to the cacheline size
3398	*/
3399	rq_size = round_up(sizeof(struct request) + set->cmd_size,
3400	cache_line_size());
3401	left = rq_size * depth;
3402
3403	for (i = 0; i < depth; ) {
3404	int this_order = max_order;
3405	struct page *page;
3406	int to_do;
3407	void *p;
3408
3409	while (this_order && left < order_to_size(order: this_order - 1))
3410	this_order--;
3411
3412	do {
3413	page = alloc_pages_node(nid: node,
3414	GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3415	order: this_order);
3416	if (page)
3417	break;
3418	if (!this_order--)
3419	break;
3420	if (order_to_size(order: this_order) < rq_size)
3421	break;
3422	} while (1);
3423
3424	if (!page)
3425	goto fail;
3426
3427	page->private = this_order;
3428	list_add_tail(new: &page->lru, head: &tags->page_list);
3429
3430	p = page_address(page);
3431	/*
3432	* Allow kmemleak to scan these pages as they contain pointers
3433	* to additional allocations like via ops->init_request().
3434	*/
3435	kmemleak_alloc(ptr: p, size: order_to_size(order: this_order), min_count: 1, GFP_NOIO);
3436	entries_per_page = order_to_size(order: this_order) / rq_size;
3437	to_do = min(entries_per_page, depth - i);
3438	left -= to_do * rq_size;
3439	for (j = 0; j < to_do; j++) {
3440	struct request *rq = p;
3441
3442	tags->static_rqs[i] = rq;
3443	if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3444	tags->static_rqs[i] = NULL;
3445	goto fail;
3446	}
3447
3448	p += rq_size;
3449	i++;
3450	}
3451	}
3452	return 0;
3453
3454	fail:
3455	blk_mq_free_rqs(set, tags, hctx_idx);
3456	return -ENOMEM;
3457	}
3458
3459	struct rq_iter_data {
3460	struct blk_mq_hw_ctx *hctx;
3461	bool has_rq;
3462	};
3463
3464	static bool blk_mq_has_request(struct request *rq, void *data)
3465	{
3466	struct rq_iter_data *iter_data = data;
3467
3468	if (rq->mq_hctx != iter_data->hctx)
3469	return true;
3470	iter_data->has_rq = true;
3471	return false;
3472	}
3473
3474	static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3475	{
3476	struct blk_mq_tags *tags = hctx->sched_tags ?
3477	hctx->sched_tags : hctx->tags;
3478	struct rq_iter_data data = {
3479	.hctx = hctx,
3480	};
3481
3482	blk_mq_all_tag_iter(tags, fn: blk_mq_has_request, priv: &data);
3483	return data.has_rq;
3484	}
3485
3486	static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3487	struct blk_mq_hw_ctx *hctx)
3488	{
3489	if (cpumask_first_and(srcp1: hctx->cpumask, cpu_online_mask) != cpu)
3490	return false;
3491	if (cpumask_next_and(n: cpu, src1p: hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3492	return false;
3493	return true;
3494	}
3495
3496	static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3497	{
3498	struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3499	struct blk_mq_hw_ctx, cpuhp_online);
3500
3501	if (!cpumask_test_cpu(cpu, cpumask: hctx->cpumask) ||
3502	!blk_mq_last_cpu_in_hctx(cpu, hctx))
3503	return 0;
3504
3505	/*
3506	* Prevent new request from being allocated on the current hctx.
3507	*
3508	* The smp_mb__after_atomic() Pairs with the implied barrier in
3509	* test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3510	* seen once we return from the tag allocator.
3511	*/
3512	set_bit(nr: BLK_MQ_S_INACTIVE, addr: &hctx->state);
3513	smp_mb__after_atomic();
3514
3515	/*
3516	* Try to grab a reference to the queue and wait for any outstanding
3517	* requests. If we could not grab a reference the queue has been
3518	* frozen and there are no requests.
3519	*/
3520	if (percpu_ref_tryget(ref: &hctx->queue->q_usage_counter)) {
3521	while (blk_mq_hctx_has_requests(hctx))
3522	msleep(msecs: 5);
3523	percpu_ref_put(ref: &hctx->queue->q_usage_counter);
3524	}
3525
3526	return 0;
3527	}
3528
3529	static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3530	{
3531	struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3532	struct blk_mq_hw_ctx, cpuhp_online);
3533
3534	if (cpumask_test_cpu(cpu, cpumask: hctx->cpumask))
3535	clear_bit(nr: BLK_MQ_S_INACTIVE, addr: &hctx->state);
3536	return 0;
3537	}
3538
3539	/*
3540	* 'cpu' is going away. splice any existing rq_list entries from this
3541	* software queue to the hw queue dispatch list, and ensure that it
3542	* gets run.
3543	*/
3544	static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3545	{
3546	struct blk_mq_hw_ctx *hctx;
3547	struct blk_mq_ctx *ctx;
3548	LIST_HEAD(tmp);
3549	enum hctx_type type;
3550
3551	hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3552	if (!cpumask_test_cpu(cpu, cpumask: hctx->cpumask))
3553	return 0;
3554
3555	ctx = __blk_mq_get_ctx(q: hctx->queue, cpu);
3556	type = hctx->type;
3557
3558	spin_lock(lock: &ctx->lock);
3559	if (!list_empty(head: &ctx->rq_lists[type])) {
3560	list_splice_init(list: &ctx->rq_lists[type], head: &tmp);
3561	blk_mq_hctx_clear_pending(hctx, ctx);
3562	}
3563	spin_unlock(lock: &ctx->lock);
3564
3565	if (list_empty(head: &tmp))
3566	return 0;
3567
3568	spin_lock(lock: &hctx->lock);
3569	list_splice_tail_init(list: &tmp, head: &hctx->dispatch);
3570	spin_unlock(lock: &hctx->lock);
3571
3572	blk_mq_run_hw_queue(hctx, true);
3573	return 0;
3574	}
3575
3576	static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3577	{
3578	if (!(hctx->flags & BLK_MQ_F_STACKING))
3579	cpuhp_state_remove_instance_nocalls(state: CPUHP_AP_BLK_MQ_ONLINE,
3580	node: &hctx->cpuhp_online);
3581	cpuhp_state_remove_instance_nocalls(state: CPUHP_BLK_MQ_DEAD,
3582	node: &hctx->cpuhp_dead);
3583	}
3584
3585	/*
3586	* Before freeing hw queue, clearing the flush request reference in
3587	* tags->rqs[] for avoiding potential UAF.
3588	*/
3589	static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3590	unsigned int queue_depth, struct request *flush_rq)
3591	{
3592	int i;
3593	unsigned long flags;
3594
3595	/* The hw queue may not be mapped yet */
3596	if (!tags)
3597	return;
3598
3599	WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3600
3601	for (i = 0; i < queue_depth; i++)
3602	cmpxchg(&tags->rqs[i], flush_rq, NULL);
3603
3604	/*
3605	* Wait until all pending iteration is done.
3606	*
3607	* Request reference is cleared and it is guaranteed to be observed
3608	* after the ->lock is released.
3609	*/
3610	spin_lock_irqsave(&tags->lock, flags);
3611	spin_unlock_irqrestore(lock: &tags->lock, flags);
3612	}
3613
3614	/* hctx->ctxs will be freed in queue's release handler */
3615	static void blk_mq_exit_hctx(struct request_queue *q,
3616	struct blk_mq_tag_set *set,
3617	struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3618	{
3619	struct request *flush_rq = hctx->fq->flush_rq;
3620
3621	if (blk_mq_hw_queue_mapped(hctx))
3622	blk_mq_tag_idle(hctx);
3623
3624	if (blk_queue_init_done(q))
3625	blk_mq_clear_flush_rq_mapping(tags: set->tags[hctx_idx],
3626	queue_depth: set->queue_depth, flush_rq);
3627	if (set->ops->exit_request)
3628	set->ops->exit_request(set, flush_rq, hctx_idx);
3629
3630	if (set->ops->exit_hctx)
3631	set->ops->exit_hctx(hctx, hctx_idx);
3632
3633	blk_mq_remove_cpuhp(hctx);
3634
3635	xa_erase(&q->hctx_table, index: hctx_idx);
3636
3637	spin_lock(lock: &q->unused_hctx_lock);
3638	list_add(new: &hctx->hctx_list, head: &q->unused_hctx_list);
3639	spin_unlock(lock: &q->unused_hctx_lock);
3640	}
3641
3642	static void blk_mq_exit_hw_queues(struct request_queue *q,
3643	struct blk_mq_tag_set *set, int nr_queue)
3644	{
3645	struct blk_mq_hw_ctx *hctx;
3646	unsigned long i;
3647
3648	queue_for_each_hw_ctx(q, hctx, i) {
3649	if (i == nr_queue)
3650	break;
3651	blk_mq_exit_hctx(q, set, hctx, hctx_idx: i);
3652	}
3653	}
3654
3655	static int blk_mq_init_hctx(struct request_queue *q,
3656	struct blk_mq_tag_set *set,
3657	struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3658	{
3659	hctx->queue_num = hctx_idx;
3660
3661	if (!(hctx->flags & BLK_MQ_F_STACKING))
3662	cpuhp_state_add_instance_nocalls(state: CPUHP_AP_BLK_MQ_ONLINE,
3663	node: &hctx->cpuhp_online);
3664	cpuhp_state_add_instance_nocalls(state: CPUHP_BLK_MQ_DEAD, node: &hctx->cpuhp_dead);
3665
3666	hctx->tags = set->tags[hctx_idx];
3667
3668	if (set->ops->init_hctx &&
3669	set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3670	goto unregister_cpu_notifier;
3671
3672	if (blk_mq_init_request(set, rq: hctx->fq->flush_rq, hctx_idx,
3673	node: hctx->numa_node))
3674	goto exit_hctx;
3675
3676	if (xa_insert(xa: &q->hctx_table, index: hctx_idx, entry: hctx, GFP_KERNEL))
3677	goto exit_flush_rq;
3678
3679	return 0;
3680
3681	exit_flush_rq:
3682	if (set->ops->exit_request)
3683	set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3684	exit_hctx:
3685	if (set->ops->exit_hctx)
3686	set->ops->exit_hctx(hctx, hctx_idx);
3687	unregister_cpu_notifier:
3688	blk_mq_remove_cpuhp(hctx);
3689	return -1;
3690	}
3691
3692	static struct blk_mq_hw_ctx *
3693	blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3694	int node)
3695	{
3696	struct blk_mq_hw_ctx *hctx;
3697	gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3698
3699	hctx = kzalloc_node(size: sizeof(struct blk_mq_hw_ctx), flags: gfp, node);
3700	if (!hctx)
3701	goto fail_alloc_hctx;
3702
3703	if (!zalloc_cpumask_var_node(mask: &hctx->cpumask, flags: gfp, node))
3704	goto free_hctx;
3705
3706	atomic_set(v: &hctx->nr_active, i: 0);
3707	if (node == NUMA_NO_NODE)
3708	node = set->numa_node;
3709	hctx->numa_node = node;
3710
3711	INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3712	spin_lock_init(&hctx->lock);
3713	INIT_LIST_HEAD(list: &hctx->dispatch);
3714	hctx->queue = q;
3715	hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3716
3717	INIT_LIST_HEAD(list: &hctx->hctx_list);
3718
3719	/*
3720	* Allocate space for all possible cpus to avoid allocation at
3721	* runtime
3722	*/
3723	hctx->ctxs = kmalloc_array_node(n: nr_cpu_ids, size: sizeof(void *),
3724	flags: gfp, node);
3725	if (!hctx->ctxs)
3726	goto free_cpumask;
3727
3728	if (sbitmap_init_node(sb: &hctx->ctx_map, depth: nr_cpu_ids, ilog2(8),
3729	flags: gfp, node, round_robin: false, alloc_hint: false))
3730	goto free_ctxs;
3731	hctx->nr_ctx = 0;
3732
3733	spin_lock_init(&hctx->dispatch_wait_lock);
3734	init_waitqueue_func_entry(wq_entry: &hctx->dispatch_wait, func: blk_mq_dispatch_wake);
3735	INIT_LIST_HEAD(list: &hctx->dispatch_wait.entry);
3736
3737	hctx->fq = blk_alloc_flush_queue(node: hctx->numa_node, cmd_size: set->cmd_size, flags: gfp);
3738	if (!hctx->fq)
3739	goto free_bitmap;
3740
3741	blk_mq_hctx_kobj_init(hctx);
3742
3743	return hctx;
3744
3745	free_bitmap:
3746	sbitmap_free(sb: &hctx->ctx_map);
3747	free_ctxs:
3748	kfree(objp: hctx->ctxs);
3749	free_cpumask:
3750	free_cpumask_var(mask: hctx->cpumask);
3751	free_hctx:
3752	kfree(objp: hctx);
3753	fail_alloc_hctx:
3754	return NULL;
3755	}
3756
3757	static void blk_mq_init_cpu_queues(struct request_queue *q,
3758	unsigned int nr_hw_queues)
3759	{
3760	struct blk_mq_tag_set *set = q->tag_set;
3761	unsigned int i, j;
3762
3763	for_each_possible_cpu(i) {
3764	struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3765	struct blk_mq_hw_ctx *hctx;
3766	int k;
3767
3768	__ctx->cpu = i;
3769	spin_lock_init(&__ctx->lock);
3770	for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3771	INIT_LIST_HEAD(list: &__ctx->rq_lists[k]);
3772
3773	__ctx->queue = q;
3774
3775	/*
3776	* Set local node, IFF we have more than one hw queue. If
3777	* not, we remain on the home node of the device
3778	*/
3779	for (j = 0; j < set->nr_maps; j++) {
3780	hctx = blk_mq_map_queue_type(q, type: j, cpu: i);
3781	if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3782	hctx->numa_node = cpu_to_node(cpu: i);
3783	}
3784	}
3785	}
3786
3787	struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3788	unsigned int hctx_idx,
3789	unsigned int depth)
3790	{
3791	struct blk_mq_tags *tags;
3792	int ret;
3793
3794	tags = blk_mq_alloc_rq_map(set, hctx_idx, nr_tags: depth, reserved_tags: set->reserved_tags);
3795	if (!tags)
3796	return NULL;
3797
3798	ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3799	if (ret) {
3800	blk_mq_free_rq_map(tags);
3801	return NULL;
3802	}
3803
3804	return tags;
3805	}
3806
3807	static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3808	int hctx_idx)
3809	{
3810	if (blk_mq_is_shared_tags(flags: set->flags)) {
3811	set->tags[hctx_idx] = set->shared_tags;
3812
3813	return true;
3814	}
3815
3816	set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3817	depth: set->queue_depth);
3818
3819	return set->tags[hctx_idx];
3820	}
3821
3822	void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3823	struct blk_mq_tags *tags,
3824	unsigned int hctx_idx)
3825	{
3826	if (tags) {
3827	blk_mq_free_rqs(set, tags, hctx_idx);
3828	blk_mq_free_rq_map(tags);
3829	}
3830	}
3831
3832	static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3833	unsigned int hctx_idx)
3834	{
3835	if (!blk_mq_is_shared_tags(flags: set->flags))
3836	blk_mq_free_map_and_rqs(set, tags: set->tags[hctx_idx], hctx_idx);
3837
3838	set->tags[hctx_idx] = NULL;
3839	}
3840
3841	static void blk_mq_map_swqueue(struct request_queue *q)
3842	{
3843	unsigned int j, hctx_idx;
3844	unsigned long i;
3845	struct blk_mq_hw_ctx *hctx;
3846	struct blk_mq_ctx *ctx;
3847	struct blk_mq_tag_set *set = q->tag_set;
3848
3849	queue_for_each_hw_ctx(q, hctx, i) {
3850	cpumask_clear(dstp: hctx->cpumask);
3851	hctx->nr_ctx = 0;
3852	hctx->dispatch_from = NULL;
3853	}
3854
3855	/*
3856	* Map software to hardware queues.
3857	*
3858	* If the cpu isn't present, the cpu is mapped to first hctx.
3859	*/
3860	for_each_possible_cpu(i) {
3861
3862	ctx = per_cpu_ptr(q->queue_ctx, i);
3863	for (j = 0; j < set->nr_maps; j++) {
3864	if (!set->map[j].nr_queues) {
3865	ctx->hctxs[j] = blk_mq_map_queue_type(q,
3866	type: HCTX_TYPE_DEFAULT, cpu: i);
3867	continue;
3868	}
3869	hctx_idx = set->map[j].mq_map[i];
3870	/* unmapped hw queue can be remapped after CPU topo changed */
3871	if (!set->tags[hctx_idx] &&
3872	!__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3873	/*
3874	* If tags initialization fail for some hctx,
3875	* that hctx won't be brought online. In this
3876	* case, remap the current ctx to hctx[0] which
3877	* is guaranteed to always have tags allocated
3878	*/
3879	set->map[j].mq_map[i] = 0;
3880	}
3881
3882	hctx = blk_mq_map_queue_type(q, type: j, cpu: i);
3883	ctx->hctxs[j] = hctx;
3884	/*
3885	* If the CPU is already set in the mask, then we've
3886	* mapped this one already. This can happen if
3887	* devices share queues across queue maps.
3888	*/
3889	if (cpumask_test_cpu(cpu: i, cpumask: hctx->cpumask))
3890	continue;
3891
3892	cpumask_set_cpu(cpu: i, dstp: hctx->cpumask);
3893	hctx->type = j;
3894	ctx->index_hw[hctx->type] = hctx->nr_ctx;
3895	hctx->ctxs[hctx->nr_ctx++] = ctx;
3896
3897	/*
3898	* If the nr_ctx type overflows, we have exceeded the
3899	* amount of sw queues we can support.
3900	*/
3901	BUG_ON(!hctx->nr_ctx);
3902	}
3903
3904	for (; j < HCTX_MAX_TYPES; j++)
3905	ctx->hctxs[j] = blk_mq_map_queue_type(q,
3906	type: HCTX_TYPE_DEFAULT, cpu: i);
3907	}
3908
3909	queue_for_each_hw_ctx(q, hctx, i) {
3910	/*
3911	* If no software queues are mapped to this hardware queue,
3912	* disable it and free the request entries.
3913	*/
3914	if (!hctx->nr_ctx) {
3915	/* Never unmap queue 0. We need it as a
3916	* fallback in case of a new remap fails
3917	* allocation
3918	*/
3919	if (i)
3920	__blk_mq_free_map_and_rqs(set, hctx_idx: i);
3921
3922	hctx->tags = NULL;
3923	continue;
3924	}
3925
3926	hctx->tags = set->tags[i];
3927	WARN_ON(!hctx->tags);
3928
3929	/*
3930	* Set the map size to the number of mapped software queues.
3931	* This is more accurate and more efficient than looping
3932	* over all possibly mapped software queues.
3933	*/
3934	sbitmap_resize(sb: &hctx->ctx_map, depth: hctx->nr_ctx);
3935
3936	/*
3937	* Initialize batch roundrobin counts
3938	*/
3939	hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3940	hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3941	}
3942	}
3943
3944	/*
3945	* Caller needs to ensure that we're either frozen/quiesced, or that
3946	* the queue isn't live yet.
3947	*/
3948	static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3949	{
3950	struct blk_mq_hw_ctx *hctx;
3951	unsigned long i;
3952
3953	queue_for_each_hw_ctx(q, hctx, i) {
3954	if (shared) {
3955	hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3956	} else {
3957	blk_mq_tag_idle(hctx);
3958	hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3959	}
3960	}
3961	}
3962
3963	static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3964	bool shared)
3965	{
3966	struct request_queue *q;
3967
3968	lockdep_assert_held(&set->tag_list_lock);
3969
3970	list_for_each_entry(q, &set->tag_list, tag_set_list) {
3971	blk_mq_freeze_queue(q);
3972	queue_set_hctx_shared(q, shared);
3973	blk_mq_unfreeze_queue(q);
3974	}
3975	}
3976
3977	static void blk_mq_del_queue_tag_set(struct request_queue *q)
3978	{
3979	struct blk_mq_tag_set *set = q->tag_set;
3980
3981	mutex_lock(&set->tag_list_lock);
3982	list_del(entry: &q->tag_set_list);
3983	if (list_is_singular(head: &set->tag_list)) {
3984	/* just transitioned to unshared */
3985	set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3986	/* update existing queue */
3987	blk_mq_update_tag_set_shared(set, shared: false);
3988	}
3989	mutex_unlock(lock: &set->tag_list_lock);
3990	INIT_LIST_HEAD(list: &q->tag_set_list);
3991	}
3992
3993	static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3994	struct request_queue *q)
3995	{
3996	mutex_lock(&set->tag_list_lock);
3997
3998	/*
3999	* Check to see if we're transitioning to shared (from 1 to 2 queues).
4000	*/
4001	if (!list_empty(head: &set->tag_list) &&
4002	!(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
4003	set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
4004	/* update existing queue */
4005	blk_mq_update_tag_set_shared(set, shared: true);
4006	}
4007	if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
4008	queue_set_hctx_shared(q, shared: true);
4009	list_add_tail(new: &q->tag_set_list, head: &set->tag_list);
4010
4011	mutex_unlock(lock: &set->tag_list_lock);
4012	}
4013
4014	/* All allocations will be freed in release handler of q->mq_kobj */
4015	static int blk_mq_alloc_ctxs(struct request_queue *q)
4016	{
4017	struct blk_mq_ctxs *ctxs;
4018	int cpu;
4019
4020	ctxs = kzalloc(size: sizeof(*ctxs), GFP_KERNEL);
4021	if (!ctxs)
4022	return -ENOMEM;
4023
4024	ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
4025	if (!ctxs->queue_ctx)
4026	goto fail;
4027
4028	for_each_possible_cpu(cpu) {
4029	struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
4030	ctx->ctxs = ctxs;
4031	}
4032
4033	q->mq_kobj = &ctxs->kobj;
4034	q->queue_ctx = ctxs->queue_ctx;
4035
4036	return 0;
4037	fail:
4038	kfree(objp: ctxs);
4039	return -ENOMEM;
4040	}
4041
4042	/*
4043	* It is the actual release handler for mq, but we do it from
4044	* request queue's release handler for avoiding use-after-free
4045	* and headache because q->mq_kobj shouldn't have been introduced,
4046	* but we can't group ctx/kctx kobj without it.
4047	*/
4048	void blk_mq_release(struct request_queue *q)
4049	{
4050	struct blk_mq_hw_ctx *hctx, *next;
4051	unsigned long i;
4052
4053	queue_for_each_hw_ctx(q, hctx, i)
4054	WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4055
4056	/* all hctx are in .unused_hctx_list now */
4057	list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4058	list_del_init(entry: &hctx->hctx_list);
4059	kobject_put(kobj: &hctx->kobj);
4060	}
4061
4062	xa_destroy(&q->hctx_table);
4063
4064	/*
4065	* release .mq_kobj and sw queue's kobject now because
4066	* both share lifetime with request queue.
4067	*/
4068	blk_mq_sysfs_deinit(q);
4069	}
4070
4071	struct request_queue *blk_mq_alloc_queue(struct blk_mq_tag_set *set,
4072	struct queue_limits *lim, void *queuedata)
4073	{
4074	struct queue_limits default_lim = { };
4075	struct request_queue *q;
4076	int ret;
4077
4078	q = blk_alloc_queue(lim: lim ? lim : &default_lim, node_id: set->numa_node);
4079	if (IS_ERR(ptr: q))
4080	return q;
4081	q->queuedata = queuedata;
4082	ret = blk_mq_init_allocated_queue(set, q);
4083	if (ret) {
4084	blk_put_queue(q);
4085	return ERR_PTR(error: ret);
4086	}
4087	return q;
4088	}
4089	EXPORT_SYMBOL(blk_mq_alloc_queue);
4090
4091	/**
4092	* blk_mq_destroy_queue - shutdown a request queue
4093	* @q: request queue to shutdown
4094	*
4095	* This shuts down a request queue allocated by blk_mq_alloc_queue(). All future
4096	* requests will be failed with -ENODEV. The caller is responsible for dropping
4097	* the reference from blk_mq_alloc_queue() by calling blk_put_queue().
4098	*
4099	* Context: can sleep
4100	*/
4101	void blk_mq_destroy_queue(struct request_queue *q)
4102	{
4103	WARN_ON_ONCE(!queue_is_mq(q));
4104	WARN_ON_ONCE(blk_queue_registered(q));
4105
4106	might_sleep();
4107
4108	blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4109	blk_queue_start_drain(q);
4110	blk_mq_freeze_queue_wait(q);
4111
4112	blk_sync_queue(q);
4113	blk_mq_cancel_work_sync(q);
4114	blk_mq_exit_queue(q);
4115	}
4116	EXPORT_SYMBOL(blk_mq_destroy_queue);
4117
4118	struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set,
4119	struct queue_limits *lim, void *queuedata,
4120	struct lock_class_key *lkclass)
4121	{
4122	struct request_queue *q;
4123	struct gendisk *disk;
4124
4125	q = blk_mq_alloc_queue(set, lim, queuedata);
4126	if (IS_ERR(ptr: q))
4127	return ERR_CAST(ptr: q);
4128
4129	disk = __alloc_disk_node(q, node_id: set->numa_node, lkclass);
4130	if (!disk) {
4131	blk_mq_destroy_queue(q);
4132	blk_put_queue(q);
4133	return ERR_PTR(error: -ENOMEM);
4134	}
4135	set_bit(GD_OWNS_QUEUE, addr: &disk->state);
4136	return disk;
4137	}
4138	EXPORT_SYMBOL(__blk_mq_alloc_disk);
4139
4140	struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4141	struct lock_class_key *lkclass)
4142	{
4143	struct gendisk *disk;
4144
4145	if (!blk_get_queue(q))
4146	return NULL;
4147	disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4148	if (!disk)
4149	blk_put_queue(q);
4150	return disk;
4151	}
4152	EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4153
4154	static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4155	struct blk_mq_tag_set *set, struct request_queue *q,
4156	int hctx_idx, int node)
4157	{
4158	struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4159
4160	/* reuse dead hctx first */
4161	spin_lock(lock: &q->unused_hctx_lock);
4162	list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4163	if (tmp->numa_node == node) {
4164	hctx = tmp;
4165	break;
4166	}
4167	}
4168	if (hctx)
4169	list_del_init(entry: &hctx->hctx_list);
4170	spin_unlock(lock: &q->unused_hctx_lock);
4171
4172	if (!hctx)
4173	hctx = blk_mq_alloc_hctx(q, set, node);
4174	if (!hctx)
4175	goto fail;
4176
4177	if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4178	goto free_hctx;
4179
4180	return hctx;
4181
4182	free_hctx:
4183	kobject_put(kobj: &hctx->kobj);
4184	fail:
4185	return NULL;
4186	}
4187
4188	static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4189	struct request_queue *q)
4190	{
4191	struct blk_mq_hw_ctx *hctx;
4192	unsigned long i, j;
4193
4194	/* protect against switching io scheduler */
4195	mutex_lock(&q->sysfs_lock);
4196	for (i = 0; i < set->nr_hw_queues; i++) {
4197	int old_node;
4198	int node = blk_mq_get_hctx_node(set, hctx_idx: i);
4199	struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, index: i);
4200
4201	if (old_hctx) {
4202	old_node = old_hctx->numa_node;
4203	blk_mq_exit_hctx(q, set, hctx: old_hctx, hctx_idx: i);
4204	}
4205
4206	if (!blk_mq_alloc_and_init_hctx(set, q, hctx_idx: i, node)) {
4207	if (!old_hctx)
4208	break;
4209	pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4210	node, old_node);
4211	hctx = blk_mq_alloc_and_init_hctx(set, q, hctx_idx: i, node: old_node);
4212	WARN_ON_ONCE(!hctx);
4213	}
4214	}
4215	/*
4216	* Increasing nr_hw_queues fails. Free the newly allocated
4217	* hctxs and keep the previous q->nr_hw_queues.
4218	*/
4219	if (i != set->nr_hw_queues) {
4220	j = q->nr_hw_queues;
4221	} else {
4222	j = i;
4223	q->nr_hw_queues = set->nr_hw_queues;
4224	}
4225
4226	xa_for_each_start(&q->hctx_table, j, hctx, j)
4227	blk_mq_exit_hctx(q, set, hctx, hctx_idx: j);
4228	mutex_unlock(lock: &q->sysfs_lock);
4229	}
4230
4231	static void blk_mq_update_poll_flag(struct request_queue *q)
4232	{
4233	struct blk_mq_tag_set *set = q->tag_set;
4234
4235	if (set->nr_maps > HCTX_TYPE_POLL &&
4236	set->map[HCTX_TYPE_POLL].nr_queues)
4237	blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4238	else
4239	blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4240	}
4241
4242	int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4243	struct request_queue *q)
4244	{
4245	/* mark the queue as mq asap */
4246	q->mq_ops = set->ops;
4247
4248	if (blk_mq_alloc_ctxs(q))
4249	goto err_exit;
4250
4251	/* init q->mq_kobj and sw queues' kobjects */
4252	blk_mq_sysfs_init(q);
4253
4254	INIT_LIST_HEAD(list: &q->unused_hctx_list);
4255	spin_lock_init(&q->unused_hctx_lock);
4256
4257	xa_init(xa: &q->hctx_table);
4258
4259	blk_mq_realloc_hw_ctxs(set, q);
4260	if (!q->nr_hw_queues)
4261	goto err_hctxs;
4262
4263	INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4264	blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4265
4266	q->tag_set = set;
4267
4268	q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4269	blk_mq_update_poll_flag(q);
4270
4271	INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4272	INIT_LIST_HEAD(list: &q->flush_list);
4273	INIT_LIST_HEAD(list: &q->requeue_list);
4274	spin_lock_init(&q->requeue_lock);
4275
4276	q->nr_requests = set->queue_depth;
4277
4278	blk_mq_init_cpu_queues(q, nr_hw_queues: set->nr_hw_queues);
4279	blk_mq_add_queue_tag_set(set, q);
4280	blk_mq_map_swqueue(q);
4281	return 0;
4282
4283	err_hctxs:
4284	blk_mq_release(q);
4285	err_exit:
4286	q->mq_ops = NULL;
4287	return -ENOMEM;
4288	}
4289	EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4290
4291	/* tags can _not_ be used after returning from blk_mq_exit_queue */
4292	void blk_mq_exit_queue(struct request_queue *q)
4293	{
4294	struct blk_mq_tag_set *set = q->tag_set;
4295
4296	/* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4297	blk_mq_exit_hw_queues(q, set, nr_queue: set->nr_hw_queues);
4298	/* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4299	blk_mq_del_queue_tag_set(q);
4300	}
4301
4302	static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4303	{
4304	int i;
4305
4306	if (blk_mq_is_shared_tags(flags: set->flags)) {
4307	set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4308	BLK_MQ_NO_HCTX_IDX,
4309	depth: set->queue_depth);
4310	if (!set->shared_tags)
4311	return -ENOMEM;
4312	}
4313
4314	for (i = 0; i < set->nr_hw_queues; i++) {
4315	if (!__blk_mq_alloc_map_and_rqs(set, hctx_idx: i))
4316	goto out_unwind;
4317	cond_resched();
4318	}
4319
4320	return 0;
4321
4322	out_unwind:
4323	while (--i >= 0)
4324	__blk_mq_free_map_and_rqs(set, hctx_idx: i);
4325
4326	if (blk_mq_is_shared_tags(flags: set->flags)) {
4327	blk_mq_free_map_and_rqs(set, tags: set->shared_tags,
4328	BLK_MQ_NO_HCTX_IDX);
4329	}
4330
4331	return -ENOMEM;
4332	}
4333
4334	/*
4335	* Allocate the request maps associated with this tag_set. Note that this
4336	* may reduce the depth asked for, if memory is tight. set->queue_depth
4337	* will be updated to reflect the allocated depth.
4338	*/
4339	static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4340	{
4341	unsigned int depth;
4342	int err;
4343
4344	depth = set->queue_depth;
4345	do {
4346	err = __blk_mq_alloc_rq_maps(set);
4347	if (!err)
4348	break;
4349
4350	set->queue_depth >>= 1;
4351	if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4352	err = -ENOMEM;
4353	break;
4354	}
4355	} while (set->queue_depth);
4356
4357	if (!set->queue_depth || err) {
4358	pr_err("blk-mq: failed to allocate request map\n");
4359	return -ENOMEM;
4360	}
4361
4362	if (depth != set->queue_depth)
4363	pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4364	depth, set->queue_depth);
4365
4366	return 0;
4367	}
4368
4369	static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4370	{
4371	/*
4372	* blk_mq_map_queues() and multiple .map_queues() implementations
4373	* expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4374	* number of hardware queues.
4375	*/
4376	if (set->nr_maps == 1)
4377	set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4378
4379	if (set->ops->map_queues) {
4380	int i;
4381
4382	/*
4383	* transport .map_queues is usually done in the following
4384	* way:
4385	*
4386	* for (queue = 0; queue < set->nr_hw_queues; queue++) {
4387	* mask = get_cpu_mask(queue)
4388	* for_each_cpu(cpu, mask)
4389	* set->map[x].mq_map[cpu] = queue;
4390	* }
4391	*
4392	* When we need to remap, the table has to be cleared for
4393	* killing stale mapping since one CPU may not be mapped
4394	* to any hw queue.
4395	*/
4396	for (i = 0; i < set->nr_maps; i++)
4397	blk_mq_clear_mq_map(qmap: &set->map[i]);
4398
4399	set->ops->map_queues(set);
4400	} else {
4401	BUG_ON(set->nr_maps > 1);
4402	blk_mq_map_queues(qmap: &set->map[HCTX_TYPE_DEFAULT]);
4403	}
4404	}
4405
4406	static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4407	int new_nr_hw_queues)
4408	{
4409	struct blk_mq_tags **new_tags;
4410	int i;
4411
4412	if (set->nr_hw_queues >= new_nr_hw_queues)
4413	goto done;
4414
4415	new_tags = kcalloc_node(n: new_nr_hw_queues, size: sizeof(struct blk_mq_tags *),
4416	GFP_KERNEL, node: set->numa_node);
4417	if (!new_tags)
4418	return -ENOMEM;
4419
4420	if (set->tags)
4421	memcpy(new_tags, set->tags, set->nr_hw_queues *
4422	sizeof(*set->tags));
4423	kfree(objp: set->tags);
4424	set->tags = new_tags;
4425
4426	for (i = set->nr_hw_queues; i < new_nr_hw_queues; i++) {
4427	if (!__blk_mq_alloc_map_and_rqs(set, hctx_idx: i)) {
4428	while (--i >= set->nr_hw_queues)
4429	__blk_mq_free_map_and_rqs(set, hctx_idx: i);
4430	return -ENOMEM;
4431	}
4432	cond_resched();
4433	}
4434
4435	done:
4436	set->nr_hw_queues = new_nr_hw_queues;
4437	return 0;
4438	}
4439
4440	/*
4441	* Alloc a tag set to be associated with one or more request queues.
4442	* May fail with EINVAL for various error conditions. May adjust the
4443	* requested depth down, if it's too large. In that case, the set
4444	* value will be stored in set->queue_depth.
4445	*/
4446	int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4447	{
4448	int i, ret;
4449
4450	BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4451
4452	if (!set->nr_hw_queues)
4453	return -EINVAL;
4454	if (!set->queue_depth)
4455	return -EINVAL;
4456	if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4457	return -EINVAL;
4458
4459	if (!set->ops->queue_rq)
4460	return -EINVAL;
4461
4462	if (!set->ops->get_budget ^ !set->ops->put_budget)
4463	return -EINVAL;
4464
4465	if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4466	pr_info("blk-mq: reduced tag depth to %u\n",
4467	BLK_MQ_MAX_DEPTH);
4468	set->queue_depth = BLK_MQ_MAX_DEPTH;
4469	}
4470
4471	if (!set->nr_maps)
4472	set->nr_maps = 1;
4473	else if (set->nr_maps > HCTX_MAX_TYPES)
4474	return -EINVAL;
4475
4476	/*
4477	* If a crashdump is active, then we are potentially in a very
4478	* memory constrained environment. Limit us to 64 tags to prevent
4479	* using too much memory.
4480	*/
4481	if (is_kdump_kernel())
4482	set->queue_depth = min(64U, set->queue_depth);
4483
4484	/*
4485	* There is no use for more h/w queues than cpus if we just have
4486	* a single map
4487	*/
4488	if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4489	set->nr_hw_queues = nr_cpu_ids;
4490
4491	if (set->flags & BLK_MQ_F_BLOCKING) {
4492	set->srcu = kmalloc(size: sizeof(*set->srcu), GFP_KERNEL);
4493	if (!set->srcu)
4494	return -ENOMEM;
4495	ret = init_srcu_struct(set->srcu);
4496	if (ret)
4497	goto out_free_srcu;
4498	}
4499
4500	ret = -ENOMEM;
4501	set->tags = kcalloc_node(n: set->nr_hw_queues,
4502	size: sizeof(struct blk_mq_tags *), GFP_KERNEL,
4503	node: set->numa_node);
4504	if (!set->tags)
4505	goto out_cleanup_srcu;
4506
4507	for (i = 0; i < set->nr_maps; i++) {
4508	set->map[i].mq_map = kcalloc_node(n: nr_cpu_ids,
4509	size: sizeof(set->map[i].mq_map[0]),
4510	GFP_KERNEL, node: set->numa_node);
4511	if (!set->map[i].mq_map)
4512	goto out_free_mq_map;
4513	set->map[i].nr_queues = set->nr_hw_queues;
4514	}
4515
4516	blk_mq_update_queue_map(set);
4517
4518	ret = blk_mq_alloc_set_map_and_rqs(set);
4519	if (ret)
4520	goto out_free_mq_map;
4521
4522	mutex_init(&set->tag_list_lock);
4523	INIT_LIST_HEAD(list: &set->tag_list);
4524
4525	return 0;
4526
4527	out_free_mq_map:
4528	for (i = 0; i < set->nr_maps; i++) {
4529	kfree(objp: set->map[i].mq_map);
4530	set->map[i].mq_map = NULL;
4531	}
4532	kfree(objp: set->tags);
4533	set->tags = NULL;
4534	out_cleanup_srcu:
4535	if (set->flags & BLK_MQ_F_BLOCKING)
4536	cleanup_srcu_struct(ssp: set->srcu);
4537	out_free_srcu:
4538	if (set->flags & BLK_MQ_F_BLOCKING)
4539	kfree(objp: set->srcu);
4540	return ret;
4541	}
4542	EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4543
4544	/* allocate and initialize a tagset for a simple single-queue device */
4545	int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4546	const struct blk_mq_ops *ops, unsigned int queue_depth,
4547	unsigned int set_flags)
4548	{
4549	memset(set, 0, sizeof(*set));
4550	set->ops = ops;
4551	set->nr_hw_queues = 1;
4552	set->nr_maps = 1;
4553	set->queue_depth = queue_depth;
4554	set->numa_node = NUMA_NO_NODE;
4555	set->flags = set_flags;
4556	return blk_mq_alloc_tag_set(set);
4557	}
4558	EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4559
4560	void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4561	{
4562	int i, j;
4563
4564	for (i = 0; i < set->nr_hw_queues; i++)
4565	__blk_mq_free_map_and_rqs(set, hctx_idx: i);
4566
4567	if (blk_mq_is_shared_tags(flags: set->flags)) {
4568	blk_mq_free_map_and_rqs(set, tags: set->shared_tags,
4569	BLK_MQ_NO_HCTX_IDX);
4570	}
4571
4572	for (j = 0; j < set->nr_maps; j++) {
4573	kfree(objp: set->map[j].mq_map);
4574	set->map[j].mq_map = NULL;
4575	}
4576
4577	kfree(objp: set->tags);
4578	set->tags = NULL;
4579	if (set->flags & BLK_MQ_F_BLOCKING) {
4580	cleanup_srcu_struct(ssp: set->srcu);
4581	kfree(objp: set->srcu);
4582	}
4583	}
4584	EXPORT_SYMBOL(blk_mq_free_tag_set);
4585
4586	int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4587	{
4588	struct blk_mq_tag_set *set = q->tag_set;
4589	struct blk_mq_hw_ctx *hctx;
4590	int ret;
4591	unsigned long i;
4592
4593	if (!set)
4594	return -EINVAL;
4595
4596	if (q->nr_requests == nr)
4597	return 0;
4598
4599	blk_mq_freeze_queue(q);
4600	blk_mq_quiesce_queue(q);
4601
4602	ret = 0;
4603	queue_for_each_hw_ctx(q, hctx, i) {
4604	if (!hctx->tags)
4605	continue;
4606	/*
4607	* If we're using an MQ scheduler, just update the scheduler
4608	* queue depth. This is similar to what the old code would do.
4609	*/
4610	if (hctx->sched_tags) {
4611	ret = blk_mq_tag_update_depth(hctx, tags: &hctx->sched_tags,
4612	depth: nr, can_grow: true);
4613	} else {
4614	ret = blk_mq_tag_update_depth(hctx, tags: &hctx->tags, depth: nr,
4615	can_grow: false);
4616	}
4617	if (ret)
4618	break;
4619	if (q->elevator && q->elevator->type->ops.depth_updated)
4620	q->elevator->type->ops.depth_updated(hctx);
4621	}
4622	if (!ret) {
4623	q->nr_requests = nr;
4624	if (blk_mq_is_shared_tags(flags: set->flags)) {
4625	if (q->elevator)
4626	blk_mq_tag_update_sched_shared_tags(q);
4627	else
4628	blk_mq_tag_resize_shared_tags(set, size: nr);
4629	}
4630	}
4631
4632	blk_mq_unquiesce_queue(q);
4633	blk_mq_unfreeze_queue(q);
4634
4635	return ret;
4636	}
4637
4638	/*
4639	* request_queue and elevator_type pair.
4640	* It is just used by __blk_mq_update_nr_hw_queues to cache
4641	* the elevator_type associated with a request_queue.
4642	*/
4643	struct blk_mq_qe_pair {
4644	struct list_head node;
4645	struct request_queue *q;
4646	struct elevator_type *type;
4647	};
4648
4649	/*
4650	* Cache the elevator_type in qe pair list and switch the
4651	* io scheduler to 'none'
4652	*/
4653	static bool blk_mq_elv_switch_none(struct list_head *head,
4654	struct request_queue *q)
4655	{
4656	struct blk_mq_qe_pair *qe;
4657
4658	qe = kmalloc(size: sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4659	if (!qe)
4660	return false;
4661
4662	/* q->elevator needs protection from ->sysfs_lock */
4663	mutex_lock(&q->sysfs_lock);
4664
4665	/* the check has to be done with holding sysfs_lock */
4666	if (!q->elevator) {
4667	kfree(objp: qe);
4668	goto unlock;
4669	}
4670
4671	INIT_LIST_HEAD(list: &qe->node);
4672	qe->q = q;
4673	qe->type = q->elevator->type;
4674	/* keep a reference to the elevator module as we'll switch back */
4675	__elevator_get(e: qe->type);
4676	list_add(new: &qe->node, head);
4677	elevator_disable(q);
4678	unlock:
4679	mutex_unlock(lock: &q->sysfs_lock);
4680
4681	return true;
4682	}
4683
4684	static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4685	struct request_queue *q)
4686	{
4687	struct blk_mq_qe_pair *qe;
4688
4689	list_for_each_entry(qe, head, node)
4690	if (qe->q == q)
4691	return qe;
4692
4693	return NULL;
4694	}
4695
4696	static void blk_mq_elv_switch_back(struct list_head *head,
4697	struct request_queue *q)
4698	{
4699	struct blk_mq_qe_pair *qe;
4700	struct elevator_type *t;
4701
4702	qe = blk_lookup_qe_pair(head, q);
4703	if (!qe)
4704	return;
4705	t = qe->type;
4706	list_del(entry: &qe->node);
4707	kfree(objp: qe);
4708
4709	mutex_lock(&q->sysfs_lock);
4710	elevator_switch(q, new_e: t);
4711	/* drop the reference acquired in blk_mq_elv_switch_none */
4712	elevator_put(e: t);
4713	mutex_unlock(lock: &q->sysfs_lock);
4714	}
4715
4716	static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4717	int nr_hw_queues)
4718	{
4719	struct request_queue *q;
4720	LIST_HEAD(head);
4721	int prev_nr_hw_queues = set->nr_hw_queues;
4722	int i;
4723
4724	lockdep_assert_held(&set->tag_list_lock);
4725
4726	if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4727	nr_hw_queues = nr_cpu_ids;
4728	if (nr_hw_queues < 1)
4729	return;
4730	if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4731	return;
4732
4733	list_for_each_entry(q, &set->tag_list, tag_set_list)
4734	blk_mq_freeze_queue(q);
4735	/*
4736	* Switch IO scheduler to 'none', cleaning up the data associated
4737	* with the previous scheduler. We will switch back once we are done
4738	* updating the new sw to hw queue mappings.
4739	*/
4740	list_for_each_entry(q, &set->tag_list, tag_set_list)
4741	if (!blk_mq_elv_switch_none(head: &head, q))
4742	goto switch_back;
4743
4744	list_for_each_entry(q, &set->tag_list, tag_set_list) {
4745	blk_mq_debugfs_unregister_hctxs(q);
4746	blk_mq_sysfs_unregister_hctxs(q);
4747	}
4748
4749	if (blk_mq_realloc_tag_set_tags(set, new_nr_hw_queues: nr_hw_queues) < 0)
4750	goto reregister;
4751
4752	fallback:
4753	blk_mq_update_queue_map(set);
4754	list_for_each_entry(q, &set->tag_list, tag_set_list) {
4755	blk_mq_realloc_hw_ctxs(set, q);
4756	blk_mq_update_poll_flag(q);
4757	if (q->nr_hw_queues != set->nr_hw_queues) {
4758	int i = prev_nr_hw_queues;
4759
4760	pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4761	nr_hw_queues, prev_nr_hw_queues);
4762	for (; i < set->nr_hw_queues; i++)
4763	__blk_mq_free_map_and_rqs(set, hctx_idx: i);
4764
4765	set->nr_hw_queues = prev_nr_hw_queues;
4766	goto fallback;
4767	}
4768	blk_mq_map_swqueue(q);
4769	}
4770
4771	reregister:
4772	list_for_each_entry(q, &set->tag_list, tag_set_list) {
4773	blk_mq_sysfs_register_hctxs(q);
4774	blk_mq_debugfs_register_hctxs(q);
4775	}
4776
4777	switch_back:
4778	list_for_each_entry(q, &set->tag_list, tag_set_list)
4779	blk_mq_elv_switch_back(head: &head, q);
4780
4781	list_for_each_entry(q, &set->tag_list, tag_set_list)
4782	blk_mq_unfreeze_queue(q);
4783
4784	/* Free the excess tags when nr_hw_queues shrink. */
4785	for (i = set->nr_hw_queues; i < prev_nr_hw_queues; i++)
4786	__blk_mq_free_map_and_rqs(set, hctx_idx: i);
4787	}
4788
4789	void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4790	{
4791	mutex_lock(&set->tag_list_lock);
4792	__blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4793	mutex_unlock(lock: &set->tag_list_lock);
4794	}
4795	EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4796
4797	static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
4798	struct io_comp_batch *iob, unsigned int flags)
4799	{
4800	long state = get_current_state();
4801	int ret;
4802
4803	do {
4804	ret = q->mq_ops->poll(hctx, iob);
4805	if (ret > 0) {
4806	__set_current_state(TASK_RUNNING);
4807	return ret;
4808	}
4809
4810	if (signal_pending_state(state, current))
4811	__set_current_state(TASK_RUNNING);
4812	if (task_is_running(current))
4813	return 1;
4814
4815	if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4816	break;
4817	cpu_relax();
4818	} while (!need_resched());
4819
4820	__set_current_state(TASK_RUNNING);
4821	return 0;
4822	}
4823
4824	int blk_mq_poll(struct request_queue *q, blk_qc_t cookie,
4825	struct io_comp_batch *iob, unsigned int flags)
4826	{
4827	struct blk_mq_hw_ctx *hctx = xa_load(&q->hctx_table, index: cookie);
4828
4829	return blk_hctx_poll(q, hctx, iob, flags);
4830	}
4831
4832	int blk_rq_poll(struct request *rq, struct io_comp_batch *iob,
4833	unsigned int poll_flags)
4834	{
4835	struct request_queue *q = rq->q;
4836	int ret;
4837
4838	if (!blk_rq_is_poll(rq))
4839	return 0;
4840	if (!percpu_ref_tryget(ref: &q->q_usage_counter))
4841	return 0;
4842
4843	ret = blk_hctx_poll(q, hctx: rq->mq_hctx, iob, flags: poll_flags);
4844	blk_queue_exit(q);
4845
4846	return ret;
4847	}
4848	EXPORT_SYMBOL_GPL(blk_rq_poll);
4849
4850	unsigned int blk_mq_rq_cpu(struct request *rq)
4851	{
4852	return rq->mq_ctx->cpu;
4853	}
4854	EXPORT_SYMBOL(blk_mq_rq_cpu);
4855
4856	void blk_mq_cancel_work_sync(struct request_queue *q)
4857	{
4858	struct blk_mq_hw_ctx *hctx;
4859	unsigned long i;
4860
4861	cancel_delayed_work_sync(dwork: &q->requeue_work);
4862
4863	queue_for_each_hw_ctx(q, hctx, i)
4864	cancel_delayed_work_sync(dwork: &hctx->run_work);
4865	}
4866
4867	static int __init blk_mq_init(void)
4868	{
4869	int i;
4870
4871	for_each_possible_cpu(i)
4872	init_llist_head(list: &per_cpu(blk_cpu_done, i));
4873	for_each_possible_cpu(i)
4874	INIT_CSD(&per_cpu(blk_cpu_csd, i),
4875	__blk_mq_complete_request_remote, NULL);
4876	open_softirq(nr: BLOCK_SOFTIRQ, action: blk_done_softirq);
4877
4878	cpuhp_setup_state_nocalls(state: CPUHP_BLOCK_SOFTIRQ_DEAD,
4879	name: "block/softirq:dead", NULL,
4880	teardown: blk_softirq_cpu_dead);
4881	cpuhp_setup_state_multi(state: CPUHP_BLK_MQ_DEAD, name: "block/mq:dead", NULL,
4882	teardown: blk_mq_hctx_notify_dead);
4883	cpuhp_setup_state_multi(state: CPUHP_AP_BLK_MQ_ONLINE, name: "block/mq:online",
4884	startup: blk_mq_hctx_notify_online,
4885	teardown: blk_mq_hctx_notify_offline);
4886	return 0;
4887	}
4888	subsys_initcall(blk_mq_init);
4889