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