Merge tag 'spi-fix-v6-6-rc4' of git://git.kernel.org/pub/scm/linux/kernel/git/broonie/spi
[platform/kernel/linux-starfive.git] / block / blk-mq.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Block multiqueue core code
4  *
5  * Copyright (C) 2013-2014 Jens Axboe
6  * Copyright (C) 2013-2014 Christoph Hellwig
7  */
8 #include <linux/kernel.h>
9 #include <linux/module.h>
10 #include <linux/backing-dev.h>
11 #include <linux/bio.h>
12 #include <linux/blkdev.h>
13 #include <linux/blk-integrity.h>
14 #include <linux/kmemleak.h>
15 #include <linux/mm.h>
16 #include <linux/init.h>
17 #include <linux/slab.h>
18 #include <linux/workqueue.h>
19 #include <linux/smp.h>
20 #include <linux/interrupt.h>
21 #include <linux/llist.h>
22 #include <linux/cpu.h>
23 #include <linux/cache.h>
24 #include <linux/sched/sysctl.h>
25 #include <linux/sched/topology.h>
26 #include <linux/sched/signal.h>
27 #include <linux/delay.h>
28 #include <linux/crash_dump.h>
29 #include <linux/prefetch.h>
30 #include <linux/blk-crypto.h>
31 #include <linux/part_stat.h>
32
33 #include <trace/events/block.h>
34
35 #include <linux/t10-pi.h>
36 #include "blk.h"
37 #include "blk-mq.h"
38 #include "blk-mq-debugfs.h"
39 #include "blk-pm.h"
40 #include "blk-stat.h"
41 #include "blk-mq-sched.h"
42 #include "blk-rq-qos.h"
43 #include "blk-ioprio.h"
44
45 static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
46 static DEFINE_PER_CPU(call_single_data_t, blk_cpu_csd);
47
48 static void blk_mq_insert_request(struct request *rq, blk_insert_t flags);
49 static void blk_mq_request_bypass_insert(struct request *rq,
50                 blk_insert_t flags);
51 static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
52                 struct list_head *list);
53 static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
54                          struct io_comp_batch *iob, unsigned int flags);
55
56 /*
57  * Check if any of the ctx, dispatch list or elevator
58  * have pending work in this hardware queue.
59  */
60 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
61 {
62         return !list_empty_careful(&hctx->dispatch) ||
63                 sbitmap_any_bit_set(&hctx->ctx_map) ||
64                         blk_mq_sched_has_work(hctx);
65 }
66
67 /*
68  * Mark this ctx as having pending work in this hardware queue
69  */
70 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
71                                      struct blk_mq_ctx *ctx)
72 {
73         const int bit = ctx->index_hw[hctx->type];
74
75         if (!sbitmap_test_bit(&hctx->ctx_map, bit))
76                 sbitmap_set_bit(&hctx->ctx_map, bit);
77 }
78
79 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
80                                       struct blk_mq_ctx *ctx)
81 {
82         const int bit = ctx->index_hw[hctx->type];
83
84         sbitmap_clear_bit(&hctx->ctx_map, bit);
85 }
86
87 struct mq_inflight {
88         struct block_device *part;
89         unsigned int inflight[2];
90 };
91
92 static bool blk_mq_check_inflight(struct request *rq, void *priv)
93 {
94         struct mq_inflight *mi = priv;
95
96         if (rq->part && blk_do_io_stat(rq) &&
97             (!mi->part->bd_partno || rq->part == mi->part) &&
98             blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
99                 mi->inflight[rq_data_dir(rq)]++;
100
101         return true;
102 }
103
104 unsigned int blk_mq_in_flight(struct request_queue *q,
105                 struct block_device *part)
106 {
107         struct mq_inflight mi = { .part = part };
108
109         blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
110
111         return mi.inflight[0] + mi.inflight[1];
112 }
113
114 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
115                 unsigned int inflight[2])
116 {
117         struct mq_inflight mi = { .part = part };
118
119         blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
120         inflight[0] = mi.inflight[0];
121         inflight[1] = mi.inflight[1];
122 }
123
124 void blk_freeze_queue_start(struct request_queue *q)
125 {
126         mutex_lock(&q->mq_freeze_lock);
127         if (++q->mq_freeze_depth == 1) {
128                 percpu_ref_kill(&q->q_usage_counter);
129                 mutex_unlock(&q->mq_freeze_lock);
130                 if (queue_is_mq(q))
131                         blk_mq_run_hw_queues(q, false);
132         } else {
133                 mutex_unlock(&q->mq_freeze_lock);
134         }
135 }
136 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
137
138 void blk_mq_freeze_queue_wait(struct request_queue *q)
139 {
140         wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
141 }
142 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
143
144 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
145                                      unsigned long timeout)
146 {
147         return wait_event_timeout(q->mq_freeze_wq,
148                                         percpu_ref_is_zero(&q->q_usage_counter),
149                                         timeout);
150 }
151 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
152
153 /*
154  * Guarantee no request is in use, so we can change any data structure of
155  * the queue afterward.
156  */
157 void blk_freeze_queue(struct request_queue *q)
158 {
159         /*
160          * In the !blk_mq case we are only calling this to kill the
161          * q_usage_counter, otherwise this increases the freeze depth
162          * and waits for it to return to zero.  For this reason there is
163          * no blk_unfreeze_queue(), and blk_freeze_queue() is not
164          * exported to drivers as the only user for unfreeze is blk_mq.
165          */
166         blk_freeze_queue_start(q);
167         blk_mq_freeze_queue_wait(q);
168 }
169
170 void blk_mq_freeze_queue(struct request_queue *q)
171 {
172         /*
173          * ...just an alias to keep freeze and unfreeze actions balanced
174          * in the blk_mq_* namespace
175          */
176         blk_freeze_queue(q);
177 }
178 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
179
180 void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
181 {
182         mutex_lock(&q->mq_freeze_lock);
183         if (force_atomic)
184                 q->q_usage_counter.data->force_atomic = true;
185         q->mq_freeze_depth--;
186         WARN_ON_ONCE(q->mq_freeze_depth < 0);
187         if (!q->mq_freeze_depth) {
188                 percpu_ref_resurrect(&q->q_usage_counter);
189                 wake_up_all(&q->mq_freeze_wq);
190         }
191         mutex_unlock(&q->mq_freeze_lock);
192 }
193
194 void blk_mq_unfreeze_queue(struct request_queue *q)
195 {
196         __blk_mq_unfreeze_queue(q, false);
197 }
198 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
199
200 /*
201  * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
202  * mpt3sas driver such that this function can be removed.
203  */
204 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
205 {
206         unsigned long flags;
207
208         spin_lock_irqsave(&q->queue_lock, flags);
209         if (!q->quiesce_depth++)
210                 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
211         spin_unlock_irqrestore(&q->queue_lock, flags);
212 }
213 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
214
215 /**
216  * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
217  * @set: tag_set to wait on
218  *
219  * Note: it is driver's responsibility for making sure that quiesce has
220  * been started on or more of the request_queues of the tag_set.  This
221  * function only waits for the quiesce on those request_queues that had
222  * the quiesce flag set using blk_mq_quiesce_queue_nowait.
223  */
224 void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set)
225 {
226         if (set->flags & BLK_MQ_F_BLOCKING)
227                 synchronize_srcu(set->srcu);
228         else
229                 synchronize_rcu();
230 }
231 EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
232
233 /**
234  * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
235  * @q: request queue.
236  *
237  * Note: this function does not prevent that the struct request end_io()
238  * callback function is invoked. Once this function is returned, we make
239  * sure no dispatch can happen until the queue is unquiesced via
240  * blk_mq_unquiesce_queue().
241  */
242 void blk_mq_quiesce_queue(struct request_queue *q)
243 {
244         blk_mq_quiesce_queue_nowait(q);
245         /* nothing to wait for non-mq queues */
246         if (queue_is_mq(q))
247                 blk_mq_wait_quiesce_done(q->tag_set);
248 }
249 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
250
251 /*
252  * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
253  * @q: request queue.
254  *
255  * This function recovers queue into the state before quiescing
256  * which is done by blk_mq_quiesce_queue.
257  */
258 void blk_mq_unquiesce_queue(struct request_queue *q)
259 {
260         unsigned long flags;
261         bool run_queue = false;
262
263         spin_lock_irqsave(&q->queue_lock, flags);
264         if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
265                 ;
266         } else if (!--q->quiesce_depth) {
267                 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
268                 run_queue = true;
269         }
270         spin_unlock_irqrestore(&q->queue_lock, flags);
271
272         /* dispatch requests which are inserted during quiescing */
273         if (run_queue)
274                 blk_mq_run_hw_queues(q, true);
275 }
276 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
277
278 void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set)
279 {
280         struct request_queue *q;
281
282         mutex_lock(&set->tag_list_lock);
283         list_for_each_entry(q, &set->tag_list, tag_set_list) {
284                 if (!blk_queue_skip_tagset_quiesce(q))
285                         blk_mq_quiesce_queue_nowait(q);
286         }
287         blk_mq_wait_quiesce_done(set);
288         mutex_unlock(&set->tag_list_lock);
289 }
290 EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset);
291
292 void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set)
293 {
294         struct request_queue *q;
295
296         mutex_lock(&set->tag_list_lock);
297         list_for_each_entry(q, &set->tag_list, tag_set_list) {
298                 if (!blk_queue_skip_tagset_quiesce(q))
299                         blk_mq_unquiesce_queue(q);
300         }
301         mutex_unlock(&set->tag_list_lock);
302 }
303 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset);
304
305 void blk_mq_wake_waiters(struct request_queue *q)
306 {
307         struct blk_mq_hw_ctx *hctx;
308         unsigned long i;
309
310         queue_for_each_hw_ctx(q, hctx, i)
311                 if (blk_mq_hw_queue_mapped(hctx))
312                         blk_mq_tag_wakeup_all(hctx->tags, true);
313 }
314
315 void blk_rq_init(struct request_queue *q, struct request *rq)
316 {
317         memset(rq, 0, sizeof(*rq));
318
319         INIT_LIST_HEAD(&rq->queuelist);
320         rq->q = q;
321         rq->__sector = (sector_t) -1;
322         INIT_HLIST_NODE(&rq->hash);
323         RB_CLEAR_NODE(&rq->rb_node);
324         rq->tag = BLK_MQ_NO_TAG;
325         rq->internal_tag = BLK_MQ_NO_TAG;
326         rq->start_time_ns = ktime_get_ns();
327         rq->part = NULL;
328         blk_crypto_rq_set_defaults(rq);
329 }
330 EXPORT_SYMBOL(blk_rq_init);
331
332 /* Set start and alloc time when the allocated request is actually used */
333 static inline void blk_mq_rq_time_init(struct request *rq, u64 alloc_time_ns)
334 {
335         if (blk_mq_need_time_stamp(rq))
336                 rq->start_time_ns = ktime_get_ns();
337         else
338                 rq->start_time_ns = 0;
339
340 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
341         if (blk_queue_rq_alloc_time(rq->q))
342                 rq->alloc_time_ns = alloc_time_ns ?: rq->start_time_ns;
343         else
344                 rq->alloc_time_ns = 0;
345 #endif
346 }
347
348 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
349                 struct blk_mq_tags *tags, unsigned int tag)
350 {
351         struct blk_mq_ctx *ctx = data->ctx;
352         struct blk_mq_hw_ctx *hctx = data->hctx;
353         struct request_queue *q = data->q;
354         struct request *rq = tags->static_rqs[tag];
355
356         rq->q = q;
357         rq->mq_ctx = ctx;
358         rq->mq_hctx = hctx;
359         rq->cmd_flags = data->cmd_flags;
360
361         if (data->flags & BLK_MQ_REQ_PM)
362                 data->rq_flags |= RQF_PM;
363         if (blk_queue_io_stat(q))
364                 data->rq_flags |= RQF_IO_STAT;
365         rq->rq_flags = data->rq_flags;
366
367         if (data->rq_flags & RQF_SCHED_TAGS) {
368                 rq->tag = BLK_MQ_NO_TAG;
369                 rq->internal_tag = tag;
370         } else {
371                 rq->tag = tag;
372                 rq->internal_tag = BLK_MQ_NO_TAG;
373         }
374         rq->timeout = 0;
375
376         rq->part = NULL;
377         rq->io_start_time_ns = 0;
378         rq->stats_sectors = 0;
379         rq->nr_phys_segments = 0;
380 #if defined(CONFIG_BLK_DEV_INTEGRITY)
381         rq->nr_integrity_segments = 0;
382 #endif
383         rq->end_io = NULL;
384         rq->end_io_data = NULL;
385
386         blk_crypto_rq_set_defaults(rq);
387         INIT_LIST_HEAD(&rq->queuelist);
388         /* tag was already set */
389         WRITE_ONCE(rq->deadline, 0);
390         req_ref_set(rq, 1);
391
392         if (rq->rq_flags & RQF_USE_SCHED) {
393                 struct elevator_queue *e = data->q->elevator;
394
395                 INIT_HLIST_NODE(&rq->hash);
396                 RB_CLEAR_NODE(&rq->rb_node);
397
398                 if (e->type->ops.prepare_request)
399                         e->type->ops.prepare_request(rq);
400         }
401
402         return rq;
403 }
404
405 static inline struct request *
406 __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data)
407 {
408         unsigned int tag, tag_offset;
409         struct blk_mq_tags *tags;
410         struct request *rq;
411         unsigned long tag_mask;
412         int i, nr = 0;
413
414         tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset);
415         if (unlikely(!tag_mask))
416                 return NULL;
417
418         tags = blk_mq_tags_from_data(data);
419         for (i = 0; tag_mask; i++) {
420                 if (!(tag_mask & (1UL << i)))
421                         continue;
422                 tag = tag_offset + i;
423                 prefetch(tags->static_rqs[tag]);
424                 tag_mask &= ~(1UL << i);
425                 rq = blk_mq_rq_ctx_init(data, tags, tag);
426                 rq_list_add(data->cached_rq, rq);
427                 nr++;
428         }
429         /* caller already holds a reference, add for remainder */
430         percpu_ref_get_many(&data->q->q_usage_counter, 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 = ktime_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(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, data->cmd_flags, data->ctx);
476         if (!(data->rq_flags & RQF_SCHED_TAGS))
477                 blk_mq_tag_busy(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(3);
510                 goto retry;
511         }
512
513         rq = blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag);
514         blk_mq_rq_time_init(rq, alloc_time_ns);
515         return rq;
516 }
517
518 static struct request *blk_mq_rq_cache_fill(struct request_queue *q,
519                                             struct blk_plug *plug,
520                                             blk_opf_t opf,
521                                             blk_mq_req_flags_t flags)
522 {
523         struct blk_mq_alloc_data data = {
524                 .q              = q,
525                 .flags          = flags,
526                 .cmd_flags      = opf,
527                 .nr_tags        = plug->nr_ios,
528                 .cached_rq      = &plug->cached_rq,
529         };
530         struct request *rq;
531
532         if (blk_queue_enter(q, flags))
533                 return NULL;
534
535         plug->nr_ios = 1;
536
537         rq = __blk_mq_alloc_requests(&data);
538         if (unlikely(!rq))
539                 blk_queue_exit(q);
540         return rq;
541 }
542
543 static struct request *blk_mq_alloc_cached_request(struct request_queue *q,
544                                                    blk_opf_t opf,
545                                                    blk_mq_req_flags_t flags)
546 {
547         struct blk_plug *plug = current->plug;
548         struct request *rq;
549
550         if (!plug)
551                 return NULL;
552
553         if (rq_list_empty(plug->cached_rq)) {
554                 if (plug->nr_ios == 1)
555                         return NULL;
556                 rq = blk_mq_rq_cache_fill(q, plug, opf, flags);
557                 if (!rq)
558                         return NULL;
559         } else {
560                 rq = rq_list_peek(&plug->cached_rq);
561                 if (!rq || rq->q != q)
562                         return NULL;
563
564                 if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type)
565                         return NULL;
566                 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
567                         return NULL;
568
569                 plug->cached_rq = rq_list_next(rq);
570                 blk_mq_rq_time_init(rq, 0);
571         }
572
573         rq->cmd_flags = opf;
574         INIT_LIST_HEAD(&rq->queuelist);
575         return rq;
576 }
577
578 struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf,
579                 blk_mq_req_flags_t flags)
580 {
581         struct request *rq;
582
583         rq = blk_mq_alloc_cached_request(q, opf, flags);
584         if (!rq) {
585                 struct blk_mq_alloc_data data = {
586                         .q              = q,
587                         .flags          = flags,
588                         .cmd_flags      = opf,
589                         .nr_tags        = 1,
590                 };
591                 int ret;
592
593                 ret = blk_queue_enter(q, flags);
594                 if (ret)
595                         return ERR_PTR(ret);
596
597                 rq = __blk_mq_alloc_requests(&data);
598                 if (!rq)
599                         goto out_queue_exit;
600         }
601         rq->__data_len = 0;
602         rq->__sector = (sector_t) -1;
603         rq->bio = rq->biotail = NULL;
604         return rq;
605 out_queue_exit:
606         blk_queue_exit(q);
607         return ERR_PTR(-EWOULDBLOCK);
608 }
609 EXPORT_SYMBOL(blk_mq_alloc_request);
610
611 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
612         blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx)
613 {
614         struct blk_mq_alloc_data data = {
615                 .q              = q,
616                 .flags          = flags,
617                 .cmd_flags      = opf,
618                 .nr_tags        = 1,
619         };
620         u64 alloc_time_ns = 0;
621         struct request *rq;
622         unsigned int cpu;
623         unsigned int tag;
624         int ret;
625
626         /* alloc_time includes depth and tag waits */
627         if (blk_queue_rq_alloc_time(q))
628                 alloc_time_ns = ktime_get_ns();
629
630         /*
631          * If the tag allocator sleeps we could get an allocation for a
632          * different hardware context.  No need to complicate the low level
633          * allocator for this for the rare use case of a command tied to
634          * a specific queue.
635          */
636         if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) ||
637             WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED)))
638                 return ERR_PTR(-EINVAL);
639
640         if (hctx_idx >= q->nr_hw_queues)
641                 return ERR_PTR(-EIO);
642
643         ret = blk_queue_enter(q, flags);
644         if (ret)
645                 return ERR_PTR(ret);
646
647         /*
648          * Check if the hardware context is actually mapped to anything.
649          * If not tell the caller that it should skip this queue.
650          */
651         ret = -EXDEV;
652         data.hctx = xa_load(&q->hctx_table, hctx_idx);
653         if (!blk_mq_hw_queue_mapped(data.hctx))
654                 goto out_queue_exit;
655         cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
656         if (cpu >= nr_cpu_ids)
657                 goto out_queue_exit;
658         data.ctx = __blk_mq_get_ctx(q, cpu);
659
660         if (q->elevator)
661                 data.rq_flags |= RQF_SCHED_TAGS;
662         else
663                 blk_mq_tag_busy(data.hctx);
664
665         if (flags & BLK_MQ_REQ_RESERVED)
666                 data.rq_flags |= RQF_RESV;
667
668         ret = -EWOULDBLOCK;
669         tag = blk_mq_get_tag(&data);
670         if (tag == BLK_MQ_NO_TAG)
671                 goto out_queue_exit;
672         rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag);
673         blk_mq_rq_time_init(rq, alloc_time_ns);
674         rq->__data_len = 0;
675         rq->__sector = (sector_t) -1;
676         rq->bio = rq->biotail = NULL;
677         return rq;
678
679 out_queue_exit:
680         blk_queue_exit(q);
681         return ERR_PTR(ret);
682 }
683 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
684
685 static void blk_mq_finish_request(struct request *rq)
686 {
687         struct request_queue *q = rq->q;
688
689         if (rq->rq_flags & RQF_USE_SCHED) {
690                 q->elevator->type->ops.finish_request(rq);
691                 /*
692                  * For postflush request that may need to be
693                  * completed twice, we should clear this flag
694                  * to avoid double finish_request() on the rq.
695                  */
696                 rq->rq_flags &= ~RQF_USE_SCHED;
697         }
698 }
699
700 static void __blk_mq_free_request(struct request *rq)
701 {
702         struct request_queue *q = rq->q;
703         struct blk_mq_ctx *ctx = rq->mq_ctx;
704         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
705         const int sched_tag = rq->internal_tag;
706
707         blk_crypto_free_request(rq);
708         blk_pm_mark_last_busy(rq);
709         rq->mq_hctx = NULL;
710
711         if (rq->rq_flags & RQF_MQ_INFLIGHT)
712                 __blk_mq_dec_active_requests(hctx);
713
714         if (rq->tag != BLK_MQ_NO_TAG)
715                 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
716         if (sched_tag != BLK_MQ_NO_TAG)
717                 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
718         blk_mq_sched_restart(hctx);
719         blk_queue_exit(q);
720 }
721
722 void blk_mq_free_request(struct request *rq)
723 {
724         struct request_queue *q = rq->q;
725
726         blk_mq_finish_request(rq);
727
728         if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
729                 laptop_io_completion(q->disk->bdi);
730
731         rq_qos_done(q, rq);
732
733         WRITE_ONCE(rq->state, MQ_RQ_IDLE);
734         if (req_ref_put_and_test(rq))
735                 __blk_mq_free_request(rq);
736 }
737 EXPORT_SYMBOL_GPL(blk_mq_free_request);
738
739 void blk_mq_free_plug_rqs(struct blk_plug *plug)
740 {
741         struct request *rq;
742
743         while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
744                 blk_mq_free_request(rq);
745 }
746
747 void blk_dump_rq_flags(struct request *rq, char *msg)
748 {
749         printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
750                 rq->q->disk ? rq->q->disk->disk_name : "?",
751                 (__force unsigned long long) rq->cmd_flags);
752
753         printk(KERN_INFO "  sector %llu, nr/cnr %u/%u\n",
754                (unsigned long long)blk_rq_pos(rq),
755                blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
756         printk(KERN_INFO "  bio %p, biotail %p, len %u\n",
757                rq->bio, rq->biotail, blk_rq_bytes(rq));
758 }
759 EXPORT_SYMBOL(blk_dump_rq_flags);
760
761 static void req_bio_endio(struct request *rq, struct bio *bio,
762                           unsigned int nbytes, blk_status_t error)
763 {
764         if (unlikely(error)) {
765                 bio->bi_status = error;
766         } else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
767                 /*
768                  * Partial zone append completions cannot be supported as the
769                  * BIO fragments may end up not being written sequentially.
770                  */
771                 if (bio->bi_iter.bi_size != nbytes)
772                         bio->bi_status = BLK_STS_IOERR;
773                 else
774                         bio->bi_iter.bi_sector = rq->__sector;
775         }
776
777         bio_advance(bio, nbytes);
778
779         if (unlikely(rq->rq_flags & RQF_QUIET))
780                 bio_set_flag(bio, BIO_QUIET);
781         /* don't actually finish bio if it's part of flush sequence */
782         if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
783                 bio_endio(bio);
784 }
785
786 static void blk_account_io_completion(struct request *req, unsigned int bytes)
787 {
788         if (req->part && blk_do_io_stat(req)) {
789                 const int sgrp = op_stat_group(req_op(req));
790
791                 part_stat_lock();
792                 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
793                 part_stat_unlock();
794         }
795 }
796
797 static void blk_print_req_error(struct request *req, blk_status_t status)
798 {
799         printk_ratelimited(KERN_ERR
800                 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
801                 "phys_seg %u prio class %u\n",
802                 blk_status_to_str(status),
803                 req->q->disk ? req->q->disk->disk_name : "?",
804                 blk_rq_pos(req), (__force u32)req_op(req),
805                 blk_op_str(req_op(req)),
806                 (__force u32)(req->cmd_flags & ~REQ_OP_MASK),
807                 req->nr_phys_segments,
808                 IOPRIO_PRIO_CLASS(req->ioprio));
809 }
810
811 /*
812  * Fully end IO on a request. Does not support partial completions, or
813  * errors.
814  */
815 static void blk_complete_request(struct request *req)
816 {
817         const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
818         int total_bytes = blk_rq_bytes(req);
819         struct bio *bio = req->bio;
820
821         trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
822
823         if (!bio)
824                 return;
825
826 #ifdef CONFIG_BLK_DEV_INTEGRITY
827         if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
828                 req->q->integrity.profile->complete_fn(req, total_bytes);
829 #endif
830
831         /*
832          * Upper layers may call blk_crypto_evict_key() anytime after the last
833          * bio_endio().  Therefore, the keyslot must be released before that.
834          */
835         blk_crypto_rq_put_keyslot(req);
836
837         blk_account_io_completion(req, total_bytes);
838
839         do {
840                 struct bio *next = bio->bi_next;
841
842                 /* Completion has already been traced */
843                 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
844
845                 if (req_op(req) == REQ_OP_ZONE_APPEND)
846                         bio->bi_iter.bi_sector = req->__sector;
847
848                 if (!is_flush)
849                         bio_endio(bio);
850                 bio = next;
851         } while (bio);
852
853         /*
854          * Reset counters so that the request stacking driver
855          * can find how many bytes remain in the request
856          * later.
857          */
858         if (!req->end_io) {
859                 req->bio = NULL;
860                 req->__data_len = 0;
861         }
862 }
863
864 /**
865  * blk_update_request - Complete multiple bytes without completing the request
866  * @req:      the request being processed
867  * @error:    block status code
868  * @nr_bytes: number of bytes to complete for @req
869  *
870  * Description:
871  *     Ends I/O on a number of bytes attached to @req, but doesn't complete
872  *     the request structure even if @req doesn't have leftover.
873  *     If @req has leftover, sets it up for the next range of segments.
874  *
875  *     Passing the result of blk_rq_bytes() as @nr_bytes guarantees
876  *     %false return from this function.
877  *
878  * Note:
879  *      The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
880  *      except in the consistency check at the end of this function.
881  *
882  * Return:
883  *     %false - this request doesn't have any more data
884  *     %true  - this request has more data
885  **/
886 bool blk_update_request(struct request *req, blk_status_t error,
887                 unsigned int nr_bytes)
888 {
889         int total_bytes;
890
891         trace_block_rq_complete(req, error, nr_bytes);
892
893         if (!req->bio)
894                 return false;
895
896 #ifdef CONFIG_BLK_DEV_INTEGRITY
897         if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
898             error == BLK_STS_OK)
899                 req->q->integrity.profile->complete_fn(req, nr_bytes);
900 #endif
901
902         /*
903          * Upper layers may call blk_crypto_evict_key() anytime after the last
904          * bio_endio().  Therefore, the keyslot must be released before that.
905          */
906         if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req))
907                 __blk_crypto_rq_put_keyslot(req);
908
909         if (unlikely(error && !blk_rq_is_passthrough(req) &&
910                      !(req->rq_flags & RQF_QUIET)) &&
911                      !test_bit(GD_DEAD, &req->q->disk->state)) {
912                 blk_print_req_error(req, error);
913                 trace_block_rq_error(req, error, nr_bytes);
914         }
915
916         blk_account_io_completion(req, nr_bytes);
917
918         total_bytes = 0;
919         while (req->bio) {
920                 struct bio *bio = req->bio;
921                 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
922
923                 if (bio_bytes == bio->bi_iter.bi_size)
924                         req->bio = bio->bi_next;
925
926                 /* Completion has already been traced */
927                 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
928                 req_bio_endio(req, bio, bio_bytes, error);
929
930                 total_bytes += bio_bytes;
931                 nr_bytes -= bio_bytes;
932
933                 if (!nr_bytes)
934                         break;
935         }
936
937         /*
938          * completely done
939          */
940         if (!req->bio) {
941                 /*
942                  * Reset counters so that the request stacking driver
943                  * can find how many bytes remain in the request
944                  * later.
945                  */
946                 req->__data_len = 0;
947                 return false;
948         }
949
950         req->__data_len -= total_bytes;
951
952         /* update sector only for requests with clear definition of sector */
953         if (!blk_rq_is_passthrough(req))
954                 req->__sector += total_bytes >> 9;
955
956         /* mixed attributes always follow the first bio */
957         if (req->rq_flags & RQF_MIXED_MERGE) {
958                 req->cmd_flags &= ~REQ_FAILFAST_MASK;
959                 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
960         }
961
962         if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
963                 /*
964                  * If total number of sectors is less than the first segment
965                  * size, something has gone terribly wrong.
966                  */
967                 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
968                         blk_dump_rq_flags(req, "request botched");
969                         req->__data_len = blk_rq_cur_bytes(req);
970                 }
971
972                 /* recalculate the number of segments */
973                 req->nr_phys_segments = blk_recalc_rq_segments(req);
974         }
975
976         return true;
977 }
978 EXPORT_SYMBOL_GPL(blk_update_request);
979
980 static inline void blk_account_io_done(struct request *req, u64 now)
981 {
982         trace_block_io_done(req);
983
984         /*
985          * Account IO completion.  flush_rq isn't accounted as a
986          * normal IO on queueing nor completion.  Accounting the
987          * containing request is enough.
988          */
989         if (blk_do_io_stat(req) && req->part &&
990             !(req->rq_flags & RQF_FLUSH_SEQ)) {
991                 const int sgrp = op_stat_group(req_op(req));
992
993                 part_stat_lock();
994                 update_io_ticks(req->part, jiffies, true);
995                 part_stat_inc(req->part, ios[sgrp]);
996                 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
997                 part_stat_unlock();
998         }
999 }
1000
1001 static inline void blk_account_io_start(struct request *req)
1002 {
1003         trace_block_io_start(req);
1004
1005         if (blk_do_io_stat(req)) {
1006                 /*
1007                  * All non-passthrough requests are created from a bio with one
1008                  * exception: when a flush command that is part of a flush sequence
1009                  * generated by the state machine in blk-flush.c is cloned onto the
1010                  * lower device by dm-multipath we can get here without a bio.
1011                  */
1012                 if (req->bio)
1013                         req->part = req->bio->bi_bdev;
1014                 else
1015                         req->part = req->q->disk->part0;
1016
1017                 part_stat_lock();
1018                 update_io_ticks(req->part, jiffies, false);
1019                 part_stat_unlock();
1020         }
1021 }
1022
1023 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
1024 {
1025         if (rq->rq_flags & RQF_STATS)
1026                 blk_stat_add(rq, now);
1027
1028         blk_mq_sched_completed_request(rq, now);
1029         blk_account_io_done(rq, now);
1030 }
1031
1032 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
1033 {
1034         if (blk_mq_need_time_stamp(rq))
1035                 __blk_mq_end_request_acct(rq, ktime_get_ns());
1036
1037         blk_mq_finish_request(rq);
1038
1039         if (rq->end_io) {
1040                 rq_qos_done(rq->q, rq);
1041                 if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1042                         blk_mq_free_request(rq);
1043         } else {
1044                 blk_mq_free_request(rq);
1045         }
1046 }
1047 EXPORT_SYMBOL(__blk_mq_end_request);
1048
1049 void blk_mq_end_request(struct request *rq, blk_status_t error)
1050 {
1051         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1052                 BUG();
1053         __blk_mq_end_request(rq, error);
1054 }
1055 EXPORT_SYMBOL(blk_mq_end_request);
1056
1057 #define TAG_COMP_BATCH          32
1058
1059 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1060                                           int *tag_array, int nr_tags)
1061 {
1062         struct request_queue *q = hctx->queue;
1063
1064         /*
1065          * All requests should have been marked as RQF_MQ_INFLIGHT, so
1066          * update hctx->nr_active in batch
1067          */
1068         if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
1069                 __blk_mq_sub_active_requests(hctx, nr_tags);
1070
1071         blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
1072         percpu_ref_put_many(&q->q_usage_counter, nr_tags);
1073 }
1074
1075 void blk_mq_end_request_batch(struct io_comp_batch *iob)
1076 {
1077         int tags[TAG_COMP_BATCH], nr_tags = 0;
1078         struct blk_mq_hw_ctx *cur_hctx = NULL;
1079         struct request *rq;
1080         u64 now = 0;
1081
1082         if (iob->need_ts)
1083                 now = ktime_get_ns();
1084
1085         while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1086                 prefetch(rq->bio);
1087                 prefetch(rq->rq_next);
1088
1089                 blk_complete_request(rq);
1090                 if (iob->need_ts)
1091                         __blk_mq_end_request_acct(rq, now);
1092
1093                 blk_mq_finish_request(rq);
1094
1095                 rq_qos_done(rq->q, rq);
1096
1097                 /*
1098                  * If end_io handler returns NONE, then it still has
1099                  * ownership of the request.
1100                  */
1101                 if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
1102                         continue;
1103
1104                 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1105                 if (!req_ref_put_and_test(rq))
1106                         continue;
1107
1108                 blk_crypto_free_request(rq);
1109                 blk_pm_mark_last_busy(rq);
1110
1111                 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1112                         if (cur_hctx)
1113                                 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1114                         nr_tags = 0;
1115                         cur_hctx = rq->mq_hctx;
1116                 }
1117                 tags[nr_tags++] = rq->tag;
1118         }
1119
1120         if (nr_tags)
1121                 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1122 }
1123 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1124
1125 static void blk_complete_reqs(struct llist_head *list)
1126 {
1127         struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1128         struct request *rq, *next;
1129
1130         llist_for_each_entry_safe(rq, next, entry, ipi_list)
1131                 rq->q->mq_ops->complete(rq);
1132 }
1133
1134 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1135 {
1136         blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1137 }
1138
1139 static int blk_softirq_cpu_dead(unsigned int cpu)
1140 {
1141         blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1142         return 0;
1143 }
1144
1145 static void __blk_mq_complete_request_remote(void *data)
1146 {
1147         __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1148 }
1149
1150 static inline bool blk_mq_complete_need_ipi(struct request *rq)
1151 {
1152         int cpu = raw_smp_processor_id();
1153
1154         if (!IS_ENABLED(CONFIG_SMP) ||
1155             !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1156                 return false;
1157         /*
1158          * With force threaded interrupts enabled, raising softirq from an SMP
1159          * function call will always result in waking the ksoftirqd thread.
1160          * This is probably worse than completing the request on a different
1161          * cache domain.
1162          */
1163         if (force_irqthreads())
1164                 return false;
1165
1166         /* same CPU or cache domain?  Complete locally */
1167         if (cpu == rq->mq_ctx->cpu ||
1168             (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1169              cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1170                 return false;
1171
1172         /* don't try to IPI to an offline CPU */
1173         return cpu_online(rq->mq_ctx->cpu);
1174 }
1175
1176 static void blk_mq_complete_send_ipi(struct request *rq)
1177 {
1178         unsigned int cpu;
1179
1180         cpu = rq->mq_ctx->cpu;
1181         if (llist_add(&rq->ipi_list, &per_cpu(blk_cpu_done, cpu)))
1182                 smp_call_function_single_async(cpu, &per_cpu(blk_cpu_csd, cpu));
1183 }
1184
1185 static void blk_mq_raise_softirq(struct request *rq)
1186 {
1187         struct llist_head *list;
1188
1189         preempt_disable();
1190         list = this_cpu_ptr(&blk_cpu_done);
1191         if (llist_add(&rq->ipi_list, list))
1192                 raise_softirq(BLOCK_SOFTIRQ);
1193         preempt_enable();
1194 }
1195
1196 bool blk_mq_complete_request_remote(struct request *rq)
1197 {
1198         WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1199
1200         /*
1201          * For request which hctx has only one ctx mapping,
1202          * or a polled request, always complete locally,
1203          * it's pointless to redirect the completion.
1204          */
1205         if ((rq->mq_hctx->nr_ctx == 1 &&
1206              rq->mq_ctx->cpu == raw_smp_processor_id()) ||
1207              rq->cmd_flags & REQ_POLLED)
1208                 return false;
1209
1210         if (blk_mq_complete_need_ipi(rq)) {
1211                 blk_mq_complete_send_ipi(rq);
1212                 return true;
1213         }
1214
1215         if (rq->q->nr_hw_queues == 1) {
1216                 blk_mq_raise_softirq(rq);
1217                 return true;
1218         }
1219         return false;
1220 }
1221 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1222
1223 /**
1224  * blk_mq_complete_request - end I/O on a request
1225  * @rq:         the request being processed
1226  *
1227  * Description:
1228  *      Complete a request by scheduling the ->complete_rq operation.
1229  **/
1230 void blk_mq_complete_request(struct request *rq)
1231 {
1232         if (!blk_mq_complete_request_remote(rq))
1233                 rq->q->mq_ops->complete(rq);
1234 }
1235 EXPORT_SYMBOL(blk_mq_complete_request);
1236
1237 /**
1238  * blk_mq_start_request - Start processing a request
1239  * @rq: Pointer to request to be started
1240  *
1241  * Function used by device drivers to notify the block layer that a request
1242  * is going to be processed now, so blk layer can do proper initializations
1243  * such as starting the timeout timer.
1244  */
1245 void blk_mq_start_request(struct request *rq)
1246 {
1247         struct request_queue *q = rq->q;
1248
1249         trace_block_rq_issue(rq);
1250
1251         if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1252                 rq->io_start_time_ns = ktime_get_ns();
1253                 rq->stats_sectors = blk_rq_sectors(rq);
1254                 rq->rq_flags |= RQF_STATS;
1255                 rq_qos_issue(q, rq);
1256         }
1257
1258         WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1259
1260         blk_add_timer(rq);
1261         WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1262
1263 #ifdef CONFIG_BLK_DEV_INTEGRITY
1264         if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1265                 q->integrity.profile->prepare_fn(rq);
1266 #endif
1267         if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1268                 WRITE_ONCE(rq->bio->bi_cookie, rq->mq_hctx->queue_num);
1269 }
1270 EXPORT_SYMBOL(blk_mq_start_request);
1271
1272 /*
1273  * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1274  * queues. This is important for md arrays to benefit from merging
1275  * requests.
1276  */
1277 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1278 {
1279         if (plug->multiple_queues)
1280                 return BLK_MAX_REQUEST_COUNT * 2;
1281         return BLK_MAX_REQUEST_COUNT;
1282 }
1283
1284 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1285 {
1286         struct request *last = rq_list_peek(&plug->mq_list);
1287
1288         if (!plug->rq_count) {
1289                 trace_block_plug(rq->q);
1290         } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1291                    (!blk_queue_nomerges(rq->q) &&
1292                     blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1293                 blk_mq_flush_plug_list(plug, false);
1294                 last = NULL;
1295                 trace_block_plug(rq->q);
1296         }
1297
1298         if (!plug->multiple_queues && last && last->q != rq->q)
1299                 plug->multiple_queues = true;
1300         /*
1301          * Any request allocated from sched tags can't be issued to
1302          * ->queue_rqs() directly
1303          */
1304         if (!plug->has_elevator && (rq->rq_flags & RQF_SCHED_TAGS))
1305                 plug->has_elevator = true;
1306         rq->rq_next = NULL;
1307         rq_list_add(&plug->mq_list, rq);
1308         plug->rq_count++;
1309 }
1310
1311 /**
1312  * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1313  * @rq:         request to insert
1314  * @at_head:    insert request at head or tail of queue
1315  *
1316  * Description:
1317  *    Insert a fully prepared request at the back of the I/O scheduler queue
1318  *    for execution.  Don't wait for completion.
1319  *
1320  * Note:
1321  *    This function will invoke @done directly if the queue is dead.
1322  */
1323 void blk_execute_rq_nowait(struct request *rq, bool at_head)
1324 {
1325         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1326
1327         WARN_ON(irqs_disabled());
1328         WARN_ON(!blk_rq_is_passthrough(rq));
1329
1330         blk_account_io_start(rq);
1331
1332         /*
1333          * As plugging can be enabled for passthrough requests on a zoned
1334          * device, directly accessing the plug instead of using blk_mq_plug()
1335          * should not have any consequences.
1336          */
1337         if (current->plug && !at_head) {
1338                 blk_add_rq_to_plug(current->plug, rq);
1339                 return;
1340         }
1341
1342         blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1343         blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);
1344 }
1345 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1346
1347 struct blk_rq_wait {
1348         struct completion done;
1349         blk_status_t ret;
1350 };
1351
1352 static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
1353 {
1354         struct blk_rq_wait *wait = rq->end_io_data;
1355
1356         wait->ret = ret;
1357         complete(&wait->done);
1358         return RQ_END_IO_NONE;
1359 }
1360
1361 bool blk_rq_is_poll(struct request *rq)
1362 {
1363         if (!rq->mq_hctx)
1364                 return false;
1365         if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1366                 return false;
1367         return true;
1368 }
1369 EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1370
1371 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1372 {
1373         do {
1374                 blk_hctx_poll(rq->q, rq->mq_hctx, NULL, 0);
1375                 cond_resched();
1376         } while (!completion_done(wait));
1377 }
1378
1379 /**
1380  * blk_execute_rq - insert a request into queue for execution
1381  * @rq:         request to insert
1382  * @at_head:    insert request at head or tail of queue
1383  *
1384  * Description:
1385  *    Insert a fully prepared request at the back of the I/O scheduler queue
1386  *    for execution and wait for completion.
1387  * Return: The blk_status_t result provided to blk_mq_end_request().
1388  */
1389 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1390 {
1391         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1392         struct blk_rq_wait wait = {
1393                 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1394         };
1395
1396         WARN_ON(irqs_disabled());
1397         WARN_ON(!blk_rq_is_passthrough(rq));
1398
1399         rq->end_io_data = &wait;
1400         rq->end_io = blk_end_sync_rq;
1401
1402         blk_account_io_start(rq);
1403         blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1404         blk_mq_run_hw_queue(hctx, false);
1405
1406         if (blk_rq_is_poll(rq)) {
1407                 blk_rq_poll_completion(rq, &wait.done);
1408         } else {
1409                 /*
1410                  * Prevent hang_check timer from firing at us during very long
1411                  * I/O
1412                  */
1413                 unsigned long hang_check = sysctl_hung_task_timeout_secs;
1414
1415                 if (hang_check)
1416                         while (!wait_for_completion_io_timeout(&wait.done,
1417                                         hang_check * (HZ/2)))
1418                                 ;
1419                 else
1420                         wait_for_completion_io(&wait.done);
1421         }
1422
1423         return wait.ret;
1424 }
1425 EXPORT_SYMBOL(blk_execute_rq);
1426
1427 static void __blk_mq_requeue_request(struct request *rq)
1428 {
1429         struct request_queue *q = rq->q;
1430
1431         blk_mq_put_driver_tag(rq);
1432
1433         trace_block_rq_requeue(rq);
1434         rq_qos_requeue(q, rq);
1435
1436         if (blk_mq_request_started(rq)) {
1437                 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1438                 rq->rq_flags &= ~RQF_TIMED_OUT;
1439         }
1440 }
1441
1442 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1443 {
1444         struct request_queue *q = rq->q;
1445         unsigned long flags;
1446
1447         __blk_mq_requeue_request(rq);
1448
1449         /* this request will be re-inserted to io scheduler queue */
1450         blk_mq_sched_requeue_request(rq);
1451
1452         spin_lock_irqsave(&q->requeue_lock, flags);
1453         list_add_tail(&rq->queuelist, &q->requeue_list);
1454         spin_unlock_irqrestore(&q->requeue_lock, flags);
1455
1456         if (kick_requeue_list)
1457                 blk_mq_kick_requeue_list(q);
1458 }
1459 EXPORT_SYMBOL(blk_mq_requeue_request);
1460
1461 static void blk_mq_requeue_work(struct work_struct *work)
1462 {
1463         struct request_queue *q =
1464                 container_of(work, struct request_queue, requeue_work.work);
1465         LIST_HEAD(rq_list);
1466         LIST_HEAD(flush_list);
1467         struct request *rq;
1468
1469         spin_lock_irq(&q->requeue_lock);
1470         list_splice_init(&q->requeue_list, &rq_list);
1471         list_splice_init(&q->flush_list, &flush_list);
1472         spin_unlock_irq(&q->requeue_lock);
1473
1474         while (!list_empty(&rq_list)) {
1475                 rq = list_entry(rq_list.next, struct request, queuelist);
1476                 /*
1477                  * If RQF_DONTPREP ist set, the request has been started by the
1478                  * driver already and might have driver-specific data allocated
1479                  * already.  Insert it into the hctx dispatch list to avoid
1480                  * block layer merges for the request.
1481                  */
1482                 if (rq->rq_flags & RQF_DONTPREP) {
1483                         list_del_init(&rq->queuelist);
1484                         blk_mq_request_bypass_insert(rq, 0);
1485                 } else {
1486                         list_del_init(&rq->queuelist);
1487                         blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD);
1488                 }
1489         }
1490
1491         while (!list_empty(&flush_list)) {
1492                 rq = list_entry(flush_list.next, struct request, queuelist);
1493                 list_del_init(&rq->queuelist);
1494                 blk_mq_insert_request(rq, 0);
1495         }
1496
1497         blk_mq_run_hw_queues(q, false);
1498 }
1499
1500 void blk_mq_kick_requeue_list(struct request_queue *q)
1501 {
1502         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1503 }
1504 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1505
1506 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1507                                     unsigned long msecs)
1508 {
1509         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1510                                     msecs_to_jiffies(msecs));
1511 }
1512 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1513
1514 static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1515 {
1516         /*
1517          * If we find a request that isn't idle we know the queue is busy
1518          * as it's checked in the iter.
1519          * Return false to stop the iteration.
1520          */
1521         if (blk_mq_request_started(rq)) {
1522                 bool *busy = priv;
1523
1524                 *busy = true;
1525                 return false;
1526         }
1527
1528         return true;
1529 }
1530
1531 bool blk_mq_queue_inflight(struct request_queue *q)
1532 {
1533         bool busy = false;
1534
1535         blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1536         return busy;
1537 }
1538 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1539
1540 static void blk_mq_rq_timed_out(struct request *req)
1541 {
1542         req->rq_flags |= RQF_TIMED_OUT;
1543         if (req->q->mq_ops->timeout) {
1544                 enum blk_eh_timer_return ret;
1545
1546                 ret = req->q->mq_ops->timeout(req);
1547                 if (ret == BLK_EH_DONE)
1548                         return;
1549                 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1550         }
1551
1552         blk_add_timer(req);
1553 }
1554
1555 struct blk_expired_data {
1556         bool has_timedout_rq;
1557         unsigned long next;
1558         unsigned long timeout_start;
1559 };
1560
1561 static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
1562 {
1563         unsigned long deadline;
1564
1565         if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1566                 return false;
1567         if (rq->rq_flags & RQF_TIMED_OUT)
1568                 return false;
1569
1570         deadline = READ_ONCE(rq->deadline);
1571         if (time_after_eq(expired->timeout_start, deadline))
1572                 return true;
1573
1574         if (expired->next == 0)
1575                 expired->next = deadline;
1576         else if (time_after(expired->next, deadline))
1577                 expired->next = deadline;
1578         return false;
1579 }
1580
1581 void blk_mq_put_rq_ref(struct request *rq)
1582 {
1583         if (is_flush_rq(rq)) {
1584                 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1585                         blk_mq_free_request(rq);
1586         } else if (req_ref_put_and_test(rq)) {
1587                 __blk_mq_free_request(rq);
1588         }
1589 }
1590
1591 static bool blk_mq_check_expired(struct request *rq, void *priv)
1592 {
1593         struct blk_expired_data *expired = priv;
1594
1595         /*
1596          * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1597          * be reallocated underneath the timeout handler's processing, then
1598          * the expire check is reliable. If the request is not expired, then
1599          * it was completed and reallocated as a new request after returning
1600          * from blk_mq_check_expired().
1601          */
1602         if (blk_mq_req_expired(rq, expired)) {
1603                 expired->has_timedout_rq = true;
1604                 return false;
1605         }
1606         return true;
1607 }
1608
1609 static bool blk_mq_handle_expired(struct request *rq, void *priv)
1610 {
1611         struct blk_expired_data *expired = priv;
1612
1613         if (blk_mq_req_expired(rq, expired))
1614                 blk_mq_rq_timed_out(rq);
1615         return true;
1616 }
1617
1618 static void blk_mq_timeout_work(struct work_struct *work)
1619 {
1620         struct request_queue *q =
1621                 container_of(work, struct request_queue, timeout_work);
1622         struct blk_expired_data expired = {
1623                 .timeout_start = jiffies,
1624         };
1625         struct blk_mq_hw_ctx *hctx;
1626         unsigned long i;
1627
1628         /* A deadlock might occur if a request is stuck requiring a
1629          * timeout at the same time a queue freeze is waiting
1630          * completion, since the timeout code would not be able to
1631          * acquire the queue reference here.
1632          *
1633          * That's why we don't use blk_queue_enter here; instead, we use
1634          * percpu_ref_tryget directly, because we need to be able to
1635          * obtain a reference even in the short window between the queue
1636          * starting to freeze, by dropping the first reference in
1637          * blk_freeze_queue_start, and the moment the last request is
1638          * consumed, marked by the instant q_usage_counter reaches
1639          * zero.
1640          */
1641         if (!percpu_ref_tryget(&q->q_usage_counter))
1642                 return;
1643
1644         /* check if there is any timed-out request */
1645         blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
1646         if (expired.has_timedout_rq) {
1647                 /*
1648                  * Before walking tags, we must ensure any submit started
1649                  * before the current time has finished. Since the submit
1650                  * uses srcu or rcu, wait for a synchronization point to
1651                  * ensure all running submits have finished
1652                  */
1653                 blk_mq_wait_quiesce_done(q->tag_set);
1654
1655                 expired.next = 0;
1656                 blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
1657         }
1658
1659         if (expired.next != 0) {
1660                 mod_timer(&q->timeout, expired.next);
1661         } else {
1662                 /*
1663                  * Request timeouts are handled as a forward rolling timer. If
1664                  * we end up here it means that no requests are pending and
1665                  * also that no request has been pending for a while. Mark
1666                  * each hctx as idle.
1667                  */
1668                 queue_for_each_hw_ctx(q, hctx, i) {
1669                         /* the hctx may be unmapped, so check it here */
1670                         if (blk_mq_hw_queue_mapped(hctx))
1671                                 blk_mq_tag_idle(hctx);
1672                 }
1673         }
1674         blk_queue_exit(q);
1675 }
1676
1677 struct flush_busy_ctx_data {
1678         struct blk_mq_hw_ctx *hctx;
1679         struct list_head *list;
1680 };
1681
1682 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1683 {
1684         struct flush_busy_ctx_data *flush_data = data;
1685         struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1686         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1687         enum hctx_type type = hctx->type;
1688
1689         spin_lock(&ctx->lock);
1690         list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1691         sbitmap_clear_bit(sb, bitnr);
1692         spin_unlock(&ctx->lock);
1693         return true;
1694 }
1695
1696 /*
1697  * Process software queues that have been marked busy, splicing them
1698  * to the for-dispatch
1699  */
1700 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1701 {
1702         struct flush_busy_ctx_data data = {
1703                 .hctx = hctx,
1704                 .list = list,
1705         };
1706
1707         sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1708 }
1709 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1710
1711 struct dispatch_rq_data {
1712         struct blk_mq_hw_ctx *hctx;
1713         struct request *rq;
1714 };
1715
1716 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1717                 void *data)
1718 {
1719         struct dispatch_rq_data *dispatch_data = data;
1720         struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1721         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1722         enum hctx_type type = hctx->type;
1723
1724         spin_lock(&ctx->lock);
1725         if (!list_empty(&ctx->rq_lists[type])) {
1726                 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1727                 list_del_init(&dispatch_data->rq->queuelist);
1728                 if (list_empty(&ctx->rq_lists[type]))
1729                         sbitmap_clear_bit(sb, bitnr);
1730         }
1731         spin_unlock(&ctx->lock);
1732
1733         return !dispatch_data->rq;
1734 }
1735
1736 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1737                                         struct blk_mq_ctx *start)
1738 {
1739         unsigned off = start ? start->index_hw[hctx->type] : 0;
1740         struct dispatch_rq_data data = {
1741                 .hctx = hctx,
1742                 .rq   = NULL,
1743         };
1744
1745         __sbitmap_for_each_set(&hctx->ctx_map, off,
1746                                dispatch_rq_from_ctx, &data);
1747
1748         return data.rq;
1749 }
1750
1751 static bool __blk_mq_alloc_driver_tag(struct request *rq)
1752 {
1753         struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1754         unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1755         int tag;
1756
1757         blk_mq_tag_busy(rq->mq_hctx);
1758
1759         if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1760                 bt = &rq->mq_hctx->tags->breserved_tags;
1761                 tag_offset = 0;
1762         } else {
1763                 if (!hctx_may_queue(rq->mq_hctx, bt))
1764                         return false;
1765         }
1766
1767         tag = __sbitmap_queue_get(bt);
1768         if (tag == BLK_MQ_NO_TAG)
1769                 return false;
1770
1771         rq->tag = tag + tag_offset;
1772         return true;
1773 }
1774
1775 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1776 {
1777         if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1778                 return false;
1779
1780         if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1781                         !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1782                 rq->rq_flags |= RQF_MQ_INFLIGHT;
1783                 __blk_mq_inc_active_requests(hctx);
1784         }
1785         hctx->tags->rqs[rq->tag] = rq;
1786         return true;
1787 }
1788
1789 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1790                                 int flags, void *key)
1791 {
1792         struct blk_mq_hw_ctx *hctx;
1793
1794         hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1795
1796         spin_lock(&hctx->dispatch_wait_lock);
1797         if (!list_empty(&wait->entry)) {
1798                 struct sbitmap_queue *sbq;
1799
1800                 list_del_init(&wait->entry);
1801                 sbq = &hctx->tags->bitmap_tags;
1802                 atomic_dec(&sbq->ws_active);
1803         }
1804         spin_unlock(&hctx->dispatch_wait_lock);
1805
1806         blk_mq_run_hw_queue(hctx, true);
1807         return 1;
1808 }
1809
1810 /*
1811  * Mark us waiting for a tag. For shared tags, this involves hooking us into
1812  * the tag wakeups. For non-shared tags, we can simply mark us needing a
1813  * restart. For both cases, take care to check the condition again after
1814  * marking us as waiting.
1815  */
1816 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1817                                  struct request *rq)
1818 {
1819         struct sbitmap_queue *sbq;
1820         struct wait_queue_head *wq;
1821         wait_queue_entry_t *wait;
1822         bool ret;
1823
1824         if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1825             !(blk_mq_is_shared_tags(hctx->flags))) {
1826                 blk_mq_sched_mark_restart_hctx(hctx);
1827
1828                 /*
1829                  * It's possible that a tag was freed in the window between the
1830                  * allocation failure and adding the hardware queue to the wait
1831                  * queue.
1832                  *
1833                  * Don't clear RESTART here, someone else could have set it.
1834                  * At most this will cost an extra queue run.
1835                  */
1836                 return blk_mq_get_driver_tag(rq);
1837         }
1838
1839         wait = &hctx->dispatch_wait;
1840         if (!list_empty_careful(&wait->entry))
1841                 return false;
1842
1843         if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag))
1844                 sbq = &hctx->tags->breserved_tags;
1845         else
1846                 sbq = &hctx->tags->bitmap_tags;
1847         wq = &bt_wait_ptr(sbq, hctx)->wait;
1848
1849         spin_lock_irq(&wq->lock);
1850         spin_lock(&hctx->dispatch_wait_lock);
1851         if (!list_empty(&wait->entry)) {
1852                 spin_unlock(&hctx->dispatch_wait_lock);
1853                 spin_unlock_irq(&wq->lock);
1854                 return false;
1855         }
1856
1857         atomic_inc(&sbq->ws_active);
1858         wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1859         __add_wait_queue(wq, wait);
1860
1861         /*
1862          * It's possible that a tag was freed in the window between the
1863          * allocation failure and adding the hardware queue to the wait
1864          * queue.
1865          */
1866         ret = blk_mq_get_driver_tag(rq);
1867         if (!ret) {
1868                 spin_unlock(&hctx->dispatch_wait_lock);
1869                 spin_unlock_irq(&wq->lock);
1870                 return false;
1871         }
1872
1873         /*
1874          * We got a tag, remove ourselves from the wait queue to ensure
1875          * someone else gets the wakeup.
1876          */
1877         list_del_init(&wait->entry);
1878         atomic_dec(&sbq->ws_active);
1879         spin_unlock(&hctx->dispatch_wait_lock);
1880         spin_unlock_irq(&wq->lock);
1881
1882         return true;
1883 }
1884
1885 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT  8
1886 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR  4
1887 /*
1888  * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1889  * - EWMA is one simple way to compute running average value
1890  * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1891  * - take 4 as factor for avoiding to get too small(0) result, and this
1892  *   factor doesn't matter because EWMA decreases exponentially
1893  */
1894 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1895 {
1896         unsigned int ewma;
1897
1898         ewma = hctx->dispatch_busy;
1899
1900         if (!ewma && !busy)
1901                 return;
1902
1903         ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1904         if (busy)
1905                 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1906         ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1907
1908         hctx->dispatch_busy = ewma;
1909 }
1910
1911 #define BLK_MQ_RESOURCE_DELAY   3               /* ms units */
1912
1913 static void blk_mq_handle_dev_resource(struct request *rq,
1914                                        struct list_head *list)
1915 {
1916         list_add(&rq->queuelist, list);
1917         __blk_mq_requeue_request(rq);
1918 }
1919
1920 static void blk_mq_handle_zone_resource(struct request *rq,
1921                                         struct list_head *zone_list)
1922 {
1923         /*
1924          * If we end up here it is because we cannot dispatch a request to a
1925          * specific zone due to LLD level zone-write locking or other zone
1926          * related resource not being available. In this case, set the request
1927          * aside in zone_list for retrying it later.
1928          */
1929         list_add(&rq->queuelist, zone_list);
1930         __blk_mq_requeue_request(rq);
1931 }
1932
1933 enum prep_dispatch {
1934         PREP_DISPATCH_OK,
1935         PREP_DISPATCH_NO_TAG,
1936         PREP_DISPATCH_NO_BUDGET,
1937 };
1938
1939 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1940                                                   bool need_budget)
1941 {
1942         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1943         int budget_token = -1;
1944
1945         if (need_budget) {
1946                 budget_token = blk_mq_get_dispatch_budget(rq->q);
1947                 if (budget_token < 0) {
1948                         blk_mq_put_driver_tag(rq);
1949                         return PREP_DISPATCH_NO_BUDGET;
1950                 }
1951                 blk_mq_set_rq_budget_token(rq, budget_token);
1952         }
1953
1954         if (!blk_mq_get_driver_tag(rq)) {
1955                 /*
1956                  * The initial allocation attempt failed, so we need to
1957                  * rerun the hardware queue when a tag is freed. The
1958                  * waitqueue takes care of that. If the queue is run
1959                  * before we add this entry back on the dispatch list,
1960                  * we'll re-run it below.
1961                  */
1962                 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1963                         /*
1964                          * All budgets not got from this function will be put
1965                          * together during handling partial dispatch
1966                          */
1967                         if (need_budget)
1968                                 blk_mq_put_dispatch_budget(rq->q, budget_token);
1969                         return PREP_DISPATCH_NO_TAG;
1970                 }
1971         }
1972
1973         return PREP_DISPATCH_OK;
1974 }
1975
1976 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1977 static void blk_mq_release_budgets(struct request_queue *q,
1978                 struct list_head *list)
1979 {
1980         struct request *rq;
1981
1982         list_for_each_entry(rq, list, queuelist) {
1983                 int budget_token = blk_mq_get_rq_budget_token(rq);
1984
1985                 if (budget_token >= 0)
1986                         blk_mq_put_dispatch_budget(q, budget_token);
1987         }
1988 }
1989
1990 /*
1991  * blk_mq_commit_rqs will notify driver using bd->last that there is no
1992  * more requests. (See comment in struct blk_mq_ops for commit_rqs for
1993  * details)
1994  * Attention, we should explicitly call this in unusual cases:
1995  *  1) did not queue everything initially scheduled to queue
1996  *  2) the last attempt to queue a request failed
1997  */
1998 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued,
1999                               bool from_schedule)
2000 {
2001         if (hctx->queue->mq_ops->commit_rqs && queued) {
2002                 trace_block_unplug(hctx->queue, queued, !from_schedule);
2003                 hctx->queue->mq_ops->commit_rqs(hctx);
2004         }
2005 }
2006
2007 /*
2008  * Returns true if we did some work AND can potentially do more.
2009  */
2010 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
2011                              unsigned int nr_budgets)
2012 {
2013         enum prep_dispatch prep;
2014         struct request_queue *q = hctx->queue;
2015         struct request *rq;
2016         int queued;
2017         blk_status_t ret = BLK_STS_OK;
2018         LIST_HEAD(zone_list);
2019         bool needs_resource = false;
2020
2021         if (list_empty(list))
2022                 return false;
2023
2024         /*
2025          * Now process all the entries, sending them to the driver.
2026          */
2027         queued = 0;
2028         do {
2029                 struct blk_mq_queue_data bd;
2030
2031                 rq = list_first_entry(list, struct request, queuelist);
2032
2033                 WARN_ON_ONCE(hctx != rq->mq_hctx);
2034                 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
2035                 if (prep != PREP_DISPATCH_OK)
2036                         break;
2037
2038                 list_del_init(&rq->queuelist);
2039
2040                 bd.rq = rq;
2041                 bd.last = list_empty(list);
2042
2043                 /*
2044                  * once the request is queued to lld, no need to cover the
2045                  * budget any more
2046                  */
2047                 if (nr_budgets)
2048                         nr_budgets--;
2049                 ret = q->mq_ops->queue_rq(hctx, &bd);
2050                 switch (ret) {
2051                 case BLK_STS_OK:
2052                         queued++;
2053                         break;
2054                 case BLK_STS_RESOURCE:
2055                         needs_resource = true;
2056                         fallthrough;
2057                 case BLK_STS_DEV_RESOURCE:
2058                         blk_mq_handle_dev_resource(rq, list);
2059                         goto out;
2060                 case BLK_STS_ZONE_RESOURCE:
2061                         /*
2062                          * Move the request to zone_list and keep going through
2063                          * the dispatch list to find more requests the drive can
2064                          * accept.
2065                          */
2066                         blk_mq_handle_zone_resource(rq, &zone_list);
2067                         needs_resource = true;
2068                         break;
2069                 default:
2070                         blk_mq_end_request(rq, ret);
2071                 }
2072         } while (!list_empty(list));
2073 out:
2074         if (!list_empty(&zone_list))
2075                 list_splice_tail_init(&zone_list, list);
2076
2077         /* If we didn't flush the entire list, we could have told the driver
2078          * there was more coming, but that turned out to be a lie.
2079          */
2080         if (!list_empty(list) || ret != BLK_STS_OK)
2081                 blk_mq_commit_rqs(hctx, queued, false);
2082
2083         /*
2084          * Any items that need requeuing? Stuff them into hctx->dispatch,
2085          * that is where we will continue on next queue run.
2086          */
2087         if (!list_empty(list)) {
2088                 bool needs_restart;
2089                 /* For non-shared tags, the RESTART check will suffice */
2090                 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2091                         ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
2092                         blk_mq_is_shared_tags(hctx->flags));
2093
2094                 if (nr_budgets)
2095                         blk_mq_release_budgets(q, list);
2096
2097                 spin_lock(&hctx->lock);
2098                 list_splice_tail_init(list, &hctx->dispatch);
2099                 spin_unlock(&hctx->lock);
2100
2101                 /*
2102                  * Order adding requests to hctx->dispatch and checking
2103                  * SCHED_RESTART flag. The pair of this smp_mb() is the one
2104                  * in blk_mq_sched_restart(). Avoid restart code path to
2105                  * miss the new added requests to hctx->dispatch, meantime
2106                  * SCHED_RESTART is observed here.
2107                  */
2108                 smp_mb();
2109
2110                 /*
2111                  * If SCHED_RESTART was set by the caller of this function and
2112                  * it is no longer set that means that it was cleared by another
2113                  * thread and hence that a queue rerun is needed.
2114                  *
2115                  * If 'no_tag' is set, that means that we failed getting
2116                  * a driver tag with an I/O scheduler attached. If our dispatch
2117                  * waitqueue is no longer active, ensure that we run the queue
2118                  * AFTER adding our entries back to the list.
2119                  *
2120                  * If no I/O scheduler has been configured it is possible that
2121                  * the hardware queue got stopped and restarted before requests
2122                  * were pushed back onto the dispatch list. Rerun the queue to
2123                  * avoid starvation. Notes:
2124                  * - blk_mq_run_hw_queue() checks whether or not a queue has
2125                  *   been stopped before rerunning a queue.
2126                  * - Some but not all block drivers stop a queue before
2127                  *   returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2128                  *   and dm-rq.
2129                  *
2130                  * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2131                  * bit is set, run queue after a delay to avoid IO stalls
2132                  * that could otherwise occur if the queue is idle.  We'll do
2133                  * similar if we couldn't get budget or couldn't lock a zone
2134                  * and SCHED_RESTART is set.
2135                  */
2136                 needs_restart = blk_mq_sched_needs_restart(hctx);
2137                 if (prep == PREP_DISPATCH_NO_BUDGET)
2138                         needs_resource = true;
2139                 if (!needs_restart ||
2140                     (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2141                         blk_mq_run_hw_queue(hctx, true);
2142                 else if (needs_resource)
2143                         blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2144
2145                 blk_mq_update_dispatch_busy(hctx, true);
2146                 return false;
2147         }
2148
2149         blk_mq_update_dispatch_busy(hctx, false);
2150         return true;
2151 }
2152
2153 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2154 {
2155         int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2156
2157         if (cpu >= nr_cpu_ids)
2158                 cpu = cpumask_first(hctx->cpumask);
2159         return cpu;
2160 }
2161
2162 /*
2163  * It'd be great if the workqueue API had a way to pass
2164  * in a mask and had some smarts for more clever placement.
2165  * For now we just round-robin here, switching for every
2166  * BLK_MQ_CPU_WORK_BATCH queued items.
2167  */
2168 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2169 {
2170         bool tried = false;
2171         int next_cpu = hctx->next_cpu;
2172
2173         if (hctx->queue->nr_hw_queues == 1)
2174                 return WORK_CPU_UNBOUND;
2175
2176         if (--hctx->next_cpu_batch <= 0) {
2177 select_cpu:
2178                 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2179                                 cpu_online_mask);
2180                 if (next_cpu >= nr_cpu_ids)
2181                         next_cpu = blk_mq_first_mapped_cpu(hctx);
2182                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2183         }
2184
2185         /*
2186          * Do unbound schedule if we can't find a online CPU for this hctx,
2187          * and it should only happen in the path of handling CPU DEAD.
2188          */
2189         if (!cpu_online(next_cpu)) {
2190                 if (!tried) {
2191                         tried = true;
2192                         goto select_cpu;
2193                 }
2194
2195                 /*
2196                  * Make sure to re-select CPU next time once after CPUs
2197                  * in hctx->cpumask become online again.
2198                  */
2199                 hctx->next_cpu = next_cpu;
2200                 hctx->next_cpu_batch = 1;
2201                 return WORK_CPU_UNBOUND;
2202         }
2203
2204         hctx->next_cpu = next_cpu;
2205         return next_cpu;
2206 }
2207
2208 /**
2209  * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2210  * @hctx: Pointer to the hardware queue to run.
2211  * @msecs: Milliseconds of delay to wait before running the queue.
2212  *
2213  * Run a hardware queue asynchronously with a delay of @msecs.
2214  */
2215 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2216 {
2217         if (unlikely(blk_mq_hctx_stopped(hctx)))
2218                 return;
2219         kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2220                                     msecs_to_jiffies(msecs));
2221 }
2222 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2223
2224 /**
2225  * blk_mq_run_hw_queue - Start to run a hardware queue.
2226  * @hctx: Pointer to the hardware queue to run.
2227  * @async: If we want to run the queue asynchronously.
2228  *
2229  * Check if the request queue is not in a quiesced state and if there are
2230  * pending requests to be sent. If this is true, run the queue to send requests
2231  * to hardware.
2232  */
2233 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2234 {
2235         bool need_run;
2236
2237         /*
2238          * We can't run the queue inline with interrupts disabled.
2239          */
2240         WARN_ON_ONCE(!async && in_interrupt());
2241
2242         might_sleep_if(!async && hctx->flags & BLK_MQ_F_BLOCKING);
2243
2244         /*
2245          * When queue is quiesced, we may be switching io scheduler, or
2246          * updating nr_hw_queues, or other things, and we can't run queue
2247          * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2248          *
2249          * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2250          * quiesced.
2251          */
2252         __blk_mq_run_dispatch_ops(hctx->queue, false,
2253                 need_run = !blk_queue_quiesced(hctx->queue) &&
2254                 blk_mq_hctx_has_pending(hctx));
2255
2256         if (!need_run)
2257                 return;
2258
2259         if (async || !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2260                 blk_mq_delay_run_hw_queue(hctx, 0);
2261                 return;
2262         }
2263
2264         blk_mq_run_dispatch_ops(hctx->queue,
2265                                 blk_mq_sched_dispatch_requests(hctx));
2266 }
2267 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2268
2269 /*
2270  * Return prefered queue to dispatch from (if any) for non-mq aware IO
2271  * scheduler.
2272  */
2273 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2274 {
2275         struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2276         /*
2277          * If the IO scheduler does not respect hardware queues when
2278          * dispatching, we just don't bother with multiple HW queues and
2279          * dispatch from hctx for the current CPU since running multiple queues
2280          * just causes lock contention inside the scheduler and pointless cache
2281          * bouncing.
2282          */
2283         struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2284
2285         if (!blk_mq_hctx_stopped(hctx))
2286                 return hctx;
2287         return NULL;
2288 }
2289
2290 /**
2291  * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2292  * @q: Pointer to the request queue to run.
2293  * @async: If we want to run the queue asynchronously.
2294  */
2295 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2296 {
2297         struct blk_mq_hw_ctx *hctx, *sq_hctx;
2298         unsigned long i;
2299
2300         sq_hctx = NULL;
2301         if (blk_queue_sq_sched(q))
2302                 sq_hctx = blk_mq_get_sq_hctx(q);
2303         queue_for_each_hw_ctx(q, hctx, i) {
2304                 if (blk_mq_hctx_stopped(hctx))
2305                         continue;
2306                 /*
2307                  * Dispatch from this hctx either if there's no hctx preferred
2308                  * by IO scheduler or if it has requests that bypass the
2309                  * scheduler.
2310                  */
2311                 if (!sq_hctx || sq_hctx == hctx ||
2312                     !list_empty_careful(&hctx->dispatch))
2313                         blk_mq_run_hw_queue(hctx, async);
2314         }
2315 }
2316 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2317
2318 /**
2319  * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2320  * @q: Pointer to the request queue to run.
2321  * @msecs: Milliseconds of delay to wait before running the queues.
2322  */
2323 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2324 {
2325         struct blk_mq_hw_ctx *hctx, *sq_hctx;
2326         unsigned long i;
2327
2328         sq_hctx = NULL;
2329         if (blk_queue_sq_sched(q))
2330                 sq_hctx = blk_mq_get_sq_hctx(q);
2331         queue_for_each_hw_ctx(q, hctx, i) {
2332                 if (blk_mq_hctx_stopped(hctx))
2333                         continue;
2334                 /*
2335                  * If there is already a run_work pending, leave the
2336                  * pending delay untouched. Otherwise, a hctx can stall
2337                  * if another hctx is re-delaying the other's work
2338                  * before the work executes.
2339                  */
2340                 if (delayed_work_pending(&hctx->run_work))
2341                         continue;
2342                 /*
2343                  * Dispatch from this hctx either if there's no hctx preferred
2344                  * by IO scheduler or if it has requests that bypass the
2345                  * scheduler.
2346                  */
2347                 if (!sq_hctx || sq_hctx == hctx ||
2348                     !list_empty_careful(&hctx->dispatch))
2349                         blk_mq_delay_run_hw_queue(hctx, msecs);
2350         }
2351 }
2352 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2353
2354 /*
2355  * This function is often used for pausing .queue_rq() by driver when
2356  * there isn't enough resource or some conditions aren't satisfied, and
2357  * BLK_STS_RESOURCE is usually returned.
2358  *
2359  * We do not guarantee that dispatch can be drained or blocked
2360  * after blk_mq_stop_hw_queue() returns. Please use
2361  * blk_mq_quiesce_queue() for that requirement.
2362  */
2363 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2364 {
2365         cancel_delayed_work(&hctx->run_work);
2366
2367         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2368 }
2369 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2370
2371 /*
2372  * This function is often used for pausing .queue_rq() by driver when
2373  * there isn't enough resource or some conditions aren't satisfied, and
2374  * BLK_STS_RESOURCE is usually returned.
2375  *
2376  * We do not guarantee that dispatch can be drained or blocked
2377  * after blk_mq_stop_hw_queues() returns. Please use
2378  * blk_mq_quiesce_queue() for that requirement.
2379  */
2380 void blk_mq_stop_hw_queues(struct request_queue *q)
2381 {
2382         struct blk_mq_hw_ctx *hctx;
2383         unsigned long i;
2384
2385         queue_for_each_hw_ctx(q, hctx, i)
2386                 blk_mq_stop_hw_queue(hctx);
2387 }
2388 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2389
2390 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2391 {
2392         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2393
2394         blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);
2395 }
2396 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2397
2398 void blk_mq_start_hw_queues(struct request_queue *q)
2399 {
2400         struct blk_mq_hw_ctx *hctx;
2401         unsigned long i;
2402
2403         queue_for_each_hw_ctx(q, hctx, i)
2404                 blk_mq_start_hw_queue(hctx);
2405 }
2406 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2407
2408 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2409 {
2410         if (!blk_mq_hctx_stopped(hctx))
2411                 return;
2412
2413         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2414         blk_mq_run_hw_queue(hctx, async);
2415 }
2416 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2417
2418 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2419 {
2420         struct blk_mq_hw_ctx *hctx;
2421         unsigned long i;
2422
2423         queue_for_each_hw_ctx(q, hctx, i)
2424                 blk_mq_start_stopped_hw_queue(hctx, async ||
2425                                         (hctx->flags & BLK_MQ_F_BLOCKING));
2426 }
2427 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2428
2429 static void blk_mq_run_work_fn(struct work_struct *work)
2430 {
2431         struct blk_mq_hw_ctx *hctx =
2432                 container_of(work, struct blk_mq_hw_ctx, run_work.work);
2433
2434         blk_mq_run_dispatch_ops(hctx->queue,
2435                                 blk_mq_sched_dispatch_requests(hctx));
2436 }
2437
2438 /**
2439  * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2440  * @rq: Pointer to request to be inserted.
2441  * @flags: BLK_MQ_INSERT_*
2442  *
2443  * Should only be used carefully, when the caller knows we want to
2444  * bypass a potential IO scheduler on the target device.
2445  */
2446 static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags)
2447 {
2448         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2449
2450         spin_lock(&hctx->lock);
2451         if (flags & BLK_MQ_INSERT_AT_HEAD)
2452                 list_add(&rq->queuelist, &hctx->dispatch);
2453         else
2454                 list_add_tail(&rq->queuelist, &hctx->dispatch);
2455         spin_unlock(&hctx->lock);
2456 }
2457
2458 static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx,
2459                 struct blk_mq_ctx *ctx, struct list_head *list,
2460                 bool run_queue_async)
2461 {
2462         struct request *rq;
2463         enum hctx_type type = hctx->type;
2464
2465         /*
2466          * Try to issue requests directly if the hw queue isn't busy to save an
2467          * extra enqueue & dequeue to the sw queue.
2468          */
2469         if (!hctx->dispatch_busy && !run_queue_async) {
2470                 blk_mq_run_dispatch_ops(hctx->queue,
2471                         blk_mq_try_issue_list_directly(hctx, list));
2472                 if (list_empty(list))
2473                         goto out;
2474         }
2475
2476         /*
2477          * preemption doesn't flush plug list, so it's possible ctx->cpu is
2478          * offline now
2479          */
2480         list_for_each_entry(rq, list, queuelist) {
2481                 BUG_ON(rq->mq_ctx != ctx);
2482                 trace_block_rq_insert(rq);
2483                 if (rq->cmd_flags & REQ_NOWAIT)
2484                         run_queue_async = true;
2485         }
2486
2487         spin_lock(&ctx->lock);
2488         list_splice_tail_init(list, &ctx->rq_lists[type]);
2489         blk_mq_hctx_mark_pending(hctx, ctx);
2490         spin_unlock(&ctx->lock);
2491 out:
2492         blk_mq_run_hw_queue(hctx, run_queue_async);
2493 }
2494
2495 static void blk_mq_insert_request(struct request *rq, blk_insert_t flags)
2496 {
2497         struct request_queue *q = rq->q;
2498         struct blk_mq_ctx *ctx = rq->mq_ctx;
2499         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2500
2501         if (blk_rq_is_passthrough(rq)) {
2502                 /*
2503                  * Passthrough request have to be added to hctx->dispatch
2504                  * directly.  The device may be in a situation where it can't
2505                  * handle FS request, and always returns BLK_STS_RESOURCE for
2506                  * them, which gets them added to hctx->dispatch.
2507                  *
2508                  * If a passthrough request is required to unblock the queues,
2509                  * and it is added to the scheduler queue, there is no chance to
2510                  * dispatch it given we prioritize requests in hctx->dispatch.
2511                  */
2512                 blk_mq_request_bypass_insert(rq, flags);
2513         } else if (req_op(rq) == REQ_OP_FLUSH) {
2514                 /*
2515                  * Firstly normal IO request is inserted to scheduler queue or
2516                  * sw queue, meantime we add flush request to dispatch queue(
2517                  * hctx->dispatch) directly and there is at most one in-flight
2518                  * flush request for each hw queue, so it doesn't matter to add
2519                  * flush request to tail or front of the dispatch queue.
2520                  *
2521                  * Secondly in case of NCQ, flush request belongs to non-NCQ
2522                  * command, and queueing it will fail when there is any
2523                  * in-flight normal IO request(NCQ command). When adding flush
2524                  * rq to the front of hctx->dispatch, it is easier to introduce
2525                  * extra time to flush rq's latency because of S_SCHED_RESTART
2526                  * compared with adding to the tail of dispatch queue, then
2527                  * chance of flush merge is increased, and less flush requests
2528                  * will be issued to controller. It is observed that ~10% time
2529                  * is saved in blktests block/004 on disk attached to AHCI/NCQ
2530                  * drive when adding flush rq to the front of hctx->dispatch.
2531                  *
2532                  * Simply queue flush rq to the front of hctx->dispatch so that
2533                  * intensive flush workloads can benefit in case of NCQ HW.
2534                  */
2535                 blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD);
2536         } else if (q->elevator) {
2537                 LIST_HEAD(list);
2538
2539                 WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG);
2540
2541                 list_add(&rq->queuelist, &list);
2542                 q->elevator->type->ops.insert_requests(hctx, &list, flags);
2543         } else {
2544                 trace_block_rq_insert(rq);
2545
2546                 spin_lock(&ctx->lock);
2547                 if (flags & BLK_MQ_INSERT_AT_HEAD)
2548                         list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]);
2549                 else
2550                         list_add_tail(&rq->queuelist,
2551                                       &ctx->rq_lists[hctx->type]);
2552                 blk_mq_hctx_mark_pending(hctx, ctx);
2553                 spin_unlock(&ctx->lock);
2554         }
2555 }
2556
2557 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2558                 unsigned int nr_segs)
2559 {
2560         int err;
2561
2562         if (bio->bi_opf & REQ_RAHEAD)
2563                 rq->cmd_flags |= REQ_FAILFAST_MASK;
2564
2565         rq->__sector = bio->bi_iter.bi_sector;
2566         blk_rq_bio_prep(rq, bio, nr_segs);
2567
2568         /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2569         err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2570         WARN_ON_ONCE(err);
2571
2572         blk_account_io_start(rq);
2573 }
2574
2575 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2576                                             struct request *rq, bool last)
2577 {
2578         struct request_queue *q = rq->q;
2579         struct blk_mq_queue_data bd = {
2580                 .rq = rq,
2581                 .last = last,
2582         };
2583         blk_status_t ret;
2584
2585         /*
2586          * For OK queue, we are done. For error, caller may kill it.
2587          * Any other error (busy), just add it to our list as we
2588          * previously would have done.
2589          */
2590         ret = q->mq_ops->queue_rq(hctx, &bd);
2591         switch (ret) {
2592         case BLK_STS_OK:
2593                 blk_mq_update_dispatch_busy(hctx, false);
2594                 break;
2595         case BLK_STS_RESOURCE:
2596         case BLK_STS_DEV_RESOURCE:
2597                 blk_mq_update_dispatch_busy(hctx, true);
2598                 __blk_mq_requeue_request(rq);
2599                 break;
2600         default:
2601                 blk_mq_update_dispatch_busy(hctx, false);
2602                 break;
2603         }
2604
2605         return ret;
2606 }
2607
2608 static bool blk_mq_get_budget_and_tag(struct request *rq)
2609 {
2610         int budget_token;
2611
2612         budget_token = blk_mq_get_dispatch_budget(rq->q);
2613         if (budget_token < 0)
2614                 return false;
2615         blk_mq_set_rq_budget_token(rq, budget_token);
2616         if (!blk_mq_get_driver_tag(rq)) {
2617                 blk_mq_put_dispatch_budget(rq->q, budget_token);
2618                 return false;
2619         }
2620         return true;
2621 }
2622
2623 /**
2624  * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2625  * @hctx: Pointer of the associated hardware queue.
2626  * @rq: Pointer to request to be sent.
2627  *
2628  * If the device has enough resources to accept a new request now, send the
2629  * request directly to device driver. Else, insert at hctx->dispatch queue, so
2630  * we can try send it another time in the future. Requests inserted at this
2631  * queue have higher priority.
2632  */
2633 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2634                 struct request *rq)
2635 {
2636         blk_status_t ret;
2637
2638         if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2639                 blk_mq_insert_request(rq, 0);
2640                 return;
2641         }
2642
2643         if ((rq->rq_flags & RQF_USE_SCHED) || !blk_mq_get_budget_and_tag(rq)) {
2644                 blk_mq_insert_request(rq, 0);
2645                 blk_mq_run_hw_queue(hctx, rq->cmd_flags & REQ_NOWAIT);
2646                 return;
2647         }
2648
2649         ret = __blk_mq_issue_directly(hctx, rq, true);
2650         switch (ret) {
2651         case BLK_STS_OK:
2652                 break;
2653         case BLK_STS_RESOURCE:
2654         case BLK_STS_DEV_RESOURCE:
2655                 blk_mq_request_bypass_insert(rq, 0);
2656                 blk_mq_run_hw_queue(hctx, false);
2657                 break;
2658         default:
2659                 blk_mq_end_request(rq, ret);
2660                 break;
2661         }
2662 }
2663
2664 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2665 {
2666         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2667
2668         if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2669                 blk_mq_insert_request(rq, 0);
2670                 return BLK_STS_OK;
2671         }
2672
2673         if (!blk_mq_get_budget_and_tag(rq))
2674                 return BLK_STS_RESOURCE;
2675         return __blk_mq_issue_directly(hctx, rq, last);
2676 }
2677
2678 static void blk_mq_plug_issue_direct(struct blk_plug *plug)
2679 {
2680         struct blk_mq_hw_ctx *hctx = NULL;
2681         struct request *rq;
2682         int queued = 0;
2683         blk_status_t ret = BLK_STS_OK;
2684
2685         while ((rq = rq_list_pop(&plug->mq_list))) {
2686                 bool last = rq_list_empty(plug->mq_list);
2687
2688                 if (hctx != rq->mq_hctx) {
2689                         if (hctx) {
2690                                 blk_mq_commit_rqs(hctx, queued, false);
2691                                 queued = 0;
2692                         }
2693                         hctx = rq->mq_hctx;
2694                 }
2695
2696                 ret = blk_mq_request_issue_directly(rq, last);
2697                 switch (ret) {
2698                 case BLK_STS_OK:
2699                         queued++;
2700                         break;
2701                 case BLK_STS_RESOURCE:
2702                 case BLK_STS_DEV_RESOURCE:
2703                         blk_mq_request_bypass_insert(rq, 0);
2704                         blk_mq_run_hw_queue(hctx, false);
2705                         goto out;
2706                 default:
2707                         blk_mq_end_request(rq, ret);
2708                         break;
2709                 }
2710         }
2711
2712 out:
2713         if (ret != BLK_STS_OK)
2714                 blk_mq_commit_rqs(hctx, queued, false);
2715 }
2716
2717 static void __blk_mq_flush_plug_list(struct request_queue *q,
2718                                      struct blk_plug *plug)
2719 {
2720         if (blk_queue_quiesced(q))
2721                 return;
2722         q->mq_ops->queue_rqs(&plug->mq_list);
2723 }
2724
2725 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2726 {
2727         struct blk_mq_hw_ctx *this_hctx = NULL;
2728         struct blk_mq_ctx *this_ctx = NULL;
2729         struct request *requeue_list = NULL;
2730         struct request **requeue_lastp = &requeue_list;
2731         unsigned int depth = 0;
2732         bool is_passthrough = false;
2733         LIST_HEAD(list);
2734
2735         do {
2736                 struct request *rq = rq_list_pop(&plug->mq_list);
2737
2738                 if (!this_hctx) {
2739                         this_hctx = rq->mq_hctx;
2740                         this_ctx = rq->mq_ctx;
2741                         is_passthrough = blk_rq_is_passthrough(rq);
2742                 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx ||
2743                            is_passthrough != blk_rq_is_passthrough(rq)) {
2744                         rq_list_add_tail(&requeue_lastp, rq);
2745                         continue;
2746                 }
2747                 list_add(&rq->queuelist, &list);
2748                 depth++;
2749         } while (!rq_list_empty(plug->mq_list));
2750
2751         plug->mq_list = requeue_list;
2752         trace_block_unplug(this_hctx->queue, depth, !from_sched);
2753
2754         percpu_ref_get(&this_hctx->queue->q_usage_counter);
2755         /* passthrough requests should never be issued to the I/O scheduler */
2756         if (is_passthrough) {
2757                 spin_lock(&this_hctx->lock);
2758                 list_splice_tail_init(&list, &this_hctx->dispatch);
2759                 spin_unlock(&this_hctx->lock);
2760                 blk_mq_run_hw_queue(this_hctx, from_sched);
2761         } else if (this_hctx->queue->elevator) {
2762                 this_hctx->queue->elevator->type->ops.insert_requests(this_hctx,
2763                                 &list, 0);
2764                 blk_mq_run_hw_queue(this_hctx, from_sched);
2765         } else {
2766                 blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched);
2767         }
2768         percpu_ref_put(&this_hctx->queue->q_usage_counter);
2769 }
2770
2771 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2772 {
2773         struct request *rq;
2774
2775         /*
2776          * We may have been called recursively midway through handling
2777          * plug->mq_list via a schedule() in the driver's queue_rq() callback.
2778          * To avoid mq_list changing under our feet, clear rq_count early and
2779          * bail out specifically if rq_count is 0 rather than checking
2780          * whether the mq_list is empty.
2781          */
2782         if (plug->rq_count == 0)
2783                 return;
2784         plug->rq_count = 0;
2785
2786         if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2787                 struct request_queue *q;
2788
2789                 rq = rq_list_peek(&plug->mq_list);
2790                 q = rq->q;
2791
2792                 /*
2793                  * Peek first request and see if we have a ->queue_rqs() hook.
2794                  * If we do, we can dispatch the whole plug list in one go. We
2795                  * already know at this point that all requests belong to the
2796                  * same queue, caller must ensure that's the case.
2797                  *
2798                  * Since we pass off the full list to the driver at this point,
2799                  * we do not increment the active request count for the queue.
2800                  * Bypass shared tags for now because of that.
2801                  */
2802                 if (q->mq_ops->queue_rqs &&
2803                     !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
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_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(list)) {
2828                 struct request *rq = list_first_entry(list, struct request,
2829                                 queuelist);
2830
2831                 list_del_init(&rq->queuelist);
2832                 ret = blk_mq_request_issue_directly(rq, list_empty(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, 0);
2840                         if (list_empty(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, 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         if (unlikely(bio_queue_enter(bio)))
2879                 return NULL;
2880
2881         if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2882                 goto queue_exit;
2883
2884         rq_qos_throttle(q, bio);
2885
2886         if (plug) {
2887                 data.nr_tags = plug->nr_ios;
2888                 plug->nr_ios = 1;
2889                 data.cached_rq = &plug->cached_rq;
2890         }
2891
2892         rq = __blk_mq_alloc_requests(&data);
2893         if (rq)
2894                 return rq;
2895         rq_qos_cleanup(q, bio);
2896         if (bio->bi_opf & REQ_NOWAIT)
2897                 bio_wouldblock_error(bio);
2898 queue_exit:
2899         blk_queue_exit(q);
2900         return NULL;
2901 }
2902
2903 static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
2904                 struct blk_plug *plug, struct bio **bio, unsigned int nsegs)
2905 {
2906         struct request *rq;
2907         enum hctx_type type, hctx_type;
2908
2909         if (!plug)
2910                 return NULL;
2911         rq = rq_list_peek(&plug->cached_rq);
2912         if (!rq || rq->q != q)
2913                 return NULL;
2914
2915         if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) {
2916                 *bio = NULL;
2917                 return NULL;
2918         }
2919
2920         type = blk_mq_get_hctx_type((*bio)->bi_opf);
2921         hctx_type = rq->mq_hctx->type;
2922         if (type != hctx_type &&
2923             !(type == HCTX_TYPE_READ && hctx_type == HCTX_TYPE_DEFAULT))
2924                 return NULL;
2925         if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf))
2926                 return NULL;
2927
2928         /*
2929          * If any qos ->throttle() end up blocking, we will have flushed the
2930          * plug and hence killed the cached_rq list as well. Pop this entry
2931          * before we throttle.
2932          */
2933         plug->cached_rq = rq_list_next(rq);
2934         rq_qos_throttle(q, *bio);
2935
2936         blk_mq_rq_time_init(rq, 0);
2937         rq->cmd_flags = (*bio)->bi_opf;
2938         INIT_LIST_HEAD(&rq->queuelist);
2939         return rq;
2940 }
2941
2942 static void bio_set_ioprio(struct bio *bio)
2943 {
2944         /* Nobody set ioprio so far? Initialize it based on task's nice value */
2945         if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE)
2946                 bio->bi_ioprio = get_current_ioprio();
2947         blkcg_set_ioprio(bio);
2948 }
2949
2950 /**
2951  * blk_mq_submit_bio - Create and send a request to block device.
2952  * @bio: Bio pointer.
2953  *
2954  * Builds up a request structure from @q and @bio and send to the device. The
2955  * request may not be queued directly to hardware if:
2956  * * This request can be merged with another one
2957  * * We want to place request at plug queue for possible future merging
2958  * * There is an IO scheduler active at this queue
2959  *
2960  * It will not queue the request if there is an error with the bio, or at the
2961  * request creation.
2962  */
2963 void blk_mq_submit_bio(struct bio *bio)
2964 {
2965         struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2966         struct blk_plug *plug = blk_mq_plug(bio);
2967         const int is_sync = op_is_sync(bio->bi_opf);
2968         struct blk_mq_hw_ctx *hctx;
2969         struct request *rq;
2970         unsigned int nr_segs = 1;
2971         blk_status_t ret;
2972
2973         bio = blk_queue_bounce(bio, q);
2974         if (bio_may_exceed_limits(bio, &q->limits)) {
2975                 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2976                 if (!bio)
2977                         return;
2978         }
2979
2980         if (!bio_integrity_prep(bio))
2981                 return;
2982
2983         bio_set_ioprio(bio);
2984
2985         rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs);
2986         if (!rq) {
2987                 if (!bio)
2988                         return;
2989                 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2990                 if (unlikely(!rq))
2991                         return;
2992         }
2993
2994         trace_block_getrq(bio);
2995
2996         rq_qos_track(q, rq, bio);
2997
2998         blk_mq_bio_to_request(rq, bio, nr_segs);
2999
3000         ret = blk_crypto_rq_get_keyslot(rq);
3001         if (ret != BLK_STS_OK) {
3002                 bio->bi_status = ret;
3003                 bio_endio(bio);
3004                 blk_mq_free_request(rq);
3005                 return;
3006         }
3007
3008         if (op_is_flush(bio->bi_opf) && blk_insert_flush(rq))
3009                 return;
3010
3011         if (plug) {
3012                 blk_add_rq_to_plug(plug, rq);
3013                 return;
3014         }
3015
3016         hctx = rq->mq_hctx;
3017         if ((rq->rq_flags & RQF_USE_SCHED) ||
3018             (hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) {
3019                 blk_mq_insert_request(rq, 0);
3020                 blk_mq_run_hw_queue(hctx, true);
3021         } else {
3022                 blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq));
3023         }
3024 }
3025
3026 #ifdef CONFIG_BLK_MQ_STACKING
3027 /**
3028  * blk_insert_cloned_request - Helper for stacking drivers to submit a request
3029  * @rq: the request being queued
3030  */
3031 blk_status_t blk_insert_cloned_request(struct request *rq)
3032 {
3033         struct request_queue *q = rq->q;
3034         unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
3035         unsigned int max_segments = blk_rq_get_max_segments(rq);
3036         blk_status_t ret;
3037
3038         if (blk_rq_sectors(rq) > max_sectors) {
3039                 /*
3040                  * SCSI device does not have a good way to return if
3041                  * Write Same/Zero is actually supported. If a device rejects
3042                  * a non-read/write command (discard, write same,etc.) the
3043                  * low-level device driver will set the relevant queue limit to
3044                  * 0 to prevent blk-lib from issuing more of the offending
3045                  * operations. Commands queued prior to the queue limit being
3046                  * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3047                  * errors being propagated to upper layers.
3048                  */
3049                 if (max_sectors == 0)
3050                         return BLK_STS_NOTSUPP;
3051
3052                 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3053                         __func__, blk_rq_sectors(rq), max_sectors);
3054                 return BLK_STS_IOERR;
3055         }
3056
3057         /*
3058          * The queue settings related to segment counting may differ from the
3059          * original queue.
3060          */
3061         rq->nr_phys_segments = blk_recalc_rq_segments(rq);
3062         if (rq->nr_phys_segments > max_segments) {
3063                 printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n",
3064                         __func__, rq->nr_phys_segments, max_segments);
3065                 return BLK_STS_IOERR;
3066         }
3067
3068         if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
3069                 return BLK_STS_IOERR;
3070
3071         ret = blk_crypto_rq_get_keyslot(rq);
3072         if (ret != BLK_STS_OK)
3073                 return ret;
3074
3075         blk_account_io_start(rq);
3076
3077         /*
3078          * Since we have a scheduler attached on the top device,
3079          * bypass a potential scheduler on the bottom device for
3080          * insert.
3081          */
3082         blk_mq_run_dispatch_ops(q,
3083                         ret = blk_mq_request_issue_directly(rq, true));
3084         if (ret)
3085                 blk_account_io_done(rq, ktime_get_ns());
3086         return ret;
3087 }
3088 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3089
3090 /**
3091  * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3092  * @rq: the clone request to be cleaned up
3093  *
3094  * Description:
3095  *     Free all bios in @rq for a cloned request.
3096  */
3097 void blk_rq_unprep_clone(struct request *rq)
3098 {
3099         struct bio *bio;
3100
3101         while ((bio = rq->bio) != NULL) {
3102                 rq->bio = bio->bi_next;
3103
3104                 bio_put(bio);
3105         }
3106 }
3107 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3108
3109 /**
3110  * blk_rq_prep_clone - Helper function to setup clone request
3111  * @rq: the request to be setup
3112  * @rq_src: original request to be cloned
3113  * @bs: bio_set that bios for clone are allocated from
3114  * @gfp_mask: memory allocation mask for bio
3115  * @bio_ctr: setup function to be called for each clone bio.
3116  *           Returns %0 for success, non %0 for failure.
3117  * @data: private data to be passed to @bio_ctr
3118  *
3119  * Description:
3120  *     Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3121  *     Also, pages which the original bios are pointing to are not copied
3122  *     and the cloned bios just point same pages.
3123  *     So cloned bios must be completed before original bios, which means
3124  *     the caller must complete @rq before @rq_src.
3125  */
3126 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3127                       struct bio_set *bs, gfp_t gfp_mask,
3128                       int (*bio_ctr)(struct bio *, struct bio *, void *),
3129                       void *data)
3130 {
3131         struct bio *bio, *bio_src;
3132
3133         if (!bs)
3134                 bs = &fs_bio_set;
3135
3136         __rq_for_each_bio(bio_src, rq_src) {
3137                 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3138                                       bs);
3139                 if (!bio)
3140                         goto free_and_out;
3141
3142                 if (bio_ctr && bio_ctr(bio, bio_src, data))
3143                         goto free_and_out;
3144
3145                 if (rq->bio) {
3146                         rq->biotail->bi_next = bio;
3147                         rq->biotail = bio;
3148                 } else {
3149                         rq->bio = rq->biotail = bio;
3150                 }
3151                 bio = NULL;
3152         }
3153
3154         /* Copy attributes of the original request to the clone request. */
3155         rq->__sector = blk_rq_pos(rq_src);
3156         rq->__data_len = blk_rq_bytes(rq_src);
3157         if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3158                 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3159                 rq->special_vec = rq_src->special_vec;
3160         }
3161         rq->nr_phys_segments = rq_src->nr_phys_segments;
3162         rq->ioprio = rq_src->ioprio;
3163
3164         if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3165                 goto free_and_out;
3166
3167         return 0;
3168
3169 free_and_out:
3170         if (bio)
3171                 bio_put(bio);
3172         blk_rq_unprep_clone(rq);
3173
3174         return -ENOMEM;
3175 }
3176 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3177 #endif /* CONFIG_BLK_MQ_STACKING */
3178
3179 /*
3180  * Steal bios from a request and add them to a bio list.
3181  * The request must not have been partially completed before.
3182  */
3183 void blk_steal_bios(struct bio_list *list, struct request *rq)
3184 {
3185         if (rq->bio) {
3186                 if (list->tail)
3187                         list->tail->bi_next = rq->bio;
3188                 else
3189                         list->head = rq->bio;
3190                 list->tail = rq->biotail;
3191
3192                 rq->bio = NULL;
3193                 rq->biotail = NULL;
3194         }
3195
3196         rq->__data_len = 0;
3197 }
3198 EXPORT_SYMBOL_GPL(blk_steal_bios);
3199
3200 static size_t order_to_size(unsigned int order)
3201 {
3202         return (size_t)PAGE_SIZE << order;
3203 }
3204
3205 /* called before freeing request pool in @tags */
3206 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3207                                     struct blk_mq_tags *tags)
3208 {
3209         struct page *page;
3210         unsigned long flags;
3211
3212         /*
3213          * There is no need to clear mapping if driver tags is not initialized
3214          * or the mapping belongs to the driver tags.
3215          */
3216         if (!drv_tags || drv_tags == tags)
3217                 return;
3218
3219         list_for_each_entry(page, &tags->page_list, lru) {
3220                 unsigned long start = (unsigned long)page_address(page);
3221                 unsigned long end = start + order_to_size(page->private);
3222                 int i;
3223
3224                 for (i = 0; i < drv_tags->nr_tags; i++) {
3225                         struct request *rq = drv_tags->rqs[i];
3226                         unsigned long rq_addr = (unsigned long)rq;
3227
3228                         if (rq_addr >= start && rq_addr < end) {
3229                                 WARN_ON_ONCE(req_ref_read(rq) != 0);
3230                                 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3231                         }
3232                 }
3233         }
3234
3235         /*
3236          * Wait until all pending iteration is done.
3237          *
3238          * Request reference is cleared and it is guaranteed to be observed
3239          * after the ->lock is released.
3240          */
3241         spin_lock_irqsave(&drv_tags->lock, flags);
3242         spin_unlock_irqrestore(&drv_tags->lock, flags);
3243 }
3244
3245 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3246                      unsigned int hctx_idx)
3247 {
3248         struct blk_mq_tags *drv_tags;
3249         struct page *page;
3250
3251         if (list_empty(&tags->page_list))
3252                 return;
3253
3254         if (blk_mq_is_shared_tags(set->flags))
3255                 drv_tags = set->shared_tags;
3256         else
3257                 drv_tags = set->tags[hctx_idx];
3258
3259         if (tags->static_rqs && set->ops->exit_request) {
3260                 int i;
3261
3262                 for (i = 0; i < tags->nr_tags; i++) {
3263                         struct request *rq = tags->static_rqs[i];
3264
3265                         if (!rq)
3266                                 continue;
3267                         set->ops->exit_request(set, rq, hctx_idx);
3268                         tags->static_rqs[i] = NULL;
3269                 }
3270         }
3271
3272         blk_mq_clear_rq_mapping(drv_tags, tags);
3273
3274         while (!list_empty(&tags->page_list)) {
3275                 page = list_first_entry(&tags->page_list, struct page, lru);
3276                 list_del_init(&page->lru);
3277                 /*
3278                  * Remove kmemleak object previously allocated in
3279                  * blk_mq_alloc_rqs().
3280                  */
3281                 kmemleak_free(page_address(page));
3282                 __free_pages(page, page->private);
3283         }
3284 }
3285
3286 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3287 {
3288         kfree(tags->rqs);
3289         tags->rqs = NULL;
3290         kfree(tags->static_rqs);
3291         tags->static_rqs = NULL;
3292
3293         blk_mq_free_tags(tags);
3294 }
3295
3296 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3297                 unsigned int hctx_idx)
3298 {
3299         int i;
3300
3301         for (i = 0; i < set->nr_maps; i++) {
3302                 unsigned int start = set->map[i].queue_offset;
3303                 unsigned int end = start + set->map[i].nr_queues;
3304
3305                 if (hctx_idx >= start && hctx_idx < end)
3306                         break;
3307         }
3308
3309         if (i >= set->nr_maps)
3310                 i = HCTX_TYPE_DEFAULT;
3311
3312         return i;
3313 }
3314
3315 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3316                 unsigned int hctx_idx)
3317 {
3318         enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3319
3320         return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3321 }
3322
3323 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3324                                                unsigned int hctx_idx,
3325                                                unsigned int nr_tags,
3326                                                unsigned int reserved_tags)
3327 {
3328         int node = blk_mq_get_hctx_node(set, hctx_idx);
3329         struct blk_mq_tags *tags;
3330
3331         if (node == NUMA_NO_NODE)
3332                 node = set->numa_node;
3333
3334         tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3335                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3336         if (!tags)
3337                 return NULL;
3338
3339         tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3340                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3341                                  node);
3342         if (!tags->rqs)
3343                 goto err_free_tags;
3344
3345         tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3346                                         GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3347                                         node);
3348         if (!tags->static_rqs)
3349                 goto err_free_rqs;
3350
3351         return tags;
3352
3353 err_free_rqs:
3354         kfree(tags->rqs);
3355 err_free_tags:
3356         blk_mq_free_tags(tags);
3357         return NULL;
3358 }
3359
3360 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3361                                unsigned int hctx_idx, int node)
3362 {
3363         int ret;
3364
3365         if (set->ops->init_request) {
3366                 ret = set->ops->init_request(set, rq, hctx_idx, node);
3367                 if (ret)
3368                         return ret;
3369         }
3370
3371         WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3372         return 0;
3373 }
3374
3375 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3376                             struct blk_mq_tags *tags,
3377                             unsigned int hctx_idx, unsigned int depth)
3378 {
3379         unsigned int i, j, entries_per_page, max_order = 4;
3380         int node = blk_mq_get_hctx_node(set, hctx_idx);
3381         size_t rq_size, left;
3382
3383         if (node == NUMA_NO_NODE)
3384                 node = set->numa_node;
3385
3386         INIT_LIST_HEAD(&tags->page_list);
3387
3388         /*
3389          * rq_size is the size of the request plus driver payload, rounded
3390          * to the cacheline size
3391          */
3392         rq_size = round_up(sizeof(struct request) + set->cmd_size,
3393                                 cache_line_size());
3394         left = rq_size * depth;
3395
3396         for (i = 0; i < depth; ) {
3397                 int this_order = max_order;
3398                 struct page *page;
3399                 int to_do;
3400                 void *p;
3401
3402                 while (this_order && left < order_to_size(this_order - 1))
3403                         this_order--;
3404
3405                 do {
3406                         page = alloc_pages_node(node,
3407                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3408                                 this_order);
3409                         if (page)
3410                                 break;
3411                         if (!this_order--)
3412                                 break;
3413                         if (order_to_size(this_order) < rq_size)
3414                                 break;
3415                 } while (1);
3416
3417                 if (!page)
3418                         goto fail;
3419
3420                 page->private = this_order;
3421                 list_add_tail(&page->lru, &tags->page_list);
3422
3423                 p = page_address(page);
3424                 /*
3425                  * Allow kmemleak to scan these pages as they contain pointers
3426                  * to additional allocations like via ops->init_request().
3427                  */
3428                 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3429                 entries_per_page = order_to_size(this_order) / rq_size;
3430                 to_do = min(entries_per_page, depth - i);
3431                 left -= to_do * rq_size;
3432                 for (j = 0; j < to_do; j++) {
3433                         struct request *rq = p;
3434
3435                         tags->static_rqs[i] = rq;
3436                         if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3437                                 tags->static_rqs[i] = NULL;
3438                                 goto fail;
3439                         }
3440
3441                         p += rq_size;
3442                         i++;
3443                 }
3444         }
3445         return 0;
3446
3447 fail:
3448         blk_mq_free_rqs(set, tags, hctx_idx);
3449         return -ENOMEM;
3450 }
3451
3452 struct rq_iter_data {
3453         struct blk_mq_hw_ctx *hctx;
3454         bool has_rq;
3455 };
3456
3457 static bool blk_mq_has_request(struct request *rq, void *data)
3458 {
3459         struct rq_iter_data *iter_data = data;
3460
3461         if (rq->mq_hctx != iter_data->hctx)
3462                 return true;
3463         iter_data->has_rq = true;
3464         return false;
3465 }
3466
3467 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3468 {
3469         struct blk_mq_tags *tags = hctx->sched_tags ?
3470                         hctx->sched_tags : hctx->tags;
3471         struct rq_iter_data data = {
3472                 .hctx   = hctx,
3473         };
3474
3475         blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3476         return data.has_rq;
3477 }
3478
3479 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3480                 struct blk_mq_hw_ctx *hctx)
3481 {
3482         if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3483                 return false;
3484         if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3485                 return false;
3486         return true;
3487 }
3488
3489 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3490 {
3491         struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3492                         struct blk_mq_hw_ctx, cpuhp_online);
3493
3494         if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3495             !blk_mq_last_cpu_in_hctx(cpu, hctx))
3496                 return 0;
3497
3498         /*
3499          * Prevent new request from being allocated on the current hctx.
3500          *
3501          * The smp_mb__after_atomic() Pairs with the implied barrier in
3502          * test_and_set_bit_lock in sbitmap_get().  Ensures the inactive flag is
3503          * seen once we return from the tag allocator.
3504          */
3505         set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3506         smp_mb__after_atomic();
3507
3508         /*
3509          * Try to grab a reference to the queue and wait for any outstanding
3510          * requests.  If we could not grab a reference the queue has been
3511          * frozen and there are no requests.
3512          */
3513         if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3514                 while (blk_mq_hctx_has_requests(hctx))
3515                         msleep(5);
3516                 percpu_ref_put(&hctx->queue->q_usage_counter);
3517         }
3518
3519         return 0;
3520 }
3521
3522 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3523 {
3524         struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3525                         struct blk_mq_hw_ctx, cpuhp_online);
3526
3527         if (cpumask_test_cpu(cpu, hctx->cpumask))
3528                 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3529         return 0;
3530 }
3531
3532 /*
3533  * 'cpu' is going away. splice any existing rq_list entries from this
3534  * software queue to the hw queue dispatch list, and ensure that it
3535  * gets run.
3536  */
3537 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3538 {
3539         struct blk_mq_hw_ctx *hctx;
3540         struct blk_mq_ctx *ctx;
3541         LIST_HEAD(tmp);
3542         enum hctx_type type;
3543
3544         hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3545         if (!cpumask_test_cpu(cpu, hctx->cpumask))
3546                 return 0;
3547
3548         ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3549         type = hctx->type;
3550
3551         spin_lock(&ctx->lock);
3552         if (!list_empty(&ctx->rq_lists[type])) {
3553                 list_splice_init(&ctx->rq_lists[type], &tmp);
3554                 blk_mq_hctx_clear_pending(hctx, ctx);
3555         }
3556         spin_unlock(&ctx->lock);
3557
3558         if (list_empty(&tmp))
3559                 return 0;
3560
3561         spin_lock(&hctx->lock);
3562         list_splice_tail_init(&tmp, &hctx->dispatch);
3563         spin_unlock(&hctx->lock);
3564
3565         blk_mq_run_hw_queue(hctx, true);
3566         return 0;
3567 }
3568
3569 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3570 {
3571         if (!(hctx->flags & BLK_MQ_F_STACKING))
3572                 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3573                                                     &hctx->cpuhp_online);
3574         cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3575                                             &hctx->cpuhp_dead);
3576 }
3577
3578 /*
3579  * Before freeing hw queue, clearing the flush request reference in
3580  * tags->rqs[] for avoiding potential UAF.
3581  */
3582 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3583                 unsigned int queue_depth, struct request *flush_rq)
3584 {
3585         int i;
3586         unsigned long flags;
3587
3588         /* The hw queue may not be mapped yet */
3589         if (!tags)
3590                 return;
3591
3592         WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3593
3594         for (i = 0; i < queue_depth; i++)
3595                 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3596
3597         /*
3598          * Wait until all pending iteration is done.
3599          *
3600          * Request reference is cleared and it is guaranteed to be observed
3601          * after the ->lock is released.
3602          */
3603         spin_lock_irqsave(&tags->lock, flags);
3604         spin_unlock_irqrestore(&tags->lock, flags);
3605 }
3606
3607 /* hctx->ctxs will be freed in queue's release handler */
3608 static void blk_mq_exit_hctx(struct request_queue *q,
3609                 struct blk_mq_tag_set *set,
3610                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3611 {
3612         struct request *flush_rq = hctx->fq->flush_rq;
3613
3614         if (blk_mq_hw_queue_mapped(hctx))
3615                 blk_mq_tag_idle(hctx);
3616
3617         if (blk_queue_init_done(q))
3618                 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3619                                 set->queue_depth, flush_rq);
3620         if (set->ops->exit_request)
3621                 set->ops->exit_request(set, flush_rq, hctx_idx);
3622
3623         if (set->ops->exit_hctx)
3624                 set->ops->exit_hctx(hctx, hctx_idx);
3625
3626         blk_mq_remove_cpuhp(hctx);
3627
3628         xa_erase(&q->hctx_table, hctx_idx);
3629
3630         spin_lock(&q->unused_hctx_lock);
3631         list_add(&hctx->hctx_list, &q->unused_hctx_list);
3632         spin_unlock(&q->unused_hctx_lock);
3633 }
3634
3635 static void blk_mq_exit_hw_queues(struct request_queue *q,
3636                 struct blk_mq_tag_set *set, int nr_queue)
3637 {
3638         struct blk_mq_hw_ctx *hctx;
3639         unsigned long i;
3640
3641         queue_for_each_hw_ctx(q, hctx, i) {
3642                 if (i == nr_queue)
3643                         break;
3644                 blk_mq_exit_hctx(q, set, hctx, i);
3645         }
3646 }
3647
3648 static int blk_mq_init_hctx(struct request_queue *q,
3649                 struct blk_mq_tag_set *set,
3650                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3651 {
3652         hctx->queue_num = hctx_idx;
3653
3654         if (!(hctx->flags & BLK_MQ_F_STACKING))
3655                 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3656                                 &hctx->cpuhp_online);
3657         cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3658
3659         hctx->tags = set->tags[hctx_idx];
3660
3661         if (set->ops->init_hctx &&
3662             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3663                 goto unregister_cpu_notifier;
3664
3665         if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3666                                 hctx->numa_node))
3667                 goto exit_hctx;
3668
3669         if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3670                 goto exit_flush_rq;
3671
3672         return 0;
3673
3674  exit_flush_rq:
3675         if (set->ops->exit_request)
3676                 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3677  exit_hctx:
3678         if (set->ops->exit_hctx)
3679                 set->ops->exit_hctx(hctx, hctx_idx);
3680  unregister_cpu_notifier:
3681         blk_mq_remove_cpuhp(hctx);
3682         return -1;
3683 }
3684
3685 static struct blk_mq_hw_ctx *
3686 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3687                 int node)
3688 {
3689         struct blk_mq_hw_ctx *hctx;
3690         gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3691
3692         hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3693         if (!hctx)
3694                 goto fail_alloc_hctx;
3695
3696         if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3697                 goto free_hctx;
3698
3699         atomic_set(&hctx->nr_active, 0);
3700         if (node == NUMA_NO_NODE)
3701                 node = set->numa_node;
3702         hctx->numa_node = node;
3703
3704         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3705         spin_lock_init(&hctx->lock);
3706         INIT_LIST_HEAD(&hctx->dispatch);
3707         hctx->queue = q;
3708         hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3709
3710         INIT_LIST_HEAD(&hctx->hctx_list);
3711
3712         /*
3713          * Allocate space for all possible cpus to avoid allocation at
3714          * runtime
3715          */
3716         hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3717                         gfp, node);
3718         if (!hctx->ctxs)
3719                 goto free_cpumask;
3720
3721         if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3722                                 gfp, node, false, false))
3723                 goto free_ctxs;
3724         hctx->nr_ctx = 0;
3725
3726         spin_lock_init(&hctx->dispatch_wait_lock);
3727         init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3728         INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3729
3730         hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3731         if (!hctx->fq)
3732                 goto free_bitmap;
3733
3734         blk_mq_hctx_kobj_init(hctx);
3735
3736         return hctx;
3737
3738  free_bitmap:
3739         sbitmap_free(&hctx->ctx_map);
3740  free_ctxs:
3741         kfree(hctx->ctxs);
3742  free_cpumask:
3743         free_cpumask_var(hctx->cpumask);
3744  free_hctx:
3745         kfree(hctx);
3746  fail_alloc_hctx:
3747         return NULL;
3748 }
3749
3750 static void blk_mq_init_cpu_queues(struct request_queue *q,
3751                                    unsigned int nr_hw_queues)
3752 {
3753         struct blk_mq_tag_set *set = q->tag_set;
3754         unsigned int i, j;
3755
3756         for_each_possible_cpu(i) {
3757                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3758                 struct blk_mq_hw_ctx *hctx;
3759                 int k;
3760
3761                 __ctx->cpu = i;
3762                 spin_lock_init(&__ctx->lock);
3763                 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3764                         INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3765
3766                 __ctx->queue = q;
3767
3768                 /*
3769                  * Set local node, IFF we have more than one hw queue. If
3770                  * not, we remain on the home node of the device
3771                  */
3772                 for (j = 0; j < set->nr_maps; j++) {
3773                         hctx = blk_mq_map_queue_type(q, j, i);
3774                         if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3775                                 hctx->numa_node = cpu_to_node(i);
3776                 }
3777         }
3778 }
3779
3780 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3781                                              unsigned int hctx_idx,
3782                                              unsigned int depth)
3783 {
3784         struct blk_mq_tags *tags;
3785         int ret;
3786
3787         tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3788         if (!tags)
3789                 return NULL;
3790
3791         ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3792         if (ret) {
3793                 blk_mq_free_rq_map(tags);
3794                 return NULL;
3795         }
3796
3797         return tags;
3798 }
3799
3800 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3801                                        int hctx_idx)
3802 {
3803         if (blk_mq_is_shared_tags(set->flags)) {
3804                 set->tags[hctx_idx] = set->shared_tags;
3805
3806                 return true;
3807         }
3808
3809         set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3810                                                        set->queue_depth);
3811
3812         return set->tags[hctx_idx];
3813 }
3814
3815 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3816                              struct blk_mq_tags *tags,
3817                              unsigned int hctx_idx)
3818 {
3819         if (tags) {
3820                 blk_mq_free_rqs(set, tags, hctx_idx);
3821                 blk_mq_free_rq_map(tags);
3822         }
3823 }
3824
3825 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3826                                       unsigned int hctx_idx)
3827 {
3828         if (!blk_mq_is_shared_tags(set->flags))
3829                 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3830
3831         set->tags[hctx_idx] = NULL;
3832 }
3833
3834 static void blk_mq_map_swqueue(struct request_queue *q)
3835 {
3836         unsigned int j, hctx_idx;
3837         unsigned long i;
3838         struct blk_mq_hw_ctx *hctx;
3839         struct blk_mq_ctx *ctx;
3840         struct blk_mq_tag_set *set = q->tag_set;
3841
3842         queue_for_each_hw_ctx(q, hctx, i) {
3843                 cpumask_clear(hctx->cpumask);
3844                 hctx->nr_ctx = 0;
3845                 hctx->dispatch_from = NULL;
3846         }
3847
3848         /*
3849          * Map software to hardware queues.
3850          *
3851          * If the cpu isn't present, the cpu is mapped to first hctx.
3852          */
3853         for_each_possible_cpu(i) {
3854
3855                 ctx = per_cpu_ptr(q->queue_ctx, i);
3856                 for (j = 0; j < set->nr_maps; j++) {
3857                         if (!set->map[j].nr_queues) {
3858                                 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3859                                                 HCTX_TYPE_DEFAULT, i);
3860                                 continue;
3861                         }
3862                         hctx_idx = set->map[j].mq_map[i];
3863                         /* unmapped hw queue can be remapped after CPU topo changed */
3864                         if (!set->tags[hctx_idx] &&
3865                             !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3866                                 /*
3867                                  * If tags initialization fail for some hctx,
3868                                  * that hctx won't be brought online.  In this
3869                                  * case, remap the current ctx to hctx[0] which
3870                                  * is guaranteed to always have tags allocated
3871                                  */
3872                                 set->map[j].mq_map[i] = 0;
3873                         }
3874
3875                         hctx = blk_mq_map_queue_type(q, j, i);
3876                         ctx->hctxs[j] = hctx;
3877                         /*
3878                          * If the CPU is already set in the mask, then we've
3879                          * mapped this one already. This can happen if
3880                          * devices share queues across queue maps.
3881                          */
3882                         if (cpumask_test_cpu(i, hctx->cpumask))
3883                                 continue;
3884
3885                         cpumask_set_cpu(i, hctx->cpumask);
3886                         hctx->type = j;
3887                         ctx->index_hw[hctx->type] = hctx->nr_ctx;
3888                         hctx->ctxs[hctx->nr_ctx++] = ctx;
3889
3890                         /*
3891                          * If the nr_ctx type overflows, we have exceeded the
3892                          * amount of sw queues we can support.
3893                          */
3894                         BUG_ON(!hctx->nr_ctx);
3895                 }
3896
3897                 for (; j < HCTX_MAX_TYPES; j++)
3898                         ctx->hctxs[j] = blk_mq_map_queue_type(q,
3899                                         HCTX_TYPE_DEFAULT, i);
3900         }
3901
3902         queue_for_each_hw_ctx(q, hctx, i) {
3903                 /*
3904                  * If no software queues are mapped to this hardware queue,
3905                  * disable it and free the request entries.
3906                  */
3907                 if (!hctx->nr_ctx) {
3908                         /* Never unmap queue 0.  We need it as a
3909                          * fallback in case of a new remap fails
3910                          * allocation
3911                          */
3912                         if (i)
3913                                 __blk_mq_free_map_and_rqs(set, i);
3914
3915                         hctx->tags = NULL;
3916                         continue;
3917                 }
3918
3919                 hctx->tags = set->tags[i];
3920                 WARN_ON(!hctx->tags);
3921
3922                 /*
3923                  * Set the map size to the number of mapped software queues.
3924                  * This is more accurate and more efficient than looping
3925                  * over all possibly mapped software queues.
3926                  */
3927                 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3928
3929                 /*
3930                  * Initialize batch roundrobin counts
3931                  */
3932                 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3933                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3934         }
3935 }
3936
3937 /*
3938  * Caller needs to ensure that we're either frozen/quiesced, or that
3939  * the queue isn't live yet.
3940  */
3941 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3942 {
3943         struct blk_mq_hw_ctx *hctx;
3944         unsigned long i;
3945
3946         queue_for_each_hw_ctx(q, hctx, i) {
3947                 if (shared) {
3948                         hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3949                 } else {
3950                         blk_mq_tag_idle(hctx);
3951                         hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3952                 }
3953         }
3954 }
3955
3956 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3957                                          bool shared)
3958 {
3959         struct request_queue *q;
3960
3961         lockdep_assert_held(&set->tag_list_lock);
3962
3963         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3964                 blk_mq_freeze_queue(q);
3965                 queue_set_hctx_shared(q, shared);
3966                 blk_mq_unfreeze_queue(q);
3967         }
3968 }
3969
3970 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3971 {
3972         struct blk_mq_tag_set *set = q->tag_set;
3973
3974         mutex_lock(&set->tag_list_lock);
3975         list_del(&q->tag_set_list);
3976         if (list_is_singular(&set->tag_list)) {
3977                 /* just transitioned to unshared */
3978                 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3979                 /* update existing queue */
3980                 blk_mq_update_tag_set_shared(set, false);
3981         }
3982         mutex_unlock(&set->tag_list_lock);
3983         INIT_LIST_HEAD(&q->tag_set_list);
3984 }
3985
3986 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3987                                      struct request_queue *q)
3988 {
3989         mutex_lock(&set->tag_list_lock);
3990
3991         /*
3992          * Check to see if we're transitioning to shared (from 1 to 2 queues).
3993          */
3994         if (!list_empty(&set->tag_list) &&
3995             !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3996                 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3997                 /* update existing queue */
3998                 blk_mq_update_tag_set_shared(set, true);
3999         }
4000         if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
4001                 queue_set_hctx_shared(q, true);
4002         list_add_tail(&q->tag_set_list, &set->tag_list);
4003
4004         mutex_unlock(&set->tag_list_lock);
4005 }
4006
4007 /* All allocations will be freed in release handler of q->mq_kobj */
4008 static int blk_mq_alloc_ctxs(struct request_queue *q)
4009 {
4010         struct blk_mq_ctxs *ctxs;
4011         int cpu;
4012
4013         ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
4014         if (!ctxs)
4015                 return -ENOMEM;
4016
4017         ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
4018         if (!ctxs->queue_ctx)
4019                 goto fail;
4020
4021         for_each_possible_cpu(cpu) {
4022                 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
4023                 ctx->ctxs = ctxs;
4024         }
4025
4026         q->mq_kobj = &ctxs->kobj;
4027         q->queue_ctx = ctxs->queue_ctx;
4028
4029         return 0;
4030  fail:
4031         kfree(ctxs);
4032         return -ENOMEM;
4033 }
4034
4035 /*
4036  * It is the actual release handler for mq, but we do it from
4037  * request queue's release handler for avoiding use-after-free
4038  * and headache because q->mq_kobj shouldn't have been introduced,
4039  * but we can't group ctx/kctx kobj without it.
4040  */
4041 void blk_mq_release(struct request_queue *q)
4042 {
4043         struct blk_mq_hw_ctx *hctx, *next;
4044         unsigned long i;
4045
4046         queue_for_each_hw_ctx(q, hctx, i)
4047                 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4048
4049         /* all hctx are in .unused_hctx_list now */
4050         list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4051                 list_del_init(&hctx->hctx_list);
4052                 kobject_put(&hctx->kobj);
4053         }
4054
4055         xa_destroy(&q->hctx_table);
4056
4057         /*
4058          * release .mq_kobj and sw queue's kobject now because
4059          * both share lifetime with request queue.
4060          */
4061         blk_mq_sysfs_deinit(q);
4062 }
4063
4064 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
4065                 void *queuedata)
4066 {
4067         struct request_queue *q;
4068         int ret;
4069
4070         q = blk_alloc_queue(set->numa_node);
4071         if (!q)
4072                 return ERR_PTR(-ENOMEM);
4073         q->queuedata = queuedata;
4074         ret = blk_mq_init_allocated_queue(set, q);
4075         if (ret) {
4076                 blk_put_queue(q);
4077                 return ERR_PTR(ret);
4078         }
4079         return q;
4080 }
4081
4082 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
4083 {
4084         return blk_mq_init_queue_data(set, NULL);
4085 }
4086 EXPORT_SYMBOL(blk_mq_init_queue);
4087
4088 /**
4089  * blk_mq_destroy_queue - shutdown a request queue
4090  * @q: request queue to shutdown
4091  *
4092  * This shuts down a request queue allocated by blk_mq_init_queue(). All future
4093  * requests will be failed with -ENODEV. The caller is responsible for dropping
4094  * the reference from blk_mq_init_queue() by calling blk_put_queue().
4095  *
4096  * Context: can sleep
4097  */
4098 void blk_mq_destroy_queue(struct request_queue *q)
4099 {
4100         WARN_ON_ONCE(!queue_is_mq(q));
4101         WARN_ON_ONCE(blk_queue_registered(q));
4102
4103         might_sleep();
4104
4105         blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4106         blk_queue_start_drain(q);
4107         blk_mq_freeze_queue_wait(q);
4108
4109         blk_sync_queue(q);
4110         blk_mq_cancel_work_sync(q);
4111         blk_mq_exit_queue(q);
4112 }
4113 EXPORT_SYMBOL(blk_mq_destroy_queue);
4114
4115 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
4116                 struct lock_class_key *lkclass)
4117 {
4118         struct request_queue *q;
4119         struct gendisk *disk;
4120
4121         q = blk_mq_init_queue_data(set, queuedata);
4122         if (IS_ERR(q))
4123                 return ERR_CAST(q);
4124
4125         disk = __alloc_disk_node(q, set->numa_node, lkclass);
4126         if (!disk) {
4127                 blk_mq_destroy_queue(q);
4128                 blk_put_queue(q);
4129                 return ERR_PTR(-ENOMEM);
4130         }
4131         set_bit(GD_OWNS_QUEUE, &disk->state);
4132         return disk;
4133 }
4134 EXPORT_SYMBOL(__blk_mq_alloc_disk);
4135
4136 struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4137                 struct lock_class_key *lkclass)
4138 {
4139         struct gendisk *disk;
4140
4141         if (!blk_get_queue(q))
4142                 return NULL;
4143         disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4144         if (!disk)
4145                 blk_put_queue(q);
4146         return disk;
4147 }
4148 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4149
4150 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4151                 struct blk_mq_tag_set *set, struct request_queue *q,
4152                 int hctx_idx, int node)
4153 {
4154         struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4155
4156         /* reuse dead hctx first */
4157         spin_lock(&q->unused_hctx_lock);
4158         list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4159                 if (tmp->numa_node == node) {
4160                         hctx = tmp;
4161                         break;
4162                 }
4163         }
4164         if (hctx)
4165                 list_del_init(&hctx->hctx_list);
4166         spin_unlock(&q->unused_hctx_lock);
4167
4168         if (!hctx)
4169                 hctx = blk_mq_alloc_hctx(q, set, node);
4170         if (!hctx)
4171                 goto fail;
4172
4173         if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4174                 goto free_hctx;
4175
4176         return hctx;
4177
4178  free_hctx:
4179         kobject_put(&hctx->kobj);
4180  fail:
4181         return NULL;
4182 }
4183
4184 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4185                                                 struct request_queue *q)
4186 {
4187         struct blk_mq_hw_ctx *hctx;
4188         unsigned long i, j;
4189
4190         /* protect against switching io scheduler  */
4191         mutex_lock(&q->sysfs_lock);
4192         for (i = 0; i < set->nr_hw_queues; i++) {
4193                 int old_node;
4194                 int node = blk_mq_get_hctx_node(set, i);
4195                 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4196
4197                 if (old_hctx) {
4198                         old_node = old_hctx->numa_node;
4199                         blk_mq_exit_hctx(q, set, old_hctx, i);
4200                 }
4201
4202                 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4203                         if (!old_hctx)
4204                                 break;
4205                         pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4206                                         node, old_node);
4207                         hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4208                         WARN_ON_ONCE(!hctx);
4209                 }
4210         }
4211         /*
4212          * Increasing nr_hw_queues fails. Free the newly allocated
4213          * hctxs and keep the previous q->nr_hw_queues.
4214          */
4215         if (i != set->nr_hw_queues) {
4216                 j = q->nr_hw_queues;
4217         } else {
4218                 j = i;
4219                 q->nr_hw_queues = set->nr_hw_queues;
4220         }
4221
4222         xa_for_each_start(&q->hctx_table, j, hctx, j)
4223                 blk_mq_exit_hctx(q, set, hctx, j);
4224         mutex_unlock(&q->sysfs_lock);
4225 }
4226
4227 static void blk_mq_update_poll_flag(struct request_queue *q)
4228 {
4229         struct blk_mq_tag_set *set = q->tag_set;
4230
4231         if (set->nr_maps > HCTX_TYPE_POLL &&
4232             set->map[HCTX_TYPE_POLL].nr_queues)
4233                 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4234         else
4235                 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4236 }
4237
4238 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4239                 struct request_queue *q)
4240 {
4241         /* mark the queue as mq asap */
4242         q->mq_ops = set->ops;
4243
4244         if (blk_mq_alloc_ctxs(q))
4245                 goto err_exit;
4246
4247         /* init q->mq_kobj and sw queues' kobjects */
4248         blk_mq_sysfs_init(q);
4249
4250         INIT_LIST_HEAD(&q->unused_hctx_list);
4251         spin_lock_init(&q->unused_hctx_lock);
4252
4253         xa_init(&q->hctx_table);
4254
4255         blk_mq_realloc_hw_ctxs(set, q);
4256         if (!q->nr_hw_queues)
4257                 goto err_hctxs;
4258
4259         INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4260         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4261
4262         q->tag_set = set;
4263
4264         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4265         blk_mq_update_poll_flag(q);
4266
4267         INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4268         INIT_LIST_HEAD(&q->flush_list);
4269         INIT_LIST_HEAD(&q->requeue_list);
4270         spin_lock_init(&q->requeue_lock);
4271
4272         q->nr_requests = set->queue_depth;
4273
4274         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4275         blk_mq_add_queue_tag_set(set, q);
4276         blk_mq_map_swqueue(q);
4277         return 0;
4278
4279 err_hctxs:
4280         blk_mq_release(q);
4281 err_exit:
4282         q->mq_ops = NULL;
4283         return -ENOMEM;
4284 }
4285 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4286
4287 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4288 void blk_mq_exit_queue(struct request_queue *q)
4289 {
4290         struct blk_mq_tag_set *set = q->tag_set;
4291
4292         /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4293         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4294         /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4295         blk_mq_del_queue_tag_set(q);
4296 }
4297
4298 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4299 {
4300         int i;
4301
4302         if (blk_mq_is_shared_tags(set->flags)) {
4303                 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4304                                                 BLK_MQ_NO_HCTX_IDX,
4305                                                 set->queue_depth);
4306                 if (!set->shared_tags)
4307                         return -ENOMEM;
4308         }
4309
4310         for (i = 0; i < set->nr_hw_queues; i++) {
4311                 if (!__blk_mq_alloc_map_and_rqs(set, i))
4312                         goto out_unwind;
4313                 cond_resched();
4314         }
4315
4316         return 0;
4317
4318 out_unwind:
4319         while (--i >= 0)
4320                 __blk_mq_free_map_and_rqs(set, i);
4321
4322         if (blk_mq_is_shared_tags(set->flags)) {
4323                 blk_mq_free_map_and_rqs(set, set->shared_tags,
4324                                         BLK_MQ_NO_HCTX_IDX);
4325         }
4326
4327         return -ENOMEM;
4328 }
4329
4330 /*
4331  * Allocate the request maps associated with this tag_set. Note that this
4332  * may reduce the depth asked for, if memory is tight. set->queue_depth
4333  * will be updated to reflect the allocated depth.
4334  */
4335 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4336 {
4337         unsigned int depth;
4338         int err;
4339
4340         depth = set->queue_depth;
4341         do {
4342                 err = __blk_mq_alloc_rq_maps(set);
4343                 if (!err)
4344                         break;
4345
4346                 set->queue_depth >>= 1;
4347                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4348                         err = -ENOMEM;
4349                         break;
4350                 }
4351         } while (set->queue_depth);
4352
4353         if (!set->queue_depth || err) {
4354                 pr_err("blk-mq: failed to allocate request map\n");
4355                 return -ENOMEM;
4356         }
4357
4358         if (depth != set->queue_depth)
4359                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4360                                                 depth, set->queue_depth);
4361
4362         return 0;
4363 }
4364
4365 static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4366 {
4367         /*
4368          * blk_mq_map_queues() and multiple .map_queues() implementations
4369          * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4370          * number of hardware queues.
4371          */
4372         if (set->nr_maps == 1)
4373                 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4374
4375         if (set->ops->map_queues && !is_kdump_kernel()) {
4376                 int i;
4377
4378                 /*
4379                  * transport .map_queues is usually done in the following
4380                  * way:
4381                  *
4382                  * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4383                  *      mask = get_cpu_mask(queue)
4384                  *      for_each_cpu(cpu, mask)
4385                  *              set->map[x].mq_map[cpu] = queue;
4386                  * }
4387                  *
4388                  * When we need to remap, the table has to be cleared for
4389                  * killing stale mapping since one CPU may not be mapped
4390                  * to any hw queue.
4391                  */
4392                 for (i = 0; i < set->nr_maps; i++)
4393                         blk_mq_clear_mq_map(&set->map[i]);
4394
4395                 set->ops->map_queues(set);
4396         } else {
4397                 BUG_ON(set->nr_maps > 1);
4398                 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4399         }
4400 }
4401
4402 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4403                                        int new_nr_hw_queues)
4404 {
4405         struct blk_mq_tags **new_tags;
4406         int i;
4407
4408         if (set->nr_hw_queues >= new_nr_hw_queues)
4409                 goto done;
4410
4411         new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4412                                 GFP_KERNEL, set->numa_node);
4413         if (!new_tags)
4414                 return -ENOMEM;
4415
4416         if (set->tags)
4417                 memcpy(new_tags, set->tags, set->nr_hw_queues *
4418                        sizeof(*set->tags));
4419         kfree(set->tags);
4420         set->tags = new_tags;
4421
4422         for (i = set->nr_hw_queues; i < new_nr_hw_queues; i++) {
4423                 if (!__blk_mq_alloc_map_and_rqs(set, i)) {
4424                         while (--i >= set->nr_hw_queues)
4425                                 __blk_mq_free_map_and_rqs(set, i);
4426                         return -ENOMEM;
4427                 }
4428                 cond_resched();
4429         }
4430
4431 done:
4432         set->nr_hw_queues = new_nr_hw_queues;
4433         return 0;
4434 }
4435
4436 /*
4437  * Alloc a tag set to be associated with one or more request queues.
4438  * May fail with EINVAL for various error conditions. May adjust the
4439  * requested depth down, if it's too large. In that case, the set
4440  * value will be stored in set->queue_depth.
4441  */
4442 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4443 {
4444         int i, ret;
4445
4446         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4447
4448         if (!set->nr_hw_queues)
4449                 return -EINVAL;
4450         if (!set->queue_depth)
4451                 return -EINVAL;
4452         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4453                 return -EINVAL;
4454
4455         if (!set->ops->queue_rq)
4456                 return -EINVAL;
4457
4458         if (!set->ops->get_budget ^ !set->ops->put_budget)
4459                 return -EINVAL;
4460
4461         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4462                 pr_info("blk-mq: reduced tag depth to %u\n",
4463                         BLK_MQ_MAX_DEPTH);
4464                 set->queue_depth = BLK_MQ_MAX_DEPTH;
4465         }
4466
4467         if (!set->nr_maps)
4468                 set->nr_maps = 1;
4469         else if (set->nr_maps > HCTX_MAX_TYPES)
4470                 return -EINVAL;
4471
4472         /*
4473          * If a crashdump is active, then we are potentially in a very
4474          * memory constrained environment. Limit us to 1 queue and
4475          * 64 tags to prevent using too much memory.
4476          */
4477         if (is_kdump_kernel()) {
4478                 set->nr_hw_queues = 1;
4479                 set->nr_maps = 1;
4480                 set->queue_depth = min(64U, set->queue_depth);
4481         }
4482         /*
4483          * There is no use for more h/w queues than cpus if we just have
4484          * a single map
4485          */
4486         if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4487                 set->nr_hw_queues = nr_cpu_ids;
4488
4489         if (set->flags & BLK_MQ_F_BLOCKING) {
4490                 set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL);
4491                 if (!set->srcu)
4492                         return -ENOMEM;
4493                 ret = init_srcu_struct(set->srcu);
4494                 if (ret)
4495                         goto out_free_srcu;
4496         }
4497
4498         ret = -ENOMEM;
4499         set->tags = kcalloc_node(set->nr_hw_queues,
4500                                  sizeof(struct blk_mq_tags *), GFP_KERNEL,
4501                                  set->numa_node);
4502         if (!set->tags)
4503                 goto out_cleanup_srcu;
4504
4505         for (i = 0; i < set->nr_maps; i++) {
4506                 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4507                                                   sizeof(set->map[i].mq_map[0]),
4508                                                   GFP_KERNEL, set->numa_node);
4509                 if (!set->map[i].mq_map)
4510                         goto out_free_mq_map;
4511                 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4512         }
4513
4514         blk_mq_update_queue_map(set);
4515
4516         ret = blk_mq_alloc_set_map_and_rqs(set);
4517         if (ret)
4518                 goto out_free_mq_map;
4519
4520         mutex_init(&set->tag_list_lock);
4521         INIT_LIST_HEAD(&set->tag_list);
4522
4523         return 0;
4524
4525 out_free_mq_map:
4526         for (i = 0; i < set->nr_maps; i++) {
4527                 kfree(set->map[i].mq_map);
4528                 set->map[i].mq_map = NULL;
4529         }
4530         kfree(set->tags);
4531         set->tags = NULL;
4532 out_cleanup_srcu:
4533         if (set->flags & BLK_MQ_F_BLOCKING)
4534                 cleanup_srcu_struct(set->srcu);
4535 out_free_srcu:
4536         if (set->flags & BLK_MQ_F_BLOCKING)
4537                 kfree(set->srcu);
4538         return ret;
4539 }
4540 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4541
4542 /* allocate and initialize a tagset for a simple single-queue device */
4543 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4544                 const struct blk_mq_ops *ops, unsigned int queue_depth,
4545                 unsigned int set_flags)
4546 {
4547         memset(set, 0, sizeof(*set));
4548         set->ops = ops;
4549         set->nr_hw_queues = 1;
4550         set->nr_maps = 1;
4551         set->queue_depth = queue_depth;
4552         set->numa_node = NUMA_NO_NODE;
4553         set->flags = set_flags;
4554         return blk_mq_alloc_tag_set(set);
4555 }
4556 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4557
4558 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4559 {
4560         int i, j;
4561
4562         for (i = 0; i < set->nr_hw_queues; i++)
4563                 __blk_mq_free_map_and_rqs(set, i);
4564
4565         if (blk_mq_is_shared_tags(set->flags)) {
4566                 blk_mq_free_map_and_rqs(set, set->shared_tags,
4567                                         BLK_MQ_NO_HCTX_IDX);
4568         }
4569
4570         for (j = 0; j < set->nr_maps; j++) {
4571                 kfree(set->map[j].mq_map);
4572                 set->map[j].mq_map = NULL;
4573         }
4574
4575         kfree(set->tags);
4576         set->tags = NULL;
4577         if (set->flags & BLK_MQ_F_BLOCKING) {
4578                 cleanup_srcu_struct(set->srcu);
4579                 kfree(set->srcu);
4580         }
4581 }
4582 EXPORT_SYMBOL(blk_mq_free_tag_set);
4583
4584 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4585 {
4586         struct blk_mq_tag_set *set = q->tag_set;
4587         struct blk_mq_hw_ctx *hctx;
4588         int ret;
4589         unsigned long i;
4590
4591         if (!set)
4592                 return -EINVAL;
4593
4594         if (q->nr_requests == nr)
4595                 return 0;
4596
4597         blk_mq_freeze_queue(q);
4598         blk_mq_quiesce_queue(q);
4599
4600         ret = 0;
4601         queue_for_each_hw_ctx(q, hctx, i) {
4602                 if (!hctx->tags)
4603                         continue;
4604                 /*
4605                  * If we're using an MQ scheduler, just update the scheduler
4606                  * queue depth. This is similar to what the old code would do.
4607                  */
4608                 if (hctx->sched_tags) {
4609                         ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4610                                                       nr, true);
4611                 } else {
4612                         ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4613                                                       false);
4614                 }
4615                 if (ret)
4616                         break;
4617                 if (q->elevator && q->elevator->type->ops.depth_updated)
4618                         q->elevator->type->ops.depth_updated(hctx);
4619         }
4620         if (!ret) {
4621                 q->nr_requests = nr;
4622                 if (blk_mq_is_shared_tags(set->flags)) {
4623                         if (q->elevator)
4624                                 blk_mq_tag_update_sched_shared_tags(q);
4625                         else
4626                                 blk_mq_tag_resize_shared_tags(set, nr);
4627                 }
4628         }
4629
4630         blk_mq_unquiesce_queue(q);
4631         blk_mq_unfreeze_queue(q);
4632
4633         return ret;
4634 }
4635
4636 /*
4637  * request_queue and elevator_type pair.
4638  * It is just used by __blk_mq_update_nr_hw_queues to cache
4639  * the elevator_type associated with a request_queue.
4640  */
4641 struct blk_mq_qe_pair {
4642         struct list_head node;
4643         struct request_queue *q;
4644         struct elevator_type *type;
4645 };
4646
4647 /*
4648  * Cache the elevator_type in qe pair list and switch the
4649  * io scheduler to 'none'
4650  */
4651 static bool blk_mq_elv_switch_none(struct list_head *head,
4652                 struct request_queue *q)
4653 {
4654         struct blk_mq_qe_pair *qe;
4655
4656         qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4657         if (!qe)
4658                 return false;
4659
4660         /* q->elevator needs protection from ->sysfs_lock */
4661         mutex_lock(&q->sysfs_lock);
4662
4663         /* the check has to be done with holding sysfs_lock */
4664         if (!q->elevator) {
4665                 kfree(qe);
4666                 goto unlock;
4667         }
4668
4669         INIT_LIST_HEAD(&qe->node);
4670         qe->q = q;
4671         qe->type = q->elevator->type;
4672         /* keep a reference to the elevator module as we'll switch back */
4673         __elevator_get(qe->type);
4674         list_add(&qe->node, head);
4675         elevator_disable(q);
4676 unlock:
4677         mutex_unlock(&q->sysfs_lock);
4678
4679         return true;
4680 }
4681
4682 static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4683                                                 struct request_queue *q)
4684 {
4685         struct blk_mq_qe_pair *qe;
4686
4687         list_for_each_entry(qe, head, node)
4688                 if (qe->q == q)
4689                         return qe;
4690
4691         return NULL;
4692 }
4693
4694 static void blk_mq_elv_switch_back(struct list_head *head,
4695                                   struct request_queue *q)
4696 {
4697         struct blk_mq_qe_pair *qe;
4698         struct elevator_type *t;
4699
4700         qe = blk_lookup_qe_pair(head, q);
4701         if (!qe)
4702                 return;
4703         t = qe->type;
4704         list_del(&qe->node);
4705         kfree(qe);
4706
4707         mutex_lock(&q->sysfs_lock);
4708         elevator_switch(q, t);
4709         /* drop the reference acquired in blk_mq_elv_switch_none */
4710         elevator_put(t);
4711         mutex_unlock(&q->sysfs_lock);
4712 }
4713
4714 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4715                                                         int nr_hw_queues)
4716 {
4717         struct request_queue *q;
4718         LIST_HEAD(head);
4719         int prev_nr_hw_queues = set->nr_hw_queues;
4720         int i;
4721
4722         lockdep_assert_held(&set->tag_list_lock);
4723
4724         if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4725                 nr_hw_queues = nr_cpu_ids;
4726         if (nr_hw_queues < 1)
4727                 return;
4728         if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4729                 return;
4730
4731         list_for_each_entry(q, &set->tag_list, tag_set_list)
4732                 blk_mq_freeze_queue(q);
4733         /*
4734          * Switch IO scheduler to 'none', cleaning up the data associated
4735          * with the previous scheduler. We will switch back once we are done
4736          * updating the new sw to hw queue mappings.
4737          */
4738         list_for_each_entry(q, &set->tag_list, tag_set_list)
4739                 if (!blk_mq_elv_switch_none(&head, q))
4740                         goto switch_back;
4741
4742         list_for_each_entry(q, &set->tag_list, tag_set_list) {
4743                 blk_mq_debugfs_unregister_hctxs(q);
4744                 blk_mq_sysfs_unregister_hctxs(q);
4745         }
4746
4747         if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0)
4748                 goto reregister;
4749
4750 fallback:
4751         blk_mq_update_queue_map(set);
4752         list_for_each_entry(q, &set->tag_list, tag_set_list) {
4753                 blk_mq_realloc_hw_ctxs(set, q);
4754                 blk_mq_update_poll_flag(q);
4755                 if (q->nr_hw_queues != set->nr_hw_queues) {
4756                         int i = prev_nr_hw_queues;
4757
4758                         pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4759                                         nr_hw_queues, prev_nr_hw_queues);
4760                         for (; i < set->nr_hw_queues; i++)
4761                                 __blk_mq_free_map_and_rqs(set, i);
4762
4763                         set->nr_hw_queues = prev_nr_hw_queues;
4764                         goto fallback;
4765                 }
4766                 blk_mq_map_swqueue(q);
4767         }
4768
4769 reregister:
4770         list_for_each_entry(q, &set->tag_list, tag_set_list) {
4771                 blk_mq_sysfs_register_hctxs(q);
4772                 blk_mq_debugfs_register_hctxs(q);
4773         }
4774
4775 switch_back:
4776         list_for_each_entry(q, &set->tag_list, tag_set_list)
4777                 blk_mq_elv_switch_back(&head, q);
4778
4779         list_for_each_entry(q, &set->tag_list, tag_set_list)
4780                 blk_mq_unfreeze_queue(q);
4781
4782         /* Free the excess tags when nr_hw_queues shrink. */
4783         for (i = set->nr_hw_queues; i < prev_nr_hw_queues; i++)
4784                 __blk_mq_free_map_and_rqs(set, i);
4785 }
4786
4787 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4788 {
4789         mutex_lock(&set->tag_list_lock);
4790         __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4791         mutex_unlock(&set->tag_list_lock);
4792 }
4793 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4794
4795 static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
4796                          struct io_comp_batch *iob, unsigned int flags)
4797 {
4798         long state = get_current_state();
4799         int ret;
4800
4801         do {
4802                 ret = q->mq_ops->poll(hctx, iob);
4803                 if (ret > 0) {
4804                         __set_current_state(TASK_RUNNING);
4805                         return ret;
4806                 }
4807
4808                 if (signal_pending_state(state, current))
4809                         __set_current_state(TASK_RUNNING);
4810                 if (task_is_running(current))
4811                         return 1;
4812
4813                 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4814                         break;
4815                 cpu_relax();
4816         } while (!need_resched());
4817
4818         __set_current_state(TASK_RUNNING);
4819         return 0;
4820 }
4821
4822 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie,
4823                 struct io_comp_batch *iob, unsigned int flags)
4824 {
4825         struct blk_mq_hw_ctx *hctx = xa_load(&q->hctx_table, cookie);
4826
4827         return blk_hctx_poll(q, hctx, iob, flags);
4828 }
4829
4830 int blk_rq_poll(struct request *rq, struct io_comp_batch *iob,
4831                 unsigned int poll_flags)
4832 {
4833         struct request_queue *q = rq->q;
4834         int ret;
4835
4836         if (!blk_rq_is_poll(rq))
4837                 return 0;
4838         if (!percpu_ref_tryget(&q->q_usage_counter))
4839                 return 0;
4840
4841         ret = blk_hctx_poll(q, rq->mq_hctx, iob, poll_flags);
4842         blk_queue_exit(q);
4843
4844         return ret;
4845 }
4846 EXPORT_SYMBOL_GPL(blk_rq_poll);
4847
4848 unsigned int blk_mq_rq_cpu(struct request *rq)
4849 {
4850         return rq->mq_ctx->cpu;
4851 }
4852 EXPORT_SYMBOL(blk_mq_rq_cpu);
4853
4854 void blk_mq_cancel_work_sync(struct request_queue *q)
4855 {
4856         struct blk_mq_hw_ctx *hctx;
4857         unsigned long i;
4858
4859         cancel_delayed_work_sync(&q->requeue_work);
4860
4861         queue_for_each_hw_ctx(q, hctx, i)
4862                 cancel_delayed_work_sync(&hctx->run_work);
4863 }
4864
4865 static int __init blk_mq_init(void)
4866 {
4867         int i;
4868
4869         for_each_possible_cpu(i)
4870                 init_llist_head(&per_cpu(blk_cpu_done, i));
4871         for_each_possible_cpu(i)
4872                 INIT_CSD(&per_cpu(blk_cpu_csd, i),
4873                          __blk_mq_complete_request_remote, NULL);
4874         open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4875
4876         cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4877                                   "block/softirq:dead", NULL,
4878                                   blk_softirq_cpu_dead);
4879         cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4880                                 blk_mq_hctx_notify_dead);
4881         cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4882                                 blk_mq_hctx_notify_online,
4883                                 blk_mq_hctx_notify_offline);
4884         return 0;
4885 }
4886 subsys_initcall(blk_mq_init);