Loading...
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/blk-mq.h>
36#include <linux/t10-pi.h>
37#include "blk.h"
38#include "blk-mq.h"
39#include "blk-mq-debugfs.h"
40#include "blk-mq-tag.h"
41#include "blk-pm.h"
42#include "blk-stat.h"
43#include "blk-mq-sched.h"
44#include "blk-rq-qos.h"
45#include "blk-ioprio.h"
46
47static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
48
49static void blk_mq_poll_stats_start(struct request_queue *q);
50static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
51
52static int blk_mq_poll_stats_bkt(const struct request *rq)
53{
54 int ddir, sectors, bucket;
55
56 ddir = rq_data_dir(rq);
57 sectors = blk_rq_stats_sectors(rq);
58
59 bucket = ddir + 2 * ilog2(sectors);
60
61 if (bucket < 0)
62 return -1;
63 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
64 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
65
66 return bucket;
67}
68
69#define BLK_QC_T_SHIFT 16
70#define BLK_QC_T_INTERNAL (1U << 31)
71
72static inline struct blk_mq_hw_ctx *blk_qc_to_hctx(struct request_queue *q,
73 blk_qc_t qc)
74{
75 return xa_load(&q->hctx_table,
76 (qc & ~BLK_QC_T_INTERNAL) >> BLK_QC_T_SHIFT);
77}
78
79static inline struct request *blk_qc_to_rq(struct blk_mq_hw_ctx *hctx,
80 blk_qc_t qc)
81{
82 unsigned int tag = qc & ((1U << BLK_QC_T_SHIFT) - 1);
83
84 if (qc & BLK_QC_T_INTERNAL)
85 return blk_mq_tag_to_rq(hctx->sched_tags, tag);
86 return blk_mq_tag_to_rq(hctx->tags, tag);
87}
88
89static inline blk_qc_t blk_rq_to_qc(struct request *rq)
90{
91 return (rq->mq_hctx->queue_num << BLK_QC_T_SHIFT) |
92 (rq->tag != -1 ?
93 rq->tag : (rq->internal_tag | BLK_QC_T_INTERNAL));
94}
95
96/*
97 * Check if any of the ctx, dispatch list or elevator
98 * have pending work in this hardware queue.
99 */
100static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
101{
102 return !list_empty_careful(&hctx->dispatch) ||
103 sbitmap_any_bit_set(&hctx->ctx_map) ||
104 blk_mq_sched_has_work(hctx);
105}
106
107/*
108 * Mark this ctx as having pending work in this hardware queue
109 */
110static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
111 struct blk_mq_ctx *ctx)
112{
113 const int bit = ctx->index_hw[hctx->type];
114
115 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
116 sbitmap_set_bit(&hctx->ctx_map, bit);
117}
118
119static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
120 struct blk_mq_ctx *ctx)
121{
122 const int bit = ctx->index_hw[hctx->type];
123
124 sbitmap_clear_bit(&hctx->ctx_map, bit);
125}
126
127struct mq_inflight {
128 struct block_device *part;
129 unsigned int inflight[2];
130};
131
132static bool blk_mq_check_inflight(struct request *rq, void *priv)
133{
134 struct mq_inflight *mi = priv;
135
136 if (rq->part && blk_do_io_stat(rq) &&
137 (!mi->part->bd_partno || rq->part == mi->part) &&
138 blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
139 mi->inflight[rq_data_dir(rq)]++;
140
141 return true;
142}
143
144unsigned int blk_mq_in_flight(struct request_queue *q,
145 struct block_device *part)
146{
147 struct mq_inflight mi = { .part = part };
148
149 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
150
151 return mi.inflight[0] + mi.inflight[1];
152}
153
154void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
155 unsigned int inflight[2])
156{
157 struct mq_inflight mi = { .part = part };
158
159 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
160 inflight[0] = mi.inflight[0];
161 inflight[1] = mi.inflight[1];
162}
163
164void blk_freeze_queue_start(struct request_queue *q)
165{
166 mutex_lock(&q->mq_freeze_lock);
167 if (++q->mq_freeze_depth == 1) {
168 percpu_ref_kill(&q->q_usage_counter);
169 mutex_unlock(&q->mq_freeze_lock);
170 if (queue_is_mq(q))
171 blk_mq_run_hw_queues(q, false);
172 } else {
173 mutex_unlock(&q->mq_freeze_lock);
174 }
175}
176EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
177
178void blk_mq_freeze_queue_wait(struct request_queue *q)
179{
180 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
181}
182EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
183
184int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
185 unsigned long timeout)
186{
187 return wait_event_timeout(q->mq_freeze_wq,
188 percpu_ref_is_zero(&q->q_usage_counter),
189 timeout);
190}
191EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
192
193/*
194 * Guarantee no request is in use, so we can change any data structure of
195 * the queue afterward.
196 */
197void blk_freeze_queue(struct request_queue *q)
198{
199 /*
200 * In the !blk_mq case we are only calling this to kill the
201 * q_usage_counter, otherwise this increases the freeze depth
202 * and waits for it to return to zero. For this reason there is
203 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
204 * exported to drivers as the only user for unfreeze is blk_mq.
205 */
206 blk_freeze_queue_start(q);
207 blk_mq_freeze_queue_wait(q);
208}
209
210void blk_mq_freeze_queue(struct request_queue *q)
211{
212 /*
213 * ...just an alias to keep freeze and unfreeze actions balanced
214 * in the blk_mq_* namespace
215 */
216 blk_freeze_queue(q);
217}
218EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
219
220void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
221{
222 mutex_lock(&q->mq_freeze_lock);
223 if (force_atomic)
224 q->q_usage_counter.data->force_atomic = true;
225 q->mq_freeze_depth--;
226 WARN_ON_ONCE(q->mq_freeze_depth < 0);
227 if (!q->mq_freeze_depth) {
228 percpu_ref_resurrect(&q->q_usage_counter);
229 wake_up_all(&q->mq_freeze_wq);
230 }
231 mutex_unlock(&q->mq_freeze_lock);
232}
233
234void blk_mq_unfreeze_queue(struct request_queue *q)
235{
236 __blk_mq_unfreeze_queue(q, false);
237}
238EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
239
240/*
241 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
242 * mpt3sas driver such that this function can be removed.
243 */
244void blk_mq_quiesce_queue_nowait(struct request_queue *q)
245{
246 unsigned long flags;
247
248 spin_lock_irqsave(&q->queue_lock, flags);
249 if (!q->quiesce_depth++)
250 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
251 spin_unlock_irqrestore(&q->queue_lock, flags);
252}
253EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
254
255/**
256 * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
257 * @set: tag_set to wait on
258 *
259 * Note: it is driver's responsibility for making sure that quiesce has
260 * been started on or more of the request_queues of the tag_set. This
261 * function only waits for the quiesce on those request_queues that had
262 * the quiesce flag set using blk_mq_quiesce_queue_nowait.
263 */
264void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set)
265{
266 if (set->flags & BLK_MQ_F_BLOCKING)
267 synchronize_srcu(set->srcu);
268 else
269 synchronize_rcu();
270}
271EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
272
273/**
274 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
275 * @q: request queue.
276 *
277 * Note: this function does not prevent that the struct request end_io()
278 * callback function is invoked. Once this function is returned, we make
279 * sure no dispatch can happen until the queue is unquiesced via
280 * blk_mq_unquiesce_queue().
281 */
282void blk_mq_quiesce_queue(struct request_queue *q)
283{
284 blk_mq_quiesce_queue_nowait(q);
285 /* nothing to wait for non-mq queues */
286 if (queue_is_mq(q))
287 blk_mq_wait_quiesce_done(q->tag_set);
288}
289EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
290
291/*
292 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
293 * @q: request queue.
294 *
295 * This function recovers queue into the state before quiescing
296 * which is done by blk_mq_quiesce_queue.
297 */
298void blk_mq_unquiesce_queue(struct request_queue *q)
299{
300 unsigned long flags;
301 bool run_queue = false;
302
303 spin_lock_irqsave(&q->queue_lock, flags);
304 if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
305 ;
306 } else if (!--q->quiesce_depth) {
307 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
308 run_queue = true;
309 }
310 spin_unlock_irqrestore(&q->queue_lock, flags);
311
312 /* dispatch requests which are inserted during quiescing */
313 if (run_queue)
314 blk_mq_run_hw_queues(q, true);
315}
316EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
317
318void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set)
319{
320 struct request_queue *q;
321
322 mutex_lock(&set->tag_list_lock);
323 list_for_each_entry(q, &set->tag_list, tag_set_list) {
324 if (!blk_queue_skip_tagset_quiesce(q))
325 blk_mq_quiesce_queue_nowait(q);
326 }
327 blk_mq_wait_quiesce_done(set);
328 mutex_unlock(&set->tag_list_lock);
329}
330EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset);
331
332void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set)
333{
334 struct request_queue *q;
335
336 mutex_lock(&set->tag_list_lock);
337 list_for_each_entry(q, &set->tag_list, tag_set_list) {
338 if (!blk_queue_skip_tagset_quiesce(q))
339 blk_mq_unquiesce_queue(q);
340 }
341 mutex_unlock(&set->tag_list_lock);
342}
343EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset);
344
345void blk_mq_wake_waiters(struct request_queue *q)
346{
347 struct blk_mq_hw_ctx *hctx;
348 unsigned long i;
349
350 queue_for_each_hw_ctx(q, hctx, i)
351 if (blk_mq_hw_queue_mapped(hctx))
352 blk_mq_tag_wakeup_all(hctx->tags, true);
353}
354
355void blk_rq_init(struct request_queue *q, struct request *rq)
356{
357 memset(rq, 0, sizeof(*rq));
358
359 INIT_LIST_HEAD(&rq->queuelist);
360 rq->q = q;
361 rq->__sector = (sector_t) -1;
362 INIT_HLIST_NODE(&rq->hash);
363 RB_CLEAR_NODE(&rq->rb_node);
364 rq->tag = BLK_MQ_NO_TAG;
365 rq->internal_tag = BLK_MQ_NO_TAG;
366 rq->start_time_ns = ktime_get_ns();
367 rq->part = NULL;
368 blk_crypto_rq_set_defaults(rq);
369}
370EXPORT_SYMBOL(blk_rq_init);
371
372static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
373 struct blk_mq_tags *tags, unsigned int tag, u64 alloc_time_ns)
374{
375 struct blk_mq_ctx *ctx = data->ctx;
376 struct blk_mq_hw_ctx *hctx = data->hctx;
377 struct request_queue *q = data->q;
378 struct request *rq = tags->static_rqs[tag];
379
380 rq->q = q;
381 rq->mq_ctx = ctx;
382 rq->mq_hctx = hctx;
383 rq->cmd_flags = data->cmd_flags;
384
385 if (data->flags & BLK_MQ_REQ_PM)
386 data->rq_flags |= RQF_PM;
387 if (blk_queue_io_stat(q))
388 data->rq_flags |= RQF_IO_STAT;
389 rq->rq_flags = data->rq_flags;
390
391 if (!(data->rq_flags & RQF_ELV)) {
392 rq->tag = tag;
393 rq->internal_tag = BLK_MQ_NO_TAG;
394 } else {
395 rq->tag = BLK_MQ_NO_TAG;
396 rq->internal_tag = tag;
397 }
398 rq->timeout = 0;
399
400 if (blk_mq_need_time_stamp(rq))
401 rq->start_time_ns = ktime_get_ns();
402 else
403 rq->start_time_ns = 0;
404 rq->part = NULL;
405#ifdef CONFIG_BLK_RQ_ALLOC_TIME
406 rq->alloc_time_ns = alloc_time_ns;
407#endif
408 rq->io_start_time_ns = 0;
409 rq->stats_sectors = 0;
410 rq->nr_phys_segments = 0;
411#if defined(CONFIG_BLK_DEV_INTEGRITY)
412 rq->nr_integrity_segments = 0;
413#endif
414 rq->end_io = NULL;
415 rq->end_io_data = NULL;
416
417 blk_crypto_rq_set_defaults(rq);
418 INIT_LIST_HEAD(&rq->queuelist);
419 /* tag was already set */
420 WRITE_ONCE(rq->deadline, 0);
421 req_ref_set(rq, 1);
422
423 if (rq->rq_flags & RQF_ELV) {
424 struct elevator_queue *e = data->q->elevator;
425
426 INIT_HLIST_NODE(&rq->hash);
427 RB_CLEAR_NODE(&rq->rb_node);
428
429 if (!op_is_flush(data->cmd_flags) &&
430 e->type->ops.prepare_request) {
431 e->type->ops.prepare_request(rq);
432 rq->rq_flags |= RQF_ELVPRIV;
433 }
434 }
435
436 return rq;
437}
438
439static inline struct request *
440__blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data,
441 u64 alloc_time_ns)
442{
443 unsigned int tag, tag_offset;
444 struct blk_mq_tags *tags;
445 struct request *rq;
446 unsigned long tag_mask;
447 int i, nr = 0;
448
449 tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset);
450 if (unlikely(!tag_mask))
451 return NULL;
452
453 tags = blk_mq_tags_from_data(data);
454 for (i = 0; tag_mask; i++) {
455 if (!(tag_mask & (1UL << i)))
456 continue;
457 tag = tag_offset + i;
458 prefetch(tags->static_rqs[tag]);
459 tag_mask &= ~(1UL << i);
460 rq = blk_mq_rq_ctx_init(data, tags, tag, alloc_time_ns);
461 rq_list_add(data->cached_rq, rq);
462 nr++;
463 }
464 /* caller already holds a reference, add for remainder */
465 percpu_ref_get_many(&data->q->q_usage_counter, nr - 1);
466 data->nr_tags -= nr;
467
468 return rq_list_pop(data->cached_rq);
469}
470
471static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
472{
473 struct request_queue *q = data->q;
474 u64 alloc_time_ns = 0;
475 struct request *rq;
476 unsigned int tag;
477
478 /* alloc_time includes depth and tag waits */
479 if (blk_queue_rq_alloc_time(q))
480 alloc_time_ns = ktime_get_ns();
481
482 if (data->cmd_flags & REQ_NOWAIT)
483 data->flags |= BLK_MQ_REQ_NOWAIT;
484
485 if (q->elevator) {
486 struct elevator_queue *e = q->elevator;
487
488 data->rq_flags |= RQF_ELV;
489
490 /*
491 * Flush/passthrough requests are special and go directly to the
492 * dispatch list. Don't include reserved tags in the
493 * limiting, as it isn't useful.
494 */
495 if (!op_is_flush(data->cmd_flags) &&
496 !blk_op_is_passthrough(data->cmd_flags) &&
497 e->type->ops.limit_depth &&
498 !(data->flags & BLK_MQ_REQ_RESERVED))
499 e->type->ops.limit_depth(data->cmd_flags, data);
500 }
501
502retry:
503 data->ctx = blk_mq_get_ctx(q);
504 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
505 if (!(data->rq_flags & RQF_ELV))
506 blk_mq_tag_busy(data->hctx);
507
508 if (data->flags & BLK_MQ_REQ_RESERVED)
509 data->rq_flags |= RQF_RESV;
510
511 /*
512 * Try batched alloc if we want more than 1 tag.
513 */
514 if (data->nr_tags > 1) {
515 rq = __blk_mq_alloc_requests_batch(data, alloc_time_ns);
516 if (rq)
517 return rq;
518 data->nr_tags = 1;
519 }
520
521 /*
522 * Waiting allocations only fail because of an inactive hctx. In that
523 * case just retry the hctx assignment and tag allocation as CPU hotplug
524 * should have migrated us to an online CPU by now.
525 */
526 tag = blk_mq_get_tag(data);
527 if (tag == BLK_MQ_NO_TAG) {
528 if (data->flags & BLK_MQ_REQ_NOWAIT)
529 return NULL;
530 /*
531 * Give up the CPU and sleep for a random short time to
532 * ensure that thread using a realtime scheduling class
533 * are migrated off the CPU, and thus off the hctx that
534 * is going away.
535 */
536 msleep(3);
537 goto retry;
538 }
539
540 return blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag,
541 alloc_time_ns);
542}
543
544static struct request *blk_mq_rq_cache_fill(struct request_queue *q,
545 struct blk_plug *plug,
546 blk_opf_t opf,
547 blk_mq_req_flags_t flags)
548{
549 struct blk_mq_alloc_data data = {
550 .q = q,
551 .flags = flags,
552 .cmd_flags = opf,
553 .nr_tags = plug->nr_ios,
554 .cached_rq = &plug->cached_rq,
555 };
556 struct request *rq;
557
558 if (blk_queue_enter(q, flags))
559 return NULL;
560
561 plug->nr_ios = 1;
562
563 rq = __blk_mq_alloc_requests(&data);
564 if (unlikely(!rq))
565 blk_queue_exit(q);
566 return rq;
567}
568
569static struct request *blk_mq_alloc_cached_request(struct request_queue *q,
570 blk_opf_t opf,
571 blk_mq_req_flags_t flags)
572{
573 struct blk_plug *plug = current->plug;
574 struct request *rq;
575
576 if (!plug)
577 return NULL;
578
579 if (rq_list_empty(plug->cached_rq)) {
580 if (plug->nr_ios == 1)
581 return NULL;
582 rq = blk_mq_rq_cache_fill(q, plug, opf, flags);
583 if (!rq)
584 return NULL;
585 } else {
586 rq = rq_list_peek(&plug->cached_rq);
587 if (!rq || rq->q != q)
588 return NULL;
589
590 if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type)
591 return NULL;
592 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
593 return NULL;
594
595 plug->cached_rq = rq_list_next(rq);
596 }
597
598 rq->cmd_flags = opf;
599 INIT_LIST_HEAD(&rq->queuelist);
600 return rq;
601}
602
603struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf,
604 blk_mq_req_flags_t flags)
605{
606 struct request *rq;
607
608 rq = blk_mq_alloc_cached_request(q, opf, flags);
609 if (!rq) {
610 struct blk_mq_alloc_data data = {
611 .q = q,
612 .flags = flags,
613 .cmd_flags = opf,
614 .nr_tags = 1,
615 };
616 int ret;
617
618 ret = blk_queue_enter(q, flags);
619 if (ret)
620 return ERR_PTR(ret);
621
622 rq = __blk_mq_alloc_requests(&data);
623 if (!rq)
624 goto out_queue_exit;
625 }
626 rq->__data_len = 0;
627 rq->__sector = (sector_t) -1;
628 rq->bio = rq->biotail = NULL;
629 return rq;
630out_queue_exit:
631 blk_queue_exit(q);
632 return ERR_PTR(-EWOULDBLOCK);
633}
634EXPORT_SYMBOL(blk_mq_alloc_request);
635
636struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
637 blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx)
638{
639 struct blk_mq_alloc_data data = {
640 .q = q,
641 .flags = flags,
642 .cmd_flags = opf,
643 .nr_tags = 1,
644 };
645 u64 alloc_time_ns = 0;
646 struct request *rq;
647 unsigned int cpu;
648 unsigned int tag;
649 int ret;
650
651 /* alloc_time includes depth and tag waits */
652 if (blk_queue_rq_alloc_time(q))
653 alloc_time_ns = ktime_get_ns();
654
655 /*
656 * If the tag allocator sleeps we could get an allocation for a
657 * different hardware context. No need to complicate the low level
658 * allocator for this for the rare use case of a command tied to
659 * a specific queue.
660 */
661 if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED))))
662 return ERR_PTR(-EINVAL);
663
664 if (hctx_idx >= q->nr_hw_queues)
665 return ERR_PTR(-EIO);
666
667 ret = blk_queue_enter(q, flags);
668 if (ret)
669 return ERR_PTR(ret);
670
671 /*
672 * Check if the hardware context is actually mapped to anything.
673 * If not tell the caller that it should skip this queue.
674 */
675 ret = -EXDEV;
676 data.hctx = xa_load(&q->hctx_table, hctx_idx);
677 if (!blk_mq_hw_queue_mapped(data.hctx))
678 goto out_queue_exit;
679 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
680 if (cpu >= nr_cpu_ids)
681 goto out_queue_exit;
682 data.ctx = __blk_mq_get_ctx(q, cpu);
683
684 if (!q->elevator)
685 blk_mq_tag_busy(data.hctx);
686 else
687 data.rq_flags |= RQF_ELV;
688
689 if (flags & BLK_MQ_REQ_RESERVED)
690 data.rq_flags |= RQF_RESV;
691
692 ret = -EWOULDBLOCK;
693 tag = blk_mq_get_tag(&data);
694 if (tag == BLK_MQ_NO_TAG)
695 goto out_queue_exit;
696 rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag,
697 alloc_time_ns);
698 rq->__data_len = 0;
699 rq->__sector = (sector_t) -1;
700 rq->bio = rq->biotail = NULL;
701 return rq;
702
703out_queue_exit:
704 blk_queue_exit(q);
705 return ERR_PTR(ret);
706}
707EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
708
709static void __blk_mq_free_request(struct request *rq)
710{
711 struct request_queue *q = rq->q;
712 struct blk_mq_ctx *ctx = rq->mq_ctx;
713 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
714 const int sched_tag = rq->internal_tag;
715
716 blk_crypto_free_request(rq);
717 blk_pm_mark_last_busy(rq);
718 rq->mq_hctx = NULL;
719 if (rq->tag != BLK_MQ_NO_TAG)
720 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
721 if (sched_tag != BLK_MQ_NO_TAG)
722 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
723 blk_mq_sched_restart(hctx);
724 blk_queue_exit(q);
725}
726
727void blk_mq_free_request(struct request *rq)
728{
729 struct request_queue *q = rq->q;
730 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
731
732 if ((rq->rq_flags & RQF_ELVPRIV) &&
733 q->elevator->type->ops.finish_request)
734 q->elevator->type->ops.finish_request(rq);
735
736 if (rq->rq_flags & RQF_MQ_INFLIGHT)
737 __blk_mq_dec_active_requests(hctx);
738
739 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
740 laptop_io_completion(q->disk->bdi);
741
742 rq_qos_done(q, rq);
743
744 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
745 if (req_ref_put_and_test(rq))
746 __blk_mq_free_request(rq);
747}
748EXPORT_SYMBOL_GPL(blk_mq_free_request);
749
750void blk_mq_free_plug_rqs(struct blk_plug *plug)
751{
752 struct request *rq;
753
754 while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
755 blk_mq_free_request(rq);
756}
757
758void blk_dump_rq_flags(struct request *rq, char *msg)
759{
760 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
761 rq->q->disk ? rq->q->disk->disk_name : "?",
762 (__force unsigned long long) rq->cmd_flags);
763
764 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
765 (unsigned long long)blk_rq_pos(rq),
766 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
767 printk(KERN_INFO " bio %p, biotail %p, len %u\n",
768 rq->bio, rq->biotail, blk_rq_bytes(rq));
769}
770EXPORT_SYMBOL(blk_dump_rq_flags);
771
772static void req_bio_endio(struct request *rq, struct bio *bio,
773 unsigned int nbytes, blk_status_t error)
774{
775 if (unlikely(error)) {
776 bio->bi_status = error;
777 } else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
778 /*
779 * Partial zone append completions cannot be supported as the
780 * BIO fragments may end up not being written sequentially.
781 */
782 if (bio->bi_iter.bi_size != nbytes)
783 bio->bi_status = BLK_STS_IOERR;
784 else
785 bio->bi_iter.bi_sector = rq->__sector;
786 }
787
788 bio_advance(bio, nbytes);
789
790 if (unlikely(rq->rq_flags & RQF_QUIET))
791 bio_set_flag(bio, BIO_QUIET);
792 /* don't actually finish bio if it's part of flush sequence */
793 if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
794 bio_endio(bio);
795}
796
797static void blk_account_io_completion(struct request *req, unsigned int bytes)
798{
799 if (req->part && blk_do_io_stat(req)) {
800 const int sgrp = op_stat_group(req_op(req));
801
802 part_stat_lock();
803 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
804 part_stat_unlock();
805 }
806}
807
808static void blk_print_req_error(struct request *req, blk_status_t status)
809{
810 printk_ratelimited(KERN_ERR
811 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
812 "phys_seg %u prio class %u\n",
813 blk_status_to_str(status),
814 req->q->disk ? req->q->disk->disk_name : "?",
815 blk_rq_pos(req), (__force u32)req_op(req),
816 blk_op_str(req_op(req)),
817 (__force u32)(req->cmd_flags & ~REQ_OP_MASK),
818 req->nr_phys_segments,
819 IOPRIO_PRIO_CLASS(req->ioprio));
820}
821
822/*
823 * Fully end IO on a request. Does not support partial completions, or
824 * errors.
825 */
826static void blk_complete_request(struct request *req)
827{
828 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
829 int total_bytes = blk_rq_bytes(req);
830 struct bio *bio = req->bio;
831
832 trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
833
834 if (!bio)
835 return;
836
837#ifdef CONFIG_BLK_DEV_INTEGRITY
838 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
839 req->q->integrity.profile->complete_fn(req, total_bytes);
840#endif
841
842 blk_account_io_completion(req, total_bytes);
843
844 do {
845 struct bio *next = bio->bi_next;
846
847 /* Completion has already been traced */
848 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
849
850 if (req_op(req) == REQ_OP_ZONE_APPEND)
851 bio->bi_iter.bi_sector = req->__sector;
852
853 if (!is_flush)
854 bio_endio(bio);
855 bio = next;
856 } while (bio);
857
858 /*
859 * Reset counters so that the request stacking driver
860 * can find how many bytes remain in the request
861 * later.
862 */
863 if (!req->end_io) {
864 req->bio = NULL;
865 req->__data_len = 0;
866 }
867}
868
869/**
870 * blk_update_request - Complete multiple bytes without completing the request
871 * @req: the request being processed
872 * @error: block status code
873 * @nr_bytes: number of bytes to complete for @req
874 *
875 * Description:
876 * Ends I/O on a number of bytes attached to @req, but doesn't complete
877 * the request structure even if @req doesn't have leftover.
878 * If @req has leftover, sets it up for the next range of segments.
879 *
880 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
881 * %false return from this function.
882 *
883 * Note:
884 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
885 * except in the consistency check at the end of this function.
886 *
887 * Return:
888 * %false - this request doesn't have any more data
889 * %true - this request has more data
890 **/
891bool blk_update_request(struct request *req, blk_status_t error,
892 unsigned int nr_bytes)
893{
894 int total_bytes;
895
896 trace_block_rq_complete(req, error, nr_bytes);
897
898 if (!req->bio)
899 return false;
900
901#ifdef CONFIG_BLK_DEV_INTEGRITY
902 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
903 error == BLK_STS_OK)
904 req->q->integrity.profile->complete_fn(req, nr_bytes);
905#endif
906
907 if (unlikely(error && !blk_rq_is_passthrough(req) &&
908 !(req->rq_flags & RQF_QUIET)) &&
909 !test_bit(GD_DEAD, &req->q->disk->state)) {
910 blk_print_req_error(req, error);
911 trace_block_rq_error(req, error, nr_bytes);
912 }
913
914 blk_account_io_completion(req, nr_bytes);
915
916 total_bytes = 0;
917 while (req->bio) {
918 struct bio *bio = req->bio;
919 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
920
921 if (bio_bytes == bio->bi_iter.bi_size)
922 req->bio = bio->bi_next;
923
924 /* Completion has already been traced */
925 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
926 req_bio_endio(req, bio, bio_bytes, error);
927
928 total_bytes += bio_bytes;
929 nr_bytes -= bio_bytes;
930
931 if (!nr_bytes)
932 break;
933 }
934
935 /*
936 * completely done
937 */
938 if (!req->bio) {
939 /*
940 * Reset counters so that the request stacking driver
941 * can find how many bytes remain in the request
942 * later.
943 */
944 req->__data_len = 0;
945 return false;
946 }
947
948 req->__data_len -= total_bytes;
949
950 /* update sector only for requests with clear definition of sector */
951 if (!blk_rq_is_passthrough(req))
952 req->__sector += total_bytes >> 9;
953
954 /* mixed attributes always follow the first bio */
955 if (req->rq_flags & RQF_MIXED_MERGE) {
956 req->cmd_flags &= ~REQ_FAILFAST_MASK;
957 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
958 }
959
960 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
961 /*
962 * If total number of sectors is less than the first segment
963 * size, something has gone terribly wrong.
964 */
965 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
966 blk_dump_rq_flags(req, "request botched");
967 req->__data_len = blk_rq_cur_bytes(req);
968 }
969
970 /* recalculate the number of segments */
971 req->nr_phys_segments = blk_recalc_rq_segments(req);
972 }
973
974 return true;
975}
976EXPORT_SYMBOL_GPL(blk_update_request);
977
978static void __blk_account_io_done(struct request *req, u64 now)
979{
980 const int sgrp = op_stat_group(req_op(req));
981
982 part_stat_lock();
983 update_io_ticks(req->part, jiffies, true);
984 part_stat_inc(req->part, ios[sgrp]);
985 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
986 part_stat_unlock();
987}
988
989static inline void blk_account_io_done(struct request *req, u64 now)
990{
991 /*
992 * Account IO completion. flush_rq isn't accounted as a
993 * normal IO on queueing nor completion. Accounting the
994 * containing request is enough.
995 */
996 if (blk_do_io_stat(req) && req->part &&
997 !(req->rq_flags & RQF_FLUSH_SEQ))
998 __blk_account_io_done(req, now);
999}
1000
1001static void __blk_account_io_start(struct request *rq)
1002{
1003 /*
1004 * All non-passthrough requests are created from a bio with one
1005 * exception: when a flush command that is part of a flush sequence
1006 * generated by the state machine in blk-flush.c is cloned onto the
1007 * lower device by dm-multipath we can get here without a bio.
1008 */
1009 if (rq->bio)
1010 rq->part = rq->bio->bi_bdev;
1011 else
1012 rq->part = rq->q->disk->part0;
1013
1014 part_stat_lock();
1015 update_io_ticks(rq->part, jiffies, false);
1016 part_stat_unlock();
1017}
1018
1019static inline void blk_account_io_start(struct request *req)
1020{
1021 if (blk_do_io_stat(req))
1022 __blk_account_io_start(req);
1023}
1024
1025static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
1026{
1027 if (rq->rq_flags & RQF_STATS) {
1028 blk_mq_poll_stats_start(rq->q);
1029 blk_stat_add(rq, now);
1030 }
1031
1032 blk_mq_sched_completed_request(rq, now);
1033 blk_account_io_done(rq, now);
1034}
1035
1036inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
1037{
1038 if (blk_mq_need_time_stamp(rq))
1039 __blk_mq_end_request_acct(rq, ktime_get_ns());
1040
1041 if (rq->end_io) {
1042 rq_qos_done(rq->q, rq);
1043 if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1044 blk_mq_free_request(rq);
1045 } else {
1046 blk_mq_free_request(rq);
1047 }
1048}
1049EXPORT_SYMBOL(__blk_mq_end_request);
1050
1051void blk_mq_end_request(struct request *rq, blk_status_t error)
1052{
1053 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1054 BUG();
1055 __blk_mq_end_request(rq, error);
1056}
1057EXPORT_SYMBOL(blk_mq_end_request);
1058
1059#define TAG_COMP_BATCH 32
1060
1061static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1062 int *tag_array, int nr_tags)
1063{
1064 struct request_queue *q = hctx->queue;
1065
1066 /*
1067 * All requests should have been marked as RQF_MQ_INFLIGHT, so
1068 * update hctx->nr_active in batch
1069 */
1070 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
1071 __blk_mq_sub_active_requests(hctx, nr_tags);
1072
1073 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
1074 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
1075}
1076
1077void blk_mq_end_request_batch(struct io_comp_batch *iob)
1078{
1079 int tags[TAG_COMP_BATCH], nr_tags = 0;
1080 struct blk_mq_hw_ctx *cur_hctx = NULL;
1081 struct request *rq;
1082 u64 now = 0;
1083
1084 if (iob->need_ts)
1085 now = ktime_get_ns();
1086
1087 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1088 prefetch(rq->bio);
1089 prefetch(rq->rq_next);
1090
1091 blk_complete_request(rq);
1092 if (iob->need_ts)
1093 __blk_mq_end_request_acct(rq, now);
1094
1095 rq_qos_done(rq->q, rq);
1096
1097 /*
1098 * If end_io handler returns NONE, then it still has
1099 * ownership of the request.
1100 */
1101 if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
1102 continue;
1103
1104 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1105 if (!req_ref_put_and_test(rq))
1106 continue;
1107
1108 blk_crypto_free_request(rq);
1109 blk_pm_mark_last_busy(rq);
1110
1111 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1112 if (cur_hctx)
1113 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1114 nr_tags = 0;
1115 cur_hctx = rq->mq_hctx;
1116 }
1117 tags[nr_tags++] = rq->tag;
1118 }
1119
1120 if (nr_tags)
1121 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1122}
1123EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1124
1125static void blk_complete_reqs(struct llist_head *list)
1126{
1127 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1128 struct request *rq, *next;
1129
1130 llist_for_each_entry_safe(rq, next, entry, ipi_list)
1131 rq->q->mq_ops->complete(rq);
1132}
1133
1134static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1135{
1136 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1137}
1138
1139static int blk_softirq_cpu_dead(unsigned int cpu)
1140{
1141 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1142 return 0;
1143}
1144
1145static void __blk_mq_complete_request_remote(void *data)
1146{
1147 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1148}
1149
1150static inline bool blk_mq_complete_need_ipi(struct request *rq)
1151{
1152 int cpu = raw_smp_processor_id();
1153
1154 if (!IS_ENABLED(CONFIG_SMP) ||
1155 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1156 return false;
1157 /*
1158 * With force threaded interrupts enabled, raising softirq from an SMP
1159 * function call will always result in waking the ksoftirqd thread.
1160 * This is probably worse than completing the request on a different
1161 * cache domain.
1162 */
1163 if (force_irqthreads())
1164 return false;
1165
1166 /* same CPU or cache domain? Complete locally */
1167 if (cpu == rq->mq_ctx->cpu ||
1168 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1169 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1170 return false;
1171
1172 /* don't try to IPI to an offline CPU */
1173 return cpu_online(rq->mq_ctx->cpu);
1174}
1175
1176static void blk_mq_complete_send_ipi(struct request *rq)
1177{
1178 struct llist_head *list;
1179 unsigned int cpu;
1180
1181 cpu = rq->mq_ctx->cpu;
1182 list = &per_cpu(blk_cpu_done, cpu);
1183 if (llist_add(&rq->ipi_list, list)) {
1184 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
1185 smp_call_function_single_async(cpu, &rq->csd);
1186 }
1187}
1188
1189static void blk_mq_raise_softirq(struct request *rq)
1190{
1191 struct llist_head *list;
1192
1193 preempt_disable();
1194 list = this_cpu_ptr(&blk_cpu_done);
1195 if (llist_add(&rq->ipi_list, list))
1196 raise_softirq(BLOCK_SOFTIRQ);
1197 preempt_enable();
1198}
1199
1200bool blk_mq_complete_request_remote(struct request *rq)
1201{
1202 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1203
1204 /*
1205 * For request which hctx has only one ctx mapping,
1206 * or a polled request, always complete locally,
1207 * it's pointless to redirect the completion.
1208 */
1209 if (rq->mq_hctx->nr_ctx == 1 ||
1210 rq->cmd_flags & REQ_POLLED)
1211 return false;
1212
1213 if (blk_mq_complete_need_ipi(rq)) {
1214 blk_mq_complete_send_ipi(rq);
1215 return true;
1216 }
1217
1218 if (rq->q->nr_hw_queues == 1) {
1219 blk_mq_raise_softirq(rq);
1220 return true;
1221 }
1222 return false;
1223}
1224EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1225
1226/**
1227 * blk_mq_complete_request - end I/O on a request
1228 * @rq: the request being processed
1229 *
1230 * Description:
1231 * Complete a request by scheduling the ->complete_rq operation.
1232 **/
1233void blk_mq_complete_request(struct request *rq)
1234{
1235 if (!blk_mq_complete_request_remote(rq))
1236 rq->q->mq_ops->complete(rq);
1237}
1238EXPORT_SYMBOL(blk_mq_complete_request);
1239
1240/**
1241 * blk_mq_start_request - Start processing a request
1242 * @rq: Pointer to request to be started
1243 *
1244 * Function used by device drivers to notify the block layer that a request
1245 * is going to be processed now, so blk layer can do proper initializations
1246 * such as starting the timeout timer.
1247 */
1248void blk_mq_start_request(struct request *rq)
1249{
1250 struct request_queue *q = rq->q;
1251
1252 trace_block_rq_issue(rq);
1253
1254 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1255 rq->io_start_time_ns = ktime_get_ns();
1256 rq->stats_sectors = blk_rq_sectors(rq);
1257 rq->rq_flags |= RQF_STATS;
1258 rq_qos_issue(q, rq);
1259 }
1260
1261 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1262
1263 blk_add_timer(rq);
1264 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1265
1266#ifdef CONFIG_BLK_DEV_INTEGRITY
1267 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1268 q->integrity.profile->prepare_fn(rq);
1269#endif
1270 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1271 WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq));
1272}
1273EXPORT_SYMBOL(blk_mq_start_request);
1274
1275/*
1276 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1277 * queues. This is important for md arrays to benefit from merging
1278 * requests.
1279 */
1280static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1281{
1282 if (plug->multiple_queues)
1283 return BLK_MAX_REQUEST_COUNT * 2;
1284 return BLK_MAX_REQUEST_COUNT;
1285}
1286
1287static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1288{
1289 struct request *last = rq_list_peek(&plug->mq_list);
1290
1291 if (!plug->rq_count) {
1292 trace_block_plug(rq->q);
1293 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1294 (!blk_queue_nomerges(rq->q) &&
1295 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1296 blk_mq_flush_plug_list(plug, false);
1297 last = NULL;
1298 trace_block_plug(rq->q);
1299 }
1300
1301 if (!plug->multiple_queues && last && last->q != rq->q)
1302 plug->multiple_queues = true;
1303 if (!plug->has_elevator && (rq->rq_flags & RQF_ELV))
1304 plug->has_elevator = true;
1305 rq->rq_next = NULL;
1306 rq_list_add(&plug->mq_list, rq);
1307 plug->rq_count++;
1308}
1309
1310/**
1311 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1312 * @rq: request to insert
1313 * @at_head: insert request at head or tail of queue
1314 *
1315 * Description:
1316 * Insert a fully prepared request at the back of the I/O scheduler queue
1317 * for execution. Don't wait for completion.
1318 *
1319 * Note:
1320 * This function will invoke @done directly if the queue is dead.
1321 */
1322void blk_execute_rq_nowait(struct request *rq, bool at_head)
1323{
1324 WARN_ON(irqs_disabled());
1325 WARN_ON(!blk_rq_is_passthrough(rq));
1326
1327 blk_account_io_start(rq);
1328
1329 /*
1330 * As plugging can be enabled for passthrough requests on a zoned
1331 * device, directly accessing the plug instead of using blk_mq_plug()
1332 * should not have any consequences.
1333 */
1334 if (current->plug)
1335 blk_add_rq_to_plug(current->plug, rq);
1336 else
1337 blk_mq_sched_insert_request(rq, at_head, true, false);
1338}
1339EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1340
1341struct blk_rq_wait {
1342 struct completion done;
1343 blk_status_t ret;
1344};
1345
1346static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
1347{
1348 struct blk_rq_wait *wait = rq->end_io_data;
1349
1350 wait->ret = ret;
1351 complete(&wait->done);
1352 return RQ_END_IO_NONE;
1353}
1354
1355bool blk_rq_is_poll(struct request *rq)
1356{
1357 if (!rq->mq_hctx)
1358 return false;
1359 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1360 return false;
1361 if (WARN_ON_ONCE(!rq->bio))
1362 return false;
1363 return true;
1364}
1365EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1366
1367static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1368{
1369 do {
1370 bio_poll(rq->bio, NULL, 0);
1371 cond_resched();
1372 } while (!completion_done(wait));
1373}
1374
1375/**
1376 * blk_execute_rq - insert a request into queue for execution
1377 * @rq: request to insert
1378 * @at_head: insert request at head or tail of queue
1379 *
1380 * Description:
1381 * Insert a fully prepared request at the back of the I/O scheduler queue
1382 * for execution and wait for completion.
1383 * Return: The blk_status_t result provided to blk_mq_end_request().
1384 */
1385blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1386{
1387 struct blk_rq_wait wait = {
1388 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1389 };
1390
1391 WARN_ON(irqs_disabled());
1392 WARN_ON(!blk_rq_is_passthrough(rq));
1393
1394 rq->end_io_data = &wait;
1395 rq->end_io = blk_end_sync_rq;
1396
1397 blk_account_io_start(rq);
1398 blk_mq_sched_insert_request(rq, at_head, true, false);
1399
1400 if (blk_rq_is_poll(rq)) {
1401 blk_rq_poll_completion(rq, &wait.done);
1402 } else {
1403 /*
1404 * Prevent hang_check timer from firing at us during very long
1405 * I/O
1406 */
1407 unsigned long hang_check = sysctl_hung_task_timeout_secs;
1408
1409 if (hang_check)
1410 while (!wait_for_completion_io_timeout(&wait.done,
1411 hang_check * (HZ/2)))
1412 ;
1413 else
1414 wait_for_completion_io(&wait.done);
1415 }
1416
1417 return wait.ret;
1418}
1419EXPORT_SYMBOL(blk_execute_rq);
1420
1421static void __blk_mq_requeue_request(struct request *rq)
1422{
1423 struct request_queue *q = rq->q;
1424
1425 blk_mq_put_driver_tag(rq);
1426
1427 trace_block_rq_requeue(rq);
1428 rq_qos_requeue(q, rq);
1429
1430 if (blk_mq_request_started(rq)) {
1431 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1432 rq->rq_flags &= ~RQF_TIMED_OUT;
1433 }
1434}
1435
1436void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1437{
1438 __blk_mq_requeue_request(rq);
1439
1440 /* this request will be re-inserted to io scheduler queue */
1441 blk_mq_sched_requeue_request(rq);
1442
1443 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
1444}
1445EXPORT_SYMBOL(blk_mq_requeue_request);
1446
1447static void blk_mq_requeue_work(struct work_struct *work)
1448{
1449 struct request_queue *q =
1450 container_of(work, struct request_queue, requeue_work.work);
1451 LIST_HEAD(rq_list);
1452 struct request *rq, *next;
1453
1454 spin_lock_irq(&q->requeue_lock);
1455 list_splice_init(&q->requeue_list, &rq_list);
1456 spin_unlock_irq(&q->requeue_lock);
1457
1458 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
1459 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
1460 continue;
1461
1462 rq->rq_flags &= ~RQF_SOFTBARRIER;
1463 list_del_init(&rq->queuelist);
1464 /*
1465 * If RQF_DONTPREP, rq has contained some driver specific
1466 * data, so insert it to hctx dispatch list to avoid any
1467 * merge.
1468 */
1469 if (rq->rq_flags & RQF_DONTPREP)
1470 blk_mq_request_bypass_insert(rq, false, false);
1471 else
1472 blk_mq_sched_insert_request(rq, true, false, false);
1473 }
1474
1475 while (!list_empty(&rq_list)) {
1476 rq = list_entry(rq_list.next, struct request, queuelist);
1477 list_del_init(&rq->queuelist);
1478 blk_mq_sched_insert_request(rq, false, false, false);
1479 }
1480
1481 blk_mq_run_hw_queues(q, false);
1482}
1483
1484void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
1485 bool kick_requeue_list)
1486{
1487 struct request_queue *q = rq->q;
1488 unsigned long flags;
1489
1490 /*
1491 * We abuse this flag that is otherwise used by the I/O scheduler to
1492 * request head insertion from the workqueue.
1493 */
1494 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
1495
1496 spin_lock_irqsave(&q->requeue_lock, flags);
1497 if (at_head) {
1498 rq->rq_flags |= RQF_SOFTBARRIER;
1499 list_add(&rq->queuelist, &q->requeue_list);
1500 } else {
1501 list_add_tail(&rq->queuelist, &q->requeue_list);
1502 }
1503 spin_unlock_irqrestore(&q->requeue_lock, flags);
1504
1505 if (kick_requeue_list)
1506 blk_mq_kick_requeue_list(q);
1507}
1508
1509void blk_mq_kick_requeue_list(struct request_queue *q)
1510{
1511 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1512}
1513EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1514
1515void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1516 unsigned long msecs)
1517{
1518 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1519 msecs_to_jiffies(msecs));
1520}
1521EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1522
1523static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1524{
1525 /*
1526 * If we find a request that isn't idle we know the queue is busy
1527 * as it's checked in the iter.
1528 * Return false to stop the iteration.
1529 */
1530 if (blk_mq_request_started(rq)) {
1531 bool *busy = priv;
1532
1533 *busy = true;
1534 return false;
1535 }
1536
1537 return true;
1538}
1539
1540bool blk_mq_queue_inflight(struct request_queue *q)
1541{
1542 bool busy = false;
1543
1544 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1545 return busy;
1546}
1547EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1548
1549static void blk_mq_rq_timed_out(struct request *req)
1550{
1551 req->rq_flags |= RQF_TIMED_OUT;
1552 if (req->q->mq_ops->timeout) {
1553 enum blk_eh_timer_return ret;
1554
1555 ret = req->q->mq_ops->timeout(req);
1556 if (ret == BLK_EH_DONE)
1557 return;
1558 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1559 }
1560
1561 blk_add_timer(req);
1562}
1563
1564struct blk_expired_data {
1565 bool has_timedout_rq;
1566 unsigned long next;
1567 unsigned long timeout_start;
1568};
1569
1570static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
1571{
1572 unsigned long deadline;
1573
1574 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1575 return false;
1576 if (rq->rq_flags & RQF_TIMED_OUT)
1577 return false;
1578
1579 deadline = READ_ONCE(rq->deadline);
1580 if (time_after_eq(expired->timeout_start, deadline))
1581 return true;
1582
1583 if (expired->next == 0)
1584 expired->next = deadline;
1585 else if (time_after(expired->next, deadline))
1586 expired->next = deadline;
1587 return false;
1588}
1589
1590void blk_mq_put_rq_ref(struct request *rq)
1591{
1592 if (is_flush_rq(rq)) {
1593 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1594 blk_mq_free_request(rq);
1595 } else if (req_ref_put_and_test(rq)) {
1596 __blk_mq_free_request(rq);
1597 }
1598}
1599
1600static bool blk_mq_check_expired(struct request *rq, void *priv)
1601{
1602 struct blk_expired_data *expired = priv;
1603
1604 /*
1605 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1606 * be reallocated underneath the timeout handler's processing, then
1607 * the expire check is reliable. If the request is not expired, then
1608 * it was completed and reallocated as a new request after returning
1609 * from blk_mq_check_expired().
1610 */
1611 if (blk_mq_req_expired(rq, expired)) {
1612 expired->has_timedout_rq = true;
1613 return false;
1614 }
1615 return true;
1616}
1617
1618static bool blk_mq_handle_expired(struct request *rq, void *priv)
1619{
1620 struct blk_expired_data *expired = priv;
1621
1622 if (blk_mq_req_expired(rq, expired))
1623 blk_mq_rq_timed_out(rq);
1624 return true;
1625}
1626
1627static void blk_mq_timeout_work(struct work_struct *work)
1628{
1629 struct request_queue *q =
1630 container_of(work, struct request_queue, timeout_work);
1631 struct blk_expired_data expired = {
1632 .timeout_start = jiffies,
1633 };
1634 struct blk_mq_hw_ctx *hctx;
1635 unsigned long i;
1636
1637 /* A deadlock might occur if a request is stuck requiring a
1638 * timeout at the same time a queue freeze is waiting
1639 * completion, since the timeout code would not be able to
1640 * acquire the queue reference here.
1641 *
1642 * That's why we don't use blk_queue_enter here; instead, we use
1643 * percpu_ref_tryget directly, because we need to be able to
1644 * obtain a reference even in the short window between the queue
1645 * starting to freeze, by dropping the first reference in
1646 * blk_freeze_queue_start, and the moment the last request is
1647 * consumed, marked by the instant q_usage_counter reaches
1648 * zero.
1649 */
1650 if (!percpu_ref_tryget(&q->q_usage_counter))
1651 return;
1652
1653 /* check if there is any timed-out request */
1654 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
1655 if (expired.has_timedout_rq) {
1656 /*
1657 * Before walking tags, we must ensure any submit started
1658 * before the current time has finished. Since the submit
1659 * uses srcu or rcu, wait for a synchronization point to
1660 * ensure all running submits have finished
1661 */
1662 blk_mq_wait_quiesce_done(q->tag_set);
1663
1664 expired.next = 0;
1665 blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
1666 }
1667
1668 if (expired.next != 0) {
1669 mod_timer(&q->timeout, expired.next);
1670 } else {
1671 /*
1672 * Request timeouts are handled as a forward rolling timer. If
1673 * we end up here it means that no requests are pending and
1674 * also that no request has been pending for a while. Mark
1675 * each hctx as idle.
1676 */
1677 queue_for_each_hw_ctx(q, hctx, i) {
1678 /* the hctx may be unmapped, so check it here */
1679 if (blk_mq_hw_queue_mapped(hctx))
1680 blk_mq_tag_idle(hctx);
1681 }
1682 }
1683 blk_queue_exit(q);
1684}
1685
1686struct flush_busy_ctx_data {
1687 struct blk_mq_hw_ctx *hctx;
1688 struct list_head *list;
1689};
1690
1691static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1692{
1693 struct flush_busy_ctx_data *flush_data = data;
1694 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1695 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1696 enum hctx_type type = hctx->type;
1697
1698 spin_lock(&ctx->lock);
1699 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1700 sbitmap_clear_bit(sb, bitnr);
1701 spin_unlock(&ctx->lock);
1702 return true;
1703}
1704
1705/*
1706 * Process software queues that have been marked busy, splicing them
1707 * to the for-dispatch
1708 */
1709void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1710{
1711 struct flush_busy_ctx_data data = {
1712 .hctx = hctx,
1713 .list = list,
1714 };
1715
1716 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1717}
1718EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1719
1720struct dispatch_rq_data {
1721 struct blk_mq_hw_ctx *hctx;
1722 struct request *rq;
1723};
1724
1725static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1726 void *data)
1727{
1728 struct dispatch_rq_data *dispatch_data = data;
1729 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1730 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1731 enum hctx_type type = hctx->type;
1732
1733 spin_lock(&ctx->lock);
1734 if (!list_empty(&ctx->rq_lists[type])) {
1735 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1736 list_del_init(&dispatch_data->rq->queuelist);
1737 if (list_empty(&ctx->rq_lists[type]))
1738 sbitmap_clear_bit(sb, bitnr);
1739 }
1740 spin_unlock(&ctx->lock);
1741
1742 return !dispatch_data->rq;
1743}
1744
1745struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1746 struct blk_mq_ctx *start)
1747{
1748 unsigned off = start ? start->index_hw[hctx->type] : 0;
1749 struct dispatch_rq_data data = {
1750 .hctx = hctx,
1751 .rq = NULL,
1752 };
1753
1754 __sbitmap_for_each_set(&hctx->ctx_map, off,
1755 dispatch_rq_from_ctx, &data);
1756
1757 return data.rq;
1758}
1759
1760static bool __blk_mq_alloc_driver_tag(struct request *rq)
1761{
1762 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1763 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1764 int tag;
1765
1766 blk_mq_tag_busy(rq->mq_hctx);
1767
1768 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1769 bt = &rq->mq_hctx->tags->breserved_tags;
1770 tag_offset = 0;
1771 } else {
1772 if (!hctx_may_queue(rq->mq_hctx, bt))
1773 return false;
1774 }
1775
1776 tag = __sbitmap_queue_get(bt);
1777 if (tag == BLK_MQ_NO_TAG)
1778 return false;
1779
1780 rq->tag = tag + tag_offset;
1781 return true;
1782}
1783
1784bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1785{
1786 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1787 return false;
1788
1789 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1790 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1791 rq->rq_flags |= RQF_MQ_INFLIGHT;
1792 __blk_mq_inc_active_requests(hctx);
1793 }
1794 hctx->tags->rqs[rq->tag] = rq;
1795 return true;
1796}
1797
1798static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1799 int flags, void *key)
1800{
1801 struct blk_mq_hw_ctx *hctx;
1802
1803 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1804
1805 spin_lock(&hctx->dispatch_wait_lock);
1806 if (!list_empty(&wait->entry)) {
1807 struct sbitmap_queue *sbq;
1808
1809 list_del_init(&wait->entry);
1810 sbq = &hctx->tags->bitmap_tags;
1811 atomic_dec(&sbq->ws_active);
1812 }
1813 spin_unlock(&hctx->dispatch_wait_lock);
1814
1815 blk_mq_run_hw_queue(hctx, true);
1816 return 1;
1817}
1818
1819/*
1820 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1821 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1822 * restart. For both cases, take care to check the condition again after
1823 * marking us as waiting.
1824 */
1825static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1826 struct request *rq)
1827{
1828 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1829 struct wait_queue_head *wq;
1830 wait_queue_entry_t *wait;
1831 bool ret;
1832
1833 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
1834 blk_mq_sched_mark_restart_hctx(hctx);
1835
1836 /*
1837 * It's possible that a tag was freed in the window between the
1838 * allocation failure and adding the hardware queue to the wait
1839 * queue.
1840 *
1841 * Don't clear RESTART here, someone else could have set it.
1842 * At most this will cost an extra queue run.
1843 */
1844 return blk_mq_get_driver_tag(rq);
1845 }
1846
1847 wait = &hctx->dispatch_wait;
1848 if (!list_empty_careful(&wait->entry))
1849 return false;
1850
1851 wq = &bt_wait_ptr(sbq, hctx)->wait;
1852
1853 spin_lock_irq(&wq->lock);
1854 spin_lock(&hctx->dispatch_wait_lock);
1855 if (!list_empty(&wait->entry)) {
1856 spin_unlock(&hctx->dispatch_wait_lock);
1857 spin_unlock_irq(&wq->lock);
1858 return false;
1859 }
1860
1861 atomic_inc(&sbq->ws_active);
1862 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1863 __add_wait_queue(wq, wait);
1864
1865 /*
1866 * It's possible that a tag was freed in the window between the
1867 * allocation failure and adding the hardware queue to the wait
1868 * queue.
1869 */
1870 ret = blk_mq_get_driver_tag(rq);
1871 if (!ret) {
1872 spin_unlock(&hctx->dispatch_wait_lock);
1873 spin_unlock_irq(&wq->lock);
1874 return false;
1875 }
1876
1877 /*
1878 * We got a tag, remove ourselves from the wait queue to ensure
1879 * someone else gets the wakeup.
1880 */
1881 list_del_init(&wait->entry);
1882 atomic_dec(&sbq->ws_active);
1883 spin_unlock(&hctx->dispatch_wait_lock);
1884 spin_unlock_irq(&wq->lock);
1885
1886 return true;
1887}
1888
1889#define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1890#define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1891/*
1892 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1893 * - EWMA is one simple way to compute running average value
1894 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1895 * - take 4 as factor for avoiding to get too small(0) result, and this
1896 * factor doesn't matter because EWMA decreases exponentially
1897 */
1898static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1899{
1900 unsigned int ewma;
1901
1902 ewma = hctx->dispatch_busy;
1903
1904 if (!ewma && !busy)
1905 return;
1906
1907 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1908 if (busy)
1909 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1910 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1911
1912 hctx->dispatch_busy = ewma;
1913}
1914
1915#define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1916
1917static void blk_mq_handle_dev_resource(struct request *rq,
1918 struct list_head *list)
1919{
1920 struct request *next =
1921 list_first_entry_or_null(list, struct request, queuelist);
1922
1923 /*
1924 * If an I/O scheduler has been configured and we got a driver tag for
1925 * the next request already, free it.
1926 */
1927 if (next)
1928 blk_mq_put_driver_tag(next);
1929
1930 list_add(&rq->queuelist, list);
1931 __blk_mq_requeue_request(rq);
1932}
1933
1934static void blk_mq_handle_zone_resource(struct request *rq,
1935 struct list_head *zone_list)
1936{
1937 /*
1938 * If we end up here it is because we cannot dispatch a request to a
1939 * specific zone due to LLD level zone-write locking or other zone
1940 * related resource not being available. In this case, set the request
1941 * aside in zone_list for retrying it later.
1942 */
1943 list_add(&rq->queuelist, zone_list);
1944 __blk_mq_requeue_request(rq);
1945}
1946
1947enum prep_dispatch {
1948 PREP_DISPATCH_OK,
1949 PREP_DISPATCH_NO_TAG,
1950 PREP_DISPATCH_NO_BUDGET,
1951};
1952
1953static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1954 bool need_budget)
1955{
1956 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1957 int budget_token = -1;
1958
1959 if (need_budget) {
1960 budget_token = blk_mq_get_dispatch_budget(rq->q);
1961 if (budget_token < 0) {
1962 blk_mq_put_driver_tag(rq);
1963 return PREP_DISPATCH_NO_BUDGET;
1964 }
1965 blk_mq_set_rq_budget_token(rq, budget_token);
1966 }
1967
1968 if (!blk_mq_get_driver_tag(rq)) {
1969 /*
1970 * The initial allocation attempt failed, so we need to
1971 * rerun the hardware queue when a tag is freed. The
1972 * waitqueue takes care of that. If the queue is run
1973 * before we add this entry back on the dispatch list,
1974 * we'll re-run it below.
1975 */
1976 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1977 /*
1978 * All budgets not got from this function will be put
1979 * together during handling partial dispatch
1980 */
1981 if (need_budget)
1982 blk_mq_put_dispatch_budget(rq->q, budget_token);
1983 return PREP_DISPATCH_NO_TAG;
1984 }
1985 }
1986
1987 return PREP_DISPATCH_OK;
1988}
1989
1990/* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1991static void blk_mq_release_budgets(struct request_queue *q,
1992 struct list_head *list)
1993{
1994 struct request *rq;
1995
1996 list_for_each_entry(rq, list, queuelist) {
1997 int budget_token = blk_mq_get_rq_budget_token(rq);
1998
1999 if (budget_token >= 0)
2000 blk_mq_put_dispatch_budget(q, budget_token);
2001 }
2002}
2003
2004/*
2005 * Returns true if we did some work AND can potentially do more.
2006 */
2007bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
2008 unsigned int nr_budgets)
2009{
2010 enum prep_dispatch prep;
2011 struct request_queue *q = hctx->queue;
2012 struct request *rq, *nxt;
2013 int errors, queued;
2014 blk_status_t ret = BLK_STS_OK;
2015 LIST_HEAD(zone_list);
2016 bool needs_resource = false;
2017
2018 if (list_empty(list))
2019 return false;
2020
2021 /*
2022 * Now process all the entries, sending them to the driver.
2023 */
2024 errors = queued = 0;
2025 do {
2026 struct blk_mq_queue_data bd;
2027
2028 rq = list_first_entry(list, struct request, queuelist);
2029
2030 WARN_ON_ONCE(hctx != rq->mq_hctx);
2031 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
2032 if (prep != PREP_DISPATCH_OK)
2033 break;
2034
2035 list_del_init(&rq->queuelist);
2036
2037 bd.rq = rq;
2038
2039 /*
2040 * Flag last if we have no more requests, or if we have more
2041 * but can't assign a driver tag to it.
2042 */
2043 if (list_empty(list))
2044 bd.last = true;
2045 else {
2046 nxt = list_first_entry(list, struct request, queuelist);
2047 bd.last = !blk_mq_get_driver_tag(nxt);
2048 }
2049
2050 /*
2051 * once the request is queued to lld, no need to cover the
2052 * budget any more
2053 */
2054 if (nr_budgets)
2055 nr_budgets--;
2056 ret = q->mq_ops->queue_rq(hctx, &bd);
2057 switch (ret) {
2058 case BLK_STS_OK:
2059 queued++;
2060 break;
2061 case BLK_STS_RESOURCE:
2062 needs_resource = true;
2063 fallthrough;
2064 case BLK_STS_DEV_RESOURCE:
2065 blk_mq_handle_dev_resource(rq, list);
2066 goto out;
2067 case BLK_STS_ZONE_RESOURCE:
2068 /*
2069 * Move the request to zone_list and keep going through
2070 * the dispatch list to find more requests the drive can
2071 * accept.
2072 */
2073 blk_mq_handle_zone_resource(rq, &zone_list);
2074 needs_resource = true;
2075 break;
2076 default:
2077 errors++;
2078 blk_mq_end_request(rq, ret);
2079 }
2080 } while (!list_empty(list));
2081out:
2082 if (!list_empty(&zone_list))
2083 list_splice_tail_init(&zone_list, list);
2084
2085 /* If we didn't flush the entire list, we could have told the driver
2086 * there was more coming, but that turned out to be a lie.
2087 */
2088 if ((!list_empty(list) || errors || needs_resource ||
2089 ret == BLK_STS_DEV_RESOURCE) && q->mq_ops->commit_rqs && queued)
2090 q->mq_ops->commit_rqs(hctx);
2091 /*
2092 * Any items that need requeuing? Stuff them into hctx->dispatch,
2093 * that is where we will continue on next queue run.
2094 */
2095 if (!list_empty(list)) {
2096 bool needs_restart;
2097 /* For non-shared tags, the RESTART check will suffice */
2098 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2099 (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
2100
2101 if (nr_budgets)
2102 blk_mq_release_budgets(q, list);
2103
2104 spin_lock(&hctx->lock);
2105 list_splice_tail_init(list, &hctx->dispatch);
2106 spin_unlock(&hctx->lock);
2107
2108 /*
2109 * Order adding requests to hctx->dispatch and checking
2110 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2111 * in blk_mq_sched_restart(). Avoid restart code path to
2112 * miss the new added requests to hctx->dispatch, meantime
2113 * SCHED_RESTART is observed here.
2114 */
2115 smp_mb();
2116
2117 /*
2118 * If SCHED_RESTART was set by the caller of this function and
2119 * it is no longer set that means that it was cleared by another
2120 * thread and hence that a queue rerun is needed.
2121 *
2122 * If 'no_tag' is set, that means that we failed getting
2123 * a driver tag with an I/O scheduler attached. If our dispatch
2124 * waitqueue is no longer active, ensure that we run the queue
2125 * AFTER adding our entries back to the list.
2126 *
2127 * If no I/O scheduler has been configured it is possible that
2128 * the hardware queue got stopped and restarted before requests
2129 * were pushed back onto the dispatch list. Rerun the queue to
2130 * avoid starvation. Notes:
2131 * - blk_mq_run_hw_queue() checks whether or not a queue has
2132 * been stopped before rerunning a queue.
2133 * - Some but not all block drivers stop a queue before
2134 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2135 * and dm-rq.
2136 *
2137 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2138 * bit is set, run queue after a delay to avoid IO stalls
2139 * that could otherwise occur if the queue is idle. We'll do
2140 * similar if we couldn't get budget or couldn't lock a zone
2141 * and SCHED_RESTART is set.
2142 */
2143 needs_restart = blk_mq_sched_needs_restart(hctx);
2144 if (prep == PREP_DISPATCH_NO_BUDGET)
2145 needs_resource = true;
2146 if (!needs_restart ||
2147 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2148 blk_mq_run_hw_queue(hctx, true);
2149 else if (needs_resource)
2150 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2151
2152 blk_mq_update_dispatch_busy(hctx, true);
2153 return false;
2154 } else
2155 blk_mq_update_dispatch_busy(hctx, false);
2156
2157 return (queued + errors) != 0;
2158}
2159
2160/**
2161 * __blk_mq_run_hw_queue - Run a hardware queue.
2162 * @hctx: Pointer to the hardware queue to run.
2163 *
2164 * Send pending requests to the hardware.
2165 */
2166static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
2167{
2168 /*
2169 * We can't run the queue inline with ints disabled. Ensure that
2170 * we catch bad users of this early.
2171 */
2172 WARN_ON_ONCE(in_interrupt());
2173
2174 blk_mq_run_dispatch_ops(hctx->queue,
2175 blk_mq_sched_dispatch_requests(hctx));
2176}
2177
2178static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2179{
2180 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2181
2182 if (cpu >= nr_cpu_ids)
2183 cpu = cpumask_first(hctx->cpumask);
2184 return cpu;
2185}
2186
2187/*
2188 * It'd be great if the workqueue API had a way to pass
2189 * in a mask and had some smarts for more clever placement.
2190 * For now we just round-robin here, switching for every
2191 * BLK_MQ_CPU_WORK_BATCH queued items.
2192 */
2193static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2194{
2195 bool tried = false;
2196 int next_cpu = hctx->next_cpu;
2197
2198 if (hctx->queue->nr_hw_queues == 1)
2199 return WORK_CPU_UNBOUND;
2200
2201 if (--hctx->next_cpu_batch <= 0) {
2202select_cpu:
2203 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2204 cpu_online_mask);
2205 if (next_cpu >= nr_cpu_ids)
2206 next_cpu = blk_mq_first_mapped_cpu(hctx);
2207 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2208 }
2209
2210 /*
2211 * Do unbound schedule if we can't find a online CPU for this hctx,
2212 * and it should only happen in the path of handling CPU DEAD.
2213 */
2214 if (!cpu_online(next_cpu)) {
2215 if (!tried) {
2216 tried = true;
2217 goto select_cpu;
2218 }
2219
2220 /*
2221 * Make sure to re-select CPU next time once after CPUs
2222 * in hctx->cpumask become online again.
2223 */
2224 hctx->next_cpu = next_cpu;
2225 hctx->next_cpu_batch = 1;
2226 return WORK_CPU_UNBOUND;
2227 }
2228
2229 hctx->next_cpu = next_cpu;
2230 return next_cpu;
2231}
2232
2233/**
2234 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
2235 * @hctx: Pointer to the hardware queue to run.
2236 * @async: If we want to run the queue asynchronously.
2237 * @msecs: Milliseconds of delay to wait before running the queue.
2238 *
2239 * If !@async, try to run the queue now. Else, run the queue asynchronously and
2240 * with a delay of @msecs.
2241 */
2242static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
2243 unsigned long msecs)
2244{
2245 if (unlikely(blk_mq_hctx_stopped(hctx)))
2246 return;
2247
2248 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
2249 if (cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2250 __blk_mq_run_hw_queue(hctx);
2251 return;
2252 }
2253 }
2254
2255 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2256 msecs_to_jiffies(msecs));
2257}
2258
2259/**
2260 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2261 * @hctx: Pointer to the hardware queue to run.
2262 * @msecs: Milliseconds of delay to wait before running the queue.
2263 *
2264 * Run a hardware queue asynchronously with a delay of @msecs.
2265 */
2266void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2267{
2268 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
2269}
2270EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2271
2272/**
2273 * blk_mq_run_hw_queue - Start to run a hardware queue.
2274 * @hctx: Pointer to the hardware queue to run.
2275 * @async: If we want to run the queue asynchronously.
2276 *
2277 * Check if the request queue is not in a quiesced state and if there are
2278 * pending requests to be sent. If this is true, run the queue to send requests
2279 * to hardware.
2280 */
2281void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2282{
2283 bool need_run;
2284
2285 /*
2286 * When queue is quiesced, we may be switching io scheduler, or
2287 * updating nr_hw_queues, or other things, and we can't run queue
2288 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2289 *
2290 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2291 * quiesced.
2292 */
2293 __blk_mq_run_dispatch_ops(hctx->queue, false,
2294 need_run = !blk_queue_quiesced(hctx->queue) &&
2295 blk_mq_hctx_has_pending(hctx));
2296
2297 if (need_run)
2298 __blk_mq_delay_run_hw_queue(hctx, async, 0);
2299}
2300EXPORT_SYMBOL(blk_mq_run_hw_queue);
2301
2302/*
2303 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2304 * scheduler.
2305 */
2306static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2307{
2308 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2309 /*
2310 * If the IO scheduler does not respect hardware queues when
2311 * dispatching, we just don't bother with multiple HW queues and
2312 * dispatch from hctx for the current CPU since running multiple queues
2313 * just causes lock contention inside the scheduler and pointless cache
2314 * bouncing.
2315 */
2316 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2317
2318 if (!blk_mq_hctx_stopped(hctx))
2319 return hctx;
2320 return NULL;
2321}
2322
2323/**
2324 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2325 * @q: Pointer to the request queue to run.
2326 * @async: If we want to run the queue asynchronously.
2327 */
2328void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2329{
2330 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2331 unsigned long i;
2332
2333 sq_hctx = NULL;
2334 if (blk_queue_sq_sched(q))
2335 sq_hctx = blk_mq_get_sq_hctx(q);
2336 queue_for_each_hw_ctx(q, hctx, i) {
2337 if (blk_mq_hctx_stopped(hctx))
2338 continue;
2339 /*
2340 * Dispatch from this hctx either if there's no hctx preferred
2341 * by IO scheduler or if it has requests that bypass the
2342 * scheduler.
2343 */
2344 if (!sq_hctx || sq_hctx == hctx ||
2345 !list_empty_careful(&hctx->dispatch))
2346 blk_mq_run_hw_queue(hctx, async);
2347 }
2348}
2349EXPORT_SYMBOL(blk_mq_run_hw_queues);
2350
2351/**
2352 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2353 * @q: Pointer to the request queue to run.
2354 * @msecs: Milliseconds of delay to wait before running the queues.
2355 */
2356void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2357{
2358 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2359 unsigned long i;
2360
2361 sq_hctx = NULL;
2362 if (blk_queue_sq_sched(q))
2363 sq_hctx = blk_mq_get_sq_hctx(q);
2364 queue_for_each_hw_ctx(q, hctx, i) {
2365 if (blk_mq_hctx_stopped(hctx))
2366 continue;
2367 /*
2368 * If there is already a run_work pending, leave the
2369 * pending delay untouched. Otherwise, a hctx can stall
2370 * if another hctx is re-delaying the other's work
2371 * before the work executes.
2372 */
2373 if (delayed_work_pending(&hctx->run_work))
2374 continue;
2375 /*
2376 * Dispatch from this hctx either if there's no hctx preferred
2377 * by IO scheduler or if it has requests that bypass the
2378 * scheduler.
2379 */
2380 if (!sq_hctx || sq_hctx == hctx ||
2381 !list_empty_careful(&hctx->dispatch))
2382 blk_mq_delay_run_hw_queue(hctx, msecs);
2383 }
2384}
2385EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2386
2387/*
2388 * This function is often used for pausing .queue_rq() by driver when
2389 * there isn't enough resource or some conditions aren't satisfied, and
2390 * BLK_STS_RESOURCE is usually returned.
2391 *
2392 * We do not guarantee that dispatch can be drained or blocked
2393 * after blk_mq_stop_hw_queue() returns. Please use
2394 * blk_mq_quiesce_queue() for that requirement.
2395 */
2396void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2397{
2398 cancel_delayed_work(&hctx->run_work);
2399
2400 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2401}
2402EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2403
2404/*
2405 * This function is often used for pausing .queue_rq() by driver when
2406 * there isn't enough resource or some conditions aren't satisfied, and
2407 * BLK_STS_RESOURCE is usually returned.
2408 *
2409 * We do not guarantee that dispatch can be drained or blocked
2410 * after blk_mq_stop_hw_queues() returns. Please use
2411 * blk_mq_quiesce_queue() for that requirement.
2412 */
2413void blk_mq_stop_hw_queues(struct request_queue *q)
2414{
2415 struct blk_mq_hw_ctx *hctx;
2416 unsigned long i;
2417
2418 queue_for_each_hw_ctx(q, hctx, i)
2419 blk_mq_stop_hw_queue(hctx);
2420}
2421EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2422
2423void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2424{
2425 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2426
2427 blk_mq_run_hw_queue(hctx, false);
2428}
2429EXPORT_SYMBOL(blk_mq_start_hw_queue);
2430
2431void blk_mq_start_hw_queues(struct request_queue *q)
2432{
2433 struct blk_mq_hw_ctx *hctx;
2434 unsigned long i;
2435
2436 queue_for_each_hw_ctx(q, hctx, i)
2437 blk_mq_start_hw_queue(hctx);
2438}
2439EXPORT_SYMBOL(blk_mq_start_hw_queues);
2440
2441void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2442{
2443 if (!blk_mq_hctx_stopped(hctx))
2444 return;
2445
2446 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2447 blk_mq_run_hw_queue(hctx, async);
2448}
2449EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2450
2451void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2452{
2453 struct blk_mq_hw_ctx *hctx;
2454 unsigned long i;
2455
2456 queue_for_each_hw_ctx(q, hctx, i)
2457 blk_mq_start_stopped_hw_queue(hctx, async);
2458}
2459EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2460
2461static void blk_mq_run_work_fn(struct work_struct *work)
2462{
2463 struct blk_mq_hw_ctx *hctx;
2464
2465 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
2466
2467 /*
2468 * If we are stopped, don't run the queue.
2469 */
2470 if (blk_mq_hctx_stopped(hctx))
2471 return;
2472
2473 __blk_mq_run_hw_queue(hctx);
2474}
2475
2476static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
2477 struct request *rq,
2478 bool at_head)
2479{
2480 struct blk_mq_ctx *ctx = rq->mq_ctx;
2481 enum hctx_type type = hctx->type;
2482
2483 lockdep_assert_held(&ctx->lock);
2484
2485 trace_block_rq_insert(rq);
2486
2487 if (at_head)
2488 list_add(&rq->queuelist, &ctx->rq_lists[type]);
2489 else
2490 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
2491}
2492
2493void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
2494 bool at_head)
2495{
2496 struct blk_mq_ctx *ctx = rq->mq_ctx;
2497
2498 lockdep_assert_held(&ctx->lock);
2499
2500 __blk_mq_insert_req_list(hctx, rq, at_head);
2501 blk_mq_hctx_mark_pending(hctx, ctx);
2502}
2503
2504/**
2505 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2506 * @rq: Pointer to request to be inserted.
2507 * @at_head: true if the request should be inserted at the head of the list.
2508 * @run_queue: If we should run the hardware queue after inserting the request.
2509 *
2510 * Should only be used carefully, when the caller knows we want to
2511 * bypass a potential IO scheduler on the target device.
2512 */
2513void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
2514 bool run_queue)
2515{
2516 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2517
2518 spin_lock(&hctx->lock);
2519 if (at_head)
2520 list_add(&rq->queuelist, &hctx->dispatch);
2521 else
2522 list_add_tail(&rq->queuelist, &hctx->dispatch);
2523 spin_unlock(&hctx->lock);
2524
2525 if (run_queue)
2526 blk_mq_run_hw_queue(hctx, false);
2527}
2528
2529void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
2530 struct list_head *list)
2531
2532{
2533 struct request *rq;
2534 enum hctx_type type = hctx->type;
2535
2536 /*
2537 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2538 * offline now
2539 */
2540 list_for_each_entry(rq, list, queuelist) {
2541 BUG_ON(rq->mq_ctx != ctx);
2542 trace_block_rq_insert(rq);
2543 }
2544
2545 spin_lock(&ctx->lock);
2546 list_splice_tail_init(list, &ctx->rq_lists[type]);
2547 blk_mq_hctx_mark_pending(hctx, ctx);
2548 spin_unlock(&ctx->lock);
2549}
2550
2551static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int *queued,
2552 bool from_schedule)
2553{
2554 if (hctx->queue->mq_ops->commit_rqs) {
2555 trace_block_unplug(hctx->queue, *queued, !from_schedule);
2556 hctx->queue->mq_ops->commit_rqs(hctx);
2557 }
2558 *queued = 0;
2559}
2560
2561static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2562 unsigned int nr_segs)
2563{
2564 int err;
2565
2566 if (bio->bi_opf & REQ_RAHEAD)
2567 rq->cmd_flags |= REQ_FAILFAST_MASK;
2568
2569 rq->__sector = bio->bi_iter.bi_sector;
2570 blk_rq_bio_prep(rq, bio, nr_segs);
2571
2572 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2573 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2574 WARN_ON_ONCE(err);
2575
2576 blk_account_io_start(rq);
2577}
2578
2579static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2580 struct request *rq, bool last)
2581{
2582 struct request_queue *q = rq->q;
2583 struct blk_mq_queue_data bd = {
2584 .rq = rq,
2585 .last = last,
2586 };
2587 blk_status_t ret;
2588
2589 /*
2590 * For OK queue, we are done. For error, caller may kill it.
2591 * Any other error (busy), just add it to our list as we
2592 * previously would have done.
2593 */
2594 ret = q->mq_ops->queue_rq(hctx, &bd);
2595 switch (ret) {
2596 case BLK_STS_OK:
2597 blk_mq_update_dispatch_busy(hctx, false);
2598 break;
2599 case BLK_STS_RESOURCE:
2600 case BLK_STS_DEV_RESOURCE:
2601 blk_mq_update_dispatch_busy(hctx, true);
2602 __blk_mq_requeue_request(rq);
2603 break;
2604 default:
2605 blk_mq_update_dispatch_busy(hctx, false);
2606 break;
2607 }
2608
2609 return ret;
2610}
2611
2612static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2613 struct request *rq,
2614 bool bypass_insert, bool last)
2615{
2616 struct request_queue *q = rq->q;
2617 bool run_queue = true;
2618 int budget_token;
2619
2620 /*
2621 * RCU or SRCU read lock is needed before checking quiesced flag.
2622 *
2623 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2624 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2625 * and avoid driver to try to dispatch again.
2626 */
2627 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2628 run_queue = false;
2629 bypass_insert = false;
2630 goto insert;
2631 }
2632
2633 if ((rq->rq_flags & RQF_ELV) && !bypass_insert)
2634 goto insert;
2635
2636 budget_token = blk_mq_get_dispatch_budget(q);
2637 if (budget_token < 0)
2638 goto insert;
2639
2640 blk_mq_set_rq_budget_token(rq, budget_token);
2641
2642 if (!blk_mq_get_driver_tag(rq)) {
2643 blk_mq_put_dispatch_budget(q, budget_token);
2644 goto insert;
2645 }
2646
2647 return __blk_mq_issue_directly(hctx, rq, last);
2648insert:
2649 if (bypass_insert)
2650 return BLK_STS_RESOURCE;
2651
2652 blk_mq_sched_insert_request(rq, false, run_queue, false);
2653
2654 return BLK_STS_OK;
2655}
2656
2657/**
2658 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2659 * @hctx: Pointer of the associated hardware queue.
2660 * @rq: Pointer to request to be sent.
2661 *
2662 * If the device has enough resources to accept a new request now, send the
2663 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2664 * we can try send it another time in the future. Requests inserted at this
2665 * queue have higher priority.
2666 */
2667static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2668 struct request *rq)
2669{
2670 blk_status_t ret =
2671 __blk_mq_try_issue_directly(hctx, rq, false, true);
2672
2673 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2674 blk_mq_request_bypass_insert(rq, false, true);
2675 else if (ret != BLK_STS_OK)
2676 blk_mq_end_request(rq, ret);
2677}
2678
2679static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2680{
2681 return __blk_mq_try_issue_directly(rq->mq_hctx, rq, true, last);
2682}
2683
2684static void blk_mq_plug_issue_direct(struct blk_plug *plug, bool from_schedule)
2685{
2686 struct blk_mq_hw_ctx *hctx = NULL;
2687 struct request *rq;
2688 int queued = 0;
2689 int errors = 0;
2690
2691 while ((rq = rq_list_pop(&plug->mq_list))) {
2692 bool last = rq_list_empty(plug->mq_list);
2693 blk_status_t ret;
2694
2695 if (hctx != rq->mq_hctx) {
2696 if (hctx)
2697 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2698 hctx = rq->mq_hctx;
2699 }
2700
2701 ret = blk_mq_request_issue_directly(rq, last);
2702 switch (ret) {
2703 case BLK_STS_OK:
2704 queued++;
2705 break;
2706 case BLK_STS_RESOURCE:
2707 case BLK_STS_DEV_RESOURCE:
2708 blk_mq_request_bypass_insert(rq, false, true);
2709 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2710 return;
2711 default:
2712 blk_mq_end_request(rq, ret);
2713 errors++;
2714 break;
2715 }
2716 }
2717
2718 /*
2719 * If we didn't flush the entire list, we could have told the driver
2720 * there was more coming, but that turned out to be a lie.
2721 */
2722 if (errors)
2723 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2724}
2725
2726static void __blk_mq_flush_plug_list(struct request_queue *q,
2727 struct blk_plug *plug)
2728{
2729 if (blk_queue_quiesced(q))
2730 return;
2731 q->mq_ops->queue_rqs(&plug->mq_list);
2732}
2733
2734static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2735{
2736 struct blk_mq_hw_ctx *this_hctx = NULL;
2737 struct blk_mq_ctx *this_ctx = NULL;
2738 struct request *requeue_list = NULL;
2739 unsigned int depth = 0;
2740 LIST_HEAD(list);
2741
2742 do {
2743 struct request *rq = rq_list_pop(&plug->mq_list);
2744
2745 if (!this_hctx) {
2746 this_hctx = rq->mq_hctx;
2747 this_ctx = rq->mq_ctx;
2748 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx) {
2749 rq_list_add(&requeue_list, rq);
2750 continue;
2751 }
2752 list_add_tail(&rq->queuelist, &list);
2753 depth++;
2754 } while (!rq_list_empty(plug->mq_list));
2755
2756 plug->mq_list = requeue_list;
2757 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2758 blk_mq_sched_insert_requests(this_hctx, this_ctx, &list, from_sched);
2759}
2760
2761void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2762{
2763 struct request *rq;
2764
2765 if (rq_list_empty(plug->mq_list))
2766 return;
2767 plug->rq_count = 0;
2768
2769 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2770 struct request_queue *q;
2771
2772 rq = rq_list_peek(&plug->mq_list);
2773 q = rq->q;
2774
2775 /*
2776 * Peek first request and see if we have a ->queue_rqs() hook.
2777 * If we do, we can dispatch the whole plug list in one go. We
2778 * already know at this point that all requests belong to the
2779 * same queue, caller must ensure that's the case.
2780 *
2781 * Since we pass off the full list to the driver at this point,
2782 * we do not increment the active request count for the queue.
2783 * Bypass shared tags for now because of that.
2784 */
2785 if (q->mq_ops->queue_rqs &&
2786 !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2787 blk_mq_run_dispatch_ops(q,
2788 __blk_mq_flush_plug_list(q, plug));
2789 if (rq_list_empty(plug->mq_list))
2790 return;
2791 }
2792
2793 blk_mq_run_dispatch_ops(q,
2794 blk_mq_plug_issue_direct(plug, false));
2795 if (rq_list_empty(plug->mq_list))
2796 return;
2797 }
2798
2799 do {
2800 blk_mq_dispatch_plug_list(plug, from_schedule);
2801 } while (!rq_list_empty(plug->mq_list));
2802}
2803
2804void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2805 struct list_head *list)
2806{
2807 int queued = 0;
2808 int errors = 0;
2809
2810 while (!list_empty(list)) {
2811 blk_status_t ret;
2812 struct request *rq = list_first_entry(list, struct request,
2813 queuelist);
2814
2815 list_del_init(&rq->queuelist);
2816 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2817 if (ret != BLK_STS_OK) {
2818 errors++;
2819 if (ret == BLK_STS_RESOURCE ||
2820 ret == BLK_STS_DEV_RESOURCE) {
2821 blk_mq_request_bypass_insert(rq, false,
2822 list_empty(list));
2823 break;
2824 }
2825 blk_mq_end_request(rq, ret);
2826 } else
2827 queued++;
2828 }
2829
2830 /*
2831 * If we didn't flush the entire list, we could have told
2832 * the driver there was more coming, but that turned out to
2833 * be a lie.
2834 */
2835 if ((!list_empty(list) || errors) &&
2836 hctx->queue->mq_ops->commit_rqs && queued)
2837 hctx->queue->mq_ops->commit_rqs(hctx);
2838}
2839
2840static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2841 struct bio *bio, unsigned int nr_segs)
2842{
2843 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2844 if (blk_attempt_plug_merge(q, bio, nr_segs))
2845 return true;
2846 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2847 return true;
2848 }
2849 return false;
2850}
2851
2852static struct request *blk_mq_get_new_requests(struct request_queue *q,
2853 struct blk_plug *plug,
2854 struct bio *bio,
2855 unsigned int nsegs)
2856{
2857 struct blk_mq_alloc_data data = {
2858 .q = q,
2859 .nr_tags = 1,
2860 .cmd_flags = bio->bi_opf,
2861 };
2862 struct request *rq;
2863
2864 if (unlikely(bio_queue_enter(bio)))
2865 return NULL;
2866
2867 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2868 goto queue_exit;
2869
2870 rq_qos_throttle(q, bio);
2871
2872 if (plug) {
2873 data.nr_tags = plug->nr_ios;
2874 plug->nr_ios = 1;
2875 data.cached_rq = &plug->cached_rq;
2876 }
2877
2878 rq = __blk_mq_alloc_requests(&data);
2879 if (rq)
2880 return rq;
2881 rq_qos_cleanup(q, bio);
2882 if (bio->bi_opf & REQ_NOWAIT)
2883 bio_wouldblock_error(bio);
2884queue_exit:
2885 blk_queue_exit(q);
2886 return NULL;
2887}
2888
2889static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
2890 struct blk_plug *plug, struct bio **bio, unsigned int nsegs)
2891{
2892 struct request *rq;
2893 enum hctx_type type, hctx_type;
2894
2895 if (!plug)
2896 return NULL;
2897 rq = rq_list_peek(&plug->cached_rq);
2898 if (!rq || rq->q != q)
2899 return NULL;
2900
2901 if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) {
2902 *bio = NULL;
2903 return NULL;
2904 }
2905
2906 type = blk_mq_get_hctx_type((*bio)->bi_opf);
2907 hctx_type = rq->mq_hctx->type;
2908 if (type != hctx_type &&
2909 !(type == HCTX_TYPE_READ && hctx_type == HCTX_TYPE_DEFAULT))
2910 return NULL;
2911 if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf))
2912 return NULL;
2913
2914 /*
2915 * If any qos ->throttle() end up blocking, we will have flushed the
2916 * plug and hence killed the cached_rq list as well. Pop this entry
2917 * before we throttle.
2918 */
2919 plug->cached_rq = rq_list_next(rq);
2920 rq_qos_throttle(q, *bio);
2921
2922 rq->cmd_flags = (*bio)->bi_opf;
2923 INIT_LIST_HEAD(&rq->queuelist);
2924 return rq;
2925}
2926
2927static void bio_set_ioprio(struct bio *bio)
2928{
2929 /* Nobody set ioprio so far? Initialize it based on task's nice value */
2930 if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE)
2931 bio->bi_ioprio = get_current_ioprio();
2932 blkcg_set_ioprio(bio);
2933}
2934
2935/**
2936 * blk_mq_submit_bio - Create and send a request to block device.
2937 * @bio: Bio pointer.
2938 *
2939 * Builds up a request structure from @q and @bio and send to the device. The
2940 * request may not be queued directly to hardware if:
2941 * * This request can be merged with another one
2942 * * We want to place request at plug queue for possible future merging
2943 * * There is an IO scheduler active at this queue
2944 *
2945 * It will not queue the request if there is an error with the bio, or at the
2946 * request creation.
2947 */
2948void blk_mq_submit_bio(struct bio *bio)
2949{
2950 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2951 struct blk_plug *plug = blk_mq_plug(bio);
2952 const int is_sync = op_is_sync(bio->bi_opf);
2953 struct request *rq;
2954 unsigned int nr_segs = 1;
2955 blk_status_t ret;
2956
2957 bio = blk_queue_bounce(bio, q);
2958 if (bio_may_exceed_limits(bio, &q->limits)) {
2959 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2960 if (!bio)
2961 return;
2962 }
2963
2964 if (!bio_integrity_prep(bio))
2965 return;
2966
2967 bio_set_ioprio(bio);
2968
2969 rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs);
2970 if (!rq) {
2971 if (!bio)
2972 return;
2973 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2974 if (unlikely(!rq))
2975 return;
2976 }
2977
2978 trace_block_getrq(bio);
2979
2980 rq_qos_track(q, rq, bio);
2981
2982 blk_mq_bio_to_request(rq, bio, nr_segs);
2983
2984 ret = blk_crypto_init_request(rq);
2985 if (ret != BLK_STS_OK) {
2986 bio->bi_status = ret;
2987 bio_endio(bio);
2988 blk_mq_free_request(rq);
2989 return;
2990 }
2991
2992 if (op_is_flush(bio->bi_opf)) {
2993 blk_insert_flush(rq);
2994 return;
2995 }
2996
2997 if (plug)
2998 blk_add_rq_to_plug(plug, rq);
2999 else if ((rq->rq_flags & RQF_ELV) ||
3000 (rq->mq_hctx->dispatch_busy &&
3001 (q->nr_hw_queues == 1 || !is_sync)))
3002 blk_mq_sched_insert_request(rq, false, true, true);
3003 else
3004 blk_mq_run_dispatch_ops(rq->q,
3005 blk_mq_try_issue_directly(rq->mq_hctx, rq));
3006}
3007
3008#ifdef CONFIG_BLK_MQ_STACKING
3009/**
3010 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
3011 * @rq: the request being queued
3012 */
3013blk_status_t blk_insert_cloned_request(struct request *rq)
3014{
3015 struct request_queue *q = rq->q;
3016 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
3017 blk_status_t ret;
3018
3019 if (blk_rq_sectors(rq) > max_sectors) {
3020 /*
3021 * SCSI device does not have a good way to return if
3022 * Write Same/Zero is actually supported. If a device rejects
3023 * a non-read/write command (discard, write same,etc.) the
3024 * low-level device driver will set the relevant queue limit to
3025 * 0 to prevent blk-lib from issuing more of the offending
3026 * operations. Commands queued prior to the queue limit being
3027 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3028 * errors being propagated to upper layers.
3029 */
3030 if (max_sectors == 0)
3031 return BLK_STS_NOTSUPP;
3032
3033 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3034 __func__, blk_rq_sectors(rq), max_sectors);
3035 return BLK_STS_IOERR;
3036 }
3037
3038 /*
3039 * The queue settings related to segment counting may differ from the
3040 * original queue.
3041 */
3042 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
3043 if (rq->nr_phys_segments > queue_max_segments(q)) {
3044 printk(KERN_ERR "%s: over max segments limit. (%hu > %hu)\n",
3045 __func__, rq->nr_phys_segments, queue_max_segments(q));
3046 return BLK_STS_IOERR;
3047 }
3048
3049 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
3050 return BLK_STS_IOERR;
3051
3052 if (blk_crypto_insert_cloned_request(rq))
3053 return BLK_STS_IOERR;
3054
3055 blk_account_io_start(rq);
3056
3057 /*
3058 * Since we have a scheduler attached on the top device,
3059 * bypass a potential scheduler on the bottom device for
3060 * insert.
3061 */
3062 blk_mq_run_dispatch_ops(q,
3063 ret = blk_mq_request_issue_directly(rq, true));
3064 if (ret)
3065 blk_account_io_done(rq, ktime_get_ns());
3066 return ret;
3067}
3068EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3069
3070/**
3071 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3072 * @rq: the clone request to be cleaned up
3073 *
3074 * Description:
3075 * Free all bios in @rq for a cloned request.
3076 */
3077void blk_rq_unprep_clone(struct request *rq)
3078{
3079 struct bio *bio;
3080
3081 while ((bio = rq->bio) != NULL) {
3082 rq->bio = bio->bi_next;
3083
3084 bio_put(bio);
3085 }
3086}
3087EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3088
3089/**
3090 * blk_rq_prep_clone - Helper function to setup clone request
3091 * @rq: the request to be setup
3092 * @rq_src: original request to be cloned
3093 * @bs: bio_set that bios for clone are allocated from
3094 * @gfp_mask: memory allocation mask for bio
3095 * @bio_ctr: setup function to be called for each clone bio.
3096 * Returns %0 for success, non %0 for failure.
3097 * @data: private data to be passed to @bio_ctr
3098 *
3099 * Description:
3100 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3101 * Also, pages which the original bios are pointing to are not copied
3102 * and the cloned bios just point same pages.
3103 * So cloned bios must be completed before original bios, which means
3104 * the caller must complete @rq before @rq_src.
3105 */
3106int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3107 struct bio_set *bs, gfp_t gfp_mask,
3108 int (*bio_ctr)(struct bio *, struct bio *, void *),
3109 void *data)
3110{
3111 struct bio *bio, *bio_src;
3112
3113 if (!bs)
3114 bs = &fs_bio_set;
3115
3116 __rq_for_each_bio(bio_src, rq_src) {
3117 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3118 bs);
3119 if (!bio)
3120 goto free_and_out;
3121
3122 if (bio_ctr && bio_ctr(bio, bio_src, data))
3123 goto free_and_out;
3124
3125 if (rq->bio) {
3126 rq->biotail->bi_next = bio;
3127 rq->biotail = bio;
3128 } else {
3129 rq->bio = rq->biotail = bio;
3130 }
3131 bio = NULL;
3132 }
3133
3134 /* Copy attributes of the original request to the clone request. */
3135 rq->__sector = blk_rq_pos(rq_src);
3136 rq->__data_len = blk_rq_bytes(rq_src);
3137 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3138 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3139 rq->special_vec = rq_src->special_vec;
3140 }
3141 rq->nr_phys_segments = rq_src->nr_phys_segments;
3142 rq->ioprio = rq_src->ioprio;
3143
3144 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3145 goto free_and_out;
3146
3147 return 0;
3148
3149free_and_out:
3150 if (bio)
3151 bio_put(bio);
3152 blk_rq_unprep_clone(rq);
3153
3154 return -ENOMEM;
3155}
3156EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3157#endif /* CONFIG_BLK_MQ_STACKING */
3158
3159/*
3160 * Steal bios from a request and add them to a bio list.
3161 * The request must not have been partially completed before.
3162 */
3163void blk_steal_bios(struct bio_list *list, struct request *rq)
3164{
3165 if (rq->bio) {
3166 if (list->tail)
3167 list->tail->bi_next = rq->bio;
3168 else
3169 list->head = rq->bio;
3170 list->tail = rq->biotail;
3171
3172 rq->bio = NULL;
3173 rq->biotail = NULL;
3174 }
3175
3176 rq->__data_len = 0;
3177}
3178EXPORT_SYMBOL_GPL(blk_steal_bios);
3179
3180static size_t order_to_size(unsigned int order)
3181{
3182 return (size_t)PAGE_SIZE << order;
3183}
3184
3185/* called before freeing request pool in @tags */
3186static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3187 struct blk_mq_tags *tags)
3188{
3189 struct page *page;
3190 unsigned long flags;
3191
3192 /*
3193 * There is no need to clear mapping if driver tags is not initialized
3194 * or the mapping belongs to the driver tags.
3195 */
3196 if (!drv_tags || drv_tags == tags)
3197 return;
3198
3199 list_for_each_entry(page, &tags->page_list, lru) {
3200 unsigned long start = (unsigned long)page_address(page);
3201 unsigned long end = start + order_to_size(page->private);
3202 int i;
3203
3204 for (i = 0; i < drv_tags->nr_tags; i++) {
3205 struct request *rq = drv_tags->rqs[i];
3206 unsigned long rq_addr = (unsigned long)rq;
3207
3208 if (rq_addr >= start && rq_addr < end) {
3209 WARN_ON_ONCE(req_ref_read(rq) != 0);
3210 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3211 }
3212 }
3213 }
3214
3215 /*
3216 * Wait until all pending iteration is done.
3217 *
3218 * Request reference is cleared and it is guaranteed to be observed
3219 * after the ->lock is released.
3220 */
3221 spin_lock_irqsave(&drv_tags->lock, flags);
3222 spin_unlock_irqrestore(&drv_tags->lock, flags);
3223}
3224
3225void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3226 unsigned int hctx_idx)
3227{
3228 struct blk_mq_tags *drv_tags;
3229 struct page *page;
3230
3231 if (list_empty(&tags->page_list))
3232 return;
3233
3234 if (blk_mq_is_shared_tags(set->flags))
3235 drv_tags = set->shared_tags;
3236 else
3237 drv_tags = set->tags[hctx_idx];
3238
3239 if (tags->static_rqs && set->ops->exit_request) {
3240 int i;
3241
3242 for (i = 0; i < tags->nr_tags; i++) {
3243 struct request *rq = tags->static_rqs[i];
3244
3245 if (!rq)
3246 continue;
3247 set->ops->exit_request(set, rq, hctx_idx);
3248 tags->static_rqs[i] = NULL;
3249 }
3250 }
3251
3252 blk_mq_clear_rq_mapping(drv_tags, tags);
3253
3254 while (!list_empty(&tags->page_list)) {
3255 page = list_first_entry(&tags->page_list, struct page, lru);
3256 list_del_init(&page->lru);
3257 /*
3258 * Remove kmemleak object previously allocated in
3259 * blk_mq_alloc_rqs().
3260 */
3261 kmemleak_free(page_address(page));
3262 __free_pages(page, page->private);
3263 }
3264}
3265
3266void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3267{
3268 kfree(tags->rqs);
3269 tags->rqs = NULL;
3270 kfree(tags->static_rqs);
3271 tags->static_rqs = NULL;
3272
3273 blk_mq_free_tags(tags);
3274}
3275
3276static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3277 unsigned int hctx_idx)
3278{
3279 int i;
3280
3281 for (i = 0; i < set->nr_maps; i++) {
3282 unsigned int start = set->map[i].queue_offset;
3283 unsigned int end = start + set->map[i].nr_queues;
3284
3285 if (hctx_idx >= start && hctx_idx < end)
3286 break;
3287 }
3288
3289 if (i >= set->nr_maps)
3290 i = HCTX_TYPE_DEFAULT;
3291
3292 return i;
3293}
3294
3295static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3296 unsigned int hctx_idx)
3297{
3298 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3299
3300 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3301}
3302
3303static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3304 unsigned int hctx_idx,
3305 unsigned int nr_tags,
3306 unsigned int reserved_tags)
3307{
3308 int node = blk_mq_get_hctx_node(set, hctx_idx);
3309 struct blk_mq_tags *tags;
3310
3311 if (node == NUMA_NO_NODE)
3312 node = set->numa_node;
3313
3314 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3315 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3316 if (!tags)
3317 return NULL;
3318
3319 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3320 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3321 node);
3322 if (!tags->rqs)
3323 goto err_free_tags;
3324
3325 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3326 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3327 node);
3328 if (!tags->static_rqs)
3329 goto err_free_rqs;
3330
3331 return tags;
3332
3333err_free_rqs:
3334 kfree(tags->rqs);
3335err_free_tags:
3336 blk_mq_free_tags(tags);
3337 return NULL;
3338}
3339
3340static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3341 unsigned int hctx_idx, int node)
3342{
3343 int ret;
3344
3345 if (set->ops->init_request) {
3346 ret = set->ops->init_request(set, rq, hctx_idx, node);
3347 if (ret)
3348 return ret;
3349 }
3350
3351 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3352 return 0;
3353}
3354
3355static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3356 struct blk_mq_tags *tags,
3357 unsigned int hctx_idx, unsigned int depth)
3358{
3359 unsigned int i, j, entries_per_page, max_order = 4;
3360 int node = blk_mq_get_hctx_node(set, hctx_idx);
3361 size_t rq_size, left;
3362
3363 if (node == NUMA_NO_NODE)
3364 node = set->numa_node;
3365
3366 INIT_LIST_HEAD(&tags->page_list);
3367
3368 /*
3369 * rq_size is the size of the request plus driver payload, rounded
3370 * to the cacheline size
3371 */
3372 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3373 cache_line_size());
3374 left = rq_size * depth;
3375
3376 for (i = 0; i < depth; ) {
3377 int this_order = max_order;
3378 struct page *page;
3379 int to_do;
3380 void *p;
3381
3382 while (this_order && left < order_to_size(this_order - 1))
3383 this_order--;
3384
3385 do {
3386 page = alloc_pages_node(node,
3387 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3388 this_order);
3389 if (page)
3390 break;
3391 if (!this_order--)
3392 break;
3393 if (order_to_size(this_order) < rq_size)
3394 break;
3395 } while (1);
3396
3397 if (!page)
3398 goto fail;
3399
3400 page->private = this_order;
3401 list_add_tail(&page->lru, &tags->page_list);
3402
3403 p = page_address(page);
3404 /*
3405 * Allow kmemleak to scan these pages as they contain pointers
3406 * to additional allocations like via ops->init_request().
3407 */
3408 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3409 entries_per_page = order_to_size(this_order) / rq_size;
3410 to_do = min(entries_per_page, depth - i);
3411 left -= to_do * rq_size;
3412 for (j = 0; j < to_do; j++) {
3413 struct request *rq = p;
3414
3415 tags->static_rqs[i] = rq;
3416 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3417 tags->static_rqs[i] = NULL;
3418 goto fail;
3419 }
3420
3421 p += rq_size;
3422 i++;
3423 }
3424 }
3425 return 0;
3426
3427fail:
3428 blk_mq_free_rqs(set, tags, hctx_idx);
3429 return -ENOMEM;
3430}
3431
3432struct rq_iter_data {
3433 struct blk_mq_hw_ctx *hctx;
3434 bool has_rq;
3435};
3436
3437static bool blk_mq_has_request(struct request *rq, void *data)
3438{
3439 struct rq_iter_data *iter_data = data;
3440
3441 if (rq->mq_hctx != iter_data->hctx)
3442 return true;
3443 iter_data->has_rq = true;
3444 return false;
3445}
3446
3447static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3448{
3449 struct blk_mq_tags *tags = hctx->sched_tags ?
3450 hctx->sched_tags : hctx->tags;
3451 struct rq_iter_data data = {
3452 .hctx = hctx,
3453 };
3454
3455 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3456 return data.has_rq;
3457}
3458
3459static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3460 struct blk_mq_hw_ctx *hctx)
3461{
3462 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3463 return false;
3464 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3465 return false;
3466 return true;
3467}
3468
3469static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3470{
3471 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3472 struct blk_mq_hw_ctx, cpuhp_online);
3473
3474 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3475 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3476 return 0;
3477
3478 /*
3479 * Prevent new request from being allocated on the current hctx.
3480 *
3481 * The smp_mb__after_atomic() Pairs with the implied barrier in
3482 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3483 * seen once we return from the tag allocator.
3484 */
3485 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3486 smp_mb__after_atomic();
3487
3488 /*
3489 * Try to grab a reference to the queue and wait for any outstanding
3490 * requests. If we could not grab a reference the queue has been
3491 * frozen and there are no requests.
3492 */
3493 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3494 while (blk_mq_hctx_has_requests(hctx))
3495 msleep(5);
3496 percpu_ref_put(&hctx->queue->q_usage_counter);
3497 }
3498
3499 return 0;
3500}
3501
3502static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3503{
3504 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3505 struct blk_mq_hw_ctx, cpuhp_online);
3506
3507 if (cpumask_test_cpu(cpu, hctx->cpumask))
3508 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3509 return 0;
3510}
3511
3512/*
3513 * 'cpu' is going away. splice any existing rq_list entries from this
3514 * software queue to the hw queue dispatch list, and ensure that it
3515 * gets run.
3516 */
3517static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3518{
3519 struct blk_mq_hw_ctx *hctx;
3520 struct blk_mq_ctx *ctx;
3521 LIST_HEAD(tmp);
3522 enum hctx_type type;
3523
3524 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3525 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3526 return 0;
3527
3528 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3529 type = hctx->type;
3530
3531 spin_lock(&ctx->lock);
3532 if (!list_empty(&ctx->rq_lists[type])) {
3533 list_splice_init(&ctx->rq_lists[type], &tmp);
3534 blk_mq_hctx_clear_pending(hctx, ctx);
3535 }
3536 spin_unlock(&ctx->lock);
3537
3538 if (list_empty(&tmp))
3539 return 0;
3540
3541 spin_lock(&hctx->lock);
3542 list_splice_tail_init(&tmp, &hctx->dispatch);
3543 spin_unlock(&hctx->lock);
3544
3545 blk_mq_run_hw_queue(hctx, true);
3546 return 0;
3547}
3548
3549static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3550{
3551 if (!(hctx->flags & BLK_MQ_F_STACKING))
3552 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3553 &hctx->cpuhp_online);
3554 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3555 &hctx->cpuhp_dead);
3556}
3557
3558/*
3559 * Before freeing hw queue, clearing the flush request reference in
3560 * tags->rqs[] for avoiding potential UAF.
3561 */
3562static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3563 unsigned int queue_depth, struct request *flush_rq)
3564{
3565 int i;
3566 unsigned long flags;
3567
3568 /* The hw queue may not be mapped yet */
3569 if (!tags)
3570 return;
3571
3572 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3573
3574 for (i = 0; i < queue_depth; i++)
3575 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3576
3577 /*
3578 * Wait until all pending iteration is done.
3579 *
3580 * Request reference is cleared and it is guaranteed to be observed
3581 * after the ->lock is released.
3582 */
3583 spin_lock_irqsave(&tags->lock, flags);
3584 spin_unlock_irqrestore(&tags->lock, flags);
3585}
3586
3587/* hctx->ctxs will be freed in queue's release handler */
3588static void blk_mq_exit_hctx(struct request_queue *q,
3589 struct blk_mq_tag_set *set,
3590 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3591{
3592 struct request *flush_rq = hctx->fq->flush_rq;
3593
3594 if (blk_mq_hw_queue_mapped(hctx))
3595 blk_mq_tag_idle(hctx);
3596
3597 if (blk_queue_init_done(q))
3598 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3599 set->queue_depth, flush_rq);
3600 if (set->ops->exit_request)
3601 set->ops->exit_request(set, flush_rq, hctx_idx);
3602
3603 if (set->ops->exit_hctx)
3604 set->ops->exit_hctx(hctx, hctx_idx);
3605
3606 blk_mq_remove_cpuhp(hctx);
3607
3608 xa_erase(&q->hctx_table, hctx_idx);
3609
3610 spin_lock(&q->unused_hctx_lock);
3611 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3612 spin_unlock(&q->unused_hctx_lock);
3613}
3614
3615static void blk_mq_exit_hw_queues(struct request_queue *q,
3616 struct blk_mq_tag_set *set, int nr_queue)
3617{
3618 struct blk_mq_hw_ctx *hctx;
3619 unsigned long i;
3620
3621 queue_for_each_hw_ctx(q, hctx, i) {
3622 if (i == nr_queue)
3623 break;
3624 blk_mq_exit_hctx(q, set, hctx, i);
3625 }
3626}
3627
3628static int blk_mq_init_hctx(struct request_queue *q,
3629 struct blk_mq_tag_set *set,
3630 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3631{
3632 hctx->queue_num = hctx_idx;
3633
3634 if (!(hctx->flags & BLK_MQ_F_STACKING))
3635 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3636 &hctx->cpuhp_online);
3637 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3638
3639 hctx->tags = set->tags[hctx_idx];
3640
3641 if (set->ops->init_hctx &&
3642 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3643 goto unregister_cpu_notifier;
3644
3645 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3646 hctx->numa_node))
3647 goto exit_hctx;
3648
3649 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3650 goto exit_flush_rq;
3651
3652 return 0;
3653
3654 exit_flush_rq:
3655 if (set->ops->exit_request)
3656 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3657 exit_hctx:
3658 if (set->ops->exit_hctx)
3659 set->ops->exit_hctx(hctx, hctx_idx);
3660 unregister_cpu_notifier:
3661 blk_mq_remove_cpuhp(hctx);
3662 return -1;
3663}
3664
3665static struct blk_mq_hw_ctx *
3666blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3667 int node)
3668{
3669 struct blk_mq_hw_ctx *hctx;
3670 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3671
3672 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3673 if (!hctx)
3674 goto fail_alloc_hctx;
3675
3676 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3677 goto free_hctx;
3678
3679 atomic_set(&hctx->nr_active, 0);
3680 if (node == NUMA_NO_NODE)
3681 node = set->numa_node;
3682 hctx->numa_node = node;
3683
3684 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3685 spin_lock_init(&hctx->lock);
3686 INIT_LIST_HEAD(&hctx->dispatch);
3687 hctx->queue = q;
3688 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3689
3690 INIT_LIST_HEAD(&hctx->hctx_list);
3691
3692 /*
3693 * Allocate space for all possible cpus to avoid allocation at
3694 * runtime
3695 */
3696 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3697 gfp, node);
3698 if (!hctx->ctxs)
3699 goto free_cpumask;
3700
3701 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3702 gfp, node, false, false))
3703 goto free_ctxs;
3704 hctx->nr_ctx = 0;
3705
3706 spin_lock_init(&hctx->dispatch_wait_lock);
3707 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3708 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3709
3710 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3711 if (!hctx->fq)
3712 goto free_bitmap;
3713
3714 blk_mq_hctx_kobj_init(hctx);
3715
3716 return hctx;
3717
3718 free_bitmap:
3719 sbitmap_free(&hctx->ctx_map);
3720 free_ctxs:
3721 kfree(hctx->ctxs);
3722 free_cpumask:
3723 free_cpumask_var(hctx->cpumask);
3724 free_hctx:
3725 kfree(hctx);
3726 fail_alloc_hctx:
3727 return NULL;
3728}
3729
3730static void blk_mq_init_cpu_queues(struct request_queue *q,
3731 unsigned int nr_hw_queues)
3732{
3733 struct blk_mq_tag_set *set = q->tag_set;
3734 unsigned int i, j;
3735
3736 for_each_possible_cpu(i) {
3737 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3738 struct blk_mq_hw_ctx *hctx;
3739 int k;
3740
3741 __ctx->cpu = i;
3742 spin_lock_init(&__ctx->lock);
3743 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3744 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3745
3746 __ctx->queue = q;
3747
3748 /*
3749 * Set local node, IFF we have more than one hw queue. If
3750 * not, we remain on the home node of the device
3751 */
3752 for (j = 0; j < set->nr_maps; j++) {
3753 hctx = blk_mq_map_queue_type(q, j, i);
3754 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3755 hctx->numa_node = cpu_to_node(i);
3756 }
3757 }
3758}
3759
3760struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3761 unsigned int hctx_idx,
3762 unsigned int depth)
3763{
3764 struct blk_mq_tags *tags;
3765 int ret;
3766
3767 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3768 if (!tags)
3769 return NULL;
3770
3771 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3772 if (ret) {
3773 blk_mq_free_rq_map(tags);
3774 return NULL;
3775 }
3776
3777 return tags;
3778}
3779
3780static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3781 int hctx_idx)
3782{
3783 if (blk_mq_is_shared_tags(set->flags)) {
3784 set->tags[hctx_idx] = set->shared_tags;
3785
3786 return true;
3787 }
3788
3789 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3790 set->queue_depth);
3791
3792 return set->tags[hctx_idx];
3793}
3794
3795void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3796 struct blk_mq_tags *tags,
3797 unsigned int hctx_idx)
3798{
3799 if (tags) {
3800 blk_mq_free_rqs(set, tags, hctx_idx);
3801 blk_mq_free_rq_map(tags);
3802 }
3803}
3804
3805static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3806 unsigned int hctx_idx)
3807{
3808 if (!blk_mq_is_shared_tags(set->flags))
3809 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3810
3811 set->tags[hctx_idx] = NULL;
3812}
3813
3814static void blk_mq_map_swqueue(struct request_queue *q)
3815{
3816 unsigned int j, hctx_idx;
3817 unsigned long i;
3818 struct blk_mq_hw_ctx *hctx;
3819 struct blk_mq_ctx *ctx;
3820 struct blk_mq_tag_set *set = q->tag_set;
3821
3822 queue_for_each_hw_ctx(q, hctx, i) {
3823 cpumask_clear(hctx->cpumask);
3824 hctx->nr_ctx = 0;
3825 hctx->dispatch_from = NULL;
3826 }
3827
3828 /*
3829 * Map software to hardware queues.
3830 *
3831 * If the cpu isn't present, the cpu is mapped to first hctx.
3832 */
3833 for_each_possible_cpu(i) {
3834
3835 ctx = per_cpu_ptr(q->queue_ctx, i);
3836 for (j = 0; j < set->nr_maps; j++) {
3837 if (!set->map[j].nr_queues) {
3838 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3839 HCTX_TYPE_DEFAULT, i);
3840 continue;
3841 }
3842 hctx_idx = set->map[j].mq_map[i];
3843 /* unmapped hw queue can be remapped after CPU topo changed */
3844 if (!set->tags[hctx_idx] &&
3845 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3846 /*
3847 * If tags initialization fail for some hctx,
3848 * that hctx won't be brought online. In this
3849 * case, remap the current ctx to hctx[0] which
3850 * is guaranteed to always have tags allocated
3851 */
3852 set->map[j].mq_map[i] = 0;
3853 }
3854
3855 hctx = blk_mq_map_queue_type(q, j, i);
3856 ctx->hctxs[j] = hctx;
3857 /*
3858 * If the CPU is already set in the mask, then we've
3859 * mapped this one already. This can happen if
3860 * devices share queues across queue maps.
3861 */
3862 if (cpumask_test_cpu(i, hctx->cpumask))
3863 continue;
3864
3865 cpumask_set_cpu(i, hctx->cpumask);
3866 hctx->type = j;
3867 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3868 hctx->ctxs[hctx->nr_ctx++] = ctx;
3869
3870 /*
3871 * If the nr_ctx type overflows, we have exceeded the
3872 * amount of sw queues we can support.
3873 */
3874 BUG_ON(!hctx->nr_ctx);
3875 }
3876
3877 for (; j < HCTX_MAX_TYPES; j++)
3878 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3879 HCTX_TYPE_DEFAULT, i);
3880 }
3881
3882 queue_for_each_hw_ctx(q, hctx, i) {
3883 /*
3884 * If no software queues are mapped to this hardware queue,
3885 * disable it and free the request entries.
3886 */
3887 if (!hctx->nr_ctx) {
3888 /* Never unmap queue 0. We need it as a
3889 * fallback in case of a new remap fails
3890 * allocation
3891 */
3892 if (i)
3893 __blk_mq_free_map_and_rqs(set, i);
3894
3895 hctx->tags = NULL;
3896 continue;
3897 }
3898
3899 hctx->tags = set->tags[i];
3900 WARN_ON(!hctx->tags);
3901
3902 /*
3903 * Set the map size to the number of mapped software queues.
3904 * This is more accurate and more efficient than looping
3905 * over all possibly mapped software queues.
3906 */
3907 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3908
3909 /*
3910 * Initialize batch roundrobin counts
3911 */
3912 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3913 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3914 }
3915}
3916
3917/*
3918 * Caller needs to ensure that we're either frozen/quiesced, or that
3919 * the queue isn't live yet.
3920 */
3921static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3922{
3923 struct blk_mq_hw_ctx *hctx;
3924 unsigned long i;
3925
3926 queue_for_each_hw_ctx(q, hctx, i) {
3927 if (shared) {
3928 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3929 } else {
3930 blk_mq_tag_idle(hctx);
3931 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3932 }
3933 }
3934}
3935
3936static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3937 bool shared)
3938{
3939 struct request_queue *q;
3940
3941 lockdep_assert_held(&set->tag_list_lock);
3942
3943 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3944 blk_mq_freeze_queue(q);
3945 queue_set_hctx_shared(q, shared);
3946 blk_mq_unfreeze_queue(q);
3947 }
3948}
3949
3950static void blk_mq_del_queue_tag_set(struct request_queue *q)
3951{
3952 struct blk_mq_tag_set *set = q->tag_set;
3953
3954 mutex_lock(&set->tag_list_lock);
3955 list_del(&q->tag_set_list);
3956 if (list_is_singular(&set->tag_list)) {
3957 /* just transitioned to unshared */
3958 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3959 /* update existing queue */
3960 blk_mq_update_tag_set_shared(set, false);
3961 }
3962 mutex_unlock(&set->tag_list_lock);
3963 INIT_LIST_HEAD(&q->tag_set_list);
3964}
3965
3966static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3967 struct request_queue *q)
3968{
3969 mutex_lock(&set->tag_list_lock);
3970
3971 /*
3972 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3973 */
3974 if (!list_empty(&set->tag_list) &&
3975 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3976 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3977 /* update existing queue */
3978 blk_mq_update_tag_set_shared(set, true);
3979 }
3980 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3981 queue_set_hctx_shared(q, true);
3982 list_add_tail(&q->tag_set_list, &set->tag_list);
3983
3984 mutex_unlock(&set->tag_list_lock);
3985}
3986
3987/* All allocations will be freed in release handler of q->mq_kobj */
3988static int blk_mq_alloc_ctxs(struct request_queue *q)
3989{
3990 struct blk_mq_ctxs *ctxs;
3991 int cpu;
3992
3993 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3994 if (!ctxs)
3995 return -ENOMEM;
3996
3997 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3998 if (!ctxs->queue_ctx)
3999 goto fail;
4000
4001 for_each_possible_cpu(cpu) {
4002 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
4003 ctx->ctxs = ctxs;
4004 }
4005
4006 q->mq_kobj = &ctxs->kobj;
4007 q->queue_ctx = ctxs->queue_ctx;
4008
4009 return 0;
4010 fail:
4011 kfree(ctxs);
4012 return -ENOMEM;
4013}
4014
4015/*
4016 * It is the actual release handler for mq, but we do it from
4017 * request queue's release handler for avoiding use-after-free
4018 * and headache because q->mq_kobj shouldn't have been introduced,
4019 * but we can't group ctx/kctx kobj without it.
4020 */
4021void blk_mq_release(struct request_queue *q)
4022{
4023 struct blk_mq_hw_ctx *hctx, *next;
4024 unsigned long i;
4025
4026 queue_for_each_hw_ctx(q, hctx, i)
4027 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4028
4029 /* all hctx are in .unused_hctx_list now */
4030 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4031 list_del_init(&hctx->hctx_list);
4032 kobject_put(&hctx->kobj);
4033 }
4034
4035 xa_destroy(&q->hctx_table);
4036
4037 /*
4038 * release .mq_kobj and sw queue's kobject now because
4039 * both share lifetime with request queue.
4040 */
4041 blk_mq_sysfs_deinit(q);
4042}
4043
4044static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
4045 void *queuedata)
4046{
4047 struct request_queue *q;
4048 int ret;
4049
4050 q = blk_alloc_queue(set->numa_node);
4051 if (!q)
4052 return ERR_PTR(-ENOMEM);
4053 q->queuedata = queuedata;
4054 ret = blk_mq_init_allocated_queue(set, q);
4055 if (ret) {
4056 blk_put_queue(q);
4057 return ERR_PTR(ret);
4058 }
4059 return q;
4060}
4061
4062struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
4063{
4064 return blk_mq_init_queue_data(set, NULL);
4065}
4066EXPORT_SYMBOL(blk_mq_init_queue);
4067
4068/**
4069 * blk_mq_destroy_queue - shutdown a request queue
4070 * @q: request queue to shutdown
4071 *
4072 * This shuts down a request queue allocated by blk_mq_init_queue(). All future
4073 * requests will be failed with -ENODEV. The caller is responsible for dropping
4074 * the reference from blk_mq_init_queue() by calling blk_put_queue().
4075 *
4076 * Context: can sleep
4077 */
4078void blk_mq_destroy_queue(struct request_queue *q)
4079{
4080 WARN_ON_ONCE(!queue_is_mq(q));
4081 WARN_ON_ONCE(blk_queue_registered(q));
4082
4083 might_sleep();
4084
4085 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4086 blk_queue_start_drain(q);
4087 blk_mq_freeze_queue_wait(q);
4088
4089 blk_sync_queue(q);
4090 blk_mq_cancel_work_sync(q);
4091 blk_mq_exit_queue(q);
4092}
4093EXPORT_SYMBOL(blk_mq_destroy_queue);
4094
4095struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
4096 struct lock_class_key *lkclass)
4097{
4098 struct request_queue *q;
4099 struct gendisk *disk;
4100
4101 q = blk_mq_init_queue_data(set, queuedata);
4102 if (IS_ERR(q))
4103 return ERR_CAST(q);
4104
4105 disk = __alloc_disk_node(q, set->numa_node, lkclass);
4106 if (!disk) {
4107 blk_mq_destroy_queue(q);
4108 blk_put_queue(q);
4109 return ERR_PTR(-ENOMEM);
4110 }
4111 set_bit(GD_OWNS_QUEUE, &disk->state);
4112 return disk;
4113}
4114EXPORT_SYMBOL(__blk_mq_alloc_disk);
4115
4116struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4117 struct lock_class_key *lkclass)
4118{
4119 struct gendisk *disk;
4120
4121 if (!blk_get_queue(q))
4122 return NULL;
4123 disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4124 if (!disk)
4125 blk_put_queue(q);
4126 return disk;
4127}
4128EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4129
4130static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4131 struct blk_mq_tag_set *set, struct request_queue *q,
4132 int hctx_idx, int node)
4133{
4134 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4135
4136 /* reuse dead hctx first */
4137 spin_lock(&q->unused_hctx_lock);
4138 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4139 if (tmp->numa_node == node) {
4140 hctx = tmp;
4141 break;
4142 }
4143 }
4144 if (hctx)
4145 list_del_init(&hctx->hctx_list);
4146 spin_unlock(&q->unused_hctx_lock);
4147
4148 if (!hctx)
4149 hctx = blk_mq_alloc_hctx(q, set, node);
4150 if (!hctx)
4151 goto fail;
4152
4153 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4154 goto free_hctx;
4155
4156 return hctx;
4157
4158 free_hctx:
4159 kobject_put(&hctx->kobj);
4160 fail:
4161 return NULL;
4162}
4163
4164static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4165 struct request_queue *q)
4166{
4167 struct blk_mq_hw_ctx *hctx;
4168 unsigned long i, j;
4169
4170 /* protect against switching io scheduler */
4171 mutex_lock(&q->sysfs_lock);
4172 for (i = 0; i < set->nr_hw_queues; i++) {
4173 int old_node;
4174 int node = blk_mq_get_hctx_node(set, i);
4175 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4176
4177 if (old_hctx) {
4178 old_node = old_hctx->numa_node;
4179 blk_mq_exit_hctx(q, set, old_hctx, i);
4180 }
4181
4182 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4183 if (!old_hctx)
4184 break;
4185 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4186 node, old_node);
4187 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4188 WARN_ON_ONCE(!hctx);
4189 }
4190 }
4191 /*
4192 * Increasing nr_hw_queues fails. Free the newly allocated
4193 * hctxs and keep the previous q->nr_hw_queues.
4194 */
4195 if (i != set->nr_hw_queues) {
4196 j = q->nr_hw_queues;
4197 } else {
4198 j = i;
4199 q->nr_hw_queues = set->nr_hw_queues;
4200 }
4201
4202 xa_for_each_start(&q->hctx_table, j, hctx, j)
4203 blk_mq_exit_hctx(q, set, hctx, j);
4204 mutex_unlock(&q->sysfs_lock);
4205}
4206
4207static void blk_mq_update_poll_flag(struct request_queue *q)
4208{
4209 struct blk_mq_tag_set *set = q->tag_set;
4210
4211 if (set->nr_maps > HCTX_TYPE_POLL &&
4212 set->map[HCTX_TYPE_POLL].nr_queues)
4213 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4214 else
4215 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4216}
4217
4218int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4219 struct request_queue *q)
4220{
4221 /* mark the queue as mq asap */
4222 q->mq_ops = set->ops;
4223
4224 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
4225 blk_mq_poll_stats_bkt,
4226 BLK_MQ_POLL_STATS_BKTS, q);
4227 if (!q->poll_cb)
4228 goto err_exit;
4229
4230 if (blk_mq_alloc_ctxs(q))
4231 goto err_poll;
4232
4233 /* init q->mq_kobj and sw queues' kobjects */
4234 blk_mq_sysfs_init(q);
4235
4236 INIT_LIST_HEAD(&q->unused_hctx_list);
4237 spin_lock_init(&q->unused_hctx_lock);
4238
4239 xa_init(&q->hctx_table);
4240
4241 blk_mq_realloc_hw_ctxs(set, q);
4242 if (!q->nr_hw_queues)
4243 goto err_hctxs;
4244
4245 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4246 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4247
4248 q->tag_set = set;
4249
4250 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4251 blk_mq_update_poll_flag(q);
4252
4253 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4254 INIT_LIST_HEAD(&q->requeue_list);
4255 spin_lock_init(&q->requeue_lock);
4256
4257 q->nr_requests = set->queue_depth;
4258
4259 /*
4260 * Default to classic polling
4261 */
4262 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
4263
4264 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4265 blk_mq_add_queue_tag_set(set, q);
4266 blk_mq_map_swqueue(q);
4267 return 0;
4268
4269err_hctxs:
4270 blk_mq_release(q);
4271err_poll:
4272 blk_stat_free_callback(q->poll_cb);
4273 q->poll_cb = NULL;
4274err_exit:
4275 q->mq_ops = NULL;
4276 return -ENOMEM;
4277}
4278EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4279
4280/* tags can _not_ be used after returning from blk_mq_exit_queue */
4281void blk_mq_exit_queue(struct request_queue *q)
4282{
4283 struct blk_mq_tag_set *set = q->tag_set;
4284
4285 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4286 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4287 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4288 blk_mq_del_queue_tag_set(q);
4289}
4290
4291static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4292{
4293 int i;
4294
4295 if (blk_mq_is_shared_tags(set->flags)) {
4296 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4297 BLK_MQ_NO_HCTX_IDX,
4298 set->queue_depth);
4299 if (!set->shared_tags)
4300 return -ENOMEM;
4301 }
4302
4303 for (i = 0; i < set->nr_hw_queues; i++) {
4304 if (!__blk_mq_alloc_map_and_rqs(set, i))
4305 goto out_unwind;
4306 cond_resched();
4307 }
4308
4309 return 0;
4310
4311out_unwind:
4312 while (--i >= 0)
4313 __blk_mq_free_map_and_rqs(set, i);
4314
4315 if (blk_mq_is_shared_tags(set->flags)) {
4316 blk_mq_free_map_and_rqs(set, set->shared_tags,
4317 BLK_MQ_NO_HCTX_IDX);
4318 }
4319
4320 return -ENOMEM;
4321}
4322
4323/*
4324 * Allocate the request maps associated with this tag_set. Note that this
4325 * may reduce the depth asked for, if memory is tight. set->queue_depth
4326 * will be updated to reflect the allocated depth.
4327 */
4328static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4329{
4330 unsigned int depth;
4331 int err;
4332
4333 depth = set->queue_depth;
4334 do {
4335 err = __blk_mq_alloc_rq_maps(set);
4336 if (!err)
4337 break;
4338
4339 set->queue_depth >>= 1;
4340 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4341 err = -ENOMEM;
4342 break;
4343 }
4344 } while (set->queue_depth);
4345
4346 if (!set->queue_depth || err) {
4347 pr_err("blk-mq: failed to allocate request map\n");
4348 return -ENOMEM;
4349 }
4350
4351 if (depth != set->queue_depth)
4352 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4353 depth, set->queue_depth);
4354
4355 return 0;
4356}
4357
4358static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4359{
4360 /*
4361 * blk_mq_map_queues() and multiple .map_queues() implementations
4362 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4363 * number of hardware queues.
4364 */
4365 if (set->nr_maps == 1)
4366 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4367
4368 if (set->ops->map_queues && !is_kdump_kernel()) {
4369 int i;
4370
4371 /*
4372 * transport .map_queues is usually done in the following
4373 * way:
4374 *
4375 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4376 * mask = get_cpu_mask(queue)
4377 * for_each_cpu(cpu, mask)
4378 * set->map[x].mq_map[cpu] = queue;
4379 * }
4380 *
4381 * When we need to remap, the table has to be cleared for
4382 * killing stale mapping since one CPU may not be mapped
4383 * to any hw queue.
4384 */
4385 for (i = 0; i < set->nr_maps; i++)
4386 blk_mq_clear_mq_map(&set->map[i]);
4387
4388 set->ops->map_queues(set);
4389 } else {
4390 BUG_ON(set->nr_maps > 1);
4391 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4392 }
4393}
4394
4395static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4396 int new_nr_hw_queues)
4397{
4398 struct blk_mq_tags **new_tags;
4399
4400 if (set->nr_hw_queues >= new_nr_hw_queues)
4401 goto done;
4402
4403 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4404 GFP_KERNEL, set->numa_node);
4405 if (!new_tags)
4406 return -ENOMEM;
4407
4408 if (set->tags)
4409 memcpy(new_tags, set->tags, set->nr_hw_queues *
4410 sizeof(*set->tags));
4411 kfree(set->tags);
4412 set->tags = new_tags;
4413done:
4414 set->nr_hw_queues = new_nr_hw_queues;
4415 return 0;
4416}
4417
4418/*
4419 * Alloc a tag set to be associated with one or more request queues.
4420 * May fail with EINVAL for various error conditions. May adjust the
4421 * requested depth down, if it's too large. In that case, the set
4422 * value will be stored in set->queue_depth.
4423 */
4424int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4425{
4426 int i, ret;
4427
4428 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4429
4430 if (!set->nr_hw_queues)
4431 return -EINVAL;
4432 if (!set->queue_depth)
4433 return -EINVAL;
4434 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4435 return -EINVAL;
4436
4437 if (!set->ops->queue_rq)
4438 return -EINVAL;
4439
4440 if (!set->ops->get_budget ^ !set->ops->put_budget)
4441 return -EINVAL;
4442
4443 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4444 pr_info("blk-mq: reduced tag depth to %u\n",
4445 BLK_MQ_MAX_DEPTH);
4446 set->queue_depth = BLK_MQ_MAX_DEPTH;
4447 }
4448
4449 if (!set->nr_maps)
4450 set->nr_maps = 1;
4451 else if (set->nr_maps > HCTX_MAX_TYPES)
4452 return -EINVAL;
4453
4454 /*
4455 * If a crashdump is active, then we are potentially in a very
4456 * memory constrained environment. Limit us to 1 queue and
4457 * 64 tags to prevent using too much memory.
4458 */
4459 if (is_kdump_kernel()) {
4460 set->nr_hw_queues = 1;
4461 set->nr_maps = 1;
4462 set->queue_depth = min(64U, set->queue_depth);
4463 }
4464 /*
4465 * There is no use for more h/w queues than cpus if we just have
4466 * a single map
4467 */
4468 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4469 set->nr_hw_queues = nr_cpu_ids;
4470
4471 if (set->flags & BLK_MQ_F_BLOCKING) {
4472 set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL);
4473 if (!set->srcu)
4474 return -ENOMEM;
4475 ret = init_srcu_struct(set->srcu);
4476 if (ret)
4477 goto out_free_srcu;
4478 }
4479
4480 ret = -ENOMEM;
4481 set->tags = kcalloc_node(set->nr_hw_queues,
4482 sizeof(struct blk_mq_tags *), GFP_KERNEL,
4483 set->numa_node);
4484 if (!set->tags)
4485 goto out_cleanup_srcu;
4486
4487 for (i = 0; i < set->nr_maps; i++) {
4488 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4489 sizeof(set->map[i].mq_map[0]),
4490 GFP_KERNEL, set->numa_node);
4491 if (!set->map[i].mq_map)
4492 goto out_free_mq_map;
4493 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4494 }
4495
4496 blk_mq_update_queue_map(set);
4497
4498 ret = blk_mq_alloc_set_map_and_rqs(set);
4499 if (ret)
4500 goto out_free_mq_map;
4501
4502 mutex_init(&set->tag_list_lock);
4503 INIT_LIST_HEAD(&set->tag_list);
4504
4505 return 0;
4506
4507out_free_mq_map:
4508 for (i = 0; i < set->nr_maps; i++) {
4509 kfree(set->map[i].mq_map);
4510 set->map[i].mq_map = NULL;
4511 }
4512 kfree(set->tags);
4513 set->tags = NULL;
4514out_cleanup_srcu:
4515 if (set->flags & BLK_MQ_F_BLOCKING)
4516 cleanup_srcu_struct(set->srcu);
4517out_free_srcu:
4518 if (set->flags & BLK_MQ_F_BLOCKING)
4519 kfree(set->srcu);
4520 return ret;
4521}
4522EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4523
4524/* allocate and initialize a tagset for a simple single-queue device */
4525int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4526 const struct blk_mq_ops *ops, unsigned int queue_depth,
4527 unsigned int set_flags)
4528{
4529 memset(set, 0, sizeof(*set));
4530 set->ops = ops;
4531 set->nr_hw_queues = 1;
4532 set->nr_maps = 1;
4533 set->queue_depth = queue_depth;
4534 set->numa_node = NUMA_NO_NODE;
4535 set->flags = set_flags;
4536 return blk_mq_alloc_tag_set(set);
4537}
4538EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4539
4540void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4541{
4542 int i, j;
4543
4544 for (i = 0; i < set->nr_hw_queues; i++)
4545 __blk_mq_free_map_and_rqs(set, i);
4546
4547 if (blk_mq_is_shared_tags(set->flags)) {
4548 blk_mq_free_map_and_rqs(set, set->shared_tags,
4549 BLK_MQ_NO_HCTX_IDX);
4550 }
4551
4552 for (j = 0; j < set->nr_maps; j++) {
4553 kfree(set->map[j].mq_map);
4554 set->map[j].mq_map = NULL;
4555 }
4556
4557 kfree(set->tags);
4558 set->tags = NULL;
4559 if (set->flags & BLK_MQ_F_BLOCKING) {
4560 cleanup_srcu_struct(set->srcu);
4561 kfree(set->srcu);
4562 }
4563}
4564EXPORT_SYMBOL(blk_mq_free_tag_set);
4565
4566int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4567{
4568 struct blk_mq_tag_set *set = q->tag_set;
4569 struct blk_mq_hw_ctx *hctx;
4570 int ret;
4571 unsigned long i;
4572
4573 if (!set)
4574 return -EINVAL;
4575
4576 if (q->nr_requests == nr)
4577 return 0;
4578
4579 blk_mq_freeze_queue(q);
4580 blk_mq_quiesce_queue(q);
4581
4582 ret = 0;
4583 queue_for_each_hw_ctx(q, hctx, i) {
4584 if (!hctx->tags)
4585 continue;
4586 /*
4587 * If we're using an MQ scheduler, just update the scheduler
4588 * queue depth. This is similar to what the old code would do.
4589 */
4590 if (hctx->sched_tags) {
4591 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4592 nr, true);
4593 } else {
4594 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4595 false);
4596 }
4597 if (ret)
4598 break;
4599 if (q->elevator && q->elevator->type->ops.depth_updated)
4600 q->elevator->type->ops.depth_updated(hctx);
4601 }
4602 if (!ret) {
4603 q->nr_requests = nr;
4604 if (blk_mq_is_shared_tags(set->flags)) {
4605 if (q->elevator)
4606 blk_mq_tag_update_sched_shared_tags(q);
4607 else
4608 blk_mq_tag_resize_shared_tags(set, nr);
4609 }
4610 }
4611
4612 blk_mq_unquiesce_queue(q);
4613 blk_mq_unfreeze_queue(q);
4614
4615 return ret;
4616}
4617
4618/*
4619 * request_queue and elevator_type pair.
4620 * It is just used by __blk_mq_update_nr_hw_queues to cache
4621 * the elevator_type associated with a request_queue.
4622 */
4623struct blk_mq_qe_pair {
4624 struct list_head node;
4625 struct request_queue *q;
4626 struct elevator_type *type;
4627};
4628
4629/*
4630 * Cache the elevator_type in qe pair list and switch the
4631 * io scheduler to 'none'
4632 */
4633static bool blk_mq_elv_switch_none(struct list_head *head,
4634 struct request_queue *q)
4635{
4636 struct blk_mq_qe_pair *qe;
4637
4638 if (!q->elevator)
4639 return true;
4640
4641 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4642 if (!qe)
4643 return false;
4644
4645 /* q->elevator needs protection from ->sysfs_lock */
4646 mutex_lock(&q->sysfs_lock);
4647
4648 INIT_LIST_HEAD(&qe->node);
4649 qe->q = q;
4650 qe->type = q->elevator->type;
4651 /* keep a reference to the elevator module as we'll switch back */
4652 __elevator_get(qe->type);
4653 list_add(&qe->node, head);
4654 elevator_disable(q);
4655 mutex_unlock(&q->sysfs_lock);
4656
4657 return true;
4658}
4659
4660static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4661 struct request_queue *q)
4662{
4663 struct blk_mq_qe_pair *qe;
4664
4665 list_for_each_entry(qe, head, node)
4666 if (qe->q == q)
4667 return qe;
4668
4669 return NULL;
4670}
4671
4672static void blk_mq_elv_switch_back(struct list_head *head,
4673 struct request_queue *q)
4674{
4675 struct blk_mq_qe_pair *qe;
4676 struct elevator_type *t;
4677
4678 qe = blk_lookup_qe_pair(head, q);
4679 if (!qe)
4680 return;
4681 t = qe->type;
4682 list_del(&qe->node);
4683 kfree(qe);
4684
4685 mutex_lock(&q->sysfs_lock);
4686 elevator_switch(q, t);
4687 /* drop the reference acquired in blk_mq_elv_switch_none */
4688 elevator_put(t);
4689 mutex_unlock(&q->sysfs_lock);
4690}
4691
4692static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4693 int nr_hw_queues)
4694{
4695 struct request_queue *q;
4696 LIST_HEAD(head);
4697 int prev_nr_hw_queues;
4698
4699 lockdep_assert_held(&set->tag_list_lock);
4700
4701 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4702 nr_hw_queues = nr_cpu_ids;
4703 if (nr_hw_queues < 1)
4704 return;
4705 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4706 return;
4707
4708 list_for_each_entry(q, &set->tag_list, tag_set_list)
4709 blk_mq_freeze_queue(q);
4710 /*
4711 * Switch IO scheduler to 'none', cleaning up the data associated
4712 * with the previous scheduler. We will switch back once we are done
4713 * updating the new sw to hw queue mappings.
4714 */
4715 list_for_each_entry(q, &set->tag_list, tag_set_list)
4716 if (!blk_mq_elv_switch_none(&head, q))
4717 goto switch_back;
4718
4719 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4720 blk_mq_debugfs_unregister_hctxs(q);
4721 blk_mq_sysfs_unregister_hctxs(q);
4722 }
4723
4724 prev_nr_hw_queues = set->nr_hw_queues;
4725 if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0)
4726 goto reregister;
4727
4728fallback:
4729 blk_mq_update_queue_map(set);
4730 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4731 blk_mq_realloc_hw_ctxs(set, q);
4732 blk_mq_update_poll_flag(q);
4733 if (q->nr_hw_queues != set->nr_hw_queues) {
4734 int i = prev_nr_hw_queues;
4735
4736 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4737 nr_hw_queues, prev_nr_hw_queues);
4738 for (; i < set->nr_hw_queues; i++)
4739 __blk_mq_free_map_and_rqs(set, i);
4740
4741 set->nr_hw_queues = prev_nr_hw_queues;
4742 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4743 goto fallback;
4744 }
4745 blk_mq_map_swqueue(q);
4746 }
4747
4748reregister:
4749 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4750 blk_mq_sysfs_register_hctxs(q);
4751 blk_mq_debugfs_register_hctxs(q);
4752 }
4753
4754switch_back:
4755 list_for_each_entry(q, &set->tag_list, tag_set_list)
4756 blk_mq_elv_switch_back(&head, q);
4757
4758 list_for_each_entry(q, &set->tag_list, tag_set_list)
4759 blk_mq_unfreeze_queue(q);
4760}
4761
4762void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4763{
4764 mutex_lock(&set->tag_list_lock);
4765 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4766 mutex_unlock(&set->tag_list_lock);
4767}
4768EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4769
4770/* Enable polling stats and return whether they were already enabled. */
4771static bool blk_poll_stats_enable(struct request_queue *q)
4772{
4773 if (q->poll_stat)
4774 return true;
4775
4776 return blk_stats_alloc_enable(q);
4777}
4778
4779static void blk_mq_poll_stats_start(struct request_queue *q)
4780{
4781 /*
4782 * We don't arm the callback if polling stats are not enabled or the
4783 * callback is already active.
4784 */
4785 if (!q->poll_stat || blk_stat_is_active(q->poll_cb))
4786 return;
4787
4788 blk_stat_activate_msecs(q->poll_cb, 100);
4789}
4790
4791static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
4792{
4793 struct request_queue *q = cb->data;
4794 int bucket;
4795
4796 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
4797 if (cb->stat[bucket].nr_samples)
4798 q->poll_stat[bucket] = cb->stat[bucket];
4799 }
4800}
4801
4802static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
4803 struct request *rq)
4804{
4805 unsigned long ret = 0;
4806 int bucket;
4807
4808 /*
4809 * If stats collection isn't on, don't sleep but turn it on for
4810 * future users
4811 */
4812 if (!blk_poll_stats_enable(q))
4813 return 0;
4814
4815 /*
4816 * As an optimistic guess, use half of the mean service time
4817 * for this type of request. We can (and should) make this smarter.
4818 * For instance, if the completion latencies are tight, we can
4819 * get closer than just half the mean. This is especially
4820 * important on devices where the completion latencies are longer
4821 * than ~10 usec. We do use the stats for the relevant IO size
4822 * if available which does lead to better estimates.
4823 */
4824 bucket = blk_mq_poll_stats_bkt(rq);
4825 if (bucket < 0)
4826 return ret;
4827
4828 if (q->poll_stat[bucket].nr_samples)
4829 ret = (q->poll_stat[bucket].mean + 1) / 2;
4830
4831 return ret;
4832}
4833
4834static bool blk_mq_poll_hybrid(struct request_queue *q, blk_qc_t qc)
4835{
4836 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, qc);
4837 struct request *rq = blk_qc_to_rq(hctx, qc);
4838 struct hrtimer_sleeper hs;
4839 enum hrtimer_mode mode;
4840 unsigned int nsecs;
4841 ktime_t kt;
4842
4843 /*
4844 * If a request has completed on queue that uses an I/O scheduler, we
4845 * won't get back a request from blk_qc_to_rq.
4846 */
4847 if (!rq || (rq->rq_flags & RQF_MQ_POLL_SLEPT))
4848 return false;
4849
4850 /*
4851 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
4852 *
4853 * 0: use half of prev avg
4854 * >0: use this specific value
4855 */
4856 if (q->poll_nsec > 0)
4857 nsecs = q->poll_nsec;
4858 else
4859 nsecs = blk_mq_poll_nsecs(q, rq);
4860
4861 if (!nsecs)
4862 return false;
4863
4864 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
4865
4866 /*
4867 * This will be replaced with the stats tracking code, using
4868 * 'avg_completion_time / 2' as the pre-sleep target.
4869 */
4870 kt = nsecs;
4871
4872 mode = HRTIMER_MODE_REL;
4873 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
4874 hrtimer_set_expires(&hs.timer, kt);
4875
4876 do {
4877 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
4878 break;
4879 set_current_state(TASK_UNINTERRUPTIBLE);
4880 hrtimer_sleeper_start_expires(&hs, mode);
4881 if (hs.task)
4882 io_schedule();
4883 hrtimer_cancel(&hs.timer);
4884 mode = HRTIMER_MODE_ABS;
4885 } while (hs.task && !signal_pending(current));
4886
4887 __set_current_state(TASK_RUNNING);
4888 destroy_hrtimer_on_stack(&hs.timer);
4889
4890 /*
4891 * If we sleep, have the caller restart the poll loop to reset the
4892 * state. Like for the other success return cases, the caller is
4893 * responsible for checking if the IO completed. If the IO isn't
4894 * complete, we'll get called again and will go straight to the busy
4895 * poll loop.
4896 */
4897 return true;
4898}
4899
4900static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie,
4901 struct io_comp_batch *iob, unsigned int flags)
4902{
4903 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie);
4904 long state = get_current_state();
4905 int ret;
4906
4907 do {
4908 ret = q->mq_ops->poll(hctx, iob);
4909 if (ret > 0) {
4910 __set_current_state(TASK_RUNNING);
4911 return ret;
4912 }
4913
4914 if (signal_pending_state(state, current))
4915 __set_current_state(TASK_RUNNING);
4916 if (task_is_running(current))
4917 return 1;
4918
4919 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4920 break;
4921 cpu_relax();
4922 } while (!need_resched());
4923
4924 __set_current_state(TASK_RUNNING);
4925 return 0;
4926}
4927
4928int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
4929 unsigned int flags)
4930{
4931 if (!(flags & BLK_POLL_NOSLEEP) &&
4932 q->poll_nsec != BLK_MQ_POLL_CLASSIC) {
4933 if (blk_mq_poll_hybrid(q, cookie))
4934 return 1;
4935 }
4936 return blk_mq_poll_classic(q, cookie, iob, flags);
4937}
4938
4939unsigned int blk_mq_rq_cpu(struct request *rq)
4940{
4941 return rq->mq_ctx->cpu;
4942}
4943EXPORT_SYMBOL(blk_mq_rq_cpu);
4944
4945void blk_mq_cancel_work_sync(struct request_queue *q)
4946{
4947 struct blk_mq_hw_ctx *hctx;
4948 unsigned long i;
4949
4950 cancel_delayed_work_sync(&q->requeue_work);
4951
4952 queue_for_each_hw_ctx(q, hctx, i)
4953 cancel_delayed_work_sync(&hctx->run_work);
4954}
4955
4956static int __init blk_mq_init(void)
4957{
4958 int i;
4959
4960 for_each_possible_cpu(i)
4961 init_llist_head(&per_cpu(blk_cpu_done, i));
4962 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4963
4964 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4965 "block/softirq:dead", NULL,
4966 blk_softirq_cpu_dead);
4967 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4968 blk_mq_hctx_notify_dead);
4969 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4970 blk_mq_hctx_notify_online,
4971 blk_mq_hctx_notify_offline);
4972 return 0;
4973}
4974subsys_initcall(blk_mq_init);
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
44static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
45static DEFINE_PER_CPU(call_single_data_t, blk_cpu_csd);
46
47static void blk_mq_insert_request(struct request *rq, blk_insert_t flags);
48static void blk_mq_request_bypass_insert(struct request *rq,
49 blk_insert_t flags);
50static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
51 struct list_head *list);
52static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
53 struct io_comp_batch *iob, unsigned int flags);
54
55/*
56 * Check if any of the ctx, dispatch list or elevator
57 * have pending work in this hardware queue.
58 */
59static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
60{
61 return !list_empty_careful(&hctx->dispatch) ||
62 sbitmap_any_bit_set(&hctx->ctx_map) ||
63 blk_mq_sched_has_work(hctx);
64}
65
66/*
67 * Mark this ctx as having pending work in this hardware queue
68 */
69static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
70 struct blk_mq_ctx *ctx)
71{
72 const int bit = ctx->index_hw[hctx->type];
73
74 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
75 sbitmap_set_bit(&hctx->ctx_map, bit);
76}
77
78static void blk_mq_hctx_clear_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 sbitmap_clear_bit(&hctx->ctx_map, bit);
84}
85
86struct mq_inflight {
87 struct block_device *part;
88 unsigned int inflight[2];
89};
90
91static bool blk_mq_check_inflight(struct request *rq, void *priv)
92{
93 struct mq_inflight *mi = priv;
94
95 if (rq->part && blk_do_io_stat(rq) &&
96 (!mi->part->bd_partno || rq->part == mi->part) &&
97 blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
98 mi->inflight[rq_data_dir(rq)]++;
99
100 return true;
101}
102
103unsigned int blk_mq_in_flight(struct request_queue *q,
104 struct block_device *part)
105{
106 struct mq_inflight mi = { .part = part };
107
108 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
109
110 return mi.inflight[0] + mi.inflight[1];
111}
112
113void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
114 unsigned int inflight[2])
115{
116 struct mq_inflight mi = { .part = part };
117
118 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
119 inflight[0] = mi.inflight[0];
120 inflight[1] = mi.inflight[1];
121}
122
123void blk_freeze_queue_start(struct request_queue *q)
124{
125 mutex_lock(&q->mq_freeze_lock);
126 if (++q->mq_freeze_depth == 1) {
127 percpu_ref_kill(&q->q_usage_counter);
128 mutex_unlock(&q->mq_freeze_lock);
129 if (queue_is_mq(q))
130 blk_mq_run_hw_queues(q, false);
131 } else {
132 mutex_unlock(&q->mq_freeze_lock);
133 }
134}
135EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
136
137void blk_mq_freeze_queue_wait(struct request_queue *q)
138{
139 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
140}
141EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
142
143int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
144 unsigned long timeout)
145{
146 return wait_event_timeout(q->mq_freeze_wq,
147 percpu_ref_is_zero(&q->q_usage_counter),
148 timeout);
149}
150EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
151
152/*
153 * Guarantee no request is in use, so we can change any data structure of
154 * the queue afterward.
155 */
156void blk_freeze_queue(struct request_queue *q)
157{
158 /*
159 * In the !blk_mq case we are only calling this to kill the
160 * q_usage_counter, otherwise this increases the freeze depth
161 * and waits for it to return to zero. For this reason there is
162 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
163 * exported to drivers as the only user for unfreeze is blk_mq.
164 */
165 blk_freeze_queue_start(q);
166 blk_mq_freeze_queue_wait(q);
167}
168
169void blk_mq_freeze_queue(struct request_queue *q)
170{
171 /*
172 * ...just an alias to keep freeze and unfreeze actions balanced
173 * in the blk_mq_* namespace
174 */
175 blk_freeze_queue(q);
176}
177EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
178
179void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
180{
181 mutex_lock(&q->mq_freeze_lock);
182 if (force_atomic)
183 q->q_usage_counter.data->force_atomic = true;
184 q->mq_freeze_depth--;
185 WARN_ON_ONCE(q->mq_freeze_depth < 0);
186 if (!q->mq_freeze_depth) {
187 percpu_ref_resurrect(&q->q_usage_counter);
188 wake_up_all(&q->mq_freeze_wq);
189 }
190 mutex_unlock(&q->mq_freeze_lock);
191}
192
193void blk_mq_unfreeze_queue(struct request_queue *q)
194{
195 __blk_mq_unfreeze_queue(q, false);
196}
197EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
198
199/*
200 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
201 * mpt3sas driver such that this function can be removed.
202 */
203void blk_mq_quiesce_queue_nowait(struct request_queue *q)
204{
205 unsigned long flags;
206
207 spin_lock_irqsave(&q->queue_lock, flags);
208 if (!q->quiesce_depth++)
209 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
210 spin_unlock_irqrestore(&q->queue_lock, flags);
211}
212EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
213
214/**
215 * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
216 * @set: tag_set to wait on
217 *
218 * Note: it is driver's responsibility for making sure that quiesce has
219 * been started on or more of the request_queues of the tag_set. This
220 * function only waits for the quiesce on those request_queues that had
221 * the quiesce flag set using blk_mq_quiesce_queue_nowait.
222 */
223void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set)
224{
225 if (set->flags & BLK_MQ_F_BLOCKING)
226 synchronize_srcu(set->srcu);
227 else
228 synchronize_rcu();
229}
230EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
231
232/**
233 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
234 * @q: request queue.
235 *
236 * Note: this function does not prevent that the struct request end_io()
237 * callback function is invoked. Once this function is returned, we make
238 * sure no dispatch can happen until the queue is unquiesced via
239 * blk_mq_unquiesce_queue().
240 */
241void blk_mq_quiesce_queue(struct request_queue *q)
242{
243 blk_mq_quiesce_queue_nowait(q);
244 /* nothing to wait for non-mq queues */
245 if (queue_is_mq(q))
246 blk_mq_wait_quiesce_done(q->tag_set);
247}
248EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
249
250/*
251 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
252 * @q: request queue.
253 *
254 * This function recovers queue into the state before quiescing
255 * which is done by blk_mq_quiesce_queue.
256 */
257void blk_mq_unquiesce_queue(struct request_queue *q)
258{
259 unsigned long flags;
260 bool run_queue = false;
261
262 spin_lock_irqsave(&q->queue_lock, flags);
263 if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
264 ;
265 } else if (!--q->quiesce_depth) {
266 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
267 run_queue = true;
268 }
269 spin_unlock_irqrestore(&q->queue_lock, flags);
270
271 /* dispatch requests which are inserted during quiescing */
272 if (run_queue)
273 blk_mq_run_hw_queues(q, true);
274}
275EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
276
277void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set)
278{
279 struct request_queue *q;
280
281 mutex_lock(&set->tag_list_lock);
282 list_for_each_entry(q, &set->tag_list, tag_set_list) {
283 if (!blk_queue_skip_tagset_quiesce(q))
284 blk_mq_quiesce_queue_nowait(q);
285 }
286 blk_mq_wait_quiesce_done(set);
287 mutex_unlock(&set->tag_list_lock);
288}
289EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset);
290
291void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set)
292{
293 struct request_queue *q;
294
295 mutex_lock(&set->tag_list_lock);
296 list_for_each_entry(q, &set->tag_list, tag_set_list) {
297 if (!blk_queue_skip_tagset_quiesce(q))
298 blk_mq_unquiesce_queue(q);
299 }
300 mutex_unlock(&set->tag_list_lock);
301}
302EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset);
303
304void blk_mq_wake_waiters(struct request_queue *q)
305{
306 struct blk_mq_hw_ctx *hctx;
307 unsigned long i;
308
309 queue_for_each_hw_ctx(q, hctx, i)
310 if (blk_mq_hw_queue_mapped(hctx))
311 blk_mq_tag_wakeup_all(hctx->tags, true);
312}
313
314void blk_rq_init(struct request_queue *q, struct request *rq)
315{
316 memset(rq, 0, sizeof(*rq));
317
318 INIT_LIST_HEAD(&rq->queuelist);
319 rq->q = q;
320 rq->__sector = (sector_t) -1;
321 INIT_HLIST_NODE(&rq->hash);
322 RB_CLEAR_NODE(&rq->rb_node);
323 rq->tag = BLK_MQ_NO_TAG;
324 rq->internal_tag = BLK_MQ_NO_TAG;
325 rq->start_time_ns = ktime_get_ns();
326 rq->part = NULL;
327 blk_crypto_rq_set_defaults(rq);
328}
329EXPORT_SYMBOL(blk_rq_init);
330
331/* Set start and alloc time when the allocated request is actually used */
332static inline void blk_mq_rq_time_init(struct request *rq, u64 alloc_time_ns)
333{
334 if (blk_mq_need_time_stamp(rq))
335 rq->start_time_ns = ktime_get_ns();
336 else
337 rq->start_time_ns = 0;
338
339#ifdef CONFIG_BLK_RQ_ALLOC_TIME
340 if (blk_queue_rq_alloc_time(rq->q))
341 rq->alloc_time_ns = alloc_time_ns ?: rq->start_time_ns;
342 else
343 rq->alloc_time_ns = 0;
344#endif
345}
346
347static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
348 struct blk_mq_tags *tags, unsigned int tag)
349{
350 struct blk_mq_ctx *ctx = data->ctx;
351 struct blk_mq_hw_ctx *hctx = data->hctx;
352 struct request_queue *q = data->q;
353 struct request *rq = tags->static_rqs[tag];
354
355 rq->q = q;
356 rq->mq_ctx = ctx;
357 rq->mq_hctx = hctx;
358 rq->cmd_flags = data->cmd_flags;
359
360 if (data->flags & BLK_MQ_REQ_PM)
361 data->rq_flags |= RQF_PM;
362 if (blk_queue_io_stat(q))
363 data->rq_flags |= RQF_IO_STAT;
364 rq->rq_flags = data->rq_flags;
365
366 if (data->rq_flags & RQF_SCHED_TAGS) {
367 rq->tag = BLK_MQ_NO_TAG;
368 rq->internal_tag = tag;
369 } else {
370 rq->tag = tag;
371 rq->internal_tag = BLK_MQ_NO_TAG;
372 }
373 rq->timeout = 0;
374
375 rq->part = NULL;
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
404static inline struct request *
405__blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data)
406{
407 unsigned int tag, tag_offset;
408 struct blk_mq_tags *tags;
409 struct request *rq;
410 unsigned long tag_mask;
411 int i, nr = 0;
412
413 tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset);
414 if (unlikely(!tag_mask))
415 return NULL;
416
417 tags = blk_mq_tags_from_data(data);
418 for (i = 0; tag_mask; i++) {
419 if (!(tag_mask & (1UL << i)))
420 continue;
421 tag = tag_offset + i;
422 prefetch(tags->static_rqs[tag]);
423 tag_mask &= ~(1UL << i);
424 rq = blk_mq_rq_ctx_init(data, tags, tag);
425 rq_list_add(data->cached_rq, rq);
426 nr++;
427 }
428 if (!(data->rq_flags & RQF_SCHED_TAGS))
429 blk_mq_add_active_requests(data->hctx, nr);
430 /* caller already holds a reference, add for remainder */
431 percpu_ref_get_many(&data->q->q_usage_counter, nr - 1);
432 data->nr_tags -= nr;
433
434 return rq_list_pop(data->cached_rq);
435}
436
437static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
438{
439 struct request_queue *q = data->q;
440 u64 alloc_time_ns = 0;
441 struct request *rq;
442 unsigned int tag;
443
444 /* alloc_time includes depth and tag waits */
445 if (blk_queue_rq_alloc_time(q))
446 alloc_time_ns = ktime_get_ns();
447
448 if (data->cmd_flags & REQ_NOWAIT)
449 data->flags |= BLK_MQ_REQ_NOWAIT;
450
451 if (q->elevator) {
452 /*
453 * All requests use scheduler tags when an I/O scheduler is
454 * enabled for the queue.
455 */
456 data->rq_flags |= RQF_SCHED_TAGS;
457
458 /*
459 * Flush/passthrough requests are special and go directly to the
460 * dispatch list.
461 */
462 if ((data->cmd_flags & REQ_OP_MASK) != REQ_OP_FLUSH &&
463 !blk_op_is_passthrough(data->cmd_flags)) {
464 struct elevator_mq_ops *ops = &q->elevator->type->ops;
465
466 WARN_ON_ONCE(data->flags & BLK_MQ_REQ_RESERVED);
467
468 data->rq_flags |= RQF_USE_SCHED;
469 if (ops->limit_depth)
470 ops->limit_depth(data->cmd_flags, data);
471 }
472 }
473
474retry:
475 data->ctx = blk_mq_get_ctx(q);
476 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
477 if (!(data->rq_flags & RQF_SCHED_TAGS))
478 blk_mq_tag_busy(data->hctx);
479
480 if (data->flags & BLK_MQ_REQ_RESERVED)
481 data->rq_flags |= RQF_RESV;
482
483 /*
484 * Try batched alloc if we want more than 1 tag.
485 */
486 if (data->nr_tags > 1) {
487 rq = __blk_mq_alloc_requests_batch(data);
488 if (rq) {
489 blk_mq_rq_time_init(rq, alloc_time_ns);
490 return rq;
491 }
492 data->nr_tags = 1;
493 }
494
495 /*
496 * Waiting allocations only fail because of an inactive hctx. In that
497 * case just retry the hctx assignment and tag allocation as CPU hotplug
498 * should have migrated us to an online CPU by now.
499 */
500 tag = blk_mq_get_tag(data);
501 if (tag == BLK_MQ_NO_TAG) {
502 if (data->flags & BLK_MQ_REQ_NOWAIT)
503 return NULL;
504 /*
505 * Give up the CPU and sleep for a random short time to
506 * ensure that thread using a realtime scheduling class
507 * are migrated off the CPU, and thus off the hctx that
508 * is going away.
509 */
510 msleep(3);
511 goto retry;
512 }
513
514 if (!(data->rq_flags & RQF_SCHED_TAGS))
515 blk_mq_inc_active_requests(data->hctx);
516 rq = blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag);
517 blk_mq_rq_time_init(rq, alloc_time_ns);
518 return rq;
519}
520
521static struct request *blk_mq_rq_cache_fill(struct request_queue *q,
522 struct blk_plug *plug,
523 blk_opf_t opf,
524 blk_mq_req_flags_t flags)
525{
526 struct blk_mq_alloc_data data = {
527 .q = q,
528 .flags = flags,
529 .cmd_flags = opf,
530 .nr_tags = plug->nr_ios,
531 .cached_rq = &plug->cached_rq,
532 };
533 struct request *rq;
534
535 if (blk_queue_enter(q, flags))
536 return NULL;
537
538 plug->nr_ios = 1;
539
540 rq = __blk_mq_alloc_requests(&data);
541 if (unlikely(!rq))
542 blk_queue_exit(q);
543 return rq;
544}
545
546static struct request *blk_mq_alloc_cached_request(struct request_queue *q,
547 blk_opf_t opf,
548 blk_mq_req_flags_t flags)
549{
550 struct blk_plug *plug = current->plug;
551 struct request *rq;
552
553 if (!plug)
554 return NULL;
555
556 if (rq_list_empty(plug->cached_rq)) {
557 if (plug->nr_ios == 1)
558 return NULL;
559 rq = blk_mq_rq_cache_fill(q, plug, opf, flags);
560 if (!rq)
561 return NULL;
562 } else {
563 rq = rq_list_peek(&plug->cached_rq);
564 if (!rq || rq->q != q)
565 return NULL;
566
567 if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type)
568 return NULL;
569 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
570 return NULL;
571
572 plug->cached_rq = rq_list_next(rq);
573 blk_mq_rq_time_init(rq, 0);
574 }
575
576 rq->cmd_flags = opf;
577 INIT_LIST_HEAD(&rq->queuelist);
578 return rq;
579}
580
581struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf,
582 blk_mq_req_flags_t flags)
583{
584 struct request *rq;
585
586 rq = blk_mq_alloc_cached_request(q, opf, flags);
587 if (!rq) {
588 struct blk_mq_alloc_data data = {
589 .q = q,
590 .flags = flags,
591 .cmd_flags = opf,
592 .nr_tags = 1,
593 };
594 int ret;
595
596 ret = blk_queue_enter(q, flags);
597 if (ret)
598 return ERR_PTR(ret);
599
600 rq = __blk_mq_alloc_requests(&data);
601 if (!rq)
602 goto out_queue_exit;
603 }
604 rq->__data_len = 0;
605 rq->__sector = (sector_t) -1;
606 rq->bio = rq->biotail = NULL;
607 return rq;
608out_queue_exit:
609 blk_queue_exit(q);
610 return ERR_PTR(-EWOULDBLOCK);
611}
612EXPORT_SYMBOL(blk_mq_alloc_request);
613
614struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
615 blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx)
616{
617 struct blk_mq_alloc_data data = {
618 .q = q,
619 .flags = flags,
620 .cmd_flags = opf,
621 .nr_tags = 1,
622 };
623 u64 alloc_time_ns = 0;
624 struct request *rq;
625 unsigned int cpu;
626 unsigned int tag;
627 int ret;
628
629 /* alloc_time includes depth and tag waits */
630 if (blk_queue_rq_alloc_time(q))
631 alloc_time_ns = ktime_get_ns();
632
633 /*
634 * If the tag allocator sleeps we could get an allocation for a
635 * different hardware context. No need to complicate the low level
636 * allocator for this for the rare use case of a command tied to
637 * a specific queue.
638 */
639 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) ||
640 WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED)))
641 return ERR_PTR(-EINVAL);
642
643 if (hctx_idx >= q->nr_hw_queues)
644 return ERR_PTR(-EIO);
645
646 ret = blk_queue_enter(q, flags);
647 if (ret)
648 return ERR_PTR(ret);
649
650 /*
651 * Check if the hardware context is actually mapped to anything.
652 * If not tell the caller that it should skip this queue.
653 */
654 ret = -EXDEV;
655 data.hctx = xa_load(&q->hctx_table, hctx_idx);
656 if (!blk_mq_hw_queue_mapped(data.hctx))
657 goto out_queue_exit;
658 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
659 if (cpu >= nr_cpu_ids)
660 goto out_queue_exit;
661 data.ctx = __blk_mq_get_ctx(q, cpu);
662
663 if (q->elevator)
664 data.rq_flags |= RQF_SCHED_TAGS;
665 else
666 blk_mq_tag_busy(data.hctx);
667
668 if (flags & BLK_MQ_REQ_RESERVED)
669 data.rq_flags |= RQF_RESV;
670
671 ret = -EWOULDBLOCK;
672 tag = blk_mq_get_tag(&data);
673 if (tag == BLK_MQ_NO_TAG)
674 goto out_queue_exit;
675 if (!(data.rq_flags & RQF_SCHED_TAGS))
676 blk_mq_inc_active_requests(data.hctx);
677 rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag);
678 blk_mq_rq_time_init(rq, alloc_time_ns);
679 rq->__data_len = 0;
680 rq->__sector = (sector_t) -1;
681 rq->bio = rq->biotail = NULL;
682 return rq;
683
684out_queue_exit:
685 blk_queue_exit(q);
686 return ERR_PTR(ret);
687}
688EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
689
690static void blk_mq_finish_request(struct request *rq)
691{
692 struct request_queue *q = rq->q;
693
694 if (rq->rq_flags & RQF_USE_SCHED) {
695 q->elevator->type->ops.finish_request(rq);
696 /*
697 * For postflush request that may need to be
698 * completed twice, we should clear this flag
699 * to avoid double finish_request() on the rq.
700 */
701 rq->rq_flags &= ~RQF_USE_SCHED;
702 }
703}
704
705static void __blk_mq_free_request(struct request *rq)
706{
707 struct request_queue *q = rq->q;
708 struct blk_mq_ctx *ctx = rq->mq_ctx;
709 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
710 const int sched_tag = rq->internal_tag;
711
712 blk_crypto_free_request(rq);
713 blk_pm_mark_last_busy(rq);
714 rq->mq_hctx = NULL;
715
716 if (rq->tag != BLK_MQ_NO_TAG) {
717 blk_mq_dec_active_requests(hctx);
718 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
719 }
720 if (sched_tag != BLK_MQ_NO_TAG)
721 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
722 blk_mq_sched_restart(hctx);
723 blk_queue_exit(q);
724}
725
726void blk_mq_free_request(struct request *rq)
727{
728 struct request_queue *q = rq->q;
729
730 blk_mq_finish_request(rq);
731
732 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
733 laptop_io_completion(q->disk->bdi);
734
735 rq_qos_done(q, rq);
736
737 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
738 if (req_ref_put_and_test(rq))
739 __blk_mq_free_request(rq);
740}
741EXPORT_SYMBOL_GPL(blk_mq_free_request);
742
743void blk_mq_free_plug_rqs(struct blk_plug *plug)
744{
745 struct request *rq;
746
747 while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
748 blk_mq_free_request(rq);
749}
750
751void blk_dump_rq_flags(struct request *rq, char *msg)
752{
753 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
754 rq->q->disk ? rq->q->disk->disk_name : "?",
755 (__force unsigned long long) rq->cmd_flags);
756
757 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
758 (unsigned long long)blk_rq_pos(rq),
759 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
760 printk(KERN_INFO " bio %p, biotail %p, len %u\n",
761 rq->bio, rq->biotail, blk_rq_bytes(rq));
762}
763EXPORT_SYMBOL(blk_dump_rq_flags);
764
765static void req_bio_endio(struct request *rq, struct bio *bio,
766 unsigned int nbytes, blk_status_t error)
767{
768 if (unlikely(error)) {
769 bio->bi_status = error;
770 } else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
771 /*
772 * Partial zone append completions cannot be supported as the
773 * BIO fragments may end up not being written sequentially.
774 * For such case, force the completed nbytes to be equal to
775 * the BIO size so that bio_advance() sets the BIO remaining
776 * size to 0 and we end up calling bio_endio() before returning.
777 */
778 if (bio->bi_iter.bi_size != nbytes) {
779 bio->bi_status = BLK_STS_IOERR;
780 nbytes = bio->bi_iter.bi_size;
781 } else {
782 bio->bi_iter.bi_sector = rq->__sector;
783 }
784 }
785
786 bio_advance(bio, nbytes);
787
788 if (unlikely(rq->rq_flags & RQF_QUIET))
789 bio_set_flag(bio, BIO_QUIET);
790 /* don't actually finish bio if it's part of flush sequence */
791 if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
792 bio_endio(bio);
793}
794
795static void blk_account_io_completion(struct request *req, unsigned int bytes)
796{
797 if (req->part && blk_do_io_stat(req)) {
798 const int sgrp = op_stat_group(req_op(req));
799
800 part_stat_lock();
801 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
802 part_stat_unlock();
803 }
804}
805
806static void blk_print_req_error(struct request *req, blk_status_t status)
807{
808 printk_ratelimited(KERN_ERR
809 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
810 "phys_seg %u prio class %u\n",
811 blk_status_to_str(status),
812 req->q->disk ? req->q->disk->disk_name : "?",
813 blk_rq_pos(req), (__force u32)req_op(req),
814 blk_op_str(req_op(req)),
815 (__force u32)(req->cmd_flags & ~REQ_OP_MASK),
816 req->nr_phys_segments,
817 IOPRIO_PRIO_CLASS(req->ioprio));
818}
819
820/*
821 * Fully end IO on a request. Does not support partial completions, or
822 * errors.
823 */
824static void blk_complete_request(struct request *req)
825{
826 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
827 int total_bytes = blk_rq_bytes(req);
828 struct bio *bio = req->bio;
829
830 trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
831
832 if (!bio)
833 return;
834
835#ifdef CONFIG_BLK_DEV_INTEGRITY
836 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
837 req->q->integrity.profile->complete_fn(req, total_bytes);
838#endif
839
840 /*
841 * Upper layers may call blk_crypto_evict_key() anytime after the last
842 * bio_endio(). Therefore, the keyslot must be released before that.
843 */
844 blk_crypto_rq_put_keyslot(req);
845
846 blk_account_io_completion(req, total_bytes);
847
848 do {
849 struct bio *next = bio->bi_next;
850
851 /* Completion has already been traced */
852 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
853
854 if (req_op(req) == REQ_OP_ZONE_APPEND)
855 bio->bi_iter.bi_sector = req->__sector;
856
857 if (!is_flush)
858 bio_endio(bio);
859 bio = next;
860 } while (bio);
861
862 /*
863 * Reset counters so that the request stacking driver
864 * can find how many bytes remain in the request
865 * later.
866 */
867 if (!req->end_io) {
868 req->bio = NULL;
869 req->__data_len = 0;
870 }
871}
872
873/**
874 * blk_update_request - Complete multiple bytes without completing the request
875 * @req: the request being processed
876 * @error: block status code
877 * @nr_bytes: number of bytes to complete for @req
878 *
879 * Description:
880 * Ends I/O on a number of bytes attached to @req, but doesn't complete
881 * the request structure even if @req doesn't have leftover.
882 * If @req has leftover, sets it up for the next range of segments.
883 *
884 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
885 * %false return from this function.
886 *
887 * Note:
888 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
889 * except in the consistency check at the end of this function.
890 *
891 * Return:
892 * %false - this request doesn't have any more data
893 * %true - this request has more data
894 **/
895bool blk_update_request(struct request *req, blk_status_t error,
896 unsigned int nr_bytes)
897{
898 int total_bytes;
899
900 trace_block_rq_complete(req, error, nr_bytes);
901
902 if (!req->bio)
903 return false;
904
905#ifdef CONFIG_BLK_DEV_INTEGRITY
906 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
907 error == BLK_STS_OK)
908 req->q->integrity.profile->complete_fn(req, nr_bytes);
909#endif
910
911 /*
912 * Upper layers may call blk_crypto_evict_key() anytime after the last
913 * bio_endio(). Therefore, the keyslot must be released before that.
914 */
915 if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req))
916 __blk_crypto_rq_put_keyslot(req);
917
918 if (unlikely(error && !blk_rq_is_passthrough(req) &&
919 !(req->rq_flags & RQF_QUIET)) &&
920 !test_bit(GD_DEAD, &req->q->disk->state)) {
921 blk_print_req_error(req, error);
922 trace_block_rq_error(req, error, nr_bytes);
923 }
924
925 blk_account_io_completion(req, nr_bytes);
926
927 total_bytes = 0;
928 while (req->bio) {
929 struct bio *bio = req->bio;
930 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
931
932 if (bio_bytes == bio->bi_iter.bi_size)
933 req->bio = bio->bi_next;
934
935 /* Completion has already been traced */
936 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
937 req_bio_endio(req, bio, bio_bytes, error);
938
939 total_bytes += bio_bytes;
940 nr_bytes -= bio_bytes;
941
942 if (!nr_bytes)
943 break;
944 }
945
946 /*
947 * completely done
948 */
949 if (!req->bio) {
950 /*
951 * Reset counters so that the request stacking driver
952 * can find how many bytes remain in the request
953 * later.
954 */
955 req->__data_len = 0;
956 return false;
957 }
958
959 req->__data_len -= total_bytes;
960
961 /* update sector only for requests with clear definition of sector */
962 if (!blk_rq_is_passthrough(req))
963 req->__sector += total_bytes >> 9;
964
965 /* mixed attributes always follow the first bio */
966 if (req->rq_flags & RQF_MIXED_MERGE) {
967 req->cmd_flags &= ~REQ_FAILFAST_MASK;
968 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
969 }
970
971 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
972 /*
973 * If total number of sectors is less than the first segment
974 * size, something has gone terribly wrong.
975 */
976 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
977 blk_dump_rq_flags(req, "request botched");
978 req->__data_len = blk_rq_cur_bytes(req);
979 }
980
981 /* recalculate the number of segments */
982 req->nr_phys_segments = blk_recalc_rq_segments(req);
983 }
984
985 return true;
986}
987EXPORT_SYMBOL_GPL(blk_update_request);
988
989static inline void blk_account_io_done(struct request *req, u64 now)
990{
991 trace_block_io_done(req);
992
993 /*
994 * Account IO completion. flush_rq isn't accounted as a
995 * normal IO on queueing nor completion. Accounting the
996 * containing request is enough.
997 */
998 if (blk_do_io_stat(req) && req->part &&
999 !(req->rq_flags & RQF_FLUSH_SEQ)) {
1000 const int sgrp = op_stat_group(req_op(req));
1001
1002 part_stat_lock();
1003 update_io_ticks(req->part, jiffies, true);
1004 part_stat_inc(req->part, ios[sgrp]);
1005 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
1006 part_stat_unlock();
1007 }
1008}
1009
1010static inline void blk_account_io_start(struct request *req)
1011{
1012 trace_block_io_start(req);
1013
1014 if (blk_do_io_stat(req)) {
1015 /*
1016 * All non-passthrough requests are created from a bio with one
1017 * exception: when a flush command that is part of a flush sequence
1018 * generated by the state machine in blk-flush.c is cloned onto the
1019 * lower device by dm-multipath we can get here without a bio.
1020 */
1021 if (req->bio)
1022 req->part = req->bio->bi_bdev;
1023 else
1024 req->part = req->q->disk->part0;
1025
1026 part_stat_lock();
1027 update_io_ticks(req->part, jiffies, false);
1028 part_stat_unlock();
1029 }
1030}
1031
1032static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
1033{
1034 if (rq->rq_flags & RQF_STATS)
1035 blk_stat_add(rq, now);
1036
1037 blk_mq_sched_completed_request(rq, now);
1038 blk_account_io_done(rq, now);
1039}
1040
1041inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
1042{
1043 if (blk_mq_need_time_stamp(rq))
1044 __blk_mq_end_request_acct(rq, ktime_get_ns());
1045
1046 blk_mq_finish_request(rq);
1047
1048 if (rq->end_io) {
1049 rq_qos_done(rq->q, rq);
1050 if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1051 blk_mq_free_request(rq);
1052 } else {
1053 blk_mq_free_request(rq);
1054 }
1055}
1056EXPORT_SYMBOL(__blk_mq_end_request);
1057
1058void blk_mq_end_request(struct request *rq, blk_status_t error)
1059{
1060 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1061 BUG();
1062 __blk_mq_end_request(rq, error);
1063}
1064EXPORT_SYMBOL(blk_mq_end_request);
1065
1066#define TAG_COMP_BATCH 32
1067
1068static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1069 int *tag_array, int nr_tags)
1070{
1071 struct request_queue *q = hctx->queue;
1072
1073 blk_mq_sub_active_requests(hctx, nr_tags);
1074
1075 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
1076 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
1077}
1078
1079void blk_mq_end_request_batch(struct io_comp_batch *iob)
1080{
1081 int tags[TAG_COMP_BATCH], nr_tags = 0;
1082 struct blk_mq_hw_ctx *cur_hctx = NULL;
1083 struct request *rq;
1084 u64 now = 0;
1085
1086 if (iob->need_ts)
1087 now = ktime_get_ns();
1088
1089 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1090 prefetch(rq->bio);
1091 prefetch(rq->rq_next);
1092
1093 blk_complete_request(rq);
1094 if (iob->need_ts)
1095 __blk_mq_end_request_acct(rq, now);
1096
1097 blk_mq_finish_request(rq);
1098
1099 rq_qos_done(rq->q, rq);
1100
1101 /*
1102 * If end_io handler returns NONE, then it still has
1103 * ownership of the request.
1104 */
1105 if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
1106 continue;
1107
1108 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1109 if (!req_ref_put_and_test(rq))
1110 continue;
1111
1112 blk_crypto_free_request(rq);
1113 blk_pm_mark_last_busy(rq);
1114
1115 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1116 if (cur_hctx)
1117 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1118 nr_tags = 0;
1119 cur_hctx = rq->mq_hctx;
1120 }
1121 tags[nr_tags++] = rq->tag;
1122 }
1123
1124 if (nr_tags)
1125 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1126}
1127EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1128
1129static void blk_complete_reqs(struct llist_head *list)
1130{
1131 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1132 struct request *rq, *next;
1133
1134 llist_for_each_entry_safe(rq, next, entry, ipi_list)
1135 rq->q->mq_ops->complete(rq);
1136}
1137
1138static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1139{
1140 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1141}
1142
1143static int blk_softirq_cpu_dead(unsigned int cpu)
1144{
1145 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1146 return 0;
1147}
1148
1149static void __blk_mq_complete_request_remote(void *data)
1150{
1151 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1152}
1153
1154static inline bool blk_mq_complete_need_ipi(struct request *rq)
1155{
1156 int cpu = raw_smp_processor_id();
1157
1158 if (!IS_ENABLED(CONFIG_SMP) ||
1159 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1160 return false;
1161 /*
1162 * With force threaded interrupts enabled, raising softirq from an SMP
1163 * function call will always result in waking the ksoftirqd thread.
1164 * This is probably worse than completing the request on a different
1165 * cache domain.
1166 */
1167 if (force_irqthreads())
1168 return false;
1169
1170 /* same CPU or cache domain? Complete locally */
1171 if (cpu == rq->mq_ctx->cpu ||
1172 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1173 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1174 return false;
1175
1176 /* don't try to IPI to an offline CPU */
1177 return cpu_online(rq->mq_ctx->cpu);
1178}
1179
1180static void blk_mq_complete_send_ipi(struct request *rq)
1181{
1182 unsigned int cpu;
1183
1184 cpu = rq->mq_ctx->cpu;
1185 if (llist_add(&rq->ipi_list, &per_cpu(blk_cpu_done, cpu)))
1186 smp_call_function_single_async(cpu, &per_cpu(blk_cpu_csd, cpu));
1187}
1188
1189static void blk_mq_raise_softirq(struct request *rq)
1190{
1191 struct llist_head *list;
1192
1193 preempt_disable();
1194 list = this_cpu_ptr(&blk_cpu_done);
1195 if (llist_add(&rq->ipi_list, list))
1196 raise_softirq(BLOCK_SOFTIRQ);
1197 preempt_enable();
1198}
1199
1200bool blk_mq_complete_request_remote(struct request *rq)
1201{
1202 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1203
1204 /*
1205 * For request which hctx has only one ctx mapping,
1206 * or a polled request, always complete locally,
1207 * it's pointless to redirect the completion.
1208 */
1209 if ((rq->mq_hctx->nr_ctx == 1 &&
1210 rq->mq_ctx->cpu == raw_smp_processor_id()) ||
1211 rq->cmd_flags & REQ_POLLED)
1212 return false;
1213
1214 if (blk_mq_complete_need_ipi(rq)) {
1215 blk_mq_complete_send_ipi(rq);
1216 return true;
1217 }
1218
1219 if (rq->q->nr_hw_queues == 1) {
1220 blk_mq_raise_softirq(rq);
1221 return true;
1222 }
1223 return false;
1224}
1225EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1226
1227/**
1228 * blk_mq_complete_request - end I/O on a request
1229 * @rq: the request being processed
1230 *
1231 * Description:
1232 * Complete a request by scheduling the ->complete_rq operation.
1233 **/
1234void blk_mq_complete_request(struct request *rq)
1235{
1236 if (!blk_mq_complete_request_remote(rq))
1237 rq->q->mq_ops->complete(rq);
1238}
1239EXPORT_SYMBOL(blk_mq_complete_request);
1240
1241/**
1242 * blk_mq_start_request - Start processing a request
1243 * @rq: Pointer to request to be started
1244 *
1245 * Function used by device drivers to notify the block layer that a request
1246 * is going to be processed now, so blk layer can do proper initializations
1247 * such as starting the timeout timer.
1248 */
1249void blk_mq_start_request(struct request *rq)
1250{
1251 struct request_queue *q = rq->q;
1252
1253 trace_block_rq_issue(rq);
1254
1255 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags) &&
1256 !blk_rq_is_passthrough(rq)) {
1257 rq->io_start_time_ns = ktime_get_ns();
1258 rq->stats_sectors = blk_rq_sectors(rq);
1259 rq->rq_flags |= RQF_STATS;
1260 rq_qos_issue(q, rq);
1261 }
1262
1263 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1264
1265 blk_add_timer(rq);
1266 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1267 rq->mq_hctx->tags->rqs[rq->tag] = rq;
1268
1269#ifdef CONFIG_BLK_DEV_INTEGRITY
1270 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1271 q->integrity.profile->prepare_fn(rq);
1272#endif
1273 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1274 WRITE_ONCE(rq->bio->bi_cookie, rq->mq_hctx->queue_num);
1275}
1276EXPORT_SYMBOL(blk_mq_start_request);
1277
1278/*
1279 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1280 * queues. This is important for md arrays to benefit from merging
1281 * requests.
1282 */
1283static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1284{
1285 if (plug->multiple_queues)
1286 return BLK_MAX_REQUEST_COUNT * 2;
1287 return BLK_MAX_REQUEST_COUNT;
1288}
1289
1290static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1291{
1292 struct request *last = rq_list_peek(&plug->mq_list);
1293
1294 if (!plug->rq_count) {
1295 trace_block_plug(rq->q);
1296 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1297 (!blk_queue_nomerges(rq->q) &&
1298 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1299 blk_mq_flush_plug_list(plug, false);
1300 last = NULL;
1301 trace_block_plug(rq->q);
1302 }
1303
1304 if (!plug->multiple_queues && last && last->q != rq->q)
1305 plug->multiple_queues = true;
1306 /*
1307 * Any request allocated from sched tags can't be issued to
1308 * ->queue_rqs() directly
1309 */
1310 if (!plug->has_elevator && (rq->rq_flags & RQF_SCHED_TAGS))
1311 plug->has_elevator = true;
1312 rq->rq_next = NULL;
1313 rq_list_add(&plug->mq_list, rq);
1314 plug->rq_count++;
1315}
1316
1317/**
1318 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1319 * @rq: request to insert
1320 * @at_head: insert request at head or tail of queue
1321 *
1322 * Description:
1323 * Insert a fully prepared request at the back of the I/O scheduler queue
1324 * for execution. Don't wait for completion.
1325 *
1326 * Note:
1327 * This function will invoke @done directly if the queue is dead.
1328 */
1329void blk_execute_rq_nowait(struct request *rq, bool at_head)
1330{
1331 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1332
1333 WARN_ON(irqs_disabled());
1334 WARN_ON(!blk_rq_is_passthrough(rq));
1335
1336 blk_account_io_start(rq);
1337
1338 /*
1339 * As plugging can be enabled for passthrough requests on a zoned
1340 * device, directly accessing the plug instead of using blk_mq_plug()
1341 * should not have any consequences.
1342 */
1343 if (current->plug && !at_head) {
1344 blk_add_rq_to_plug(current->plug, rq);
1345 return;
1346 }
1347
1348 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1349 blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);
1350}
1351EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1352
1353struct blk_rq_wait {
1354 struct completion done;
1355 blk_status_t ret;
1356};
1357
1358static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
1359{
1360 struct blk_rq_wait *wait = rq->end_io_data;
1361
1362 wait->ret = ret;
1363 complete(&wait->done);
1364 return RQ_END_IO_NONE;
1365}
1366
1367bool blk_rq_is_poll(struct request *rq)
1368{
1369 if (!rq->mq_hctx)
1370 return false;
1371 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1372 return false;
1373 return true;
1374}
1375EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1376
1377static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1378{
1379 do {
1380 blk_hctx_poll(rq->q, rq->mq_hctx, NULL, 0);
1381 cond_resched();
1382 } while (!completion_done(wait));
1383}
1384
1385/**
1386 * blk_execute_rq - insert a request into queue for execution
1387 * @rq: request to insert
1388 * @at_head: insert request at head or tail of queue
1389 *
1390 * Description:
1391 * Insert a fully prepared request at the back of the I/O scheduler queue
1392 * for execution and wait for completion.
1393 * Return: The blk_status_t result provided to blk_mq_end_request().
1394 */
1395blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1396{
1397 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1398 struct blk_rq_wait wait = {
1399 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1400 };
1401
1402 WARN_ON(irqs_disabled());
1403 WARN_ON(!blk_rq_is_passthrough(rq));
1404
1405 rq->end_io_data = &wait;
1406 rq->end_io = blk_end_sync_rq;
1407
1408 blk_account_io_start(rq);
1409 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1410 blk_mq_run_hw_queue(hctx, false);
1411
1412 if (blk_rq_is_poll(rq)) {
1413 blk_rq_poll_completion(rq, &wait.done);
1414 } else {
1415 /*
1416 * Prevent hang_check timer from firing at us during very long
1417 * I/O
1418 */
1419 unsigned long hang_check = sysctl_hung_task_timeout_secs;
1420
1421 if (hang_check)
1422 while (!wait_for_completion_io_timeout(&wait.done,
1423 hang_check * (HZ/2)))
1424 ;
1425 else
1426 wait_for_completion_io(&wait.done);
1427 }
1428
1429 return wait.ret;
1430}
1431EXPORT_SYMBOL(blk_execute_rq);
1432
1433static void __blk_mq_requeue_request(struct request *rq)
1434{
1435 struct request_queue *q = rq->q;
1436
1437 blk_mq_put_driver_tag(rq);
1438
1439 trace_block_rq_requeue(rq);
1440 rq_qos_requeue(q, rq);
1441
1442 if (blk_mq_request_started(rq)) {
1443 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1444 rq->rq_flags &= ~RQF_TIMED_OUT;
1445 }
1446}
1447
1448void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1449{
1450 struct request_queue *q = rq->q;
1451 unsigned long flags;
1452
1453 __blk_mq_requeue_request(rq);
1454
1455 /* this request will be re-inserted to io scheduler queue */
1456 blk_mq_sched_requeue_request(rq);
1457
1458 spin_lock_irqsave(&q->requeue_lock, flags);
1459 list_add_tail(&rq->queuelist, &q->requeue_list);
1460 spin_unlock_irqrestore(&q->requeue_lock, flags);
1461
1462 if (kick_requeue_list)
1463 blk_mq_kick_requeue_list(q);
1464}
1465EXPORT_SYMBOL(blk_mq_requeue_request);
1466
1467static void blk_mq_requeue_work(struct work_struct *work)
1468{
1469 struct request_queue *q =
1470 container_of(work, struct request_queue, requeue_work.work);
1471 LIST_HEAD(rq_list);
1472 LIST_HEAD(flush_list);
1473 struct request *rq;
1474
1475 spin_lock_irq(&q->requeue_lock);
1476 list_splice_init(&q->requeue_list, &rq_list);
1477 list_splice_init(&q->flush_list, &flush_list);
1478 spin_unlock_irq(&q->requeue_lock);
1479
1480 while (!list_empty(&rq_list)) {
1481 rq = list_entry(rq_list.next, struct request, queuelist);
1482 /*
1483 * If RQF_DONTPREP ist set, the request has been started by the
1484 * driver already and might have driver-specific data allocated
1485 * already. Insert it into the hctx dispatch list to avoid
1486 * block layer merges for the request.
1487 */
1488 if (rq->rq_flags & RQF_DONTPREP) {
1489 list_del_init(&rq->queuelist);
1490 blk_mq_request_bypass_insert(rq, 0);
1491 } else {
1492 list_del_init(&rq->queuelist);
1493 blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD);
1494 }
1495 }
1496
1497 while (!list_empty(&flush_list)) {
1498 rq = list_entry(flush_list.next, struct request, queuelist);
1499 list_del_init(&rq->queuelist);
1500 blk_mq_insert_request(rq, 0);
1501 }
1502
1503 blk_mq_run_hw_queues(q, false);
1504}
1505
1506void blk_mq_kick_requeue_list(struct request_queue *q)
1507{
1508 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1509}
1510EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1511
1512void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1513 unsigned long msecs)
1514{
1515 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1516 msecs_to_jiffies(msecs));
1517}
1518EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1519
1520static bool blk_is_flush_data_rq(struct request *rq)
1521{
1522 return (rq->rq_flags & RQF_FLUSH_SEQ) && !is_flush_rq(rq);
1523}
1524
1525static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1526{
1527 /*
1528 * If we find a request that isn't idle we know the queue is busy
1529 * as it's checked in the iter.
1530 * Return false to stop the iteration.
1531 *
1532 * In case of queue quiesce, if one flush data request is completed,
1533 * don't count it as inflight given the flush sequence is suspended,
1534 * and the original flush data request is invisible to driver, just
1535 * like other pending requests because of quiesce
1536 */
1537 if (blk_mq_request_started(rq) && !(blk_queue_quiesced(rq->q) &&
1538 blk_is_flush_data_rq(rq) &&
1539 blk_mq_request_completed(rq))) {
1540 bool *busy = priv;
1541
1542 *busy = true;
1543 return false;
1544 }
1545
1546 return true;
1547}
1548
1549bool blk_mq_queue_inflight(struct request_queue *q)
1550{
1551 bool busy = false;
1552
1553 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1554 return busy;
1555}
1556EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1557
1558static void blk_mq_rq_timed_out(struct request *req)
1559{
1560 req->rq_flags |= RQF_TIMED_OUT;
1561 if (req->q->mq_ops->timeout) {
1562 enum blk_eh_timer_return ret;
1563
1564 ret = req->q->mq_ops->timeout(req);
1565 if (ret == BLK_EH_DONE)
1566 return;
1567 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1568 }
1569
1570 blk_add_timer(req);
1571}
1572
1573struct blk_expired_data {
1574 bool has_timedout_rq;
1575 unsigned long next;
1576 unsigned long timeout_start;
1577};
1578
1579static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
1580{
1581 unsigned long deadline;
1582
1583 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1584 return false;
1585 if (rq->rq_flags & RQF_TIMED_OUT)
1586 return false;
1587
1588 deadline = READ_ONCE(rq->deadline);
1589 if (time_after_eq(expired->timeout_start, deadline))
1590 return true;
1591
1592 if (expired->next == 0)
1593 expired->next = deadline;
1594 else if (time_after(expired->next, deadline))
1595 expired->next = deadline;
1596 return false;
1597}
1598
1599void blk_mq_put_rq_ref(struct request *rq)
1600{
1601 if (is_flush_rq(rq)) {
1602 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1603 blk_mq_free_request(rq);
1604 } else if (req_ref_put_and_test(rq)) {
1605 __blk_mq_free_request(rq);
1606 }
1607}
1608
1609static bool blk_mq_check_expired(struct request *rq, void *priv)
1610{
1611 struct blk_expired_data *expired = priv;
1612
1613 /*
1614 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1615 * be reallocated underneath the timeout handler's processing, then
1616 * the expire check is reliable. If the request is not expired, then
1617 * it was completed and reallocated as a new request after returning
1618 * from blk_mq_check_expired().
1619 */
1620 if (blk_mq_req_expired(rq, expired)) {
1621 expired->has_timedout_rq = true;
1622 return false;
1623 }
1624 return true;
1625}
1626
1627static bool blk_mq_handle_expired(struct request *rq, void *priv)
1628{
1629 struct blk_expired_data *expired = priv;
1630
1631 if (blk_mq_req_expired(rq, expired))
1632 blk_mq_rq_timed_out(rq);
1633 return true;
1634}
1635
1636static void blk_mq_timeout_work(struct work_struct *work)
1637{
1638 struct request_queue *q =
1639 container_of(work, struct request_queue, timeout_work);
1640 struct blk_expired_data expired = {
1641 .timeout_start = jiffies,
1642 };
1643 struct blk_mq_hw_ctx *hctx;
1644 unsigned long i;
1645
1646 /* A deadlock might occur if a request is stuck requiring a
1647 * timeout at the same time a queue freeze is waiting
1648 * completion, since the timeout code would not be able to
1649 * acquire the queue reference here.
1650 *
1651 * That's why we don't use blk_queue_enter here; instead, we use
1652 * percpu_ref_tryget directly, because we need to be able to
1653 * obtain a reference even in the short window between the queue
1654 * starting to freeze, by dropping the first reference in
1655 * blk_freeze_queue_start, and the moment the last request is
1656 * consumed, marked by the instant q_usage_counter reaches
1657 * zero.
1658 */
1659 if (!percpu_ref_tryget(&q->q_usage_counter))
1660 return;
1661
1662 /* check if there is any timed-out request */
1663 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
1664 if (expired.has_timedout_rq) {
1665 /*
1666 * Before walking tags, we must ensure any submit started
1667 * before the current time has finished. Since the submit
1668 * uses srcu or rcu, wait for a synchronization point to
1669 * ensure all running submits have finished
1670 */
1671 blk_mq_wait_quiesce_done(q->tag_set);
1672
1673 expired.next = 0;
1674 blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
1675 }
1676
1677 if (expired.next != 0) {
1678 mod_timer(&q->timeout, expired.next);
1679 } else {
1680 /*
1681 * Request timeouts are handled as a forward rolling timer. If
1682 * we end up here it means that no requests are pending and
1683 * also that no request has been pending for a while. Mark
1684 * each hctx as idle.
1685 */
1686 queue_for_each_hw_ctx(q, hctx, i) {
1687 /* the hctx may be unmapped, so check it here */
1688 if (blk_mq_hw_queue_mapped(hctx))
1689 blk_mq_tag_idle(hctx);
1690 }
1691 }
1692 blk_queue_exit(q);
1693}
1694
1695struct flush_busy_ctx_data {
1696 struct blk_mq_hw_ctx *hctx;
1697 struct list_head *list;
1698};
1699
1700static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1701{
1702 struct flush_busy_ctx_data *flush_data = data;
1703 struct blk_mq_hw_ctx *hctx = flush_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 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1709 sbitmap_clear_bit(sb, bitnr);
1710 spin_unlock(&ctx->lock);
1711 return true;
1712}
1713
1714/*
1715 * Process software queues that have been marked busy, splicing them
1716 * to the for-dispatch
1717 */
1718void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1719{
1720 struct flush_busy_ctx_data data = {
1721 .hctx = hctx,
1722 .list = list,
1723 };
1724
1725 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1726}
1727EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1728
1729struct dispatch_rq_data {
1730 struct blk_mq_hw_ctx *hctx;
1731 struct request *rq;
1732};
1733
1734static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1735 void *data)
1736{
1737 struct dispatch_rq_data *dispatch_data = data;
1738 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1739 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1740 enum hctx_type type = hctx->type;
1741
1742 spin_lock(&ctx->lock);
1743 if (!list_empty(&ctx->rq_lists[type])) {
1744 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1745 list_del_init(&dispatch_data->rq->queuelist);
1746 if (list_empty(&ctx->rq_lists[type]))
1747 sbitmap_clear_bit(sb, bitnr);
1748 }
1749 spin_unlock(&ctx->lock);
1750
1751 return !dispatch_data->rq;
1752}
1753
1754struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1755 struct blk_mq_ctx *start)
1756{
1757 unsigned off = start ? start->index_hw[hctx->type] : 0;
1758 struct dispatch_rq_data data = {
1759 .hctx = hctx,
1760 .rq = NULL,
1761 };
1762
1763 __sbitmap_for_each_set(&hctx->ctx_map, off,
1764 dispatch_rq_from_ctx, &data);
1765
1766 return data.rq;
1767}
1768
1769bool __blk_mq_alloc_driver_tag(struct request *rq)
1770{
1771 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1772 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1773 int tag;
1774
1775 blk_mq_tag_busy(rq->mq_hctx);
1776
1777 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1778 bt = &rq->mq_hctx->tags->breserved_tags;
1779 tag_offset = 0;
1780 } else {
1781 if (!hctx_may_queue(rq->mq_hctx, bt))
1782 return false;
1783 }
1784
1785 tag = __sbitmap_queue_get(bt);
1786 if (tag == BLK_MQ_NO_TAG)
1787 return false;
1788
1789 rq->tag = tag + tag_offset;
1790 blk_mq_inc_active_requests(rq->mq_hctx);
1791 return true;
1792}
1793
1794static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1795 int flags, void *key)
1796{
1797 struct blk_mq_hw_ctx *hctx;
1798
1799 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1800
1801 spin_lock(&hctx->dispatch_wait_lock);
1802 if (!list_empty(&wait->entry)) {
1803 struct sbitmap_queue *sbq;
1804
1805 list_del_init(&wait->entry);
1806 sbq = &hctx->tags->bitmap_tags;
1807 atomic_dec(&sbq->ws_active);
1808 }
1809 spin_unlock(&hctx->dispatch_wait_lock);
1810
1811 blk_mq_run_hw_queue(hctx, true);
1812 return 1;
1813}
1814
1815/*
1816 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1817 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1818 * restart. For both cases, take care to check the condition again after
1819 * marking us as waiting.
1820 */
1821static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1822 struct request *rq)
1823{
1824 struct sbitmap_queue *sbq;
1825 struct wait_queue_head *wq;
1826 wait_queue_entry_t *wait;
1827 bool ret;
1828
1829 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1830 !(blk_mq_is_shared_tags(hctx->flags))) {
1831 blk_mq_sched_mark_restart_hctx(hctx);
1832
1833 /*
1834 * It's possible that a tag was freed in the window between the
1835 * allocation failure and adding the hardware queue to the wait
1836 * queue.
1837 *
1838 * Don't clear RESTART here, someone else could have set it.
1839 * At most this will cost an extra queue run.
1840 */
1841 return blk_mq_get_driver_tag(rq);
1842 }
1843
1844 wait = &hctx->dispatch_wait;
1845 if (!list_empty_careful(&wait->entry))
1846 return false;
1847
1848 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag))
1849 sbq = &hctx->tags->breserved_tags;
1850 else
1851 sbq = &hctx->tags->bitmap_tags;
1852 wq = &bt_wait_ptr(sbq, hctx)->wait;
1853
1854 spin_lock_irq(&wq->lock);
1855 spin_lock(&hctx->dispatch_wait_lock);
1856 if (!list_empty(&wait->entry)) {
1857 spin_unlock(&hctx->dispatch_wait_lock);
1858 spin_unlock_irq(&wq->lock);
1859 return false;
1860 }
1861
1862 atomic_inc(&sbq->ws_active);
1863 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1864 __add_wait_queue(wq, wait);
1865
1866 /*
1867 * Add one explicit barrier since blk_mq_get_driver_tag() may
1868 * not imply barrier in case of failure.
1869 *
1870 * Order adding us to wait queue and allocating driver tag.
1871 *
1872 * The pair is the one implied in sbitmap_queue_wake_up() which
1873 * orders clearing sbitmap tag bits and waitqueue_active() in
1874 * __sbitmap_queue_wake_up(), since waitqueue_active() is lockless
1875 *
1876 * Otherwise, re-order of adding wait queue and getting driver tag
1877 * may cause __sbitmap_queue_wake_up() to wake up nothing because
1878 * the waitqueue_active() may not observe us in wait queue.
1879 */
1880 smp_mb();
1881
1882 /*
1883 * It's possible that a tag was freed in the window between the
1884 * allocation failure and adding the hardware queue to the wait
1885 * queue.
1886 */
1887 ret = blk_mq_get_driver_tag(rq);
1888 if (!ret) {
1889 spin_unlock(&hctx->dispatch_wait_lock);
1890 spin_unlock_irq(&wq->lock);
1891 return false;
1892 }
1893
1894 /*
1895 * We got a tag, remove ourselves from the wait queue to ensure
1896 * someone else gets the wakeup.
1897 */
1898 list_del_init(&wait->entry);
1899 atomic_dec(&sbq->ws_active);
1900 spin_unlock(&hctx->dispatch_wait_lock);
1901 spin_unlock_irq(&wq->lock);
1902
1903 return true;
1904}
1905
1906#define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1907#define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1908/*
1909 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1910 * - EWMA is one simple way to compute running average value
1911 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1912 * - take 4 as factor for avoiding to get too small(0) result, and this
1913 * factor doesn't matter because EWMA decreases exponentially
1914 */
1915static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1916{
1917 unsigned int ewma;
1918
1919 ewma = hctx->dispatch_busy;
1920
1921 if (!ewma && !busy)
1922 return;
1923
1924 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1925 if (busy)
1926 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1927 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1928
1929 hctx->dispatch_busy = ewma;
1930}
1931
1932#define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1933
1934static void blk_mq_handle_dev_resource(struct request *rq,
1935 struct list_head *list)
1936{
1937 list_add(&rq->queuelist, list);
1938 __blk_mq_requeue_request(rq);
1939}
1940
1941static void blk_mq_handle_zone_resource(struct request *rq,
1942 struct list_head *zone_list)
1943{
1944 /*
1945 * If we end up here it is because we cannot dispatch a request to a
1946 * specific zone due to LLD level zone-write locking or other zone
1947 * related resource not being available. In this case, set the request
1948 * aside in zone_list for retrying it later.
1949 */
1950 list_add(&rq->queuelist, zone_list);
1951 __blk_mq_requeue_request(rq);
1952}
1953
1954enum prep_dispatch {
1955 PREP_DISPATCH_OK,
1956 PREP_DISPATCH_NO_TAG,
1957 PREP_DISPATCH_NO_BUDGET,
1958};
1959
1960static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1961 bool need_budget)
1962{
1963 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1964 int budget_token = -1;
1965
1966 if (need_budget) {
1967 budget_token = blk_mq_get_dispatch_budget(rq->q);
1968 if (budget_token < 0) {
1969 blk_mq_put_driver_tag(rq);
1970 return PREP_DISPATCH_NO_BUDGET;
1971 }
1972 blk_mq_set_rq_budget_token(rq, budget_token);
1973 }
1974
1975 if (!blk_mq_get_driver_tag(rq)) {
1976 /*
1977 * The initial allocation attempt failed, so we need to
1978 * rerun the hardware queue when a tag is freed. The
1979 * waitqueue takes care of that. If the queue is run
1980 * before we add this entry back on the dispatch list,
1981 * we'll re-run it below.
1982 */
1983 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1984 /*
1985 * All budgets not got from this function will be put
1986 * together during handling partial dispatch
1987 */
1988 if (need_budget)
1989 blk_mq_put_dispatch_budget(rq->q, budget_token);
1990 return PREP_DISPATCH_NO_TAG;
1991 }
1992 }
1993
1994 return PREP_DISPATCH_OK;
1995}
1996
1997/* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1998static void blk_mq_release_budgets(struct request_queue *q,
1999 struct list_head *list)
2000{
2001 struct request *rq;
2002
2003 list_for_each_entry(rq, list, queuelist) {
2004 int budget_token = blk_mq_get_rq_budget_token(rq);
2005
2006 if (budget_token >= 0)
2007 blk_mq_put_dispatch_budget(q, budget_token);
2008 }
2009}
2010
2011/*
2012 * blk_mq_commit_rqs will notify driver using bd->last that there is no
2013 * more requests. (See comment in struct blk_mq_ops for commit_rqs for
2014 * details)
2015 * Attention, we should explicitly call this in unusual cases:
2016 * 1) did not queue everything initially scheduled to queue
2017 * 2) the last attempt to queue a request failed
2018 */
2019static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued,
2020 bool from_schedule)
2021{
2022 if (hctx->queue->mq_ops->commit_rqs && queued) {
2023 trace_block_unplug(hctx->queue, queued, !from_schedule);
2024 hctx->queue->mq_ops->commit_rqs(hctx);
2025 }
2026}
2027
2028/*
2029 * Returns true if we did some work AND can potentially do more.
2030 */
2031bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
2032 unsigned int nr_budgets)
2033{
2034 enum prep_dispatch prep;
2035 struct request_queue *q = hctx->queue;
2036 struct request *rq;
2037 int queued;
2038 blk_status_t ret = BLK_STS_OK;
2039 LIST_HEAD(zone_list);
2040 bool needs_resource = false;
2041
2042 if (list_empty(list))
2043 return false;
2044
2045 /*
2046 * Now process all the entries, sending them to the driver.
2047 */
2048 queued = 0;
2049 do {
2050 struct blk_mq_queue_data bd;
2051
2052 rq = list_first_entry(list, struct request, queuelist);
2053
2054 WARN_ON_ONCE(hctx != rq->mq_hctx);
2055 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
2056 if (prep != PREP_DISPATCH_OK)
2057 break;
2058
2059 list_del_init(&rq->queuelist);
2060
2061 bd.rq = rq;
2062 bd.last = list_empty(list);
2063
2064 /*
2065 * once the request is queued to lld, no need to cover the
2066 * budget any more
2067 */
2068 if (nr_budgets)
2069 nr_budgets--;
2070 ret = q->mq_ops->queue_rq(hctx, &bd);
2071 switch (ret) {
2072 case BLK_STS_OK:
2073 queued++;
2074 break;
2075 case BLK_STS_RESOURCE:
2076 needs_resource = true;
2077 fallthrough;
2078 case BLK_STS_DEV_RESOURCE:
2079 blk_mq_handle_dev_resource(rq, list);
2080 goto out;
2081 case BLK_STS_ZONE_RESOURCE:
2082 /*
2083 * Move the request to zone_list and keep going through
2084 * the dispatch list to find more requests the drive can
2085 * accept.
2086 */
2087 blk_mq_handle_zone_resource(rq, &zone_list);
2088 needs_resource = true;
2089 break;
2090 default:
2091 blk_mq_end_request(rq, ret);
2092 }
2093 } while (!list_empty(list));
2094out:
2095 if (!list_empty(&zone_list))
2096 list_splice_tail_init(&zone_list, list);
2097
2098 /* If we didn't flush the entire list, we could have told the driver
2099 * there was more coming, but that turned out to be a lie.
2100 */
2101 if (!list_empty(list) || ret != BLK_STS_OK)
2102 blk_mq_commit_rqs(hctx, queued, false);
2103
2104 /*
2105 * Any items that need requeuing? Stuff them into hctx->dispatch,
2106 * that is where we will continue on next queue run.
2107 */
2108 if (!list_empty(list)) {
2109 bool needs_restart;
2110 /* For non-shared tags, the RESTART check will suffice */
2111 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2112 ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
2113 blk_mq_is_shared_tags(hctx->flags));
2114
2115 if (nr_budgets)
2116 blk_mq_release_budgets(q, list);
2117
2118 spin_lock(&hctx->lock);
2119 list_splice_tail_init(list, &hctx->dispatch);
2120 spin_unlock(&hctx->lock);
2121
2122 /*
2123 * Order adding requests to hctx->dispatch and checking
2124 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2125 * in blk_mq_sched_restart(). Avoid restart code path to
2126 * miss the new added requests to hctx->dispatch, meantime
2127 * SCHED_RESTART is observed here.
2128 */
2129 smp_mb();
2130
2131 /*
2132 * If SCHED_RESTART was set by the caller of this function and
2133 * it is no longer set that means that it was cleared by another
2134 * thread and hence that a queue rerun is needed.
2135 *
2136 * If 'no_tag' is set, that means that we failed getting
2137 * a driver tag with an I/O scheduler attached. If our dispatch
2138 * waitqueue is no longer active, ensure that we run the queue
2139 * AFTER adding our entries back to the list.
2140 *
2141 * If no I/O scheduler has been configured it is possible that
2142 * the hardware queue got stopped and restarted before requests
2143 * were pushed back onto the dispatch list. Rerun the queue to
2144 * avoid starvation. Notes:
2145 * - blk_mq_run_hw_queue() checks whether or not a queue has
2146 * been stopped before rerunning a queue.
2147 * - Some but not all block drivers stop a queue before
2148 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2149 * and dm-rq.
2150 *
2151 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2152 * bit is set, run queue after a delay to avoid IO stalls
2153 * that could otherwise occur if the queue is idle. We'll do
2154 * similar if we couldn't get budget or couldn't lock a zone
2155 * and SCHED_RESTART is set.
2156 */
2157 needs_restart = blk_mq_sched_needs_restart(hctx);
2158 if (prep == PREP_DISPATCH_NO_BUDGET)
2159 needs_resource = true;
2160 if (!needs_restart ||
2161 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2162 blk_mq_run_hw_queue(hctx, true);
2163 else if (needs_resource)
2164 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2165
2166 blk_mq_update_dispatch_busy(hctx, true);
2167 return false;
2168 }
2169
2170 blk_mq_update_dispatch_busy(hctx, false);
2171 return true;
2172}
2173
2174static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2175{
2176 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2177
2178 if (cpu >= nr_cpu_ids)
2179 cpu = cpumask_first(hctx->cpumask);
2180 return cpu;
2181}
2182
2183/*
2184 * It'd be great if the workqueue API had a way to pass
2185 * in a mask and had some smarts for more clever placement.
2186 * For now we just round-robin here, switching for every
2187 * BLK_MQ_CPU_WORK_BATCH queued items.
2188 */
2189static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2190{
2191 bool tried = false;
2192 int next_cpu = hctx->next_cpu;
2193
2194 if (hctx->queue->nr_hw_queues == 1)
2195 return WORK_CPU_UNBOUND;
2196
2197 if (--hctx->next_cpu_batch <= 0) {
2198select_cpu:
2199 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2200 cpu_online_mask);
2201 if (next_cpu >= nr_cpu_ids)
2202 next_cpu = blk_mq_first_mapped_cpu(hctx);
2203 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2204 }
2205
2206 /*
2207 * Do unbound schedule if we can't find a online CPU for this hctx,
2208 * and it should only happen in the path of handling CPU DEAD.
2209 */
2210 if (!cpu_online(next_cpu)) {
2211 if (!tried) {
2212 tried = true;
2213 goto select_cpu;
2214 }
2215
2216 /*
2217 * Make sure to re-select CPU next time once after CPUs
2218 * in hctx->cpumask become online again.
2219 */
2220 hctx->next_cpu = next_cpu;
2221 hctx->next_cpu_batch = 1;
2222 return WORK_CPU_UNBOUND;
2223 }
2224
2225 hctx->next_cpu = next_cpu;
2226 return next_cpu;
2227}
2228
2229/**
2230 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2231 * @hctx: Pointer to the hardware queue to run.
2232 * @msecs: Milliseconds of delay to wait before running the queue.
2233 *
2234 * Run a hardware queue asynchronously with a delay of @msecs.
2235 */
2236void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2237{
2238 if (unlikely(blk_mq_hctx_stopped(hctx)))
2239 return;
2240 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2241 msecs_to_jiffies(msecs));
2242}
2243EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2244
2245/**
2246 * blk_mq_run_hw_queue - Start to run a hardware queue.
2247 * @hctx: Pointer to the hardware queue to run.
2248 * @async: If we want to run the queue asynchronously.
2249 *
2250 * Check if the request queue is not in a quiesced state and if there are
2251 * pending requests to be sent. If this is true, run the queue to send requests
2252 * to hardware.
2253 */
2254void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2255{
2256 bool need_run;
2257
2258 /*
2259 * We can't run the queue inline with interrupts disabled.
2260 */
2261 WARN_ON_ONCE(!async && in_interrupt());
2262
2263 might_sleep_if(!async && hctx->flags & BLK_MQ_F_BLOCKING);
2264
2265 /*
2266 * When queue is quiesced, we may be switching io scheduler, or
2267 * updating nr_hw_queues, or other things, and we can't run queue
2268 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2269 *
2270 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2271 * quiesced.
2272 */
2273 __blk_mq_run_dispatch_ops(hctx->queue, false,
2274 need_run = !blk_queue_quiesced(hctx->queue) &&
2275 blk_mq_hctx_has_pending(hctx));
2276
2277 if (!need_run)
2278 return;
2279
2280 if (async || !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2281 blk_mq_delay_run_hw_queue(hctx, 0);
2282 return;
2283 }
2284
2285 blk_mq_run_dispatch_ops(hctx->queue,
2286 blk_mq_sched_dispatch_requests(hctx));
2287}
2288EXPORT_SYMBOL(blk_mq_run_hw_queue);
2289
2290/*
2291 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2292 * scheduler.
2293 */
2294static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2295{
2296 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2297 /*
2298 * If the IO scheduler does not respect hardware queues when
2299 * dispatching, we just don't bother with multiple HW queues and
2300 * dispatch from hctx for the current CPU since running multiple queues
2301 * just causes lock contention inside the scheduler and pointless cache
2302 * bouncing.
2303 */
2304 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2305
2306 if (!blk_mq_hctx_stopped(hctx))
2307 return hctx;
2308 return NULL;
2309}
2310
2311/**
2312 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2313 * @q: Pointer to the request queue to run.
2314 * @async: If we want to run the queue asynchronously.
2315 */
2316void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2317{
2318 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2319 unsigned long i;
2320
2321 sq_hctx = NULL;
2322 if (blk_queue_sq_sched(q))
2323 sq_hctx = blk_mq_get_sq_hctx(q);
2324 queue_for_each_hw_ctx(q, hctx, i) {
2325 if (blk_mq_hctx_stopped(hctx))
2326 continue;
2327 /*
2328 * Dispatch from this hctx either if there's no hctx preferred
2329 * by IO scheduler or if it has requests that bypass the
2330 * scheduler.
2331 */
2332 if (!sq_hctx || sq_hctx == hctx ||
2333 !list_empty_careful(&hctx->dispatch))
2334 blk_mq_run_hw_queue(hctx, async);
2335 }
2336}
2337EXPORT_SYMBOL(blk_mq_run_hw_queues);
2338
2339/**
2340 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2341 * @q: Pointer to the request queue to run.
2342 * @msecs: Milliseconds of delay to wait before running the queues.
2343 */
2344void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2345{
2346 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2347 unsigned long i;
2348
2349 sq_hctx = NULL;
2350 if (blk_queue_sq_sched(q))
2351 sq_hctx = blk_mq_get_sq_hctx(q);
2352 queue_for_each_hw_ctx(q, hctx, i) {
2353 if (blk_mq_hctx_stopped(hctx))
2354 continue;
2355 /*
2356 * If there is already a run_work pending, leave the
2357 * pending delay untouched. Otherwise, a hctx can stall
2358 * if another hctx is re-delaying the other's work
2359 * before the work executes.
2360 */
2361 if (delayed_work_pending(&hctx->run_work))
2362 continue;
2363 /*
2364 * Dispatch from this hctx either if there's no hctx preferred
2365 * by IO scheduler or if it has requests that bypass the
2366 * scheduler.
2367 */
2368 if (!sq_hctx || sq_hctx == hctx ||
2369 !list_empty_careful(&hctx->dispatch))
2370 blk_mq_delay_run_hw_queue(hctx, msecs);
2371 }
2372}
2373EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2374
2375/*
2376 * This function is often used for pausing .queue_rq() by driver when
2377 * there isn't enough resource or some conditions aren't satisfied, and
2378 * BLK_STS_RESOURCE is usually returned.
2379 *
2380 * We do not guarantee that dispatch can be drained or blocked
2381 * after blk_mq_stop_hw_queue() returns. Please use
2382 * blk_mq_quiesce_queue() for that requirement.
2383 */
2384void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2385{
2386 cancel_delayed_work(&hctx->run_work);
2387
2388 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2389}
2390EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2391
2392/*
2393 * This function is often used for pausing .queue_rq() by driver when
2394 * there isn't enough resource or some conditions aren't satisfied, and
2395 * BLK_STS_RESOURCE is usually returned.
2396 *
2397 * We do not guarantee that dispatch can be drained or blocked
2398 * after blk_mq_stop_hw_queues() returns. Please use
2399 * blk_mq_quiesce_queue() for that requirement.
2400 */
2401void blk_mq_stop_hw_queues(struct request_queue *q)
2402{
2403 struct blk_mq_hw_ctx *hctx;
2404 unsigned long i;
2405
2406 queue_for_each_hw_ctx(q, hctx, i)
2407 blk_mq_stop_hw_queue(hctx);
2408}
2409EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2410
2411void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2412{
2413 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2414
2415 blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);
2416}
2417EXPORT_SYMBOL(blk_mq_start_hw_queue);
2418
2419void blk_mq_start_hw_queues(struct request_queue *q)
2420{
2421 struct blk_mq_hw_ctx *hctx;
2422 unsigned long i;
2423
2424 queue_for_each_hw_ctx(q, hctx, i)
2425 blk_mq_start_hw_queue(hctx);
2426}
2427EXPORT_SYMBOL(blk_mq_start_hw_queues);
2428
2429void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2430{
2431 if (!blk_mq_hctx_stopped(hctx))
2432 return;
2433
2434 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2435 blk_mq_run_hw_queue(hctx, async);
2436}
2437EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2438
2439void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2440{
2441 struct blk_mq_hw_ctx *hctx;
2442 unsigned long i;
2443
2444 queue_for_each_hw_ctx(q, hctx, i)
2445 blk_mq_start_stopped_hw_queue(hctx, async ||
2446 (hctx->flags & BLK_MQ_F_BLOCKING));
2447}
2448EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2449
2450static void blk_mq_run_work_fn(struct work_struct *work)
2451{
2452 struct blk_mq_hw_ctx *hctx =
2453 container_of(work, struct blk_mq_hw_ctx, run_work.work);
2454
2455 blk_mq_run_dispatch_ops(hctx->queue,
2456 blk_mq_sched_dispatch_requests(hctx));
2457}
2458
2459/**
2460 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2461 * @rq: Pointer to request to be inserted.
2462 * @flags: BLK_MQ_INSERT_*
2463 *
2464 * Should only be used carefully, when the caller knows we want to
2465 * bypass a potential IO scheduler on the target device.
2466 */
2467static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags)
2468{
2469 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2470
2471 spin_lock(&hctx->lock);
2472 if (flags & BLK_MQ_INSERT_AT_HEAD)
2473 list_add(&rq->queuelist, &hctx->dispatch);
2474 else
2475 list_add_tail(&rq->queuelist, &hctx->dispatch);
2476 spin_unlock(&hctx->lock);
2477}
2478
2479static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx,
2480 struct blk_mq_ctx *ctx, struct list_head *list,
2481 bool run_queue_async)
2482{
2483 struct request *rq;
2484 enum hctx_type type = hctx->type;
2485
2486 /*
2487 * Try to issue requests directly if the hw queue isn't busy to save an
2488 * extra enqueue & dequeue to the sw queue.
2489 */
2490 if (!hctx->dispatch_busy && !run_queue_async) {
2491 blk_mq_run_dispatch_ops(hctx->queue,
2492 blk_mq_try_issue_list_directly(hctx, list));
2493 if (list_empty(list))
2494 goto out;
2495 }
2496
2497 /*
2498 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2499 * offline now
2500 */
2501 list_for_each_entry(rq, list, queuelist) {
2502 BUG_ON(rq->mq_ctx != ctx);
2503 trace_block_rq_insert(rq);
2504 if (rq->cmd_flags & REQ_NOWAIT)
2505 run_queue_async = true;
2506 }
2507
2508 spin_lock(&ctx->lock);
2509 list_splice_tail_init(list, &ctx->rq_lists[type]);
2510 blk_mq_hctx_mark_pending(hctx, ctx);
2511 spin_unlock(&ctx->lock);
2512out:
2513 blk_mq_run_hw_queue(hctx, run_queue_async);
2514}
2515
2516static void blk_mq_insert_request(struct request *rq, blk_insert_t flags)
2517{
2518 struct request_queue *q = rq->q;
2519 struct blk_mq_ctx *ctx = rq->mq_ctx;
2520 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2521
2522 if (blk_rq_is_passthrough(rq)) {
2523 /*
2524 * Passthrough request have to be added to hctx->dispatch
2525 * directly. The device may be in a situation where it can't
2526 * handle FS request, and always returns BLK_STS_RESOURCE for
2527 * them, which gets them added to hctx->dispatch.
2528 *
2529 * If a passthrough request is required to unblock the queues,
2530 * and it is added to the scheduler queue, there is no chance to
2531 * dispatch it given we prioritize requests in hctx->dispatch.
2532 */
2533 blk_mq_request_bypass_insert(rq, flags);
2534 } else if (req_op(rq) == REQ_OP_FLUSH) {
2535 /*
2536 * Firstly normal IO request is inserted to scheduler queue or
2537 * sw queue, meantime we add flush request to dispatch queue(
2538 * hctx->dispatch) directly and there is at most one in-flight
2539 * flush request for each hw queue, so it doesn't matter to add
2540 * flush request to tail or front of the dispatch queue.
2541 *
2542 * Secondly in case of NCQ, flush request belongs to non-NCQ
2543 * command, and queueing it will fail when there is any
2544 * in-flight normal IO request(NCQ command). When adding flush
2545 * rq to the front of hctx->dispatch, it is easier to introduce
2546 * extra time to flush rq's latency because of S_SCHED_RESTART
2547 * compared with adding to the tail of dispatch queue, then
2548 * chance of flush merge is increased, and less flush requests
2549 * will be issued to controller. It is observed that ~10% time
2550 * is saved in blktests block/004 on disk attached to AHCI/NCQ
2551 * drive when adding flush rq to the front of hctx->dispatch.
2552 *
2553 * Simply queue flush rq to the front of hctx->dispatch so that
2554 * intensive flush workloads can benefit in case of NCQ HW.
2555 */
2556 blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD);
2557 } else if (q->elevator) {
2558 LIST_HEAD(list);
2559
2560 WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG);
2561
2562 list_add(&rq->queuelist, &list);
2563 q->elevator->type->ops.insert_requests(hctx, &list, flags);
2564 } else {
2565 trace_block_rq_insert(rq);
2566
2567 spin_lock(&ctx->lock);
2568 if (flags & BLK_MQ_INSERT_AT_HEAD)
2569 list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]);
2570 else
2571 list_add_tail(&rq->queuelist,
2572 &ctx->rq_lists[hctx->type]);
2573 blk_mq_hctx_mark_pending(hctx, ctx);
2574 spin_unlock(&ctx->lock);
2575 }
2576}
2577
2578static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2579 unsigned int nr_segs)
2580{
2581 int err;
2582
2583 if (bio->bi_opf & REQ_RAHEAD)
2584 rq->cmd_flags |= REQ_FAILFAST_MASK;
2585
2586 rq->__sector = bio->bi_iter.bi_sector;
2587 blk_rq_bio_prep(rq, bio, nr_segs);
2588
2589 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2590 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2591 WARN_ON_ONCE(err);
2592
2593 blk_account_io_start(rq);
2594}
2595
2596static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2597 struct request *rq, bool last)
2598{
2599 struct request_queue *q = rq->q;
2600 struct blk_mq_queue_data bd = {
2601 .rq = rq,
2602 .last = last,
2603 };
2604 blk_status_t ret;
2605
2606 /*
2607 * For OK queue, we are done. For error, caller may kill it.
2608 * Any other error (busy), just add it to our list as we
2609 * previously would have done.
2610 */
2611 ret = q->mq_ops->queue_rq(hctx, &bd);
2612 switch (ret) {
2613 case BLK_STS_OK:
2614 blk_mq_update_dispatch_busy(hctx, false);
2615 break;
2616 case BLK_STS_RESOURCE:
2617 case BLK_STS_DEV_RESOURCE:
2618 blk_mq_update_dispatch_busy(hctx, true);
2619 __blk_mq_requeue_request(rq);
2620 break;
2621 default:
2622 blk_mq_update_dispatch_busy(hctx, false);
2623 break;
2624 }
2625
2626 return ret;
2627}
2628
2629static bool blk_mq_get_budget_and_tag(struct request *rq)
2630{
2631 int budget_token;
2632
2633 budget_token = blk_mq_get_dispatch_budget(rq->q);
2634 if (budget_token < 0)
2635 return false;
2636 blk_mq_set_rq_budget_token(rq, budget_token);
2637 if (!blk_mq_get_driver_tag(rq)) {
2638 blk_mq_put_dispatch_budget(rq->q, budget_token);
2639 return false;
2640 }
2641 return true;
2642}
2643
2644/**
2645 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2646 * @hctx: Pointer of the associated hardware queue.
2647 * @rq: Pointer to request to be sent.
2648 *
2649 * If the device has enough resources to accept a new request now, send the
2650 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2651 * we can try send it another time in the future. Requests inserted at this
2652 * queue have higher priority.
2653 */
2654static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2655 struct request *rq)
2656{
2657 blk_status_t ret;
2658
2659 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2660 blk_mq_insert_request(rq, 0);
2661 return;
2662 }
2663
2664 if ((rq->rq_flags & RQF_USE_SCHED) || !blk_mq_get_budget_and_tag(rq)) {
2665 blk_mq_insert_request(rq, 0);
2666 blk_mq_run_hw_queue(hctx, rq->cmd_flags & REQ_NOWAIT);
2667 return;
2668 }
2669
2670 ret = __blk_mq_issue_directly(hctx, rq, true);
2671 switch (ret) {
2672 case BLK_STS_OK:
2673 break;
2674 case BLK_STS_RESOURCE:
2675 case BLK_STS_DEV_RESOURCE:
2676 blk_mq_request_bypass_insert(rq, 0);
2677 blk_mq_run_hw_queue(hctx, false);
2678 break;
2679 default:
2680 blk_mq_end_request(rq, ret);
2681 break;
2682 }
2683}
2684
2685static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2686{
2687 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2688
2689 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2690 blk_mq_insert_request(rq, 0);
2691 return BLK_STS_OK;
2692 }
2693
2694 if (!blk_mq_get_budget_and_tag(rq))
2695 return BLK_STS_RESOURCE;
2696 return __blk_mq_issue_directly(hctx, rq, last);
2697}
2698
2699static void blk_mq_plug_issue_direct(struct blk_plug *plug)
2700{
2701 struct blk_mq_hw_ctx *hctx = NULL;
2702 struct request *rq;
2703 int queued = 0;
2704 blk_status_t ret = BLK_STS_OK;
2705
2706 while ((rq = rq_list_pop(&plug->mq_list))) {
2707 bool last = rq_list_empty(plug->mq_list);
2708
2709 if (hctx != rq->mq_hctx) {
2710 if (hctx) {
2711 blk_mq_commit_rqs(hctx, queued, false);
2712 queued = 0;
2713 }
2714 hctx = rq->mq_hctx;
2715 }
2716
2717 ret = blk_mq_request_issue_directly(rq, last);
2718 switch (ret) {
2719 case BLK_STS_OK:
2720 queued++;
2721 break;
2722 case BLK_STS_RESOURCE:
2723 case BLK_STS_DEV_RESOURCE:
2724 blk_mq_request_bypass_insert(rq, 0);
2725 blk_mq_run_hw_queue(hctx, false);
2726 goto out;
2727 default:
2728 blk_mq_end_request(rq, ret);
2729 break;
2730 }
2731 }
2732
2733out:
2734 if (ret != BLK_STS_OK)
2735 blk_mq_commit_rqs(hctx, queued, false);
2736}
2737
2738static void __blk_mq_flush_plug_list(struct request_queue *q,
2739 struct blk_plug *plug)
2740{
2741 if (blk_queue_quiesced(q))
2742 return;
2743 q->mq_ops->queue_rqs(&plug->mq_list);
2744}
2745
2746static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2747{
2748 struct blk_mq_hw_ctx *this_hctx = NULL;
2749 struct blk_mq_ctx *this_ctx = NULL;
2750 struct request *requeue_list = NULL;
2751 struct request **requeue_lastp = &requeue_list;
2752 unsigned int depth = 0;
2753 bool is_passthrough = false;
2754 LIST_HEAD(list);
2755
2756 do {
2757 struct request *rq = rq_list_pop(&plug->mq_list);
2758
2759 if (!this_hctx) {
2760 this_hctx = rq->mq_hctx;
2761 this_ctx = rq->mq_ctx;
2762 is_passthrough = blk_rq_is_passthrough(rq);
2763 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx ||
2764 is_passthrough != blk_rq_is_passthrough(rq)) {
2765 rq_list_add_tail(&requeue_lastp, rq);
2766 continue;
2767 }
2768 list_add(&rq->queuelist, &list);
2769 depth++;
2770 } while (!rq_list_empty(plug->mq_list));
2771
2772 plug->mq_list = requeue_list;
2773 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2774
2775 percpu_ref_get(&this_hctx->queue->q_usage_counter);
2776 /* passthrough requests should never be issued to the I/O scheduler */
2777 if (is_passthrough) {
2778 spin_lock(&this_hctx->lock);
2779 list_splice_tail_init(&list, &this_hctx->dispatch);
2780 spin_unlock(&this_hctx->lock);
2781 blk_mq_run_hw_queue(this_hctx, from_sched);
2782 } else if (this_hctx->queue->elevator) {
2783 this_hctx->queue->elevator->type->ops.insert_requests(this_hctx,
2784 &list, 0);
2785 blk_mq_run_hw_queue(this_hctx, from_sched);
2786 } else {
2787 blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched);
2788 }
2789 percpu_ref_put(&this_hctx->queue->q_usage_counter);
2790}
2791
2792void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2793{
2794 struct request *rq;
2795
2796 /*
2797 * We may have been called recursively midway through handling
2798 * plug->mq_list via a schedule() in the driver's queue_rq() callback.
2799 * To avoid mq_list changing under our feet, clear rq_count early and
2800 * bail out specifically if rq_count is 0 rather than checking
2801 * whether the mq_list is empty.
2802 */
2803 if (plug->rq_count == 0)
2804 return;
2805 plug->rq_count = 0;
2806
2807 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2808 struct request_queue *q;
2809
2810 rq = rq_list_peek(&plug->mq_list);
2811 q = rq->q;
2812
2813 /*
2814 * Peek first request and see if we have a ->queue_rqs() hook.
2815 * If we do, we can dispatch the whole plug list in one go. We
2816 * already know at this point that all requests belong to the
2817 * same queue, caller must ensure that's the case.
2818 */
2819 if (q->mq_ops->queue_rqs) {
2820 blk_mq_run_dispatch_ops(q,
2821 __blk_mq_flush_plug_list(q, plug));
2822 if (rq_list_empty(plug->mq_list))
2823 return;
2824 }
2825
2826 blk_mq_run_dispatch_ops(q,
2827 blk_mq_plug_issue_direct(plug));
2828 if (rq_list_empty(plug->mq_list))
2829 return;
2830 }
2831
2832 do {
2833 blk_mq_dispatch_plug_list(plug, from_schedule);
2834 } while (!rq_list_empty(plug->mq_list));
2835}
2836
2837static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2838 struct list_head *list)
2839{
2840 int queued = 0;
2841 blk_status_t ret = BLK_STS_OK;
2842
2843 while (!list_empty(list)) {
2844 struct request *rq = list_first_entry(list, struct request,
2845 queuelist);
2846
2847 list_del_init(&rq->queuelist);
2848 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2849 switch (ret) {
2850 case BLK_STS_OK:
2851 queued++;
2852 break;
2853 case BLK_STS_RESOURCE:
2854 case BLK_STS_DEV_RESOURCE:
2855 blk_mq_request_bypass_insert(rq, 0);
2856 if (list_empty(list))
2857 blk_mq_run_hw_queue(hctx, false);
2858 goto out;
2859 default:
2860 blk_mq_end_request(rq, ret);
2861 break;
2862 }
2863 }
2864
2865out:
2866 if (ret != BLK_STS_OK)
2867 blk_mq_commit_rqs(hctx, queued, false);
2868}
2869
2870static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2871 struct bio *bio, unsigned int nr_segs)
2872{
2873 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2874 if (blk_attempt_plug_merge(q, bio, nr_segs))
2875 return true;
2876 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2877 return true;
2878 }
2879 return false;
2880}
2881
2882static struct request *blk_mq_get_new_requests(struct request_queue *q,
2883 struct blk_plug *plug,
2884 struct bio *bio,
2885 unsigned int nsegs)
2886{
2887 struct blk_mq_alloc_data data = {
2888 .q = q,
2889 .nr_tags = 1,
2890 .cmd_flags = bio->bi_opf,
2891 };
2892 struct request *rq;
2893
2894 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2895 return NULL;
2896
2897 rq_qos_throttle(q, bio);
2898
2899 if (plug) {
2900 data.nr_tags = plug->nr_ios;
2901 plug->nr_ios = 1;
2902 data.cached_rq = &plug->cached_rq;
2903 }
2904
2905 rq = __blk_mq_alloc_requests(&data);
2906 if (rq)
2907 return rq;
2908 rq_qos_cleanup(q, bio);
2909 if (bio->bi_opf & REQ_NOWAIT)
2910 bio_wouldblock_error(bio);
2911 return NULL;
2912}
2913
2914/*
2915 * Check if we can use the passed on request for submitting the passed in bio,
2916 * and remove it from the request list if it can be used.
2917 */
2918static bool blk_mq_use_cached_rq(struct request *rq, struct blk_plug *plug,
2919 struct bio *bio)
2920{
2921 enum hctx_type type = blk_mq_get_hctx_type(bio->bi_opf);
2922 enum hctx_type hctx_type = rq->mq_hctx->type;
2923
2924 WARN_ON_ONCE(rq_list_peek(&plug->cached_rq) != rq);
2925
2926 if (type != hctx_type &&
2927 !(type == HCTX_TYPE_READ && hctx_type == HCTX_TYPE_DEFAULT))
2928 return false;
2929 if (op_is_flush(rq->cmd_flags) != op_is_flush(bio->bi_opf))
2930 return false;
2931
2932 /*
2933 * If any qos ->throttle() end up blocking, we will have flushed the
2934 * plug and hence killed the cached_rq list as well. Pop this entry
2935 * before we throttle.
2936 */
2937 plug->cached_rq = rq_list_next(rq);
2938 rq_qos_throttle(rq->q, bio);
2939
2940 blk_mq_rq_time_init(rq, 0);
2941 rq->cmd_flags = bio->bi_opf;
2942 INIT_LIST_HEAD(&rq->queuelist);
2943 return true;
2944}
2945
2946/**
2947 * blk_mq_submit_bio - Create and send a request to block device.
2948 * @bio: Bio pointer.
2949 *
2950 * Builds up a request structure from @q and @bio and send to the device. The
2951 * request may not be queued directly to hardware if:
2952 * * This request can be merged with another one
2953 * * We want to place request at plug queue for possible future merging
2954 * * There is an IO scheduler active at this queue
2955 *
2956 * It will not queue the request if there is an error with the bio, or at the
2957 * request creation.
2958 */
2959void blk_mq_submit_bio(struct bio *bio)
2960{
2961 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2962 struct blk_plug *plug = blk_mq_plug(bio);
2963 const int is_sync = op_is_sync(bio->bi_opf);
2964 struct blk_mq_hw_ctx *hctx;
2965 struct request *rq = NULL;
2966 unsigned int nr_segs = 1;
2967 blk_status_t ret;
2968
2969 bio = blk_queue_bounce(bio, q);
2970
2971 if (plug) {
2972 rq = rq_list_peek(&plug->cached_rq);
2973 if (rq && rq->q != q)
2974 rq = NULL;
2975 }
2976 if (rq) {
2977 if (unlikely(bio_may_exceed_limits(bio, &q->limits))) {
2978 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2979 if (!bio)
2980 return;
2981 }
2982 if (!bio_integrity_prep(bio))
2983 return;
2984 if (blk_mq_attempt_bio_merge(q, bio, nr_segs))
2985 return;
2986 if (blk_mq_use_cached_rq(rq, plug, bio))
2987 goto done;
2988 percpu_ref_get(&q->q_usage_counter);
2989 } else {
2990 if (unlikely(bio_queue_enter(bio)))
2991 return;
2992 if (unlikely(bio_may_exceed_limits(bio, &q->limits))) {
2993 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2994 if (!bio)
2995 goto fail;
2996 }
2997 if (!bio_integrity_prep(bio))
2998 goto fail;
2999 }
3000
3001 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
3002 if (unlikely(!rq)) {
3003fail:
3004 blk_queue_exit(q);
3005 return;
3006 }
3007
3008done:
3009 trace_block_getrq(bio);
3010
3011 rq_qos_track(q, rq, bio);
3012
3013 blk_mq_bio_to_request(rq, bio, nr_segs);
3014
3015 ret = blk_crypto_rq_get_keyslot(rq);
3016 if (ret != BLK_STS_OK) {
3017 bio->bi_status = ret;
3018 bio_endio(bio);
3019 blk_mq_free_request(rq);
3020 return;
3021 }
3022
3023 if (op_is_flush(bio->bi_opf) && blk_insert_flush(rq))
3024 return;
3025
3026 if (plug) {
3027 blk_add_rq_to_plug(plug, rq);
3028 return;
3029 }
3030
3031 hctx = rq->mq_hctx;
3032 if ((rq->rq_flags & RQF_USE_SCHED) ||
3033 (hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) {
3034 blk_mq_insert_request(rq, 0);
3035 blk_mq_run_hw_queue(hctx, true);
3036 } else {
3037 blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq));
3038 }
3039}
3040
3041#ifdef CONFIG_BLK_MQ_STACKING
3042/**
3043 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
3044 * @rq: the request being queued
3045 */
3046blk_status_t blk_insert_cloned_request(struct request *rq)
3047{
3048 struct request_queue *q = rq->q;
3049 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
3050 unsigned int max_segments = blk_rq_get_max_segments(rq);
3051 blk_status_t ret;
3052
3053 if (blk_rq_sectors(rq) > max_sectors) {
3054 /*
3055 * SCSI device does not have a good way to return if
3056 * Write Same/Zero is actually supported. If a device rejects
3057 * a non-read/write command (discard, write same,etc.) the
3058 * low-level device driver will set the relevant queue limit to
3059 * 0 to prevent blk-lib from issuing more of the offending
3060 * operations. Commands queued prior to the queue limit being
3061 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3062 * errors being propagated to upper layers.
3063 */
3064 if (max_sectors == 0)
3065 return BLK_STS_NOTSUPP;
3066
3067 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3068 __func__, blk_rq_sectors(rq), max_sectors);
3069 return BLK_STS_IOERR;
3070 }
3071
3072 /*
3073 * The queue settings related to segment counting may differ from the
3074 * original queue.
3075 */
3076 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
3077 if (rq->nr_phys_segments > max_segments) {
3078 printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n",
3079 __func__, rq->nr_phys_segments, max_segments);
3080 return BLK_STS_IOERR;
3081 }
3082
3083 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
3084 return BLK_STS_IOERR;
3085
3086 ret = blk_crypto_rq_get_keyslot(rq);
3087 if (ret != BLK_STS_OK)
3088 return ret;
3089
3090 blk_account_io_start(rq);
3091
3092 /*
3093 * Since we have a scheduler attached on the top device,
3094 * bypass a potential scheduler on the bottom device for
3095 * insert.
3096 */
3097 blk_mq_run_dispatch_ops(q,
3098 ret = blk_mq_request_issue_directly(rq, true));
3099 if (ret)
3100 blk_account_io_done(rq, ktime_get_ns());
3101 return ret;
3102}
3103EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3104
3105/**
3106 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3107 * @rq: the clone request to be cleaned up
3108 *
3109 * Description:
3110 * Free all bios in @rq for a cloned request.
3111 */
3112void blk_rq_unprep_clone(struct request *rq)
3113{
3114 struct bio *bio;
3115
3116 while ((bio = rq->bio) != NULL) {
3117 rq->bio = bio->bi_next;
3118
3119 bio_put(bio);
3120 }
3121}
3122EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3123
3124/**
3125 * blk_rq_prep_clone - Helper function to setup clone request
3126 * @rq: the request to be setup
3127 * @rq_src: original request to be cloned
3128 * @bs: bio_set that bios for clone are allocated from
3129 * @gfp_mask: memory allocation mask for bio
3130 * @bio_ctr: setup function to be called for each clone bio.
3131 * Returns %0 for success, non %0 for failure.
3132 * @data: private data to be passed to @bio_ctr
3133 *
3134 * Description:
3135 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3136 * Also, pages which the original bios are pointing to are not copied
3137 * and the cloned bios just point same pages.
3138 * So cloned bios must be completed before original bios, which means
3139 * the caller must complete @rq before @rq_src.
3140 */
3141int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3142 struct bio_set *bs, gfp_t gfp_mask,
3143 int (*bio_ctr)(struct bio *, struct bio *, void *),
3144 void *data)
3145{
3146 struct bio *bio, *bio_src;
3147
3148 if (!bs)
3149 bs = &fs_bio_set;
3150
3151 __rq_for_each_bio(bio_src, rq_src) {
3152 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3153 bs);
3154 if (!bio)
3155 goto free_and_out;
3156
3157 if (bio_ctr && bio_ctr(bio, bio_src, data))
3158 goto free_and_out;
3159
3160 if (rq->bio) {
3161 rq->biotail->bi_next = bio;
3162 rq->biotail = bio;
3163 } else {
3164 rq->bio = rq->biotail = bio;
3165 }
3166 bio = NULL;
3167 }
3168
3169 /* Copy attributes of the original request to the clone request. */
3170 rq->__sector = blk_rq_pos(rq_src);
3171 rq->__data_len = blk_rq_bytes(rq_src);
3172 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3173 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3174 rq->special_vec = rq_src->special_vec;
3175 }
3176 rq->nr_phys_segments = rq_src->nr_phys_segments;
3177 rq->ioprio = rq_src->ioprio;
3178
3179 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3180 goto free_and_out;
3181
3182 return 0;
3183
3184free_and_out:
3185 if (bio)
3186 bio_put(bio);
3187 blk_rq_unprep_clone(rq);
3188
3189 return -ENOMEM;
3190}
3191EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3192#endif /* CONFIG_BLK_MQ_STACKING */
3193
3194/*
3195 * Steal bios from a request and add them to a bio list.
3196 * The request must not have been partially completed before.
3197 */
3198void blk_steal_bios(struct bio_list *list, struct request *rq)
3199{
3200 if (rq->bio) {
3201 if (list->tail)
3202 list->tail->bi_next = rq->bio;
3203 else
3204 list->head = rq->bio;
3205 list->tail = rq->biotail;
3206
3207 rq->bio = NULL;
3208 rq->biotail = NULL;
3209 }
3210
3211 rq->__data_len = 0;
3212}
3213EXPORT_SYMBOL_GPL(blk_steal_bios);
3214
3215static size_t order_to_size(unsigned int order)
3216{
3217 return (size_t)PAGE_SIZE << order;
3218}
3219
3220/* called before freeing request pool in @tags */
3221static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3222 struct blk_mq_tags *tags)
3223{
3224 struct page *page;
3225 unsigned long flags;
3226
3227 /*
3228 * There is no need to clear mapping if driver tags is not initialized
3229 * or the mapping belongs to the driver tags.
3230 */
3231 if (!drv_tags || drv_tags == tags)
3232 return;
3233
3234 list_for_each_entry(page, &tags->page_list, lru) {
3235 unsigned long start = (unsigned long)page_address(page);
3236 unsigned long end = start + order_to_size(page->private);
3237 int i;
3238
3239 for (i = 0; i < drv_tags->nr_tags; i++) {
3240 struct request *rq = drv_tags->rqs[i];
3241 unsigned long rq_addr = (unsigned long)rq;
3242
3243 if (rq_addr >= start && rq_addr < end) {
3244 WARN_ON_ONCE(req_ref_read(rq) != 0);
3245 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3246 }
3247 }
3248 }
3249
3250 /*
3251 * Wait until all pending iteration is done.
3252 *
3253 * Request reference is cleared and it is guaranteed to be observed
3254 * after the ->lock is released.
3255 */
3256 spin_lock_irqsave(&drv_tags->lock, flags);
3257 spin_unlock_irqrestore(&drv_tags->lock, flags);
3258}
3259
3260void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3261 unsigned int hctx_idx)
3262{
3263 struct blk_mq_tags *drv_tags;
3264 struct page *page;
3265
3266 if (list_empty(&tags->page_list))
3267 return;
3268
3269 if (blk_mq_is_shared_tags(set->flags))
3270 drv_tags = set->shared_tags;
3271 else
3272 drv_tags = set->tags[hctx_idx];
3273
3274 if (tags->static_rqs && set->ops->exit_request) {
3275 int i;
3276
3277 for (i = 0; i < tags->nr_tags; i++) {
3278 struct request *rq = tags->static_rqs[i];
3279
3280 if (!rq)
3281 continue;
3282 set->ops->exit_request(set, rq, hctx_idx);
3283 tags->static_rqs[i] = NULL;
3284 }
3285 }
3286
3287 blk_mq_clear_rq_mapping(drv_tags, tags);
3288
3289 while (!list_empty(&tags->page_list)) {
3290 page = list_first_entry(&tags->page_list, struct page, lru);
3291 list_del_init(&page->lru);
3292 /*
3293 * Remove kmemleak object previously allocated in
3294 * blk_mq_alloc_rqs().
3295 */
3296 kmemleak_free(page_address(page));
3297 __free_pages(page, page->private);
3298 }
3299}
3300
3301void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3302{
3303 kfree(tags->rqs);
3304 tags->rqs = NULL;
3305 kfree(tags->static_rqs);
3306 tags->static_rqs = NULL;
3307
3308 blk_mq_free_tags(tags);
3309}
3310
3311static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3312 unsigned int hctx_idx)
3313{
3314 int i;
3315
3316 for (i = 0; i < set->nr_maps; i++) {
3317 unsigned int start = set->map[i].queue_offset;
3318 unsigned int end = start + set->map[i].nr_queues;
3319
3320 if (hctx_idx >= start && hctx_idx < end)
3321 break;
3322 }
3323
3324 if (i >= set->nr_maps)
3325 i = HCTX_TYPE_DEFAULT;
3326
3327 return i;
3328}
3329
3330static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3331 unsigned int hctx_idx)
3332{
3333 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3334
3335 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3336}
3337
3338static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3339 unsigned int hctx_idx,
3340 unsigned int nr_tags,
3341 unsigned int reserved_tags)
3342{
3343 int node = blk_mq_get_hctx_node(set, hctx_idx);
3344 struct blk_mq_tags *tags;
3345
3346 if (node == NUMA_NO_NODE)
3347 node = set->numa_node;
3348
3349 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3350 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3351 if (!tags)
3352 return NULL;
3353
3354 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3355 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3356 node);
3357 if (!tags->rqs)
3358 goto err_free_tags;
3359
3360 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3361 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3362 node);
3363 if (!tags->static_rqs)
3364 goto err_free_rqs;
3365
3366 return tags;
3367
3368err_free_rqs:
3369 kfree(tags->rqs);
3370err_free_tags:
3371 blk_mq_free_tags(tags);
3372 return NULL;
3373}
3374
3375static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3376 unsigned int hctx_idx, int node)
3377{
3378 int ret;
3379
3380 if (set->ops->init_request) {
3381 ret = set->ops->init_request(set, rq, hctx_idx, node);
3382 if (ret)
3383 return ret;
3384 }
3385
3386 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3387 return 0;
3388}
3389
3390static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3391 struct blk_mq_tags *tags,
3392 unsigned int hctx_idx, unsigned int depth)
3393{
3394 unsigned int i, j, entries_per_page, max_order = 4;
3395 int node = blk_mq_get_hctx_node(set, hctx_idx);
3396 size_t rq_size, left;
3397
3398 if (node == NUMA_NO_NODE)
3399 node = set->numa_node;
3400
3401 INIT_LIST_HEAD(&tags->page_list);
3402
3403 /*
3404 * rq_size is the size of the request plus driver payload, rounded
3405 * to the cacheline size
3406 */
3407 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3408 cache_line_size());
3409 left = rq_size * depth;
3410
3411 for (i = 0; i < depth; ) {
3412 int this_order = max_order;
3413 struct page *page;
3414 int to_do;
3415 void *p;
3416
3417 while (this_order && left < order_to_size(this_order - 1))
3418 this_order--;
3419
3420 do {
3421 page = alloc_pages_node(node,
3422 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3423 this_order);
3424 if (page)
3425 break;
3426 if (!this_order--)
3427 break;
3428 if (order_to_size(this_order) < rq_size)
3429 break;
3430 } while (1);
3431
3432 if (!page)
3433 goto fail;
3434
3435 page->private = this_order;
3436 list_add_tail(&page->lru, &tags->page_list);
3437
3438 p = page_address(page);
3439 /*
3440 * Allow kmemleak to scan these pages as they contain pointers
3441 * to additional allocations like via ops->init_request().
3442 */
3443 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3444 entries_per_page = order_to_size(this_order) / rq_size;
3445 to_do = min(entries_per_page, depth - i);
3446 left -= to_do * rq_size;
3447 for (j = 0; j < to_do; j++) {
3448 struct request *rq = p;
3449
3450 tags->static_rqs[i] = rq;
3451 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3452 tags->static_rqs[i] = NULL;
3453 goto fail;
3454 }
3455
3456 p += rq_size;
3457 i++;
3458 }
3459 }
3460 return 0;
3461
3462fail:
3463 blk_mq_free_rqs(set, tags, hctx_idx);
3464 return -ENOMEM;
3465}
3466
3467struct rq_iter_data {
3468 struct blk_mq_hw_ctx *hctx;
3469 bool has_rq;
3470};
3471
3472static bool blk_mq_has_request(struct request *rq, void *data)
3473{
3474 struct rq_iter_data *iter_data = data;
3475
3476 if (rq->mq_hctx != iter_data->hctx)
3477 return true;
3478 iter_data->has_rq = true;
3479 return false;
3480}
3481
3482static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3483{
3484 struct blk_mq_tags *tags = hctx->sched_tags ?
3485 hctx->sched_tags : hctx->tags;
3486 struct rq_iter_data data = {
3487 .hctx = hctx,
3488 };
3489
3490 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3491 return data.has_rq;
3492}
3493
3494static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3495 struct blk_mq_hw_ctx *hctx)
3496{
3497 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3498 return false;
3499 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3500 return false;
3501 return true;
3502}
3503
3504static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3505{
3506 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3507 struct blk_mq_hw_ctx, cpuhp_online);
3508
3509 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3510 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3511 return 0;
3512
3513 /*
3514 * Prevent new request from being allocated on the current hctx.
3515 *
3516 * The smp_mb__after_atomic() Pairs with the implied barrier in
3517 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3518 * seen once we return from the tag allocator.
3519 */
3520 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3521 smp_mb__after_atomic();
3522
3523 /*
3524 * Try to grab a reference to the queue and wait for any outstanding
3525 * requests. If we could not grab a reference the queue has been
3526 * frozen and there are no requests.
3527 */
3528 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3529 while (blk_mq_hctx_has_requests(hctx))
3530 msleep(5);
3531 percpu_ref_put(&hctx->queue->q_usage_counter);
3532 }
3533
3534 return 0;
3535}
3536
3537static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3538{
3539 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3540 struct blk_mq_hw_ctx, cpuhp_online);
3541
3542 if (cpumask_test_cpu(cpu, hctx->cpumask))
3543 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3544 return 0;
3545}
3546
3547/*
3548 * 'cpu' is going away. splice any existing rq_list entries from this
3549 * software queue to the hw queue dispatch list, and ensure that it
3550 * gets run.
3551 */
3552static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3553{
3554 struct blk_mq_hw_ctx *hctx;
3555 struct blk_mq_ctx *ctx;
3556 LIST_HEAD(tmp);
3557 enum hctx_type type;
3558
3559 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3560 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3561 return 0;
3562
3563 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3564 type = hctx->type;
3565
3566 spin_lock(&ctx->lock);
3567 if (!list_empty(&ctx->rq_lists[type])) {
3568 list_splice_init(&ctx->rq_lists[type], &tmp);
3569 blk_mq_hctx_clear_pending(hctx, ctx);
3570 }
3571 spin_unlock(&ctx->lock);
3572
3573 if (list_empty(&tmp))
3574 return 0;
3575
3576 spin_lock(&hctx->lock);
3577 list_splice_tail_init(&tmp, &hctx->dispatch);
3578 spin_unlock(&hctx->lock);
3579
3580 blk_mq_run_hw_queue(hctx, true);
3581 return 0;
3582}
3583
3584static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3585{
3586 if (!(hctx->flags & BLK_MQ_F_STACKING))
3587 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3588 &hctx->cpuhp_online);
3589 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3590 &hctx->cpuhp_dead);
3591}
3592
3593/*
3594 * Before freeing hw queue, clearing the flush request reference in
3595 * tags->rqs[] for avoiding potential UAF.
3596 */
3597static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3598 unsigned int queue_depth, struct request *flush_rq)
3599{
3600 int i;
3601 unsigned long flags;
3602
3603 /* The hw queue may not be mapped yet */
3604 if (!tags)
3605 return;
3606
3607 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3608
3609 for (i = 0; i < queue_depth; i++)
3610 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3611
3612 /*
3613 * Wait until all pending iteration is done.
3614 *
3615 * Request reference is cleared and it is guaranteed to be observed
3616 * after the ->lock is released.
3617 */
3618 spin_lock_irqsave(&tags->lock, flags);
3619 spin_unlock_irqrestore(&tags->lock, flags);
3620}
3621
3622/* hctx->ctxs will be freed in queue's release handler */
3623static void blk_mq_exit_hctx(struct request_queue *q,
3624 struct blk_mq_tag_set *set,
3625 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3626{
3627 struct request *flush_rq = hctx->fq->flush_rq;
3628
3629 if (blk_mq_hw_queue_mapped(hctx))
3630 blk_mq_tag_idle(hctx);
3631
3632 if (blk_queue_init_done(q))
3633 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3634 set->queue_depth, flush_rq);
3635 if (set->ops->exit_request)
3636 set->ops->exit_request(set, flush_rq, hctx_idx);
3637
3638 if (set->ops->exit_hctx)
3639 set->ops->exit_hctx(hctx, hctx_idx);
3640
3641 blk_mq_remove_cpuhp(hctx);
3642
3643 xa_erase(&q->hctx_table, hctx_idx);
3644
3645 spin_lock(&q->unused_hctx_lock);
3646 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3647 spin_unlock(&q->unused_hctx_lock);
3648}
3649
3650static void blk_mq_exit_hw_queues(struct request_queue *q,
3651 struct blk_mq_tag_set *set, int nr_queue)
3652{
3653 struct blk_mq_hw_ctx *hctx;
3654 unsigned long i;
3655
3656 queue_for_each_hw_ctx(q, hctx, i) {
3657 if (i == nr_queue)
3658 break;
3659 blk_mq_exit_hctx(q, set, hctx, i);
3660 }
3661}
3662
3663static int blk_mq_init_hctx(struct request_queue *q,
3664 struct blk_mq_tag_set *set,
3665 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3666{
3667 hctx->queue_num = hctx_idx;
3668
3669 if (!(hctx->flags & BLK_MQ_F_STACKING))
3670 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3671 &hctx->cpuhp_online);
3672 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3673
3674 hctx->tags = set->tags[hctx_idx];
3675
3676 if (set->ops->init_hctx &&
3677 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3678 goto unregister_cpu_notifier;
3679
3680 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3681 hctx->numa_node))
3682 goto exit_hctx;
3683
3684 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3685 goto exit_flush_rq;
3686
3687 return 0;
3688
3689 exit_flush_rq:
3690 if (set->ops->exit_request)
3691 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3692 exit_hctx:
3693 if (set->ops->exit_hctx)
3694 set->ops->exit_hctx(hctx, hctx_idx);
3695 unregister_cpu_notifier:
3696 blk_mq_remove_cpuhp(hctx);
3697 return -1;
3698}
3699
3700static struct blk_mq_hw_ctx *
3701blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3702 int node)
3703{
3704 struct blk_mq_hw_ctx *hctx;
3705 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3706
3707 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3708 if (!hctx)
3709 goto fail_alloc_hctx;
3710
3711 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3712 goto free_hctx;
3713
3714 atomic_set(&hctx->nr_active, 0);
3715 if (node == NUMA_NO_NODE)
3716 node = set->numa_node;
3717 hctx->numa_node = node;
3718
3719 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3720 spin_lock_init(&hctx->lock);
3721 INIT_LIST_HEAD(&hctx->dispatch);
3722 hctx->queue = q;
3723 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3724
3725 INIT_LIST_HEAD(&hctx->hctx_list);
3726
3727 /*
3728 * Allocate space for all possible cpus to avoid allocation at
3729 * runtime
3730 */
3731 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3732 gfp, node);
3733 if (!hctx->ctxs)
3734 goto free_cpumask;
3735
3736 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3737 gfp, node, false, false))
3738 goto free_ctxs;
3739 hctx->nr_ctx = 0;
3740
3741 spin_lock_init(&hctx->dispatch_wait_lock);
3742 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3743 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3744
3745 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3746 if (!hctx->fq)
3747 goto free_bitmap;
3748
3749 blk_mq_hctx_kobj_init(hctx);
3750
3751 return hctx;
3752
3753 free_bitmap:
3754 sbitmap_free(&hctx->ctx_map);
3755 free_ctxs:
3756 kfree(hctx->ctxs);
3757 free_cpumask:
3758 free_cpumask_var(hctx->cpumask);
3759 free_hctx:
3760 kfree(hctx);
3761 fail_alloc_hctx:
3762 return NULL;
3763}
3764
3765static void blk_mq_init_cpu_queues(struct request_queue *q,
3766 unsigned int nr_hw_queues)
3767{
3768 struct blk_mq_tag_set *set = q->tag_set;
3769 unsigned int i, j;
3770
3771 for_each_possible_cpu(i) {
3772 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3773 struct blk_mq_hw_ctx *hctx;
3774 int k;
3775
3776 __ctx->cpu = i;
3777 spin_lock_init(&__ctx->lock);
3778 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3779 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3780
3781 __ctx->queue = q;
3782
3783 /*
3784 * Set local node, IFF we have more than one hw queue. If
3785 * not, we remain on the home node of the device
3786 */
3787 for (j = 0; j < set->nr_maps; j++) {
3788 hctx = blk_mq_map_queue_type(q, j, i);
3789 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3790 hctx->numa_node = cpu_to_node(i);
3791 }
3792 }
3793}
3794
3795struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3796 unsigned int hctx_idx,
3797 unsigned int depth)
3798{
3799 struct blk_mq_tags *tags;
3800 int ret;
3801
3802 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3803 if (!tags)
3804 return NULL;
3805
3806 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3807 if (ret) {
3808 blk_mq_free_rq_map(tags);
3809 return NULL;
3810 }
3811
3812 return tags;
3813}
3814
3815static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3816 int hctx_idx)
3817{
3818 if (blk_mq_is_shared_tags(set->flags)) {
3819 set->tags[hctx_idx] = set->shared_tags;
3820
3821 return true;
3822 }
3823
3824 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3825 set->queue_depth);
3826
3827 return set->tags[hctx_idx];
3828}
3829
3830void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3831 struct blk_mq_tags *tags,
3832 unsigned int hctx_idx)
3833{
3834 if (tags) {
3835 blk_mq_free_rqs(set, tags, hctx_idx);
3836 blk_mq_free_rq_map(tags);
3837 }
3838}
3839
3840static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3841 unsigned int hctx_idx)
3842{
3843 if (!blk_mq_is_shared_tags(set->flags))
3844 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3845
3846 set->tags[hctx_idx] = NULL;
3847}
3848
3849static void blk_mq_map_swqueue(struct request_queue *q)
3850{
3851 unsigned int j, hctx_idx;
3852 unsigned long i;
3853 struct blk_mq_hw_ctx *hctx;
3854 struct blk_mq_ctx *ctx;
3855 struct blk_mq_tag_set *set = q->tag_set;
3856
3857 queue_for_each_hw_ctx(q, hctx, i) {
3858 cpumask_clear(hctx->cpumask);
3859 hctx->nr_ctx = 0;
3860 hctx->dispatch_from = NULL;
3861 }
3862
3863 /*
3864 * Map software to hardware queues.
3865 *
3866 * If the cpu isn't present, the cpu is mapped to first hctx.
3867 */
3868 for_each_possible_cpu(i) {
3869
3870 ctx = per_cpu_ptr(q->queue_ctx, i);
3871 for (j = 0; j < set->nr_maps; j++) {
3872 if (!set->map[j].nr_queues) {
3873 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3874 HCTX_TYPE_DEFAULT, i);
3875 continue;
3876 }
3877 hctx_idx = set->map[j].mq_map[i];
3878 /* unmapped hw queue can be remapped after CPU topo changed */
3879 if (!set->tags[hctx_idx] &&
3880 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3881 /*
3882 * If tags initialization fail for some hctx,
3883 * that hctx won't be brought online. In this
3884 * case, remap the current ctx to hctx[0] which
3885 * is guaranteed to always have tags allocated
3886 */
3887 set->map[j].mq_map[i] = 0;
3888 }
3889
3890 hctx = blk_mq_map_queue_type(q, j, i);
3891 ctx->hctxs[j] = hctx;
3892 /*
3893 * If the CPU is already set in the mask, then we've
3894 * mapped this one already. This can happen if
3895 * devices share queues across queue maps.
3896 */
3897 if (cpumask_test_cpu(i, hctx->cpumask))
3898 continue;
3899
3900 cpumask_set_cpu(i, hctx->cpumask);
3901 hctx->type = j;
3902 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3903 hctx->ctxs[hctx->nr_ctx++] = ctx;
3904
3905 /*
3906 * If the nr_ctx type overflows, we have exceeded the
3907 * amount of sw queues we can support.
3908 */
3909 BUG_ON(!hctx->nr_ctx);
3910 }
3911
3912 for (; j < HCTX_MAX_TYPES; j++)
3913 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3914 HCTX_TYPE_DEFAULT, i);
3915 }
3916
3917 queue_for_each_hw_ctx(q, hctx, i) {
3918 /*
3919 * If no software queues are mapped to this hardware queue,
3920 * disable it and free the request entries.
3921 */
3922 if (!hctx->nr_ctx) {
3923 /* Never unmap queue 0. We need it as a
3924 * fallback in case of a new remap fails
3925 * allocation
3926 */
3927 if (i)
3928 __blk_mq_free_map_and_rqs(set, i);
3929
3930 hctx->tags = NULL;
3931 continue;
3932 }
3933
3934 hctx->tags = set->tags[i];
3935 WARN_ON(!hctx->tags);
3936
3937 /*
3938 * Set the map size to the number of mapped software queues.
3939 * This is more accurate and more efficient than looping
3940 * over all possibly mapped software queues.
3941 */
3942 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3943
3944 /*
3945 * Initialize batch roundrobin counts
3946 */
3947 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3948 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3949 }
3950}
3951
3952/*
3953 * Caller needs to ensure that we're either frozen/quiesced, or that
3954 * the queue isn't live yet.
3955 */
3956static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3957{
3958 struct blk_mq_hw_ctx *hctx;
3959 unsigned long i;
3960
3961 queue_for_each_hw_ctx(q, hctx, i) {
3962 if (shared) {
3963 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3964 } else {
3965 blk_mq_tag_idle(hctx);
3966 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3967 }
3968 }
3969}
3970
3971static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3972 bool shared)
3973{
3974 struct request_queue *q;
3975
3976 lockdep_assert_held(&set->tag_list_lock);
3977
3978 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3979 blk_mq_freeze_queue(q);
3980 queue_set_hctx_shared(q, shared);
3981 blk_mq_unfreeze_queue(q);
3982 }
3983}
3984
3985static void blk_mq_del_queue_tag_set(struct request_queue *q)
3986{
3987 struct blk_mq_tag_set *set = q->tag_set;
3988
3989 mutex_lock(&set->tag_list_lock);
3990 list_del(&q->tag_set_list);
3991 if (list_is_singular(&set->tag_list)) {
3992 /* just transitioned to unshared */
3993 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3994 /* update existing queue */
3995 blk_mq_update_tag_set_shared(set, false);
3996 }
3997 mutex_unlock(&set->tag_list_lock);
3998 INIT_LIST_HEAD(&q->tag_set_list);
3999}
4000
4001static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
4002 struct request_queue *q)
4003{
4004 mutex_lock(&set->tag_list_lock);
4005
4006 /*
4007 * Check to see if we're transitioning to shared (from 1 to 2 queues).
4008 */
4009 if (!list_empty(&set->tag_list) &&
4010 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
4011 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
4012 /* update existing queue */
4013 blk_mq_update_tag_set_shared(set, true);
4014 }
4015 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
4016 queue_set_hctx_shared(q, true);
4017 list_add_tail(&q->tag_set_list, &set->tag_list);
4018
4019 mutex_unlock(&set->tag_list_lock);
4020}
4021
4022/* All allocations will be freed in release handler of q->mq_kobj */
4023static int blk_mq_alloc_ctxs(struct request_queue *q)
4024{
4025 struct blk_mq_ctxs *ctxs;
4026 int cpu;
4027
4028 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
4029 if (!ctxs)
4030 return -ENOMEM;
4031
4032 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
4033 if (!ctxs->queue_ctx)
4034 goto fail;
4035
4036 for_each_possible_cpu(cpu) {
4037 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
4038 ctx->ctxs = ctxs;
4039 }
4040
4041 q->mq_kobj = &ctxs->kobj;
4042 q->queue_ctx = ctxs->queue_ctx;
4043
4044 return 0;
4045 fail:
4046 kfree(ctxs);
4047 return -ENOMEM;
4048}
4049
4050/*
4051 * It is the actual release handler for mq, but we do it from
4052 * request queue's release handler for avoiding use-after-free
4053 * and headache because q->mq_kobj shouldn't have been introduced,
4054 * but we can't group ctx/kctx kobj without it.
4055 */
4056void blk_mq_release(struct request_queue *q)
4057{
4058 struct blk_mq_hw_ctx *hctx, *next;
4059 unsigned long i;
4060
4061 queue_for_each_hw_ctx(q, hctx, i)
4062 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4063
4064 /* all hctx are in .unused_hctx_list now */
4065 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4066 list_del_init(&hctx->hctx_list);
4067 kobject_put(&hctx->kobj);
4068 }
4069
4070 xa_destroy(&q->hctx_table);
4071
4072 /*
4073 * release .mq_kobj and sw queue's kobject now because
4074 * both share lifetime with request queue.
4075 */
4076 blk_mq_sysfs_deinit(q);
4077}
4078
4079static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
4080 void *queuedata)
4081{
4082 struct request_queue *q;
4083 int ret;
4084
4085 q = blk_alloc_queue(set->numa_node);
4086 if (!q)
4087 return ERR_PTR(-ENOMEM);
4088 q->queuedata = queuedata;
4089 ret = blk_mq_init_allocated_queue(set, q);
4090 if (ret) {
4091 blk_put_queue(q);
4092 return ERR_PTR(ret);
4093 }
4094 return q;
4095}
4096
4097struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
4098{
4099 return blk_mq_init_queue_data(set, NULL);
4100}
4101EXPORT_SYMBOL(blk_mq_init_queue);
4102
4103/**
4104 * blk_mq_destroy_queue - shutdown a request queue
4105 * @q: request queue to shutdown
4106 *
4107 * This shuts down a request queue allocated by blk_mq_init_queue(). All future
4108 * requests will be failed with -ENODEV. The caller is responsible for dropping
4109 * the reference from blk_mq_init_queue() by calling blk_put_queue().
4110 *
4111 * Context: can sleep
4112 */
4113void blk_mq_destroy_queue(struct request_queue *q)
4114{
4115 WARN_ON_ONCE(!queue_is_mq(q));
4116 WARN_ON_ONCE(blk_queue_registered(q));
4117
4118 might_sleep();
4119
4120 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4121 blk_queue_start_drain(q);
4122 blk_mq_freeze_queue_wait(q);
4123
4124 blk_sync_queue(q);
4125 blk_mq_cancel_work_sync(q);
4126 blk_mq_exit_queue(q);
4127}
4128EXPORT_SYMBOL(blk_mq_destroy_queue);
4129
4130struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
4131 struct lock_class_key *lkclass)
4132{
4133 struct request_queue *q;
4134 struct gendisk *disk;
4135
4136 q = blk_mq_init_queue_data(set, queuedata);
4137 if (IS_ERR(q))
4138 return ERR_CAST(q);
4139
4140 disk = __alloc_disk_node(q, set->numa_node, lkclass);
4141 if (!disk) {
4142 blk_mq_destroy_queue(q);
4143 blk_put_queue(q);
4144 return ERR_PTR(-ENOMEM);
4145 }
4146 set_bit(GD_OWNS_QUEUE, &disk->state);
4147 return disk;
4148}
4149EXPORT_SYMBOL(__blk_mq_alloc_disk);
4150
4151struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4152 struct lock_class_key *lkclass)
4153{
4154 struct gendisk *disk;
4155
4156 if (!blk_get_queue(q))
4157 return NULL;
4158 disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4159 if (!disk)
4160 blk_put_queue(q);
4161 return disk;
4162}
4163EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4164
4165static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4166 struct blk_mq_tag_set *set, struct request_queue *q,
4167 int hctx_idx, int node)
4168{
4169 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4170
4171 /* reuse dead hctx first */
4172 spin_lock(&q->unused_hctx_lock);
4173 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4174 if (tmp->numa_node == node) {
4175 hctx = tmp;
4176 break;
4177 }
4178 }
4179 if (hctx)
4180 list_del_init(&hctx->hctx_list);
4181 spin_unlock(&q->unused_hctx_lock);
4182
4183 if (!hctx)
4184 hctx = blk_mq_alloc_hctx(q, set, node);
4185 if (!hctx)
4186 goto fail;
4187
4188 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4189 goto free_hctx;
4190
4191 return hctx;
4192
4193 free_hctx:
4194 kobject_put(&hctx->kobj);
4195 fail:
4196 return NULL;
4197}
4198
4199static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4200 struct request_queue *q)
4201{
4202 struct blk_mq_hw_ctx *hctx;
4203 unsigned long i, j;
4204
4205 /* protect against switching io scheduler */
4206 mutex_lock(&q->sysfs_lock);
4207 for (i = 0; i < set->nr_hw_queues; i++) {
4208 int old_node;
4209 int node = blk_mq_get_hctx_node(set, i);
4210 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4211
4212 if (old_hctx) {
4213 old_node = old_hctx->numa_node;
4214 blk_mq_exit_hctx(q, set, old_hctx, i);
4215 }
4216
4217 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4218 if (!old_hctx)
4219 break;
4220 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4221 node, old_node);
4222 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4223 WARN_ON_ONCE(!hctx);
4224 }
4225 }
4226 /*
4227 * Increasing nr_hw_queues fails. Free the newly allocated
4228 * hctxs and keep the previous q->nr_hw_queues.
4229 */
4230 if (i != set->nr_hw_queues) {
4231 j = q->nr_hw_queues;
4232 } else {
4233 j = i;
4234 q->nr_hw_queues = set->nr_hw_queues;
4235 }
4236
4237 xa_for_each_start(&q->hctx_table, j, hctx, j)
4238 blk_mq_exit_hctx(q, set, hctx, j);
4239 mutex_unlock(&q->sysfs_lock);
4240}
4241
4242static void blk_mq_update_poll_flag(struct request_queue *q)
4243{
4244 struct blk_mq_tag_set *set = q->tag_set;
4245
4246 if (set->nr_maps > HCTX_TYPE_POLL &&
4247 set->map[HCTX_TYPE_POLL].nr_queues)
4248 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4249 else
4250 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4251}
4252
4253int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4254 struct request_queue *q)
4255{
4256 /* mark the queue as mq asap */
4257 q->mq_ops = set->ops;
4258
4259 if (blk_mq_alloc_ctxs(q))
4260 goto err_exit;
4261
4262 /* init q->mq_kobj and sw queues' kobjects */
4263 blk_mq_sysfs_init(q);
4264
4265 INIT_LIST_HEAD(&q->unused_hctx_list);
4266 spin_lock_init(&q->unused_hctx_lock);
4267
4268 xa_init(&q->hctx_table);
4269
4270 blk_mq_realloc_hw_ctxs(set, q);
4271 if (!q->nr_hw_queues)
4272 goto err_hctxs;
4273
4274 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4275 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4276
4277 q->tag_set = set;
4278
4279 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4280 blk_mq_update_poll_flag(q);
4281
4282 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4283 INIT_LIST_HEAD(&q->flush_list);
4284 INIT_LIST_HEAD(&q->requeue_list);
4285 spin_lock_init(&q->requeue_lock);
4286
4287 q->nr_requests = set->queue_depth;
4288
4289 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4290 blk_mq_add_queue_tag_set(set, q);
4291 blk_mq_map_swqueue(q);
4292 return 0;
4293
4294err_hctxs:
4295 blk_mq_release(q);
4296err_exit:
4297 q->mq_ops = NULL;
4298 return -ENOMEM;
4299}
4300EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4301
4302/* tags can _not_ be used after returning from blk_mq_exit_queue */
4303void blk_mq_exit_queue(struct request_queue *q)
4304{
4305 struct blk_mq_tag_set *set = q->tag_set;
4306
4307 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4308 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4309 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4310 blk_mq_del_queue_tag_set(q);
4311}
4312
4313static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4314{
4315 int i;
4316
4317 if (blk_mq_is_shared_tags(set->flags)) {
4318 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4319 BLK_MQ_NO_HCTX_IDX,
4320 set->queue_depth);
4321 if (!set->shared_tags)
4322 return -ENOMEM;
4323 }
4324
4325 for (i = 0; i < set->nr_hw_queues; i++) {
4326 if (!__blk_mq_alloc_map_and_rqs(set, i))
4327 goto out_unwind;
4328 cond_resched();
4329 }
4330
4331 return 0;
4332
4333out_unwind:
4334 while (--i >= 0)
4335 __blk_mq_free_map_and_rqs(set, i);
4336
4337 if (blk_mq_is_shared_tags(set->flags)) {
4338 blk_mq_free_map_and_rqs(set, set->shared_tags,
4339 BLK_MQ_NO_HCTX_IDX);
4340 }
4341
4342 return -ENOMEM;
4343}
4344
4345/*
4346 * Allocate the request maps associated with this tag_set. Note that this
4347 * may reduce the depth asked for, if memory is tight. set->queue_depth
4348 * will be updated to reflect the allocated depth.
4349 */
4350static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4351{
4352 unsigned int depth;
4353 int err;
4354
4355 depth = set->queue_depth;
4356 do {
4357 err = __blk_mq_alloc_rq_maps(set);
4358 if (!err)
4359 break;
4360
4361 set->queue_depth >>= 1;
4362 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4363 err = -ENOMEM;
4364 break;
4365 }
4366 } while (set->queue_depth);
4367
4368 if (!set->queue_depth || err) {
4369 pr_err("blk-mq: failed to allocate request map\n");
4370 return -ENOMEM;
4371 }
4372
4373 if (depth != set->queue_depth)
4374 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4375 depth, set->queue_depth);
4376
4377 return 0;
4378}
4379
4380static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4381{
4382 /*
4383 * blk_mq_map_queues() and multiple .map_queues() implementations
4384 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4385 * number of hardware queues.
4386 */
4387 if (set->nr_maps == 1)
4388 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4389
4390 if (set->ops->map_queues && !is_kdump_kernel()) {
4391 int i;
4392
4393 /*
4394 * transport .map_queues is usually done in the following
4395 * way:
4396 *
4397 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4398 * mask = get_cpu_mask(queue)
4399 * for_each_cpu(cpu, mask)
4400 * set->map[x].mq_map[cpu] = queue;
4401 * }
4402 *
4403 * When we need to remap, the table has to be cleared for
4404 * killing stale mapping since one CPU may not be mapped
4405 * to any hw queue.
4406 */
4407 for (i = 0; i < set->nr_maps; i++)
4408 blk_mq_clear_mq_map(&set->map[i]);
4409
4410 set->ops->map_queues(set);
4411 } else {
4412 BUG_ON(set->nr_maps > 1);
4413 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4414 }
4415}
4416
4417static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4418 int new_nr_hw_queues)
4419{
4420 struct blk_mq_tags **new_tags;
4421 int i;
4422
4423 if (set->nr_hw_queues >= new_nr_hw_queues)
4424 goto done;
4425
4426 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4427 GFP_KERNEL, set->numa_node);
4428 if (!new_tags)
4429 return -ENOMEM;
4430
4431 if (set->tags)
4432 memcpy(new_tags, set->tags, set->nr_hw_queues *
4433 sizeof(*set->tags));
4434 kfree(set->tags);
4435 set->tags = new_tags;
4436
4437 for (i = set->nr_hw_queues; i < new_nr_hw_queues; i++) {
4438 if (!__blk_mq_alloc_map_and_rqs(set, i)) {
4439 while (--i >= set->nr_hw_queues)
4440 __blk_mq_free_map_and_rqs(set, i);
4441 return -ENOMEM;
4442 }
4443 cond_resched();
4444 }
4445
4446done:
4447 set->nr_hw_queues = new_nr_hw_queues;
4448 return 0;
4449}
4450
4451/*
4452 * Alloc a tag set to be associated with one or more request queues.
4453 * May fail with EINVAL for various error conditions. May adjust the
4454 * requested depth down, if it's too large. In that case, the set
4455 * value will be stored in set->queue_depth.
4456 */
4457int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4458{
4459 int i, ret;
4460
4461 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4462
4463 if (!set->nr_hw_queues)
4464 return -EINVAL;
4465 if (!set->queue_depth)
4466 return -EINVAL;
4467 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4468 return -EINVAL;
4469
4470 if (!set->ops->queue_rq)
4471 return -EINVAL;
4472
4473 if (!set->ops->get_budget ^ !set->ops->put_budget)
4474 return -EINVAL;
4475
4476 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4477 pr_info("blk-mq: reduced tag depth to %u\n",
4478 BLK_MQ_MAX_DEPTH);
4479 set->queue_depth = BLK_MQ_MAX_DEPTH;
4480 }
4481
4482 if (!set->nr_maps)
4483 set->nr_maps = 1;
4484 else if (set->nr_maps > HCTX_MAX_TYPES)
4485 return -EINVAL;
4486
4487 /*
4488 * If a crashdump is active, then we are potentially in a very
4489 * memory constrained environment. Limit us to 1 queue and
4490 * 64 tags to prevent using too much memory.
4491 */
4492 if (is_kdump_kernel()) {
4493 set->nr_hw_queues = 1;
4494 set->nr_maps = 1;
4495 set->queue_depth = min(64U, set->queue_depth);
4496 }
4497 /*
4498 * There is no use for more h/w queues than cpus if we just have
4499 * a single map
4500 */
4501 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4502 set->nr_hw_queues = nr_cpu_ids;
4503
4504 if (set->flags & BLK_MQ_F_BLOCKING) {
4505 set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL);
4506 if (!set->srcu)
4507 return -ENOMEM;
4508 ret = init_srcu_struct(set->srcu);
4509 if (ret)
4510 goto out_free_srcu;
4511 }
4512
4513 ret = -ENOMEM;
4514 set->tags = kcalloc_node(set->nr_hw_queues,
4515 sizeof(struct blk_mq_tags *), GFP_KERNEL,
4516 set->numa_node);
4517 if (!set->tags)
4518 goto out_cleanup_srcu;
4519
4520 for (i = 0; i < set->nr_maps; i++) {
4521 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4522 sizeof(set->map[i].mq_map[0]),
4523 GFP_KERNEL, set->numa_node);
4524 if (!set->map[i].mq_map)
4525 goto out_free_mq_map;
4526 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4527 }
4528
4529 blk_mq_update_queue_map(set);
4530
4531 ret = blk_mq_alloc_set_map_and_rqs(set);
4532 if (ret)
4533 goto out_free_mq_map;
4534
4535 mutex_init(&set->tag_list_lock);
4536 INIT_LIST_HEAD(&set->tag_list);
4537
4538 return 0;
4539
4540out_free_mq_map:
4541 for (i = 0; i < set->nr_maps; i++) {
4542 kfree(set->map[i].mq_map);
4543 set->map[i].mq_map = NULL;
4544 }
4545 kfree(set->tags);
4546 set->tags = NULL;
4547out_cleanup_srcu:
4548 if (set->flags & BLK_MQ_F_BLOCKING)
4549 cleanup_srcu_struct(set->srcu);
4550out_free_srcu:
4551 if (set->flags & BLK_MQ_F_BLOCKING)
4552 kfree(set->srcu);
4553 return ret;
4554}
4555EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4556
4557/* allocate and initialize a tagset for a simple single-queue device */
4558int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4559 const struct blk_mq_ops *ops, unsigned int queue_depth,
4560 unsigned int set_flags)
4561{
4562 memset(set, 0, sizeof(*set));
4563 set->ops = ops;
4564 set->nr_hw_queues = 1;
4565 set->nr_maps = 1;
4566 set->queue_depth = queue_depth;
4567 set->numa_node = NUMA_NO_NODE;
4568 set->flags = set_flags;
4569 return blk_mq_alloc_tag_set(set);
4570}
4571EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4572
4573void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4574{
4575 int i, j;
4576
4577 for (i = 0; i < set->nr_hw_queues; i++)
4578 __blk_mq_free_map_and_rqs(set, i);
4579
4580 if (blk_mq_is_shared_tags(set->flags)) {
4581 blk_mq_free_map_and_rqs(set, set->shared_tags,
4582 BLK_MQ_NO_HCTX_IDX);
4583 }
4584
4585 for (j = 0; j < set->nr_maps; j++) {
4586 kfree(set->map[j].mq_map);
4587 set->map[j].mq_map = NULL;
4588 }
4589
4590 kfree(set->tags);
4591 set->tags = NULL;
4592 if (set->flags & BLK_MQ_F_BLOCKING) {
4593 cleanup_srcu_struct(set->srcu);
4594 kfree(set->srcu);
4595 }
4596}
4597EXPORT_SYMBOL(blk_mq_free_tag_set);
4598
4599int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4600{
4601 struct blk_mq_tag_set *set = q->tag_set;
4602 struct blk_mq_hw_ctx *hctx;
4603 int ret;
4604 unsigned long i;
4605
4606 if (!set)
4607 return -EINVAL;
4608
4609 if (q->nr_requests == nr)
4610 return 0;
4611
4612 blk_mq_freeze_queue(q);
4613 blk_mq_quiesce_queue(q);
4614
4615 ret = 0;
4616 queue_for_each_hw_ctx(q, hctx, i) {
4617 if (!hctx->tags)
4618 continue;
4619 /*
4620 * If we're using an MQ scheduler, just update the scheduler
4621 * queue depth. This is similar to what the old code would do.
4622 */
4623 if (hctx->sched_tags) {
4624 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4625 nr, true);
4626 } else {
4627 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4628 false);
4629 }
4630 if (ret)
4631 break;
4632 if (q->elevator && q->elevator->type->ops.depth_updated)
4633 q->elevator->type->ops.depth_updated(hctx);
4634 }
4635 if (!ret) {
4636 q->nr_requests = nr;
4637 if (blk_mq_is_shared_tags(set->flags)) {
4638 if (q->elevator)
4639 blk_mq_tag_update_sched_shared_tags(q);
4640 else
4641 blk_mq_tag_resize_shared_tags(set, nr);
4642 }
4643 }
4644
4645 blk_mq_unquiesce_queue(q);
4646 blk_mq_unfreeze_queue(q);
4647
4648 return ret;
4649}
4650
4651/*
4652 * request_queue and elevator_type pair.
4653 * It is just used by __blk_mq_update_nr_hw_queues to cache
4654 * the elevator_type associated with a request_queue.
4655 */
4656struct blk_mq_qe_pair {
4657 struct list_head node;
4658 struct request_queue *q;
4659 struct elevator_type *type;
4660};
4661
4662/*
4663 * Cache the elevator_type in qe pair list and switch the
4664 * io scheduler to 'none'
4665 */
4666static bool blk_mq_elv_switch_none(struct list_head *head,
4667 struct request_queue *q)
4668{
4669 struct blk_mq_qe_pair *qe;
4670
4671 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4672 if (!qe)
4673 return false;
4674
4675 /* q->elevator needs protection from ->sysfs_lock */
4676 mutex_lock(&q->sysfs_lock);
4677
4678 /* the check has to be done with holding sysfs_lock */
4679 if (!q->elevator) {
4680 kfree(qe);
4681 goto unlock;
4682 }
4683
4684 INIT_LIST_HEAD(&qe->node);
4685 qe->q = q;
4686 qe->type = q->elevator->type;
4687 /* keep a reference to the elevator module as we'll switch back */
4688 __elevator_get(qe->type);
4689 list_add(&qe->node, head);
4690 elevator_disable(q);
4691unlock:
4692 mutex_unlock(&q->sysfs_lock);
4693
4694 return true;
4695}
4696
4697static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4698 struct request_queue *q)
4699{
4700 struct blk_mq_qe_pair *qe;
4701
4702 list_for_each_entry(qe, head, node)
4703 if (qe->q == q)
4704 return qe;
4705
4706 return NULL;
4707}
4708
4709static void blk_mq_elv_switch_back(struct list_head *head,
4710 struct request_queue *q)
4711{
4712 struct blk_mq_qe_pair *qe;
4713 struct elevator_type *t;
4714
4715 qe = blk_lookup_qe_pair(head, q);
4716 if (!qe)
4717 return;
4718 t = qe->type;
4719 list_del(&qe->node);
4720 kfree(qe);
4721
4722 mutex_lock(&q->sysfs_lock);
4723 elevator_switch(q, t);
4724 /* drop the reference acquired in blk_mq_elv_switch_none */
4725 elevator_put(t);
4726 mutex_unlock(&q->sysfs_lock);
4727}
4728
4729static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4730 int nr_hw_queues)
4731{
4732 struct request_queue *q;
4733 LIST_HEAD(head);
4734 int prev_nr_hw_queues = set->nr_hw_queues;
4735 int i;
4736
4737 lockdep_assert_held(&set->tag_list_lock);
4738
4739 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4740 nr_hw_queues = nr_cpu_ids;
4741 if (nr_hw_queues < 1)
4742 return;
4743 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4744 return;
4745
4746 list_for_each_entry(q, &set->tag_list, tag_set_list)
4747 blk_mq_freeze_queue(q);
4748 /*
4749 * Switch IO scheduler to 'none', cleaning up the data associated
4750 * with the previous scheduler. We will switch back once we are done
4751 * updating the new sw to hw queue mappings.
4752 */
4753 list_for_each_entry(q, &set->tag_list, tag_set_list)
4754 if (!blk_mq_elv_switch_none(&head, q))
4755 goto switch_back;
4756
4757 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4758 blk_mq_debugfs_unregister_hctxs(q);
4759 blk_mq_sysfs_unregister_hctxs(q);
4760 }
4761
4762 if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0)
4763 goto reregister;
4764
4765fallback:
4766 blk_mq_update_queue_map(set);
4767 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4768 blk_mq_realloc_hw_ctxs(set, q);
4769 blk_mq_update_poll_flag(q);
4770 if (q->nr_hw_queues != set->nr_hw_queues) {
4771 int i = prev_nr_hw_queues;
4772
4773 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4774 nr_hw_queues, prev_nr_hw_queues);
4775 for (; i < set->nr_hw_queues; i++)
4776 __blk_mq_free_map_and_rqs(set, i);
4777
4778 set->nr_hw_queues = prev_nr_hw_queues;
4779 goto fallback;
4780 }
4781 blk_mq_map_swqueue(q);
4782 }
4783
4784reregister:
4785 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4786 blk_mq_sysfs_register_hctxs(q);
4787 blk_mq_debugfs_register_hctxs(q);
4788 }
4789
4790switch_back:
4791 list_for_each_entry(q, &set->tag_list, tag_set_list)
4792 blk_mq_elv_switch_back(&head, q);
4793
4794 list_for_each_entry(q, &set->tag_list, tag_set_list)
4795 blk_mq_unfreeze_queue(q);
4796
4797 /* Free the excess tags when nr_hw_queues shrink. */
4798 for (i = set->nr_hw_queues; i < prev_nr_hw_queues; i++)
4799 __blk_mq_free_map_and_rqs(set, i);
4800}
4801
4802void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4803{
4804 mutex_lock(&set->tag_list_lock);
4805 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4806 mutex_unlock(&set->tag_list_lock);
4807}
4808EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4809
4810static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
4811 struct io_comp_batch *iob, unsigned int flags)
4812{
4813 long state = get_current_state();
4814 int ret;
4815
4816 do {
4817 ret = q->mq_ops->poll(hctx, iob);
4818 if (ret > 0) {
4819 __set_current_state(TASK_RUNNING);
4820 return ret;
4821 }
4822
4823 if (signal_pending_state(state, current))
4824 __set_current_state(TASK_RUNNING);
4825 if (task_is_running(current))
4826 return 1;
4827
4828 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4829 break;
4830 cpu_relax();
4831 } while (!need_resched());
4832
4833 __set_current_state(TASK_RUNNING);
4834 return 0;
4835}
4836
4837int blk_mq_poll(struct request_queue *q, blk_qc_t cookie,
4838 struct io_comp_batch *iob, unsigned int flags)
4839{
4840 struct blk_mq_hw_ctx *hctx = xa_load(&q->hctx_table, cookie);
4841
4842 return blk_hctx_poll(q, hctx, iob, flags);
4843}
4844
4845int blk_rq_poll(struct request *rq, struct io_comp_batch *iob,
4846 unsigned int poll_flags)
4847{
4848 struct request_queue *q = rq->q;
4849 int ret;
4850
4851 if (!blk_rq_is_poll(rq))
4852 return 0;
4853 if (!percpu_ref_tryget(&q->q_usage_counter))
4854 return 0;
4855
4856 ret = blk_hctx_poll(q, rq->mq_hctx, iob, poll_flags);
4857 blk_queue_exit(q);
4858
4859 return ret;
4860}
4861EXPORT_SYMBOL_GPL(blk_rq_poll);
4862
4863unsigned int blk_mq_rq_cpu(struct request *rq)
4864{
4865 return rq->mq_ctx->cpu;
4866}
4867EXPORT_SYMBOL(blk_mq_rq_cpu);
4868
4869void blk_mq_cancel_work_sync(struct request_queue *q)
4870{
4871 struct blk_mq_hw_ctx *hctx;
4872 unsigned long i;
4873
4874 cancel_delayed_work_sync(&q->requeue_work);
4875
4876 queue_for_each_hw_ctx(q, hctx, i)
4877 cancel_delayed_work_sync(&hctx->run_work);
4878}
4879
4880static int __init blk_mq_init(void)
4881{
4882 int i;
4883
4884 for_each_possible_cpu(i)
4885 init_llist_head(&per_cpu(blk_cpu_done, i));
4886 for_each_possible_cpu(i)
4887 INIT_CSD(&per_cpu(blk_cpu_csd, i),
4888 __blk_mq_complete_request_remote, NULL);
4889 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4890
4891 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4892 "block/softirq:dead", NULL,
4893 blk_softirq_cpu_dead);
4894 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4895 blk_mq_hctx_notify_dead);
4896 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4897 blk_mq_hctx_notify_online,
4898 blk_mq_hctx_notify_offline);
4899 return 0;
4900}
4901subsys_initcall(blk_mq_init);