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/topology.h>
25#include <linux/sched/signal.h>
26#include <linux/delay.h>
27#include <linux/crash_dump.h>
28#include <linux/prefetch.h>
29#include <linux/blk-crypto.h>
30#include <linux/part_stat.h>
31#include <linux/sched/isolation.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);
46static DEFINE_MUTEX(blk_mq_cpuhp_lock);
47
48static void blk_mq_insert_request(struct request *rq, blk_insert_t flags);
49static void blk_mq_request_bypass_insert(struct request *rq,
50 blk_insert_t flags);
51static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
52 struct list_head *list);
53static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
54 struct io_comp_batch *iob, unsigned int flags);
55
56/*
57 * Check if any of the ctx, dispatch list or elevator
58 * have pending work in this hardware queue.
59 */
60static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
61{
62 return !list_empty_careful(&hctx->dispatch) ||
63 sbitmap_any_bit_set(&hctx->ctx_map) ||
64 blk_mq_sched_has_work(hctx);
65}
66
67/*
68 * Mark this ctx as having pending work in this hardware queue
69 */
70static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
71 struct blk_mq_ctx *ctx)
72{
73 const int bit = ctx->index_hw[hctx->type];
74
75 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
76 sbitmap_set_bit(&hctx->ctx_map, bit);
77}
78
79static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
80 struct blk_mq_ctx *ctx)
81{
82 const int bit = ctx->index_hw[hctx->type];
83
84 sbitmap_clear_bit(&hctx->ctx_map, bit);
85}
86
87struct mq_inflight {
88 struct block_device *part;
89 unsigned int inflight[2];
90};
91
92static bool blk_mq_check_inflight(struct request *rq, void *priv)
93{
94 struct mq_inflight *mi = priv;
95
96 if (rq->rq_flags & RQF_IO_STAT &&
97 (!bdev_is_partition(mi->part) || rq->part == mi->part) &&
98 blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
99 mi->inflight[rq_data_dir(rq)]++;
100
101 return true;
102}
103
104unsigned int blk_mq_in_flight(struct request_queue *q,
105 struct block_device *part)
106{
107 struct mq_inflight mi = { .part = part };
108
109 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
110
111 return mi.inflight[0] + mi.inflight[1];
112}
113
114void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
115 unsigned int inflight[2])
116{
117 struct mq_inflight mi = { .part = part };
118
119 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
120 inflight[0] = mi.inflight[0];
121 inflight[1] = mi.inflight[1];
122}
123
124#ifdef CONFIG_LOCKDEP
125static bool blk_freeze_set_owner(struct request_queue *q,
126 struct task_struct *owner)
127{
128 if (!owner)
129 return false;
130
131 if (!q->mq_freeze_depth) {
132 q->mq_freeze_owner = owner;
133 q->mq_freeze_owner_depth = 1;
134 return true;
135 }
136
137 if (owner == q->mq_freeze_owner)
138 q->mq_freeze_owner_depth += 1;
139 return false;
140}
141
142/* verify the last unfreeze in owner context */
143static bool blk_unfreeze_check_owner(struct request_queue *q)
144{
145 if (!q->mq_freeze_owner)
146 return false;
147 if (q->mq_freeze_owner != current)
148 return false;
149 if (--q->mq_freeze_owner_depth == 0) {
150 q->mq_freeze_owner = NULL;
151 return true;
152 }
153 return false;
154}
155
156#else
157
158static bool blk_freeze_set_owner(struct request_queue *q,
159 struct task_struct *owner)
160{
161 return false;
162}
163
164static bool blk_unfreeze_check_owner(struct request_queue *q)
165{
166 return false;
167}
168#endif
169
170bool __blk_freeze_queue_start(struct request_queue *q,
171 struct task_struct *owner)
172{
173 bool freeze;
174
175 mutex_lock(&q->mq_freeze_lock);
176 freeze = blk_freeze_set_owner(q, owner);
177 if (++q->mq_freeze_depth == 1) {
178 percpu_ref_kill(&q->q_usage_counter);
179 mutex_unlock(&q->mq_freeze_lock);
180 if (queue_is_mq(q))
181 blk_mq_run_hw_queues(q, false);
182 } else {
183 mutex_unlock(&q->mq_freeze_lock);
184 }
185
186 return freeze;
187}
188
189void blk_freeze_queue_start(struct request_queue *q)
190{
191 if (__blk_freeze_queue_start(q, current))
192 blk_freeze_acquire_lock(q, false, false);
193}
194EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
195
196void blk_mq_freeze_queue_wait(struct request_queue *q)
197{
198 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
199}
200EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
201
202int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
203 unsigned long timeout)
204{
205 return wait_event_timeout(q->mq_freeze_wq,
206 percpu_ref_is_zero(&q->q_usage_counter),
207 timeout);
208}
209EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
210
211void blk_mq_freeze_queue(struct request_queue *q)
212{
213 blk_freeze_queue_start(q);
214 blk_mq_freeze_queue_wait(q);
215}
216EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
217
218bool __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
219{
220 bool unfreeze;
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 unfreeze = blk_unfreeze_check_owner(q);
232 mutex_unlock(&q->mq_freeze_lock);
233
234 return unfreeze;
235}
236
237void blk_mq_unfreeze_queue(struct request_queue *q)
238{
239 if (__blk_mq_unfreeze_queue(q, false))
240 blk_unfreeze_release_lock(q, false, false);
241}
242EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
243
244/*
245 * non_owner variant of blk_freeze_queue_start
246 *
247 * Unlike blk_freeze_queue_start, the queue doesn't need to be unfrozen
248 * by the same task. This is fragile and should not be used if at all
249 * possible.
250 */
251void blk_freeze_queue_start_non_owner(struct request_queue *q)
252{
253 __blk_freeze_queue_start(q, NULL);
254}
255EXPORT_SYMBOL_GPL(blk_freeze_queue_start_non_owner);
256
257/* non_owner variant of blk_mq_unfreeze_queue */
258void blk_mq_unfreeze_queue_non_owner(struct request_queue *q)
259{
260 __blk_mq_unfreeze_queue(q, false);
261}
262EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue_non_owner);
263
264/*
265 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
266 * mpt3sas driver such that this function can be removed.
267 */
268void blk_mq_quiesce_queue_nowait(struct request_queue *q)
269{
270 unsigned long flags;
271
272 spin_lock_irqsave(&q->queue_lock, flags);
273 if (!q->quiesce_depth++)
274 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
275 spin_unlock_irqrestore(&q->queue_lock, flags);
276}
277EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
278
279/**
280 * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
281 * @set: tag_set to wait on
282 *
283 * Note: it is driver's responsibility for making sure that quiesce has
284 * been started on or more of the request_queues of the tag_set. This
285 * function only waits for the quiesce on those request_queues that had
286 * the quiesce flag set using blk_mq_quiesce_queue_nowait.
287 */
288void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set)
289{
290 if (set->flags & BLK_MQ_F_BLOCKING)
291 synchronize_srcu(set->srcu);
292 else
293 synchronize_rcu();
294}
295EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
296
297/**
298 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
299 * @q: request queue.
300 *
301 * Note: this function does not prevent that the struct request end_io()
302 * callback function is invoked. Once this function is returned, we make
303 * sure no dispatch can happen until the queue is unquiesced via
304 * blk_mq_unquiesce_queue().
305 */
306void blk_mq_quiesce_queue(struct request_queue *q)
307{
308 blk_mq_quiesce_queue_nowait(q);
309 /* nothing to wait for non-mq queues */
310 if (queue_is_mq(q))
311 blk_mq_wait_quiesce_done(q->tag_set);
312}
313EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
314
315/*
316 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
317 * @q: request queue.
318 *
319 * This function recovers queue into the state before quiescing
320 * which is done by blk_mq_quiesce_queue.
321 */
322void blk_mq_unquiesce_queue(struct request_queue *q)
323{
324 unsigned long flags;
325 bool run_queue = false;
326
327 spin_lock_irqsave(&q->queue_lock, flags);
328 if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
329 ;
330 } else if (!--q->quiesce_depth) {
331 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
332 run_queue = true;
333 }
334 spin_unlock_irqrestore(&q->queue_lock, flags);
335
336 /* dispatch requests which are inserted during quiescing */
337 if (run_queue)
338 blk_mq_run_hw_queues(q, true);
339}
340EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
341
342void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set)
343{
344 struct request_queue *q;
345
346 mutex_lock(&set->tag_list_lock);
347 list_for_each_entry(q, &set->tag_list, tag_set_list) {
348 if (!blk_queue_skip_tagset_quiesce(q))
349 blk_mq_quiesce_queue_nowait(q);
350 }
351 mutex_unlock(&set->tag_list_lock);
352
353 blk_mq_wait_quiesce_done(set);
354}
355EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset);
356
357void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set)
358{
359 struct request_queue *q;
360
361 mutex_lock(&set->tag_list_lock);
362 list_for_each_entry(q, &set->tag_list, tag_set_list) {
363 if (!blk_queue_skip_tagset_quiesce(q))
364 blk_mq_unquiesce_queue(q);
365 }
366 mutex_unlock(&set->tag_list_lock);
367}
368EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset);
369
370void blk_mq_wake_waiters(struct request_queue *q)
371{
372 struct blk_mq_hw_ctx *hctx;
373 unsigned long i;
374
375 queue_for_each_hw_ctx(q, hctx, i)
376 if (blk_mq_hw_queue_mapped(hctx))
377 blk_mq_tag_wakeup_all(hctx->tags, true);
378}
379
380void blk_rq_init(struct request_queue *q, struct request *rq)
381{
382 memset(rq, 0, sizeof(*rq));
383
384 INIT_LIST_HEAD(&rq->queuelist);
385 rq->q = q;
386 rq->__sector = (sector_t) -1;
387 INIT_HLIST_NODE(&rq->hash);
388 RB_CLEAR_NODE(&rq->rb_node);
389 rq->tag = BLK_MQ_NO_TAG;
390 rq->internal_tag = BLK_MQ_NO_TAG;
391 rq->start_time_ns = blk_time_get_ns();
392 blk_crypto_rq_set_defaults(rq);
393}
394EXPORT_SYMBOL(blk_rq_init);
395
396/* Set start and alloc time when the allocated request is actually used */
397static inline void blk_mq_rq_time_init(struct request *rq, u64 alloc_time_ns)
398{
399#ifdef CONFIG_BLK_RQ_ALLOC_TIME
400 if (blk_queue_rq_alloc_time(rq->q))
401 rq->alloc_time_ns = alloc_time_ns;
402 else
403 rq->alloc_time_ns = 0;
404#endif
405}
406
407static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
408 struct blk_mq_tags *tags, unsigned int tag)
409{
410 struct blk_mq_ctx *ctx = data->ctx;
411 struct blk_mq_hw_ctx *hctx = data->hctx;
412 struct request_queue *q = data->q;
413 struct request *rq = tags->static_rqs[tag];
414
415 rq->q = q;
416 rq->mq_ctx = ctx;
417 rq->mq_hctx = hctx;
418 rq->cmd_flags = data->cmd_flags;
419
420 if (data->flags & BLK_MQ_REQ_PM)
421 data->rq_flags |= RQF_PM;
422 rq->rq_flags = data->rq_flags;
423
424 if (data->rq_flags & RQF_SCHED_TAGS) {
425 rq->tag = BLK_MQ_NO_TAG;
426 rq->internal_tag = tag;
427 } else {
428 rq->tag = tag;
429 rq->internal_tag = BLK_MQ_NO_TAG;
430 }
431 rq->timeout = 0;
432
433 rq->part = NULL;
434 rq->io_start_time_ns = 0;
435 rq->stats_sectors = 0;
436 rq->nr_phys_segments = 0;
437 rq->nr_integrity_segments = 0;
438 rq->end_io = NULL;
439 rq->end_io_data = NULL;
440
441 blk_crypto_rq_set_defaults(rq);
442 INIT_LIST_HEAD(&rq->queuelist);
443 /* tag was already set */
444 WRITE_ONCE(rq->deadline, 0);
445 req_ref_set(rq, 1);
446
447 if (rq->rq_flags & RQF_USE_SCHED) {
448 struct elevator_queue *e = data->q->elevator;
449
450 INIT_HLIST_NODE(&rq->hash);
451 RB_CLEAR_NODE(&rq->rb_node);
452
453 if (e->type->ops.prepare_request)
454 e->type->ops.prepare_request(rq);
455 }
456
457 return rq;
458}
459
460static inline struct request *
461__blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data)
462{
463 unsigned int tag, tag_offset;
464 struct blk_mq_tags *tags;
465 struct request *rq;
466 unsigned long tag_mask;
467 int i, nr = 0;
468
469 tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset);
470 if (unlikely(!tag_mask))
471 return NULL;
472
473 tags = blk_mq_tags_from_data(data);
474 for (i = 0; tag_mask; i++) {
475 if (!(tag_mask & (1UL << i)))
476 continue;
477 tag = tag_offset + i;
478 prefetch(tags->static_rqs[tag]);
479 tag_mask &= ~(1UL << i);
480 rq = blk_mq_rq_ctx_init(data, tags, tag);
481 rq_list_add_head(data->cached_rqs, rq);
482 nr++;
483 }
484 if (!(data->rq_flags & RQF_SCHED_TAGS))
485 blk_mq_add_active_requests(data->hctx, nr);
486 /* caller already holds a reference, add for remainder */
487 percpu_ref_get_many(&data->q->q_usage_counter, nr - 1);
488 data->nr_tags -= nr;
489
490 return rq_list_pop(data->cached_rqs);
491}
492
493static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
494{
495 struct request_queue *q = data->q;
496 u64 alloc_time_ns = 0;
497 struct request *rq;
498 unsigned int tag;
499
500 /* alloc_time includes depth and tag waits */
501 if (blk_queue_rq_alloc_time(q))
502 alloc_time_ns = blk_time_get_ns();
503
504 if (data->cmd_flags & REQ_NOWAIT)
505 data->flags |= BLK_MQ_REQ_NOWAIT;
506
507retry:
508 data->ctx = blk_mq_get_ctx(q);
509 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
510
511 if (q->elevator) {
512 /*
513 * All requests use scheduler tags when an I/O scheduler is
514 * enabled for the queue.
515 */
516 data->rq_flags |= RQF_SCHED_TAGS;
517
518 /*
519 * Flush/passthrough requests are special and go directly to the
520 * dispatch list.
521 */
522 if ((data->cmd_flags & REQ_OP_MASK) != REQ_OP_FLUSH &&
523 !blk_op_is_passthrough(data->cmd_flags)) {
524 struct elevator_mq_ops *ops = &q->elevator->type->ops;
525
526 WARN_ON_ONCE(data->flags & BLK_MQ_REQ_RESERVED);
527
528 data->rq_flags |= RQF_USE_SCHED;
529 if (ops->limit_depth)
530 ops->limit_depth(data->cmd_flags, data);
531 }
532 } else {
533 blk_mq_tag_busy(data->hctx);
534 }
535
536 if (data->flags & BLK_MQ_REQ_RESERVED)
537 data->rq_flags |= RQF_RESV;
538
539 /*
540 * Try batched alloc if we want more than 1 tag.
541 */
542 if (data->nr_tags > 1) {
543 rq = __blk_mq_alloc_requests_batch(data);
544 if (rq) {
545 blk_mq_rq_time_init(rq, alloc_time_ns);
546 return rq;
547 }
548 data->nr_tags = 1;
549 }
550
551 /*
552 * Waiting allocations only fail because of an inactive hctx. In that
553 * case just retry the hctx assignment and tag allocation as CPU hotplug
554 * should have migrated us to an online CPU by now.
555 */
556 tag = blk_mq_get_tag(data);
557 if (tag == BLK_MQ_NO_TAG) {
558 if (data->flags & BLK_MQ_REQ_NOWAIT)
559 return NULL;
560 /*
561 * Give up the CPU and sleep for a random short time to
562 * ensure that thread using a realtime scheduling class
563 * are migrated off the CPU, and thus off the hctx that
564 * is going away.
565 */
566 msleep(3);
567 goto retry;
568 }
569
570 if (!(data->rq_flags & RQF_SCHED_TAGS))
571 blk_mq_inc_active_requests(data->hctx);
572 rq = blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag);
573 blk_mq_rq_time_init(rq, alloc_time_ns);
574 return rq;
575}
576
577static struct request *blk_mq_rq_cache_fill(struct request_queue *q,
578 struct blk_plug *plug,
579 blk_opf_t opf,
580 blk_mq_req_flags_t flags)
581{
582 struct blk_mq_alloc_data data = {
583 .q = q,
584 .flags = flags,
585 .cmd_flags = opf,
586 .nr_tags = plug->nr_ios,
587 .cached_rqs = &plug->cached_rqs,
588 };
589 struct request *rq;
590
591 if (blk_queue_enter(q, flags))
592 return NULL;
593
594 plug->nr_ios = 1;
595
596 rq = __blk_mq_alloc_requests(&data);
597 if (unlikely(!rq))
598 blk_queue_exit(q);
599 return rq;
600}
601
602static struct request *blk_mq_alloc_cached_request(struct request_queue *q,
603 blk_opf_t opf,
604 blk_mq_req_flags_t flags)
605{
606 struct blk_plug *plug = current->plug;
607 struct request *rq;
608
609 if (!plug)
610 return NULL;
611
612 if (rq_list_empty(&plug->cached_rqs)) {
613 if (plug->nr_ios == 1)
614 return NULL;
615 rq = blk_mq_rq_cache_fill(q, plug, opf, flags);
616 if (!rq)
617 return NULL;
618 } else {
619 rq = rq_list_peek(&plug->cached_rqs);
620 if (!rq || rq->q != q)
621 return NULL;
622
623 if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type)
624 return NULL;
625 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
626 return NULL;
627
628 rq_list_pop(&plug->cached_rqs);
629 blk_mq_rq_time_init(rq, blk_time_get_ns());
630 }
631
632 rq->cmd_flags = opf;
633 INIT_LIST_HEAD(&rq->queuelist);
634 return rq;
635}
636
637struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf,
638 blk_mq_req_flags_t flags)
639{
640 struct request *rq;
641
642 rq = blk_mq_alloc_cached_request(q, opf, flags);
643 if (!rq) {
644 struct blk_mq_alloc_data data = {
645 .q = q,
646 .flags = flags,
647 .cmd_flags = opf,
648 .nr_tags = 1,
649 };
650 int ret;
651
652 ret = blk_queue_enter(q, flags);
653 if (ret)
654 return ERR_PTR(ret);
655
656 rq = __blk_mq_alloc_requests(&data);
657 if (!rq)
658 goto out_queue_exit;
659 }
660 rq->__data_len = 0;
661 rq->__sector = (sector_t) -1;
662 rq->bio = rq->biotail = NULL;
663 return rq;
664out_queue_exit:
665 blk_queue_exit(q);
666 return ERR_PTR(-EWOULDBLOCK);
667}
668EXPORT_SYMBOL(blk_mq_alloc_request);
669
670struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
671 blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx)
672{
673 struct blk_mq_alloc_data data = {
674 .q = q,
675 .flags = flags,
676 .cmd_flags = opf,
677 .nr_tags = 1,
678 };
679 u64 alloc_time_ns = 0;
680 struct request *rq;
681 unsigned int cpu;
682 unsigned int tag;
683 int ret;
684
685 /* alloc_time includes depth and tag waits */
686 if (blk_queue_rq_alloc_time(q))
687 alloc_time_ns = blk_time_get_ns();
688
689 /*
690 * If the tag allocator sleeps we could get an allocation for a
691 * different hardware context. No need to complicate the low level
692 * allocator for this for the rare use case of a command tied to
693 * a specific queue.
694 */
695 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) ||
696 WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED)))
697 return ERR_PTR(-EINVAL);
698
699 if (hctx_idx >= q->nr_hw_queues)
700 return ERR_PTR(-EIO);
701
702 ret = blk_queue_enter(q, flags);
703 if (ret)
704 return ERR_PTR(ret);
705
706 /*
707 * Check if the hardware context is actually mapped to anything.
708 * If not tell the caller that it should skip this queue.
709 */
710 ret = -EXDEV;
711 data.hctx = xa_load(&q->hctx_table, hctx_idx);
712 if (!blk_mq_hw_queue_mapped(data.hctx))
713 goto out_queue_exit;
714 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
715 if (cpu >= nr_cpu_ids)
716 goto out_queue_exit;
717 data.ctx = __blk_mq_get_ctx(q, cpu);
718
719 if (q->elevator)
720 data.rq_flags |= RQF_SCHED_TAGS;
721 else
722 blk_mq_tag_busy(data.hctx);
723
724 if (flags & BLK_MQ_REQ_RESERVED)
725 data.rq_flags |= RQF_RESV;
726
727 ret = -EWOULDBLOCK;
728 tag = blk_mq_get_tag(&data);
729 if (tag == BLK_MQ_NO_TAG)
730 goto out_queue_exit;
731 if (!(data.rq_flags & RQF_SCHED_TAGS))
732 blk_mq_inc_active_requests(data.hctx);
733 rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag);
734 blk_mq_rq_time_init(rq, alloc_time_ns);
735 rq->__data_len = 0;
736 rq->__sector = (sector_t) -1;
737 rq->bio = rq->biotail = NULL;
738 return rq;
739
740out_queue_exit:
741 blk_queue_exit(q);
742 return ERR_PTR(ret);
743}
744EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
745
746static void blk_mq_finish_request(struct request *rq)
747{
748 struct request_queue *q = rq->q;
749
750 blk_zone_finish_request(rq);
751
752 if (rq->rq_flags & RQF_USE_SCHED) {
753 q->elevator->type->ops.finish_request(rq);
754 /*
755 * For postflush request that may need to be
756 * completed twice, we should clear this flag
757 * to avoid double finish_request() on the rq.
758 */
759 rq->rq_flags &= ~RQF_USE_SCHED;
760 }
761}
762
763static void __blk_mq_free_request(struct request *rq)
764{
765 struct request_queue *q = rq->q;
766 struct blk_mq_ctx *ctx = rq->mq_ctx;
767 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
768 const int sched_tag = rq->internal_tag;
769
770 blk_crypto_free_request(rq);
771 blk_pm_mark_last_busy(rq);
772 rq->mq_hctx = NULL;
773
774 if (rq->tag != BLK_MQ_NO_TAG) {
775 blk_mq_dec_active_requests(hctx);
776 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
777 }
778 if (sched_tag != BLK_MQ_NO_TAG)
779 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
780 blk_mq_sched_restart(hctx);
781 blk_queue_exit(q);
782}
783
784void blk_mq_free_request(struct request *rq)
785{
786 struct request_queue *q = rq->q;
787
788 blk_mq_finish_request(rq);
789
790 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
791 laptop_io_completion(q->disk->bdi);
792
793 rq_qos_done(q, rq);
794
795 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
796 if (req_ref_put_and_test(rq))
797 __blk_mq_free_request(rq);
798}
799EXPORT_SYMBOL_GPL(blk_mq_free_request);
800
801void blk_mq_free_plug_rqs(struct blk_plug *plug)
802{
803 struct request *rq;
804
805 while ((rq = rq_list_pop(&plug->cached_rqs)) != NULL)
806 blk_mq_free_request(rq);
807}
808
809void blk_dump_rq_flags(struct request *rq, char *msg)
810{
811 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
812 rq->q->disk ? rq->q->disk->disk_name : "?",
813 (__force unsigned long long) rq->cmd_flags);
814
815 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
816 (unsigned long long)blk_rq_pos(rq),
817 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
818 printk(KERN_INFO " bio %p, biotail %p, len %u\n",
819 rq->bio, rq->biotail, blk_rq_bytes(rq));
820}
821EXPORT_SYMBOL(blk_dump_rq_flags);
822
823static void blk_account_io_completion(struct request *req, unsigned int bytes)
824{
825 if (req->rq_flags & RQF_IO_STAT) {
826 const int sgrp = op_stat_group(req_op(req));
827
828 part_stat_lock();
829 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
830 part_stat_unlock();
831 }
832}
833
834static void blk_print_req_error(struct request *req, blk_status_t status)
835{
836 printk_ratelimited(KERN_ERR
837 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
838 "phys_seg %u prio class %u\n",
839 blk_status_to_str(status),
840 req->q->disk ? req->q->disk->disk_name : "?",
841 blk_rq_pos(req), (__force u32)req_op(req),
842 blk_op_str(req_op(req)),
843 (__force u32)(req->cmd_flags & ~REQ_OP_MASK),
844 req->nr_phys_segments,
845 IOPRIO_PRIO_CLASS(req_get_ioprio(req)));
846}
847
848/*
849 * Fully end IO on a request. Does not support partial completions, or
850 * errors.
851 */
852static void blk_complete_request(struct request *req)
853{
854 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
855 int total_bytes = blk_rq_bytes(req);
856 struct bio *bio = req->bio;
857
858 trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
859
860 if (!bio)
861 return;
862
863 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
864 blk_integrity_complete(req, total_bytes);
865
866 /*
867 * Upper layers may call blk_crypto_evict_key() anytime after the last
868 * bio_endio(). Therefore, the keyslot must be released before that.
869 */
870 blk_crypto_rq_put_keyslot(req);
871
872 blk_account_io_completion(req, total_bytes);
873
874 do {
875 struct bio *next = bio->bi_next;
876
877 /* Completion has already been traced */
878 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
879
880 blk_zone_update_request_bio(req, bio);
881
882 if (!is_flush)
883 bio_endio(bio);
884 bio = next;
885 } while (bio);
886
887 /*
888 * Reset counters so that the request stacking driver
889 * can find how many bytes remain in the request
890 * later.
891 */
892 if (!req->end_io) {
893 req->bio = NULL;
894 req->__data_len = 0;
895 }
896}
897
898/**
899 * blk_update_request - Complete multiple bytes without completing the request
900 * @req: the request being processed
901 * @error: block status code
902 * @nr_bytes: number of bytes to complete for @req
903 *
904 * Description:
905 * Ends I/O on a number of bytes attached to @req, but doesn't complete
906 * the request structure even if @req doesn't have leftover.
907 * If @req has leftover, sets it up for the next range of segments.
908 *
909 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
910 * %false return from this function.
911 *
912 * Note:
913 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
914 * except in the consistency check at the end of this function.
915 *
916 * Return:
917 * %false - this request doesn't have any more data
918 * %true - this request has more data
919 **/
920bool blk_update_request(struct request *req, blk_status_t error,
921 unsigned int nr_bytes)
922{
923 bool is_flush = req->rq_flags & RQF_FLUSH_SEQ;
924 bool quiet = req->rq_flags & RQF_QUIET;
925 int total_bytes;
926
927 trace_block_rq_complete(req, error, nr_bytes);
928
929 if (!req->bio)
930 return false;
931
932 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
933 error == BLK_STS_OK)
934 blk_integrity_complete(req, nr_bytes);
935
936 /*
937 * Upper layers may call blk_crypto_evict_key() anytime after the last
938 * bio_endio(). Therefore, the keyslot must be released before that.
939 */
940 if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req))
941 __blk_crypto_rq_put_keyslot(req);
942
943 if (unlikely(error && !blk_rq_is_passthrough(req) && !quiet) &&
944 !test_bit(GD_DEAD, &req->q->disk->state)) {
945 blk_print_req_error(req, error);
946 trace_block_rq_error(req, error, nr_bytes);
947 }
948
949 blk_account_io_completion(req, nr_bytes);
950
951 total_bytes = 0;
952 while (req->bio) {
953 struct bio *bio = req->bio;
954 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
955
956 if (unlikely(error))
957 bio->bi_status = error;
958
959 if (bio_bytes == bio->bi_iter.bi_size) {
960 req->bio = bio->bi_next;
961 } else if (bio_is_zone_append(bio) && error == BLK_STS_OK) {
962 /*
963 * Partial zone append completions cannot be supported
964 * as the BIO fragments may end up not being written
965 * sequentially.
966 */
967 bio->bi_status = BLK_STS_IOERR;
968 }
969
970 /* Completion has already been traced */
971 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
972 if (unlikely(quiet))
973 bio_set_flag(bio, BIO_QUIET);
974
975 bio_advance(bio, bio_bytes);
976
977 /* Don't actually finish bio if it's part of flush sequence */
978 if (!bio->bi_iter.bi_size) {
979 blk_zone_update_request_bio(req, bio);
980 if (!is_flush)
981 bio_endio(bio);
982 }
983
984 total_bytes += bio_bytes;
985 nr_bytes -= bio_bytes;
986
987 if (!nr_bytes)
988 break;
989 }
990
991 /*
992 * completely done
993 */
994 if (!req->bio) {
995 /*
996 * Reset counters so that the request stacking driver
997 * can find how many bytes remain in the request
998 * later.
999 */
1000 req->__data_len = 0;
1001 return false;
1002 }
1003
1004 req->__data_len -= total_bytes;
1005
1006 /* update sector only for requests with clear definition of sector */
1007 if (!blk_rq_is_passthrough(req))
1008 req->__sector += total_bytes >> 9;
1009
1010 /* mixed attributes always follow the first bio */
1011 if (req->rq_flags & RQF_MIXED_MERGE) {
1012 req->cmd_flags &= ~REQ_FAILFAST_MASK;
1013 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
1014 }
1015
1016 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
1017 /*
1018 * If total number of sectors is less than the first segment
1019 * size, something has gone terribly wrong.
1020 */
1021 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
1022 blk_dump_rq_flags(req, "request botched");
1023 req->__data_len = blk_rq_cur_bytes(req);
1024 }
1025
1026 /* recalculate the number of segments */
1027 req->nr_phys_segments = blk_recalc_rq_segments(req);
1028 }
1029
1030 return true;
1031}
1032EXPORT_SYMBOL_GPL(blk_update_request);
1033
1034static inline void blk_account_io_done(struct request *req, u64 now)
1035{
1036 trace_block_io_done(req);
1037
1038 /*
1039 * Account IO completion. flush_rq isn't accounted as a
1040 * normal IO on queueing nor completion. Accounting the
1041 * containing request is enough.
1042 */
1043 if ((req->rq_flags & (RQF_IO_STAT|RQF_FLUSH_SEQ)) == RQF_IO_STAT) {
1044 const int sgrp = op_stat_group(req_op(req));
1045
1046 part_stat_lock();
1047 update_io_ticks(req->part, jiffies, true);
1048 part_stat_inc(req->part, ios[sgrp]);
1049 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
1050 part_stat_local_dec(req->part,
1051 in_flight[op_is_write(req_op(req))]);
1052 part_stat_unlock();
1053 }
1054}
1055
1056static inline bool blk_rq_passthrough_stats(struct request *req)
1057{
1058 struct bio *bio = req->bio;
1059
1060 if (!blk_queue_passthrough_stat(req->q))
1061 return false;
1062
1063 /* Requests without a bio do not transfer data. */
1064 if (!bio)
1065 return false;
1066
1067 /*
1068 * Stats are accumulated in the bdev, so must have one attached to a
1069 * bio to track stats. Most drivers do not set the bdev for passthrough
1070 * requests, but nvme is one that will set it.
1071 */
1072 if (!bio->bi_bdev)
1073 return false;
1074
1075 /*
1076 * We don't know what a passthrough command does, but we know the
1077 * payload size and data direction. Ensuring the size is aligned to the
1078 * block size filters out most commands with payloads that don't
1079 * represent sector access.
1080 */
1081 if (blk_rq_bytes(req) & (bdev_logical_block_size(bio->bi_bdev) - 1))
1082 return false;
1083 return true;
1084}
1085
1086static inline void blk_account_io_start(struct request *req)
1087{
1088 trace_block_io_start(req);
1089
1090 if (!blk_queue_io_stat(req->q))
1091 return;
1092 if (blk_rq_is_passthrough(req) && !blk_rq_passthrough_stats(req))
1093 return;
1094
1095 req->rq_flags |= RQF_IO_STAT;
1096 req->start_time_ns = blk_time_get_ns();
1097
1098 /*
1099 * All non-passthrough requests are created from a bio with one
1100 * exception: when a flush command that is part of a flush sequence
1101 * generated by the state machine in blk-flush.c is cloned onto the
1102 * lower device by dm-multipath we can get here without a bio.
1103 */
1104 if (req->bio)
1105 req->part = req->bio->bi_bdev;
1106 else
1107 req->part = req->q->disk->part0;
1108
1109 part_stat_lock();
1110 update_io_ticks(req->part, jiffies, false);
1111 part_stat_local_inc(req->part, in_flight[op_is_write(req_op(req))]);
1112 part_stat_unlock();
1113}
1114
1115static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
1116{
1117 if (rq->rq_flags & RQF_STATS)
1118 blk_stat_add(rq, now);
1119
1120 blk_mq_sched_completed_request(rq, now);
1121 blk_account_io_done(rq, now);
1122}
1123
1124inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
1125{
1126 if (blk_mq_need_time_stamp(rq))
1127 __blk_mq_end_request_acct(rq, blk_time_get_ns());
1128
1129 blk_mq_finish_request(rq);
1130
1131 if (rq->end_io) {
1132 rq_qos_done(rq->q, rq);
1133 if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1134 blk_mq_free_request(rq);
1135 } else {
1136 blk_mq_free_request(rq);
1137 }
1138}
1139EXPORT_SYMBOL(__blk_mq_end_request);
1140
1141void blk_mq_end_request(struct request *rq, blk_status_t error)
1142{
1143 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1144 BUG();
1145 __blk_mq_end_request(rq, error);
1146}
1147EXPORT_SYMBOL(blk_mq_end_request);
1148
1149#define TAG_COMP_BATCH 32
1150
1151static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1152 int *tag_array, int nr_tags)
1153{
1154 struct request_queue *q = hctx->queue;
1155
1156 blk_mq_sub_active_requests(hctx, nr_tags);
1157
1158 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
1159 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
1160}
1161
1162void blk_mq_end_request_batch(struct io_comp_batch *iob)
1163{
1164 int tags[TAG_COMP_BATCH], nr_tags = 0;
1165 struct blk_mq_hw_ctx *cur_hctx = NULL;
1166 struct request *rq;
1167 u64 now = 0;
1168
1169 if (iob->need_ts)
1170 now = blk_time_get_ns();
1171
1172 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1173 prefetch(rq->bio);
1174 prefetch(rq->rq_next);
1175
1176 blk_complete_request(rq);
1177 if (iob->need_ts)
1178 __blk_mq_end_request_acct(rq, now);
1179
1180 blk_mq_finish_request(rq);
1181
1182 rq_qos_done(rq->q, rq);
1183
1184 /*
1185 * If end_io handler returns NONE, then it still has
1186 * ownership of the request.
1187 */
1188 if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
1189 continue;
1190
1191 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1192 if (!req_ref_put_and_test(rq))
1193 continue;
1194
1195 blk_crypto_free_request(rq);
1196 blk_pm_mark_last_busy(rq);
1197
1198 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1199 if (cur_hctx)
1200 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1201 nr_tags = 0;
1202 cur_hctx = rq->mq_hctx;
1203 }
1204 tags[nr_tags++] = rq->tag;
1205 }
1206
1207 if (nr_tags)
1208 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1209}
1210EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1211
1212static void blk_complete_reqs(struct llist_head *list)
1213{
1214 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1215 struct request *rq, *next;
1216
1217 llist_for_each_entry_safe(rq, next, entry, ipi_list)
1218 rq->q->mq_ops->complete(rq);
1219}
1220
1221static __latent_entropy void blk_done_softirq(void)
1222{
1223 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1224}
1225
1226static int blk_softirq_cpu_dead(unsigned int cpu)
1227{
1228 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1229 return 0;
1230}
1231
1232static void __blk_mq_complete_request_remote(void *data)
1233{
1234 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1235}
1236
1237static inline bool blk_mq_complete_need_ipi(struct request *rq)
1238{
1239 int cpu = raw_smp_processor_id();
1240
1241 if (!IS_ENABLED(CONFIG_SMP) ||
1242 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1243 return false;
1244 /*
1245 * With force threaded interrupts enabled, raising softirq from an SMP
1246 * function call will always result in waking the ksoftirqd thread.
1247 * This is probably worse than completing the request on a different
1248 * cache domain.
1249 */
1250 if (force_irqthreads())
1251 return false;
1252
1253 /* same CPU or cache domain and capacity? Complete locally */
1254 if (cpu == rq->mq_ctx->cpu ||
1255 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1256 cpus_share_cache(cpu, rq->mq_ctx->cpu) &&
1257 cpus_equal_capacity(cpu, rq->mq_ctx->cpu)))
1258 return false;
1259
1260 /* don't try to IPI to an offline CPU */
1261 return cpu_online(rq->mq_ctx->cpu);
1262}
1263
1264static void blk_mq_complete_send_ipi(struct request *rq)
1265{
1266 unsigned int cpu;
1267
1268 cpu = rq->mq_ctx->cpu;
1269 if (llist_add(&rq->ipi_list, &per_cpu(blk_cpu_done, cpu)))
1270 smp_call_function_single_async(cpu, &per_cpu(blk_cpu_csd, cpu));
1271}
1272
1273static void blk_mq_raise_softirq(struct request *rq)
1274{
1275 struct llist_head *list;
1276
1277 preempt_disable();
1278 list = this_cpu_ptr(&blk_cpu_done);
1279 if (llist_add(&rq->ipi_list, list))
1280 raise_softirq(BLOCK_SOFTIRQ);
1281 preempt_enable();
1282}
1283
1284bool blk_mq_complete_request_remote(struct request *rq)
1285{
1286 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1287
1288 /*
1289 * For request which hctx has only one ctx mapping,
1290 * or a polled request, always complete locally,
1291 * it's pointless to redirect the completion.
1292 */
1293 if ((rq->mq_hctx->nr_ctx == 1 &&
1294 rq->mq_ctx->cpu == raw_smp_processor_id()) ||
1295 rq->cmd_flags & REQ_POLLED)
1296 return false;
1297
1298 if (blk_mq_complete_need_ipi(rq)) {
1299 blk_mq_complete_send_ipi(rq);
1300 return true;
1301 }
1302
1303 if (rq->q->nr_hw_queues == 1) {
1304 blk_mq_raise_softirq(rq);
1305 return true;
1306 }
1307 return false;
1308}
1309EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1310
1311/**
1312 * blk_mq_complete_request - end I/O on a request
1313 * @rq: the request being processed
1314 *
1315 * Description:
1316 * Complete a request by scheduling the ->complete_rq operation.
1317 **/
1318void blk_mq_complete_request(struct request *rq)
1319{
1320 if (!blk_mq_complete_request_remote(rq))
1321 rq->q->mq_ops->complete(rq);
1322}
1323EXPORT_SYMBOL(blk_mq_complete_request);
1324
1325/**
1326 * blk_mq_start_request - Start processing a request
1327 * @rq: Pointer to request to be started
1328 *
1329 * Function used by device drivers to notify the block layer that a request
1330 * is going to be processed now, so blk layer can do proper initializations
1331 * such as starting the timeout timer.
1332 */
1333void blk_mq_start_request(struct request *rq)
1334{
1335 struct request_queue *q = rq->q;
1336
1337 trace_block_rq_issue(rq);
1338
1339 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags) &&
1340 !blk_rq_is_passthrough(rq)) {
1341 rq->io_start_time_ns = blk_time_get_ns();
1342 rq->stats_sectors = blk_rq_sectors(rq);
1343 rq->rq_flags |= RQF_STATS;
1344 rq_qos_issue(q, rq);
1345 }
1346
1347 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1348
1349 blk_add_timer(rq);
1350 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1351 rq->mq_hctx->tags->rqs[rq->tag] = rq;
1352
1353 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1354 blk_integrity_prepare(rq);
1355
1356 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1357 WRITE_ONCE(rq->bio->bi_cookie, rq->mq_hctx->queue_num);
1358}
1359EXPORT_SYMBOL(blk_mq_start_request);
1360
1361/*
1362 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1363 * queues. This is important for md arrays to benefit from merging
1364 * requests.
1365 */
1366static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1367{
1368 if (plug->multiple_queues)
1369 return BLK_MAX_REQUEST_COUNT * 2;
1370 return BLK_MAX_REQUEST_COUNT;
1371}
1372
1373static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1374{
1375 struct request *last = rq_list_peek(&plug->mq_list);
1376
1377 if (!plug->rq_count) {
1378 trace_block_plug(rq->q);
1379 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1380 (!blk_queue_nomerges(rq->q) &&
1381 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1382 blk_mq_flush_plug_list(plug, false);
1383 last = NULL;
1384 trace_block_plug(rq->q);
1385 }
1386
1387 if (!plug->multiple_queues && last && last->q != rq->q)
1388 plug->multiple_queues = true;
1389 /*
1390 * Any request allocated from sched tags can't be issued to
1391 * ->queue_rqs() directly
1392 */
1393 if (!plug->has_elevator && (rq->rq_flags & RQF_SCHED_TAGS))
1394 plug->has_elevator = true;
1395 rq_list_add_tail(&plug->mq_list, rq);
1396 plug->rq_count++;
1397}
1398
1399/**
1400 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1401 * @rq: request to insert
1402 * @at_head: insert request at head or tail of queue
1403 *
1404 * Description:
1405 * Insert a fully prepared request at the back of the I/O scheduler queue
1406 * for execution. Don't wait for completion.
1407 *
1408 * Note:
1409 * This function will invoke @done directly if the queue is dead.
1410 */
1411void blk_execute_rq_nowait(struct request *rq, bool at_head)
1412{
1413 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1414
1415 WARN_ON(irqs_disabled());
1416 WARN_ON(!blk_rq_is_passthrough(rq));
1417
1418 blk_account_io_start(rq);
1419
1420 if (current->plug && !at_head) {
1421 blk_add_rq_to_plug(current->plug, rq);
1422 return;
1423 }
1424
1425 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1426 blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);
1427}
1428EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1429
1430struct blk_rq_wait {
1431 struct completion done;
1432 blk_status_t ret;
1433};
1434
1435static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
1436{
1437 struct blk_rq_wait *wait = rq->end_io_data;
1438
1439 wait->ret = ret;
1440 complete(&wait->done);
1441 return RQ_END_IO_NONE;
1442}
1443
1444bool blk_rq_is_poll(struct request *rq)
1445{
1446 if (!rq->mq_hctx)
1447 return false;
1448 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1449 return false;
1450 return true;
1451}
1452EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1453
1454static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1455{
1456 do {
1457 blk_hctx_poll(rq->q, rq->mq_hctx, NULL, 0);
1458 cond_resched();
1459 } while (!completion_done(wait));
1460}
1461
1462/**
1463 * blk_execute_rq - insert a request into queue for execution
1464 * @rq: request to insert
1465 * @at_head: insert request at head or tail of queue
1466 *
1467 * Description:
1468 * Insert a fully prepared request at the back of the I/O scheduler queue
1469 * for execution and wait for completion.
1470 * Return: The blk_status_t result provided to blk_mq_end_request().
1471 */
1472blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1473{
1474 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1475 struct blk_rq_wait wait = {
1476 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1477 };
1478
1479 WARN_ON(irqs_disabled());
1480 WARN_ON(!blk_rq_is_passthrough(rq));
1481
1482 rq->end_io_data = &wait;
1483 rq->end_io = blk_end_sync_rq;
1484
1485 blk_account_io_start(rq);
1486 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1487 blk_mq_run_hw_queue(hctx, false);
1488
1489 if (blk_rq_is_poll(rq))
1490 blk_rq_poll_completion(rq, &wait.done);
1491 else
1492 blk_wait_io(&wait.done);
1493
1494 return wait.ret;
1495}
1496EXPORT_SYMBOL(blk_execute_rq);
1497
1498static void __blk_mq_requeue_request(struct request *rq)
1499{
1500 struct request_queue *q = rq->q;
1501
1502 blk_mq_put_driver_tag(rq);
1503
1504 trace_block_rq_requeue(rq);
1505 rq_qos_requeue(q, rq);
1506
1507 if (blk_mq_request_started(rq)) {
1508 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1509 rq->rq_flags &= ~RQF_TIMED_OUT;
1510 }
1511}
1512
1513void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1514{
1515 struct request_queue *q = rq->q;
1516 unsigned long flags;
1517
1518 __blk_mq_requeue_request(rq);
1519
1520 /* this request will be re-inserted to io scheduler queue */
1521 blk_mq_sched_requeue_request(rq);
1522
1523 spin_lock_irqsave(&q->requeue_lock, flags);
1524 list_add_tail(&rq->queuelist, &q->requeue_list);
1525 spin_unlock_irqrestore(&q->requeue_lock, flags);
1526
1527 if (kick_requeue_list)
1528 blk_mq_kick_requeue_list(q);
1529}
1530EXPORT_SYMBOL(blk_mq_requeue_request);
1531
1532static void blk_mq_requeue_work(struct work_struct *work)
1533{
1534 struct request_queue *q =
1535 container_of(work, struct request_queue, requeue_work.work);
1536 LIST_HEAD(rq_list);
1537 LIST_HEAD(flush_list);
1538 struct request *rq;
1539
1540 spin_lock_irq(&q->requeue_lock);
1541 list_splice_init(&q->requeue_list, &rq_list);
1542 list_splice_init(&q->flush_list, &flush_list);
1543 spin_unlock_irq(&q->requeue_lock);
1544
1545 while (!list_empty(&rq_list)) {
1546 rq = list_entry(rq_list.next, struct request, queuelist);
1547 list_del_init(&rq->queuelist);
1548 /*
1549 * If RQF_DONTPREP is set, the request has been started by the
1550 * driver already and might have driver-specific data allocated
1551 * already. Insert it into the hctx dispatch list to avoid
1552 * block layer merges for the request.
1553 */
1554 if (rq->rq_flags & RQF_DONTPREP)
1555 blk_mq_request_bypass_insert(rq, 0);
1556 else
1557 blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD);
1558 }
1559
1560 while (!list_empty(&flush_list)) {
1561 rq = list_entry(flush_list.next, struct request, queuelist);
1562 list_del_init(&rq->queuelist);
1563 blk_mq_insert_request(rq, 0);
1564 }
1565
1566 blk_mq_run_hw_queues(q, false);
1567}
1568
1569void blk_mq_kick_requeue_list(struct request_queue *q)
1570{
1571 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1572}
1573EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1574
1575void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1576 unsigned long msecs)
1577{
1578 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1579 msecs_to_jiffies(msecs));
1580}
1581EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1582
1583static bool blk_is_flush_data_rq(struct request *rq)
1584{
1585 return (rq->rq_flags & RQF_FLUSH_SEQ) && !is_flush_rq(rq);
1586}
1587
1588static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1589{
1590 /*
1591 * If we find a request that isn't idle we know the queue is busy
1592 * as it's checked in the iter.
1593 * Return false to stop the iteration.
1594 *
1595 * In case of queue quiesce, if one flush data request is completed,
1596 * don't count it as inflight given the flush sequence is suspended,
1597 * and the original flush data request is invisible to driver, just
1598 * like other pending requests because of quiesce
1599 */
1600 if (blk_mq_request_started(rq) && !(blk_queue_quiesced(rq->q) &&
1601 blk_is_flush_data_rq(rq) &&
1602 blk_mq_request_completed(rq))) {
1603 bool *busy = priv;
1604
1605 *busy = true;
1606 return false;
1607 }
1608
1609 return true;
1610}
1611
1612bool blk_mq_queue_inflight(struct request_queue *q)
1613{
1614 bool busy = false;
1615
1616 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1617 return busy;
1618}
1619EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1620
1621static void blk_mq_rq_timed_out(struct request *req)
1622{
1623 req->rq_flags |= RQF_TIMED_OUT;
1624 if (req->q->mq_ops->timeout) {
1625 enum blk_eh_timer_return ret;
1626
1627 ret = req->q->mq_ops->timeout(req);
1628 if (ret == BLK_EH_DONE)
1629 return;
1630 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1631 }
1632
1633 blk_add_timer(req);
1634}
1635
1636struct blk_expired_data {
1637 bool has_timedout_rq;
1638 unsigned long next;
1639 unsigned long timeout_start;
1640};
1641
1642static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
1643{
1644 unsigned long deadline;
1645
1646 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1647 return false;
1648 if (rq->rq_flags & RQF_TIMED_OUT)
1649 return false;
1650
1651 deadline = READ_ONCE(rq->deadline);
1652 if (time_after_eq(expired->timeout_start, deadline))
1653 return true;
1654
1655 if (expired->next == 0)
1656 expired->next = deadline;
1657 else if (time_after(expired->next, deadline))
1658 expired->next = deadline;
1659 return false;
1660}
1661
1662void blk_mq_put_rq_ref(struct request *rq)
1663{
1664 if (is_flush_rq(rq)) {
1665 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1666 blk_mq_free_request(rq);
1667 } else if (req_ref_put_and_test(rq)) {
1668 __blk_mq_free_request(rq);
1669 }
1670}
1671
1672static bool blk_mq_check_expired(struct request *rq, void *priv)
1673{
1674 struct blk_expired_data *expired = priv;
1675
1676 /*
1677 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1678 * be reallocated underneath the timeout handler's processing, then
1679 * the expire check is reliable. If the request is not expired, then
1680 * it was completed and reallocated as a new request after returning
1681 * from blk_mq_check_expired().
1682 */
1683 if (blk_mq_req_expired(rq, expired)) {
1684 expired->has_timedout_rq = true;
1685 return false;
1686 }
1687 return true;
1688}
1689
1690static bool blk_mq_handle_expired(struct request *rq, void *priv)
1691{
1692 struct blk_expired_data *expired = priv;
1693
1694 if (blk_mq_req_expired(rq, expired))
1695 blk_mq_rq_timed_out(rq);
1696 return true;
1697}
1698
1699static void blk_mq_timeout_work(struct work_struct *work)
1700{
1701 struct request_queue *q =
1702 container_of(work, struct request_queue, timeout_work);
1703 struct blk_expired_data expired = {
1704 .timeout_start = jiffies,
1705 };
1706 struct blk_mq_hw_ctx *hctx;
1707 unsigned long i;
1708
1709 /* A deadlock might occur if a request is stuck requiring a
1710 * timeout at the same time a queue freeze is waiting
1711 * completion, since the timeout code would not be able to
1712 * acquire the queue reference here.
1713 *
1714 * That's why we don't use blk_queue_enter here; instead, we use
1715 * percpu_ref_tryget directly, because we need to be able to
1716 * obtain a reference even in the short window between the queue
1717 * starting to freeze, by dropping the first reference in
1718 * blk_freeze_queue_start, and the moment the last request is
1719 * consumed, marked by the instant q_usage_counter reaches
1720 * zero.
1721 */
1722 if (!percpu_ref_tryget(&q->q_usage_counter))
1723 return;
1724
1725 /* check if there is any timed-out request */
1726 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
1727 if (expired.has_timedout_rq) {
1728 /*
1729 * Before walking tags, we must ensure any submit started
1730 * before the current time has finished. Since the submit
1731 * uses srcu or rcu, wait for a synchronization point to
1732 * ensure all running submits have finished
1733 */
1734 blk_mq_wait_quiesce_done(q->tag_set);
1735
1736 expired.next = 0;
1737 blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
1738 }
1739
1740 if (expired.next != 0) {
1741 mod_timer(&q->timeout, expired.next);
1742 } else {
1743 /*
1744 * Request timeouts are handled as a forward rolling timer. If
1745 * we end up here it means that no requests are pending and
1746 * also that no request has been pending for a while. Mark
1747 * each hctx as idle.
1748 */
1749 queue_for_each_hw_ctx(q, hctx, i) {
1750 /* the hctx may be unmapped, so check it here */
1751 if (blk_mq_hw_queue_mapped(hctx))
1752 blk_mq_tag_idle(hctx);
1753 }
1754 }
1755 blk_queue_exit(q);
1756}
1757
1758struct flush_busy_ctx_data {
1759 struct blk_mq_hw_ctx *hctx;
1760 struct list_head *list;
1761};
1762
1763static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1764{
1765 struct flush_busy_ctx_data *flush_data = data;
1766 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1767 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1768 enum hctx_type type = hctx->type;
1769
1770 spin_lock(&ctx->lock);
1771 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1772 sbitmap_clear_bit(sb, bitnr);
1773 spin_unlock(&ctx->lock);
1774 return true;
1775}
1776
1777/*
1778 * Process software queues that have been marked busy, splicing them
1779 * to the for-dispatch
1780 */
1781void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1782{
1783 struct flush_busy_ctx_data data = {
1784 .hctx = hctx,
1785 .list = list,
1786 };
1787
1788 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1789}
1790
1791struct dispatch_rq_data {
1792 struct blk_mq_hw_ctx *hctx;
1793 struct request *rq;
1794};
1795
1796static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1797 void *data)
1798{
1799 struct dispatch_rq_data *dispatch_data = data;
1800 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1801 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1802 enum hctx_type type = hctx->type;
1803
1804 spin_lock(&ctx->lock);
1805 if (!list_empty(&ctx->rq_lists[type])) {
1806 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1807 list_del_init(&dispatch_data->rq->queuelist);
1808 if (list_empty(&ctx->rq_lists[type]))
1809 sbitmap_clear_bit(sb, bitnr);
1810 }
1811 spin_unlock(&ctx->lock);
1812
1813 return !dispatch_data->rq;
1814}
1815
1816struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1817 struct blk_mq_ctx *start)
1818{
1819 unsigned off = start ? start->index_hw[hctx->type] : 0;
1820 struct dispatch_rq_data data = {
1821 .hctx = hctx,
1822 .rq = NULL,
1823 };
1824
1825 __sbitmap_for_each_set(&hctx->ctx_map, off,
1826 dispatch_rq_from_ctx, &data);
1827
1828 return data.rq;
1829}
1830
1831bool __blk_mq_alloc_driver_tag(struct request *rq)
1832{
1833 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1834 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1835 int tag;
1836
1837 blk_mq_tag_busy(rq->mq_hctx);
1838
1839 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1840 bt = &rq->mq_hctx->tags->breserved_tags;
1841 tag_offset = 0;
1842 } else {
1843 if (!hctx_may_queue(rq->mq_hctx, bt))
1844 return false;
1845 }
1846
1847 tag = __sbitmap_queue_get(bt);
1848 if (tag == BLK_MQ_NO_TAG)
1849 return false;
1850
1851 rq->tag = tag + tag_offset;
1852 blk_mq_inc_active_requests(rq->mq_hctx);
1853 return true;
1854}
1855
1856static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1857 int flags, void *key)
1858{
1859 struct blk_mq_hw_ctx *hctx;
1860
1861 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1862
1863 spin_lock(&hctx->dispatch_wait_lock);
1864 if (!list_empty(&wait->entry)) {
1865 struct sbitmap_queue *sbq;
1866
1867 list_del_init(&wait->entry);
1868 sbq = &hctx->tags->bitmap_tags;
1869 atomic_dec(&sbq->ws_active);
1870 }
1871 spin_unlock(&hctx->dispatch_wait_lock);
1872
1873 blk_mq_run_hw_queue(hctx, true);
1874 return 1;
1875}
1876
1877/*
1878 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1879 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1880 * restart. For both cases, take care to check the condition again after
1881 * marking us as waiting.
1882 */
1883static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1884 struct request *rq)
1885{
1886 struct sbitmap_queue *sbq;
1887 struct wait_queue_head *wq;
1888 wait_queue_entry_t *wait;
1889 bool ret;
1890
1891 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1892 !(blk_mq_is_shared_tags(hctx->flags))) {
1893 blk_mq_sched_mark_restart_hctx(hctx);
1894
1895 /*
1896 * It's possible that a tag was freed in the window between the
1897 * allocation failure and adding the hardware queue to the wait
1898 * queue.
1899 *
1900 * Don't clear RESTART here, someone else could have set it.
1901 * At most this will cost an extra queue run.
1902 */
1903 return blk_mq_get_driver_tag(rq);
1904 }
1905
1906 wait = &hctx->dispatch_wait;
1907 if (!list_empty_careful(&wait->entry))
1908 return false;
1909
1910 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag))
1911 sbq = &hctx->tags->breserved_tags;
1912 else
1913 sbq = &hctx->tags->bitmap_tags;
1914 wq = &bt_wait_ptr(sbq, hctx)->wait;
1915
1916 spin_lock_irq(&wq->lock);
1917 spin_lock(&hctx->dispatch_wait_lock);
1918 if (!list_empty(&wait->entry)) {
1919 spin_unlock(&hctx->dispatch_wait_lock);
1920 spin_unlock_irq(&wq->lock);
1921 return false;
1922 }
1923
1924 atomic_inc(&sbq->ws_active);
1925 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1926 __add_wait_queue(wq, wait);
1927
1928 /*
1929 * Add one explicit barrier since blk_mq_get_driver_tag() may
1930 * not imply barrier in case of failure.
1931 *
1932 * Order adding us to wait queue and allocating driver tag.
1933 *
1934 * The pair is the one implied in sbitmap_queue_wake_up() which
1935 * orders clearing sbitmap tag bits and waitqueue_active() in
1936 * __sbitmap_queue_wake_up(), since waitqueue_active() is lockless
1937 *
1938 * Otherwise, re-order of adding wait queue and getting driver tag
1939 * may cause __sbitmap_queue_wake_up() to wake up nothing because
1940 * the waitqueue_active() may not observe us in wait queue.
1941 */
1942 smp_mb();
1943
1944 /*
1945 * It's possible that a tag was freed in the window between the
1946 * allocation failure and adding the hardware queue to the wait
1947 * queue.
1948 */
1949 ret = blk_mq_get_driver_tag(rq);
1950 if (!ret) {
1951 spin_unlock(&hctx->dispatch_wait_lock);
1952 spin_unlock_irq(&wq->lock);
1953 return false;
1954 }
1955
1956 /*
1957 * We got a tag, remove ourselves from the wait queue to ensure
1958 * someone else gets the wakeup.
1959 */
1960 list_del_init(&wait->entry);
1961 atomic_dec(&sbq->ws_active);
1962 spin_unlock(&hctx->dispatch_wait_lock);
1963 spin_unlock_irq(&wq->lock);
1964
1965 return true;
1966}
1967
1968#define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1969#define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1970/*
1971 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1972 * - EWMA is one simple way to compute running average value
1973 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1974 * - take 4 as factor for avoiding to get too small(0) result, and this
1975 * factor doesn't matter because EWMA decreases exponentially
1976 */
1977static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1978{
1979 unsigned int ewma;
1980
1981 ewma = hctx->dispatch_busy;
1982
1983 if (!ewma && !busy)
1984 return;
1985
1986 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1987 if (busy)
1988 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1989 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1990
1991 hctx->dispatch_busy = ewma;
1992}
1993
1994#define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1995
1996static void blk_mq_handle_dev_resource(struct request *rq,
1997 struct list_head *list)
1998{
1999 list_add(&rq->queuelist, list);
2000 __blk_mq_requeue_request(rq);
2001}
2002
2003enum prep_dispatch {
2004 PREP_DISPATCH_OK,
2005 PREP_DISPATCH_NO_TAG,
2006 PREP_DISPATCH_NO_BUDGET,
2007};
2008
2009static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
2010 bool need_budget)
2011{
2012 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2013 int budget_token = -1;
2014
2015 if (need_budget) {
2016 budget_token = blk_mq_get_dispatch_budget(rq->q);
2017 if (budget_token < 0) {
2018 blk_mq_put_driver_tag(rq);
2019 return PREP_DISPATCH_NO_BUDGET;
2020 }
2021 blk_mq_set_rq_budget_token(rq, budget_token);
2022 }
2023
2024 if (!blk_mq_get_driver_tag(rq)) {
2025 /*
2026 * The initial allocation attempt failed, so we need to
2027 * rerun the hardware queue when a tag is freed. The
2028 * waitqueue takes care of that. If the queue is run
2029 * before we add this entry back on the dispatch list,
2030 * we'll re-run it below.
2031 */
2032 if (!blk_mq_mark_tag_wait(hctx, rq)) {
2033 /*
2034 * All budgets not got from this function will be put
2035 * together during handling partial dispatch
2036 */
2037 if (need_budget)
2038 blk_mq_put_dispatch_budget(rq->q, budget_token);
2039 return PREP_DISPATCH_NO_TAG;
2040 }
2041 }
2042
2043 return PREP_DISPATCH_OK;
2044}
2045
2046/* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
2047static void blk_mq_release_budgets(struct request_queue *q,
2048 struct list_head *list)
2049{
2050 struct request *rq;
2051
2052 list_for_each_entry(rq, list, queuelist) {
2053 int budget_token = blk_mq_get_rq_budget_token(rq);
2054
2055 if (budget_token >= 0)
2056 blk_mq_put_dispatch_budget(q, budget_token);
2057 }
2058}
2059
2060/*
2061 * blk_mq_commit_rqs will notify driver using bd->last that there is no
2062 * more requests. (See comment in struct blk_mq_ops for commit_rqs for
2063 * details)
2064 * Attention, we should explicitly call this in unusual cases:
2065 * 1) did not queue everything initially scheduled to queue
2066 * 2) the last attempt to queue a request failed
2067 */
2068static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued,
2069 bool from_schedule)
2070{
2071 if (hctx->queue->mq_ops->commit_rqs && queued) {
2072 trace_block_unplug(hctx->queue, queued, !from_schedule);
2073 hctx->queue->mq_ops->commit_rqs(hctx);
2074 }
2075}
2076
2077/*
2078 * Returns true if we did some work AND can potentially do more.
2079 */
2080bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
2081 unsigned int nr_budgets)
2082{
2083 enum prep_dispatch prep;
2084 struct request_queue *q = hctx->queue;
2085 struct request *rq;
2086 int queued;
2087 blk_status_t ret = BLK_STS_OK;
2088 bool needs_resource = false;
2089
2090 if (list_empty(list))
2091 return false;
2092
2093 /*
2094 * Now process all the entries, sending them to the driver.
2095 */
2096 queued = 0;
2097 do {
2098 struct blk_mq_queue_data bd;
2099
2100 rq = list_first_entry(list, struct request, queuelist);
2101
2102 WARN_ON_ONCE(hctx != rq->mq_hctx);
2103 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
2104 if (prep != PREP_DISPATCH_OK)
2105 break;
2106
2107 list_del_init(&rq->queuelist);
2108
2109 bd.rq = rq;
2110 bd.last = list_empty(list);
2111
2112 /*
2113 * once the request is queued to lld, no need to cover the
2114 * budget any more
2115 */
2116 if (nr_budgets)
2117 nr_budgets--;
2118 ret = q->mq_ops->queue_rq(hctx, &bd);
2119 switch (ret) {
2120 case BLK_STS_OK:
2121 queued++;
2122 break;
2123 case BLK_STS_RESOURCE:
2124 needs_resource = true;
2125 fallthrough;
2126 case BLK_STS_DEV_RESOURCE:
2127 blk_mq_handle_dev_resource(rq, list);
2128 goto out;
2129 default:
2130 blk_mq_end_request(rq, ret);
2131 }
2132 } while (!list_empty(list));
2133out:
2134 /* If we didn't flush the entire list, we could have told the driver
2135 * there was more coming, but that turned out to be a lie.
2136 */
2137 if (!list_empty(list) || ret != BLK_STS_OK)
2138 blk_mq_commit_rqs(hctx, queued, false);
2139
2140 /*
2141 * Any items that need requeuing? Stuff them into hctx->dispatch,
2142 * that is where we will continue on next queue run.
2143 */
2144 if (!list_empty(list)) {
2145 bool needs_restart;
2146 /* For non-shared tags, the RESTART check will suffice */
2147 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2148 ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
2149 blk_mq_is_shared_tags(hctx->flags));
2150
2151 if (nr_budgets)
2152 blk_mq_release_budgets(q, list);
2153
2154 spin_lock(&hctx->lock);
2155 list_splice_tail_init(list, &hctx->dispatch);
2156 spin_unlock(&hctx->lock);
2157
2158 /*
2159 * Order adding requests to hctx->dispatch and checking
2160 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2161 * in blk_mq_sched_restart(). Avoid restart code path to
2162 * miss the new added requests to hctx->dispatch, meantime
2163 * SCHED_RESTART is observed here.
2164 */
2165 smp_mb();
2166
2167 /*
2168 * If SCHED_RESTART was set by the caller of this function and
2169 * it is no longer set that means that it was cleared by another
2170 * thread and hence that a queue rerun is needed.
2171 *
2172 * If 'no_tag' is set, that means that we failed getting
2173 * a driver tag with an I/O scheduler attached. If our dispatch
2174 * waitqueue is no longer active, ensure that we run the queue
2175 * AFTER adding our entries back to the list.
2176 *
2177 * If no I/O scheduler has been configured it is possible that
2178 * the hardware queue got stopped and restarted before requests
2179 * were pushed back onto the dispatch list. Rerun the queue to
2180 * avoid starvation. Notes:
2181 * - blk_mq_run_hw_queue() checks whether or not a queue has
2182 * been stopped before rerunning a queue.
2183 * - Some but not all block drivers stop a queue before
2184 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2185 * and dm-rq.
2186 *
2187 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2188 * bit is set, run queue after a delay to avoid IO stalls
2189 * that could otherwise occur if the queue is idle. We'll do
2190 * similar if we couldn't get budget or couldn't lock a zone
2191 * and SCHED_RESTART is set.
2192 */
2193 needs_restart = blk_mq_sched_needs_restart(hctx);
2194 if (prep == PREP_DISPATCH_NO_BUDGET)
2195 needs_resource = true;
2196 if (!needs_restart ||
2197 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2198 blk_mq_run_hw_queue(hctx, true);
2199 else if (needs_resource)
2200 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2201
2202 blk_mq_update_dispatch_busy(hctx, true);
2203 return false;
2204 }
2205
2206 blk_mq_update_dispatch_busy(hctx, false);
2207 return true;
2208}
2209
2210static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2211{
2212 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2213
2214 if (cpu >= nr_cpu_ids)
2215 cpu = cpumask_first(hctx->cpumask);
2216 return cpu;
2217}
2218
2219/*
2220 * ->next_cpu is always calculated from hctx->cpumask, so simply use
2221 * it for speeding up the check
2222 */
2223static bool blk_mq_hctx_empty_cpumask(struct blk_mq_hw_ctx *hctx)
2224{
2225 return hctx->next_cpu >= nr_cpu_ids;
2226}
2227
2228/*
2229 * It'd be great if the workqueue API had a way to pass
2230 * in a mask and had some smarts for more clever placement.
2231 * For now we just round-robin here, switching for every
2232 * BLK_MQ_CPU_WORK_BATCH queued items.
2233 */
2234static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2235{
2236 bool tried = false;
2237 int next_cpu = hctx->next_cpu;
2238
2239 /* Switch to unbound if no allowable CPUs in this hctx */
2240 if (hctx->queue->nr_hw_queues == 1 || blk_mq_hctx_empty_cpumask(hctx))
2241 return WORK_CPU_UNBOUND;
2242
2243 if (--hctx->next_cpu_batch <= 0) {
2244select_cpu:
2245 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2246 cpu_online_mask);
2247 if (next_cpu >= nr_cpu_ids)
2248 next_cpu = blk_mq_first_mapped_cpu(hctx);
2249 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2250 }
2251
2252 /*
2253 * Do unbound schedule if we can't find a online CPU for this hctx,
2254 * and it should only happen in the path of handling CPU DEAD.
2255 */
2256 if (!cpu_online(next_cpu)) {
2257 if (!tried) {
2258 tried = true;
2259 goto select_cpu;
2260 }
2261
2262 /*
2263 * Make sure to re-select CPU next time once after CPUs
2264 * in hctx->cpumask become online again.
2265 */
2266 hctx->next_cpu = next_cpu;
2267 hctx->next_cpu_batch = 1;
2268 return WORK_CPU_UNBOUND;
2269 }
2270
2271 hctx->next_cpu = next_cpu;
2272 return next_cpu;
2273}
2274
2275/**
2276 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2277 * @hctx: Pointer to the hardware queue to run.
2278 * @msecs: Milliseconds of delay to wait before running the queue.
2279 *
2280 * Run a hardware queue asynchronously with a delay of @msecs.
2281 */
2282void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2283{
2284 if (unlikely(blk_mq_hctx_stopped(hctx)))
2285 return;
2286 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2287 msecs_to_jiffies(msecs));
2288}
2289EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2290
2291static inline bool blk_mq_hw_queue_need_run(struct blk_mq_hw_ctx *hctx)
2292{
2293 bool need_run;
2294
2295 /*
2296 * When queue is quiesced, we may be switching io scheduler, or
2297 * updating nr_hw_queues, or other things, and we can't run queue
2298 * any more, even blk_mq_hctx_has_pending() can't be called safely.
2299 *
2300 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2301 * quiesced.
2302 */
2303 __blk_mq_run_dispatch_ops(hctx->queue, false,
2304 need_run = !blk_queue_quiesced(hctx->queue) &&
2305 blk_mq_hctx_has_pending(hctx));
2306 return need_run;
2307}
2308
2309/**
2310 * blk_mq_run_hw_queue - Start to run a hardware queue.
2311 * @hctx: Pointer to the hardware queue to run.
2312 * @async: If we want to run the queue asynchronously.
2313 *
2314 * Check if the request queue is not in a quiesced state and if there are
2315 * pending requests to be sent. If this is true, run the queue to send requests
2316 * to hardware.
2317 */
2318void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2319{
2320 bool need_run;
2321
2322 /*
2323 * We can't run the queue inline with interrupts disabled.
2324 */
2325 WARN_ON_ONCE(!async && in_interrupt());
2326
2327 might_sleep_if(!async && hctx->flags & BLK_MQ_F_BLOCKING);
2328
2329 need_run = blk_mq_hw_queue_need_run(hctx);
2330 if (!need_run) {
2331 unsigned long flags;
2332
2333 /*
2334 * Synchronize with blk_mq_unquiesce_queue(), because we check
2335 * if hw queue is quiesced locklessly above, we need the use
2336 * ->queue_lock to make sure we see the up-to-date status to
2337 * not miss rerunning the hw queue.
2338 */
2339 spin_lock_irqsave(&hctx->queue->queue_lock, flags);
2340 need_run = blk_mq_hw_queue_need_run(hctx);
2341 spin_unlock_irqrestore(&hctx->queue->queue_lock, flags);
2342
2343 if (!need_run)
2344 return;
2345 }
2346
2347 if (async || !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2348 blk_mq_delay_run_hw_queue(hctx, 0);
2349 return;
2350 }
2351
2352 blk_mq_run_dispatch_ops(hctx->queue,
2353 blk_mq_sched_dispatch_requests(hctx));
2354}
2355EXPORT_SYMBOL(blk_mq_run_hw_queue);
2356
2357/*
2358 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2359 * scheduler.
2360 */
2361static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2362{
2363 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2364 /*
2365 * If the IO scheduler does not respect hardware queues when
2366 * dispatching, we just don't bother with multiple HW queues and
2367 * dispatch from hctx for the current CPU since running multiple queues
2368 * just causes lock contention inside the scheduler and pointless cache
2369 * bouncing.
2370 */
2371 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2372
2373 if (!blk_mq_hctx_stopped(hctx))
2374 return hctx;
2375 return NULL;
2376}
2377
2378/**
2379 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2380 * @q: Pointer to the request queue to run.
2381 * @async: If we want to run the queue asynchronously.
2382 */
2383void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2384{
2385 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2386 unsigned long i;
2387
2388 sq_hctx = NULL;
2389 if (blk_queue_sq_sched(q))
2390 sq_hctx = blk_mq_get_sq_hctx(q);
2391 queue_for_each_hw_ctx(q, hctx, i) {
2392 if (blk_mq_hctx_stopped(hctx))
2393 continue;
2394 /*
2395 * Dispatch from this hctx either if there's no hctx preferred
2396 * by IO scheduler or if it has requests that bypass the
2397 * scheduler.
2398 */
2399 if (!sq_hctx || sq_hctx == hctx ||
2400 !list_empty_careful(&hctx->dispatch))
2401 blk_mq_run_hw_queue(hctx, async);
2402 }
2403}
2404EXPORT_SYMBOL(blk_mq_run_hw_queues);
2405
2406/**
2407 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2408 * @q: Pointer to the request queue to run.
2409 * @msecs: Milliseconds of delay to wait before running the queues.
2410 */
2411void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2412{
2413 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2414 unsigned long i;
2415
2416 sq_hctx = NULL;
2417 if (blk_queue_sq_sched(q))
2418 sq_hctx = blk_mq_get_sq_hctx(q);
2419 queue_for_each_hw_ctx(q, hctx, i) {
2420 if (blk_mq_hctx_stopped(hctx))
2421 continue;
2422 /*
2423 * If there is already a run_work pending, leave the
2424 * pending delay untouched. Otherwise, a hctx can stall
2425 * if another hctx is re-delaying the other's work
2426 * before the work executes.
2427 */
2428 if (delayed_work_pending(&hctx->run_work))
2429 continue;
2430 /*
2431 * Dispatch from this hctx either if there's no hctx preferred
2432 * by IO scheduler or if it has requests that bypass the
2433 * scheduler.
2434 */
2435 if (!sq_hctx || sq_hctx == hctx ||
2436 !list_empty_careful(&hctx->dispatch))
2437 blk_mq_delay_run_hw_queue(hctx, msecs);
2438 }
2439}
2440EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2441
2442/*
2443 * This function is often used for pausing .queue_rq() by driver when
2444 * there isn't enough resource or some conditions aren't satisfied, and
2445 * BLK_STS_RESOURCE is usually returned.
2446 *
2447 * We do not guarantee that dispatch can be drained or blocked
2448 * after blk_mq_stop_hw_queue() returns. Please use
2449 * blk_mq_quiesce_queue() for that requirement.
2450 */
2451void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2452{
2453 cancel_delayed_work(&hctx->run_work);
2454
2455 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2456}
2457EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2458
2459/*
2460 * This function is often used for pausing .queue_rq() by driver when
2461 * there isn't enough resource or some conditions aren't satisfied, and
2462 * BLK_STS_RESOURCE is usually returned.
2463 *
2464 * We do not guarantee that dispatch can be drained or blocked
2465 * after blk_mq_stop_hw_queues() returns. Please use
2466 * blk_mq_quiesce_queue() for that requirement.
2467 */
2468void blk_mq_stop_hw_queues(struct request_queue *q)
2469{
2470 struct blk_mq_hw_ctx *hctx;
2471 unsigned long i;
2472
2473 queue_for_each_hw_ctx(q, hctx, i)
2474 blk_mq_stop_hw_queue(hctx);
2475}
2476EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2477
2478void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2479{
2480 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2481
2482 blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);
2483}
2484EXPORT_SYMBOL(blk_mq_start_hw_queue);
2485
2486void blk_mq_start_hw_queues(struct request_queue *q)
2487{
2488 struct blk_mq_hw_ctx *hctx;
2489 unsigned long i;
2490
2491 queue_for_each_hw_ctx(q, hctx, i)
2492 blk_mq_start_hw_queue(hctx);
2493}
2494EXPORT_SYMBOL(blk_mq_start_hw_queues);
2495
2496void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2497{
2498 if (!blk_mq_hctx_stopped(hctx))
2499 return;
2500
2501 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2502 /*
2503 * Pairs with the smp_mb() in blk_mq_hctx_stopped() to order the
2504 * clearing of BLK_MQ_S_STOPPED above and the checking of dispatch
2505 * list in the subsequent routine.
2506 */
2507 smp_mb__after_atomic();
2508 blk_mq_run_hw_queue(hctx, async);
2509}
2510EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2511
2512void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2513{
2514 struct blk_mq_hw_ctx *hctx;
2515 unsigned long i;
2516
2517 queue_for_each_hw_ctx(q, hctx, i)
2518 blk_mq_start_stopped_hw_queue(hctx, async ||
2519 (hctx->flags & BLK_MQ_F_BLOCKING));
2520}
2521EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2522
2523static void blk_mq_run_work_fn(struct work_struct *work)
2524{
2525 struct blk_mq_hw_ctx *hctx =
2526 container_of(work, struct blk_mq_hw_ctx, run_work.work);
2527
2528 blk_mq_run_dispatch_ops(hctx->queue,
2529 blk_mq_sched_dispatch_requests(hctx));
2530}
2531
2532/**
2533 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2534 * @rq: Pointer to request to be inserted.
2535 * @flags: BLK_MQ_INSERT_*
2536 *
2537 * Should only be used carefully, when the caller knows we want to
2538 * bypass a potential IO scheduler on the target device.
2539 */
2540static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags)
2541{
2542 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2543
2544 spin_lock(&hctx->lock);
2545 if (flags & BLK_MQ_INSERT_AT_HEAD)
2546 list_add(&rq->queuelist, &hctx->dispatch);
2547 else
2548 list_add_tail(&rq->queuelist, &hctx->dispatch);
2549 spin_unlock(&hctx->lock);
2550}
2551
2552static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx,
2553 struct blk_mq_ctx *ctx, struct list_head *list,
2554 bool run_queue_async)
2555{
2556 struct request *rq;
2557 enum hctx_type type = hctx->type;
2558
2559 /*
2560 * Try to issue requests directly if the hw queue isn't busy to save an
2561 * extra enqueue & dequeue to the sw queue.
2562 */
2563 if (!hctx->dispatch_busy && !run_queue_async) {
2564 blk_mq_run_dispatch_ops(hctx->queue,
2565 blk_mq_try_issue_list_directly(hctx, list));
2566 if (list_empty(list))
2567 goto out;
2568 }
2569
2570 /*
2571 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2572 * offline now
2573 */
2574 list_for_each_entry(rq, list, queuelist) {
2575 BUG_ON(rq->mq_ctx != ctx);
2576 trace_block_rq_insert(rq);
2577 if (rq->cmd_flags & REQ_NOWAIT)
2578 run_queue_async = true;
2579 }
2580
2581 spin_lock(&ctx->lock);
2582 list_splice_tail_init(list, &ctx->rq_lists[type]);
2583 blk_mq_hctx_mark_pending(hctx, ctx);
2584 spin_unlock(&ctx->lock);
2585out:
2586 blk_mq_run_hw_queue(hctx, run_queue_async);
2587}
2588
2589static void blk_mq_insert_request(struct request *rq, blk_insert_t flags)
2590{
2591 struct request_queue *q = rq->q;
2592 struct blk_mq_ctx *ctx = rq->mq_ctx;
2593 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2594
2595 if (blk_rq_is_passthrough(rq)) {
2596 /*
2597 * Passthrough request have to be added to hctx->dispatch
2598 * directly. The device may be in a situation where it can't
2599 * handle FS request, and always returns BLK_STS_RESOURCE for
2600 * them, which gets them added to hctx->dispatch.
2601 *
2602 * If a passthrough request is required to unblock the queues,
2603 * and it is added to the scheduler queue, there is no chance to
2604 * dispatch it given we prioritize requests in hctx->dispatch.
2605 */
2606 blk_mq_request_bypass_insert(rq, flags);
2607 } else if (req_op(rq) == REQ_OP_FLUSH) {
2608 /*
2609 * Firstly normal IO request is inserted to scheduler queue or
2610 * sw queue, meantime we add flush request to dispatch queue(
2611 * hctx->dispatch) directly and there is at most one in-flight
2612 * flush request for each hw queue, so it doesn't matter to add
2613 * flush request to tail or front of the dispatch queue.
2614 *
2615 * Secondly in case of NCQ, flush request belongs to non-NCQ
2616 * command, and queueing it will fail when there is any
2617 * in-flight normal IO request(NCQ command). When adding flush
2618 * rq to the front of hctx->dispatch, it is easier to introduce
2619 * extra time to flush rq's latency because of S_SCHED_RESTART
2620 * compared with adding to the tail of dispatch queue, then
2621 * chance of flush merge is increased, and less flush requests
2622 * will be issued to controller. It is observed that ~10% time
2623 * is saved in blktests block/004 on disk attached to AHCI/NCQ
2624 * drive when adding flush rq to the front of hctx->dispatch.
2625 *
2626 * Simply queue flush rq to the front of hctx->dispatch so that
2627 * intensive flush workloads can benefit in case of NCQ HW.
2628 */
2629 blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD);
2630 } else if (q->elevator) {
2631 LIST_HEAD(list);
2632
2633 WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG);
2634
2635 list_add(&rq->queuelist, &list);
2636 q->elevator->type->ops.insert_requests(hctx, &list, flags);
2637 } else {
2638 trace_block_rq_insert(rq);
2639
2640 spin_lock(&ctx->lock);
2641 if (flags & BLK_MQ_INSERT_AT_HEAD)
2642 list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]);
2643 else
2644 list_add_tail(&rq->queuelist,
2645 &ctx->rq_lists[hctx->type]);
2646 blk_mq_hctx_mark_pending(hctx, ctx);
2647 spin_unlock(&ctx->lock);
2648 }
2649}
2650
2651static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2652 unsigned int nr_segs)
2653{
2654 int err;
2655
2656 if (bio->bi_opf & REQ_RAHEAD)
2657 rq->cmd_flags |= REQ_FAILFAST_MASK;
2658
2659 rq->__sector = bio->bi_iter.bi_sector;
2660 blk_rq_bio_prep(rq, bio, nr_segs);
2661 if (bio_integrity(bio))
2662 rq->nr_integrity_segments = blk_rq_count_integrity_sg(rq->q,
2663 bio);
2664
2665 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2666 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2667 WARN_ON_ONCE(err);
2668
2669 blk_account_io_start(rq);
2670}
2671
2672static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2673 struct request *rq, bool last)
2674{
2675 struct request_queue *q = rq->q;
2676 struct blk_mq_queue_data bd = {
2677 .rq = rq,
2678 .last = last,
2679 };
2680 blk_status_t ret;
2681
2682 /*
2683 * For OK queue, we are done. For error, caller may kill it.
2684 * Any other error (busy), just add it to our list as we
2685 * previously would have done.
2686 */
2687 ret = q->mq_ops->queue_rq(hctx, &bd);
2688 switch (ret) {
2689 case BLK_STS_OK:
2690 blk_mq_update_dispatch_busy(hctx, false);
2691 break;
2692 case BLK_STS_RESOURCE:
2693 case BLK_STS_DEV_RESOURCE:
2694 blk_mq_update_dispatch_busy(hctx, true);
2695 __blk_mq_requeue_request(rq);
2696 break;
2697 default:
2698 blk_mq_update_dispatch_busy(hctx, false);
2699 break;
2700 }
2701
2702 return ret;
2703}
2704
2705static bool blk_mq_get_budget_and_tag(struct request *rq)
2706{
2707 int budget_token;
2708
2709 budget_token = blk_mq_get_dispatch_budget(rq->q);
2710 if (budget_token < 0)
2711 return false;
2712 blk_mq_set_rq_budget_token(rq, budget_token);
2713 if (!blk_mq_get_driver_tag(rq)) {
2714 blk_mq_put_dispatch_budget(rq->q, budget_token);
2715 return false;
2716 }
2717 return true;
2718}
2719
2720/**
2721 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2722 * @hctx: Pointer of the associated hardware queue.
2723 * @rq: Pointer to request to be sent.
2724 *
2725 * If the device has enough resources to accept a new request now, send the
2726 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2727 * we can try send it another time in the future. Requests inserted at this
2728 * queue have higher priority.
2729 */
2730static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2731 struct request *rq)
2732{
2733 blk_status_t ret;
2734
2735 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2736 blk_mq_insert_request(rq, 0);
2737 blk_mq_run_hw_queue(hctx, false);
2738 return;
2739 }
2740
2741 if ((rq->rq_flags & RQF_USE_SCHED) || !blk_mq_get_budget_and_tag(rq)) {
2742 blk_mq_insert_request(rq, 0);
2743 blk_mq_run_hw_queue(hctx, rq->cmd_flags & REQ_NOWAIT);
2744 return;
2745 }
2746
2747 ret = __blk_mq_issue_directly(hctx, rq, true);
2748 switch (ret) {
2749 case BLK_STS_OK:
2750 break;
2751 case BLK_STS_RESOURCE:
2752 case BLK_STS_DEV_RESOURCE:
2753 blk_mq_request_bypass_insert(rq, 0);
2754 blk_mq_run_hw_queue(hctx, false);
2755 break;
2756 default:
2757 blk_mq_end_request(rq, ret);
2758 break;
2759 }
2760}
2761
2762static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2763{
2764 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2765
2766 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2767 blk_mq_insert_request(rq, 0);
2768 blk_mq_run_hw_queue(hctx, false);
2769 return BLK_STS_OK;
2770 }
2771
2772 if (!blk_mq_get_budget_and_tag(rq))
2773 return BLK_STS_RESOURCE;
2774 return __blk_mq_issue_directly(hctx, rq, last);
2775}
2776
2777static void blk_mq_plug_issue_direct(struct blk_plug *plug)
2778{
2779 struct blk_mq_hw_ctx *hctx = NULL;
2780 struct request *rq;
2781 int queued = 0;
2782 blk_status_t ret = BLK_STS_OK;
2783
2784 while ((rq = rq_list_pop(&plug->mq_list))) {
2785 bool last = rq_list_empty(&plug->mq_list);
2786
2787 if (hctx != rq->mq_hctx) {
2788 if (hctx) {
2789 blk_mq_commit_rqs(hctx, queued, false);
2790 queued = 0;
2791 }
2792 hctx = rq->mq_hctx;
2793 }
2794
2795 ret = blk_mq_request_issue_directly(rq, last);
2796 switch (ret) {
2797 case BLK_STS_OK:
2798 queued++;
2799 break;
2800 case BLK_STS_RESOURCE:
2801 case BLK_STS_DEV_RESOURCE:
2802 blk_mq_request_bypass_insert(rq, 0);
2803 blk_mq_run_hw_queue(hctx, false);
2804 goto out;
2805 default:
2806 blk_mq_end_request(rq, ret);
2807 break;
2808 }
2809 }
2810
2811out:
2812 if (ret != BLK_STS_OK)
2813 blk_mq_commit_rqs(hctx, queued, false);
2814}
2815
2816static void __blk_mq_flush_plug_list(struct request_queue *q,
2817 struct blk_plug *plug)
2818{
2819 if (blk_queue_quiesced(q))
2820 return;
2821 q->mq_ops->queue_rqs(&plug->mq_list);
2822}
2823
2824static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2825{
2826 struct blk_mq_hw_ctx *this_hctx = NULL;
2827 struct blk_mq_ctx *this_ctx = NULL;
2828 struct rq_list requeue_list = {};
2829 unsigned int depth = 0;
2830 bool is_passthrough = false;
2831 LIST_HEAD(list);
2832
2833 do {
2834 struct request *rq = rq_list_pop(&plug->mq_list);
2835
2836 if (!this_hctx) {
2837 this_hctx = rq->mq_hctx;
2838 this_ctx = rq->mq_ctx;
2839 is_passthrough = blk_rq_is_passthrough(rq);
2840 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx ||
2841 is_passthrough != blk_rq_is_passthrough(rq)) {
2842 rq_list_add_tail(&requeue_list, rq);
2843 continue;
2844 }
2845 list_add_tail(&rq->queuelist, &list);
2846 depth++;
2847 } while (!rq_list_empty(&plug->mq_list));
2848
2849 plug->mq_list = requeue_list;
2850 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2851
2852 percpu_ref_get(&this_hctx->queue->q_usage_counter);
2853 /* passthrough requests should never be issued to the I/O scheduler */
2854 if (is_passthrough) {
2855 spin_lock(&this_hctx->lock);
2856 list_splice_tail_init(&list, &this_hctx->dispatch);
2857 spin_unlock(&this_hctx->lock);
2858 blk_mq_run_hw_queue(this_hctx, from_sched);
2859 } else if (this_hctx->queue->elevator) {
2860 this_hctx->queue->elevator->type->ops.insert_requests(this_hctx,
2861 &list, 0);
2862 blk_mq_run_hw_queue(this_hctx, from_sched);
2863 } else {
2864 blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched);
2865 }
2866 percpu_ref_put(&this_hctx->queue->q_usage_counter);
2867}
2868
2869void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2870{
2871 struct request *rq;
2872 unsigned int depth;
2873
2874 /*
2875 * We may have been called recursively midway through handling
2876 * plug->mq_list via a schedule() in the driver's queue_rq() callback.
2877 * To avoid mq_list changing under our feet, clear rq_count early and
2878 * bail out specifically if rq_count is 0 rather than checking
2879 * whether the mq_list is empty.
2880 */
2881 if (plug->rq_count == 0)
2882 return;
2883 depth = plug->rq_count;
2884 plug->rq_count = 0;
2885
2886 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2887 struct request_queue *q;
2888
2889 rq = rq_list_peek(&plug->mq_list);
2890 q = rq->q;
2891 trace_block_unplug(q, depth, true);
2892
2893 /*
2894 * Peek first request and see if we have a ->queue_rqs() hook.
2895 * If we do, we can dispatch the whole plug list in one go. We
2896 * already know at this point that all requests belong to the
2897 * same queue, caller must ensure that's the case.
2898 */
2899 if (q->mq_ops->queue_rqs) {
2900 blk_mq_run_dispatch_ops(q,
2901 __blk_mq_flush_plug_list(q, plug));
2902 if (rq_list_empty(&plug->mq_list))
2903 return;
2904 }
2905
2906 blk_mq_run_dispatch_ops(q,
2907 blk_mq_plug_issue_direct(plug));
2908 if (rq_list_empty(&plug->mq_list))
2909 return;
2910 }
2911
2912 do {
2913 blk_mq_dispatch_plug_list(plug, from_schedule);
2914 } while (!rq_list_empty(&plug->mq_list));
2915}
2916
2917static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2918 struct list_head *list)
2919{
2920 int queued = 0;
2921 blk_status_t ret = BLK_STS_OK;
2922
2923 while (!list_empty(list)) {
2924 struct request *rq = list_first_entry(list, struct request,
2925 queuelist);
2926
2927 list_del_init(&rq->queuelist);
2928 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2929 switch (ret) {
2930 case BLK_STS_OK:
2931 queued++;
2932 break;
2933 case BLK_STS_RESOURCE:
2934 case BLK_STS_DEV_RESOURCE:
2935 blk_mq_request_bypass_insert(rq, 0);
2936 if (list_empty(list))
2937 blk_mq_run_hw_queue(hctx, false);
2938 goto out;
2939 default:
2940 blk_mq_end_request(rq, ret);
2941 break;
2942 }
2943 }
2944
2945out:
2946 if (ret != BLK_STS_OK)
2947 blk_mq_commit_rqs(hctx, queued, false);
2948}
2949
2950static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2951 struct bio *bio, unsigned int nr_segs)
2952{
2953 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2954 if (blk_attempt_plug_merge(q, bio, nr_segs))
2955 return true;
2956 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2957 return true;
2958 }
2959 return false;
2960}
2961
2962static struct request *blk_mq_get_new_requests(struct request_queue *q,
2963 struct blk_plug *plug,
2964 struct bio *bio,
2965 unsigned int nsegs)
2966{
2967 struct blk_mq_alloc_data data = {
2968 .q = q,
2969 .nr_tags = 1,
2970 .cmd_flags = bio->bi_opf,
2971 };
2972 struct request *rq;
2973
2974 rq_qos_throttle(q, bio);
2975
2976 if (plug) {
2977 data.nr_tags = plug->nr_ios;
2978 plug->nr_ios = 1;
2979 data.cached_rqs = &plug->cached_rqs;
2980 }
2981
2982 rq = __blk_mq_alloc_requests(&data);
2983 if (rq)
2984 return rq;
2985 rq_qos_cleanup(q, bio);
2986 if (bio->bi_opf & REQ_NOWAIT)
2987 bio_wouldblock_error(bio);
2988 return NULL;
2989}
2990
2991/*
2992 * Check if there is a suitable cached request and return it.
2993 */
2994static struct request *blk_mq_peek_cached_request(struct blk_plug *plug,
2995 struct request_queue *q, blk_opf_t opf)
2996{
2997 enum hctx_type type = blk_mq_get_hctx_type(opf);
2998 struct request *rq;
2999
3000 if (!plug)
3001 return NULL;
3002 rq = rq_list_peek(&plug->cached_rqs);
3003 if (!rq || rq->q != q)
3004 return NULL;
3005 if (type != rq->mq_hctx->type &&
3006 (type != HCTX_TYPE_READ || rq->mq_hctx->type != HCTX_TYPE_DEFAULT))
3007 return NULL;
3008 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
3009 return NULL;
3010 return rq;
3011}
3012
3013static void blk_mq_use_cached_rq(struct request *rq, struct blk_plug *plug,
3014 struct bio *bio)
3015{
3016 if (rq_list_pop(&plug->cached_rqs) != rq)
3017 WARN_ON_ONCE(1);
3018
3019 /*
3020 * If any qos ->throttle() end up blocking, we will have flushed the
3021 * plug and hence killed the cached_rq list as well. Pop this entry
3022 * before we throttle.
3023 */
3024 rq_qos_throttle(rq->q, bio);
3025
3026 blk_mq_rq_time_init(rq, blk_time_get_ns());
3027 rq->cmd_flags = bio->bi_opf;
3028 INIT_LIST_HEAD(&rq->queuelist);
3029}
3030
3031static bool bio_unaligned(const struct bio *bio, struct request_queue *q)
3032{
3033 unsigned int bs_mask = queue_logical_block_size(q) - 1;
3034
3035 /* .bi_sector of any zero sized bio need to be initialized */
3036 if ((bio->bi_iter.bi_size & bs_mask) ||
3037 ((bio->bi_iter.bi_sector << SECTOR_SHIFT) & bs_mask))
3038 return true;
3039 return false;
3040}
3041
3042/**
3043 * blk_mq_submit_bio - Create and send a request to block device.
3044 * @bio: Bio pointer.
3045 *
3046 * Builds up a request structure from @q and @bio and send to the device. The
3047 * request may not be queued directly to hardware if:
3048 * * This request can be merged with another one
3049 * * We want to place request at plug queue for possible future merging
3050 * * There is an IO scheduler active at this queue
3051 *
3052 * It will not queue the request if there is an error with the bio, or at the
3053 * request creation.
3054 */
3055void blk_mq_submit_bio(struct bio *bio)
3056{
3057 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
3058 struct blk_plug *plug = current->plug;
3059 const int is_sync = op_is_sync(bio->bi_opf);
3060 struct blk_mq_hw_ctx *hctx;
3061 unsigned int nr_segs;
3062 struct request *rq;
3063 blk_status_t ret;
3064
3065 /*
3066 * If the plug has a cached request for this queue, try to use it.
3067 */
3068 rq = blk_mq_peek_cached_request(plug, q, bio->bi_opf);
3069
3070 /*
3071 * A BIO that was released from a zone write plug has already been
3072 * through the preparation in this function, already holds a reference
3073 * on the queue usage counter, and is the only write BIO in-flight for
3074 * the target zone. Go straight to preparing a request for it.
3075 */
3076 if (bio_zone_write_plugging(bio)) {
3077 nr_segs = bio->__bi_nr_segments;
3078 if (rq)
3079 blk_queue_exit(q);
3080 goto new_request;
3081 }
3082
3083 bio = blk_queue_bounce(bio, q);
3084
3085 /*
3086 * The cached request already holds a q_usage_counter reference and we
3087 * don't have to acquire a new one if we use it.
3088 */
3089 if (!rq) {
3090 if (unlikely(bio_queue_enter(bio)))
3091 return;
3092 }
3093
3094 /*
3095 * Device reconfiguration may change logical block size or reduce the
3096 * number of poll queues, so the checks for alignment and poll support
3097 * have to be done with queue usage counter held.
3098 */
3099 if (unlikely(bio_unaligned(bio, q))) {
3100 bio_io_error(bio);
3101 goto queue_exit;
3102 }
3103
3104 if ((bio->bi_opf & REQ_POLLED) && !blk_mq_can_poll(q)) {
3105 bio->bi_status = BLK_STS_NOTSUPP;
3106 bio_endio(bio);
3107 goto queue_exit;
3108 }
3109
3110 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
3111 if (!bio)
3112 goto queue_exit;
3113
3114 if (!bio_integrity_prep(bio))
3115 goto queue_exit;
3116
3117 if (blk_mq_attempt_bio_merge(q, bio, nr_segs))
3118 goto queue_exit;
3119
3120 if (blk_queue_is_zoned(q) && blk_zone_plug_bio(bio, nr_segs))
3121 goto queue_exit;
3122
3123new_request:
3124 if (!rq) {
3125 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
3126 if (unlikely(!rq))
3127 goto queue_exit;
3128 } else {
3129 blk_mq_use_cached_rq(rq, plug, bio);
3130 }
3131
3132 trace_block_getrq(bio);
3133
3134 rq_qos_track(q, rq, bio);
3135
3136 blk_mq_bio_to_request(rq, bio, nr_segs);
3137
3138 ret = blk_crypto_rq_get_keyslot(rq);
3139 if (ret != BLK_STS_OK) {
3140 bio->bi_status = ret;
3141 bio_endio(bio);
3142 blk_mq_free_request(rq);
3143 return;
3144 }
3145
3146 if (bio_zone_write_plugging(bio))
3147 blk_zone_write_plug_init_request(rq);
3148
3149 if (op_is_flush(bio->bi_opf) && blk_insert_flush(rq))
3150 return;
3151
3152 if (plug) {
3153 blk_add_rq_to_plug(plug, rq);
3154 return;
3155 }
3156
3157 hctx = rq->mq_hctx;
3158 if ((rq->rq_flags & RQF_USE_SCHED) ||
3159 (hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) {
3160 blk_mq_insert_request(rq, 0);
3161 blk_mq_run_hw_queue(hctx, true);
3162 } else {
3163 blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq));
3164 }
3165 return;
3166
3167queue_exit:
3168 /*
3169 * Don't drop the queue reference if we were trying to use a cached
3170 * request and thus didn't acquire one.
3171 */
3172 if (!rq)
3173 blk_queue_exit(q);
3174}
3175
3176#ifdef CONFIG_BLK_MQ_STACKING
3177/**
3178 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
3179 * @rq: the request being queued
3180 */
3181blk_status_t blk_insert_cloned_request(struct request *rq)
3182{
3183 struct request_queue *q = rq->q;
3184 unsigned int max_sectors = blk_queue_get_max_sectors(rq);
3185 unsigned int max_segments = blk_rq_get_max_segments(rq);
3186 blk_status_t ret;
3187
3188 if (blk_rq_sectors(rq) > max_sectors) {
3189 /*
3190 * SCSI device does not have a good way to return if
3191 * Write Same/Zero is actually supported. If a device rejects
3192 * a non-read/write command (discard, write same,etc.) the
3193 * low-level device driver will set the relevant queue limit to
3194 * 0 to prevent blk-lib from issuing more of the offending
3195 * operations. Commands queued prior to the queue limit being
3196 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3197 * errors being propagated to upper layers.
3198 */
3199 if (max_sectors == 0)
3200 return BLK_STS_NOTSUPP;
3201
3202 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3203 __func__, blk_rq_sectors(rq), max_sectors);
3204 return BLK_STS_IOERR;
3205 }
3206
3207 /*
3208 * The queue settings related to segment counting may differ from the
3209 * original queue.
3210 */
3211 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
3212 if (rq->nr_phys_segments > max_segments) {
3213 printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n",
3214 __func__, rq->nr_phys_segments, max_segments);
3215 return BLK_STS_IOERR;
3216 }
3217
3218 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
3219 return BLK_STS_IOERR;
3220
3221 ret = blk_crypto_rq_get_keyslot(rq);
3222 if (ret != BLK_STS_OK)
3223 return ret;
3224
3225 blk_account_io_start(rq);
3226
3227 /*
3228 * Since we have a scheduler attached on the top device,
3229 * bypass a potential scheduler on the bottom device for
3230 * insert.
3231 */
3232 blk_mq_run_dispatch_ops(q,
3233 ret = blk_mq_request_issue_directly(rq, true));
3234 if (ret)
3235 blk_account_io_done(rq, blk_time_get_ns());
3236 return ret;
3237}
3238EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3239
3240/**
3241 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3242 * @rq: the clone request to be cleaned up
3243 *
3244 * Description:
3245 * Free all bios in @rq for a cloned request.
3246 */
3247void blk_rq_unprep_clone(struct request *rq)
3248{
3249 struct bio *bio;
3250
3251 while ((bio = rq->bio) != NULL) {
3252 rq->bio = bio->bi_next;
3253
3254 bio_put(bio);
3255 }
3256}
3257EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3258
3259/**
3260 * blk_rq_prep_clone - Helper function to setup clone request
3261 * @rq: the request to be setup
3262 * @rq_src: original request to be cloned
3263 * @bs: bio_set that bios for clone are allocated from
3264 * @gfp_mask: memory allocation mask for bio
3265 * @bio_ctr: setup function to be called for each clone bio.
3266 * Returns %0 for success, non %0 for failure.
3267 * @data: private data to be passed to @bio_ctr
3268 *
3269 * Description:
3270 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3271 * Also, pages which the original bios are pointing to are not copied
3272 * and the cloned bios just point same pages.
3273 * So cloned bios must be completed before original bios, which means
3274 * the caller must complete @rq before @rq_src.
3275 */
3276int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3277 struct bio_set *bs, gfp_t gfp_mask,
3278 int (*bio_ctr)(struct bio *, struct bio *, void *),
3279 void *data)
3280{
3281 struct bio *bio_src;
3282
3283 if (!bs)
3284 bs = &fs_bio_set;
3285
3286 __rq_for_each_bio(bio_src, rq_src) {
3287 struct bio *bio = bio_alloc_clone(rq->q->disk->part0, bio_src,
3288 gfp_mask, bs);
3289 if (!bio)
3290 goto free_and_out;
3291
3292 if (bio_ctr && bio_ctr(bio, bio_src, data)) {
3293 bio_put(bio);
3294 goto free_and_out;
3295 }
3296
3297 if (rq->bio) {
3298 rq->biotail->bi_next = bio;
3299 rq->biotail = bio;
3300 } else {
3301 rq->bio = rq->biotail = bio;
3302 }
3303 }
3304
3305 /* Copy attributes of the original request to the clone request. */
3306 rq->__sector = blk_rq_pos(rq_src);
3307 rq->__data_len = blk_rq_bytes(rq_src);
3308 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3309 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3310 rq->special_vec = rq_src->special_vec;
3311 }
3312 rq->nr_phys_segments = rq_src->nr_phys_segments;
3313
3314 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3315 goto free_and_out;
3316
3317 return 0;
3318
3319free_and_out:
3320 blk_rq_unprep_clone(rq);
3321
3322 return -ENOMEM;
3323}
3324EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3325#endif /* CONFIG_BLK_MQ_STACKING */
3326
3327/*
3328 * Steal bios from a request and add them to a bio list.
3329 * The request must not have been partially completed before.
3330 */
3331void blk_steal_bios(struct bio_list *list, struct request *rq)
3332{
3333 if (rq->bio) {
3334 if (list->tail)
3335 list->tail->bi_next = rq->bio;
3336 else
3337 list->head = rq->bio;
3338 list->tail = rq->biotail;
3339
3340 rq->bio = NULL;
3341 rq->biotail = NULL;
3342 }
3343
3344 rq->__data_len = 0;
3345}
3346EXPORT_SYMBOL_GPL(blk_steal_bios);
3347
3348static size_t order_to_size(unsigned int order)
3349{
3350 return (size_t)PAGE_SIZE << order;
3351}
3352
3353/* called before freeing request pool in @tags */
3354static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3355 struct blk_mq_tags *tags)
3356{
3357 struct page *page;
3358 unsigned long flags;
3359
3360 /*
3361 * There is no need to clear mapping if driver tags is not initialized
3362 * or the mapping belongs to the driver tags.
3363 */
3364 if (!drv_tags || drv_tags == tags)
3365 return;
3366
3367 list_for_each_entry(page, &tags->page_list, lru) {
3368 unsigned long start = (unsigned long)page_address(page);
3369 unsigned long end = start + order_to_size(page->private);
3370 int i;
3371
3372 for (i = 0; i < drv_tags->nr_tags; i++) {
3373 struct request *rq = drv_tags->rqs[i];
3374 unsigned long rq_addr = (unsigned long)rq;
3375
3376 if (rq_addr >= start && rq_addr < end) {
3377 WARN_ON_ONCE(req_ref_read(rq) != 0);
3378 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3379 }
3380 }
3381 }
3382
3383 /*
3384 * Wait until all pending iteration is done.
3385 *
3386 * Request reference is cleared and it is guaranteed to be observed
3387 * after the ->lock is released.
3388 */
3389 spin_lock_irqsave(&drv_tags->lock, flags);
3390 spin_unlock_irqrestore(&drv_tags->lock, flags);
3391}
3392
3393void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3394 unsigned int hctx_idx)
3395{
3396 struct blk_mq_tags *drv_tags;
3397 struct page *page;
3398
3399 if (list_empty(&tags->page_list))
3400 return;
3401
3402 if (blk_mq_is_shared_tags(set->flags))
3403 drv_tags = set->shared_tags;
3404 else
3405 drv_tags = set->tags[hctx_idx];
3406
3407 if (tags->static_rqs && set->ops->exit_request) {
3408 int i;
3409
3410 for (i = 0; i < tags->nr_tags; i++) {
3411 struct request *rq = tags->static_rqs[i];
3412
3413 if (!rq)
3414 continue;
3415 set->ops->exit_request(set, rq, hctx_idx);
3416 tags->static_rqs[i] = NULL;
3417 }
3418 }
3419
3420 blk_mq_clear_rq_mapping(drv_tags, tags);
3421
3422 while (!list_empty(&tags->page_list)) {
3423 page = list_first_entry(&tags->page_list, struct page, lru);
3424 list_del_init(&page->lru);
3425 /*
3426 * Remove kmemleak object previously allocated in
3427 * blk_mq_alloc_rqs().
3428 */
3429 kmemleak_free(page_address(page));
3430 __free_pages(page, page->private);
3431 }
3432}
3433
3434void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3435{
3436 kfree(tags->rqs);
3437 tags->rqs = NULL;
3438 kfree(tags->static_rqs);
3439 tags->static_rqs = NULL;
3440
3441 blk_mq_free_tags(tags);
3442}
3443
3444static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3445 unsigned int hctx_idx)
3446{
3447 int i;
3448
3449 for (i = 0; i < set->nr_maps; i++) {
3450 unsigned int start = set->map[i].queue_offset;
3451 unsigned int end = start + set->map[i].nr_queues;
3452
3453 if (hctx_idx >= start && hctx_idx < end)
3454 break;
3455 }
3456
3457 if (i >= set->nr_maps)
3458 i = HCTX_TYPE_DEFAULT;
3459
3460 return i;
3461}
3462
3463static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3464 unsigned int hctx_idx)
3465{
3466 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3467
3468 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3469}
3470
3471static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3472 unsigned int hctx_idx,
3473 unsigned int nr_tags,
3474 unsigned int reserved_tags)
3475{
3476 int node = blk_mq_get_hctx_node(set, hctx_idx);
3477 struct blk_mq_tags *tags;
3478
3479 if (node == NUMA_NO_NODE)
3480 node = set->numa_node;
3481
3482 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3483 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3484 if (!tags)
3485 return NULL;
3486
3487 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3488 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3489 node);
3490 if (!tags->rqs)
3491 goto err_free_tags;
3492
3493 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3494 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3495 node);
3496 if (!tags->static_rqs)
3497 goto err_free_rqs;
3498
3499 return tags;
3500
3501err_free_rqs:
3502 kfree(tags->rqs);
3503err_free_tags:
3504 blk_mq_free_tags(tags);
3505 return NULL;
3506}
3507
3508static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3509 unsigned int hctx_idx, int node)
3510{
3511 int ret;
3512
3513 if (set->ops->init_request) {
3514 ret = set->ops->init_request(set, rq, hctx_idx, node);
3515 if (ret)
3516 return ret;
3517 }
3518
3519 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3520 return 0;
3521}
3522
3523static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3524 struct blk_mq_tags *tags,
3525 unsigned int hctx_idx, unsigned int depth)
3526{
3527 unsigned int i, j, entries_per_page, max_order = 4;
3528 int node = blk_mq_get_hctx_node(set, hctx_idx);
3529 size_t rq_size, left;
3530
3531 if (node == NUMA_NO_NODE)
3532 node = set->numa_node;
3533
3534 INIT_LIST_HEAD(&tags->page_list);
3535
3536 /*
3537 * rq_size is the size of the request plus driver payload, rounded
3538 * to the cacheline size
3539 */
3540 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3541 cache_line_size());
3542 left = rq_size * depth;
3543
3544 for (i = 0; i < depth; ) {
3545 int this_order = max_order;
3546 struct page *page;
3547 int to_do;
3548 void *p;
3549
3550 while (this_order && left < order_to_size(this_order - 1))
3551 this_order--;
3552
3553 do {
3554 page = alloc_pages_node(node,
3555 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3556 this_order);
3557 if (page)
3558 break;
3559 if (!this_order--)
3560 break;
3561 if (order_to_size(this_order) < rq_size)
3562 break;
3563 } while (1);
3564
3565 if (!page)
3566 goto fail;
3567
3568 page->private = this_order;
3569 list_add_tail(&page->lru, &tags->page_list);
3570
3571 p = page_address(page);
3572 /*
3573 * Allow kmemleak to scan these pages as they contain pointers
3574 * to additional allocations like via ops->init_request().
3575 */
3576 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3577 entries_per_page = order_to_size(this_order) / rq_size;
3578 to_do = min(entries_per_page, depth - i);
3579 left -= to_do * rq_size;
3580 for (j = 0; j < to_do; j++) {
3581 struct request *rq = p;
3582
3583 tags->static_rqs[i] = rq;
3584 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3585 tags->static_rqs[i] = NULL;
3586 goto fail;
3587 }
3588
3589 p += rq_size;
3590 i++;
3591 }
3592 }
3593 return 0;
3594
3595fail:
3596 blk_mq_free_rqs(set, tags, hctx_idx);
3597 return -ENOMEM;
3598}
3599
3600struct rq_iter_data {
3601 struct blk_mq_hw_ctx *hctx;
3602 bool has_rq;
3603};
3604
3605static bool blk_mq_has_request(struct request *rq, void *data)
3606{
3607 struct rq_iter_data *iter_data = data;
3608
3609 if (rq->mq_hctx != iter_data->hctx)
3610 return true;
3611 iter_data->has_rq = true;
3612 return false;
3613}
3614
3615static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3616{
3617 struct blk_mq_tags *tags = hctx->sched_tags ?
3618 hctx->sched_tags : hctx->tags;
3619 struct rq_iter_data data = {
3620 .hctx = hctx,
3621 };
3622
3623 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3624 return data.has_rq;
3625}
3626
3627static bool blk_mq_hctx_has_online_cpu(struct blk_mq_hw_ctx *hctx,
3628 unsigned int this_cpu)
3629{
3630 enum hctx_type type = hctx->type;
3631 int cpu;
3632
3633 /*
3634 * hctx->cpumask has to rule out isolated CPUs, but userspace still
3635 * might submit IOs on these isolated CPUs, so use the queue map to
3636 * check if all CPUs mapped to this hctx are offline
3637 */
3638 for_each_online_cpu(cpu) {
3639 struct blk_mq_hw_ctx *h = blk_mq_map_queue_type(hctx->queue,
3640 type, cpu);
3641
3642 if (h != hctx)
3643 continue;
3644
3645 /* this hctx has at least one online CPU */
3646 if (this_cpu != cpu)
3647 return true;
3648 }
3649
3650 return false;
3651}
3652
3653static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3654{
3655 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3656 struct blk_mq_hw_ctx, cpuhp_online);
3657
3658 if (blk_mq_hctx_has_online_cpu(hctx, cpu))
3659 return 0;
3660
3661 /*
3662 * Prevent new request from being allocated on the current hctx.
3663 *
3664 * The smp_mb__after_atomic() Pairs with the implied barrier in
3665 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3666 * seen once we return from the tag allocator.
3667 */
3668 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3669 smp_mb__after_atomic();
3670
3671 /*
3672 * Try to grab a reference to the queue and wait for any outstanding
3673 * requests. If we could not grab a reference the queue has been
3674 * frozen and there are no requests.
3675 */
3676 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3677 while (blk_mq_hctx_has_requests(hctx))
3678 msleep(5);
3679 percpu_ref_put(&hctx->queue->q_usage_counter);
3680 }
3681
3682 return 0;
3683}
3684
3685/*
3686 * Check if one CPU is mapped to the specified hctx
3687 *
3688 * Isolated CPUs have been ruled out from hctx->cpumask, which is supposed
3689 * to be used for scheduling kworker only. For other usage, please call this
3690 * helper for checking if one CPU belongs to the specified hctx
3691 */
3692static bool blk_mq_cpu_mapped_to_hctx(unsigned int cpu,
3693 const struct blk_mq_hw_ctx *hctx)
3694{
3695 struct blk_mq_hw_ctx *mapped_hctx = blk_mq_map_queue_type(hctx->queue,
3696 hctx->type, cpu);
3697
3698 return mapped_hctx == hctx;
3699}
3700
3701static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3702{
3703 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3704 struct blk_mq_hw_ctx, cpuhp_online);
3705
3706 if (blk_mq_cpu_mapped_to_hctx(cpu, hctx))
3707 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3708 return 0;
3709}
3710
3711/*
3712 * 'cpu' is going away. splice any existing rq_list entries from this
3713 * software queue to the hw queue dispatch list, and ensure that it
3714 * gets run.
3715 */
3716static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3717{
3718 struct blk_mq_hw_ctx *hctx;
3719 struct blk_mq_ctx *ctx;
3720 LIST_HEAD(tmp);
3721 enum hctx_type type;
3722
3723 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3724 if (!blk_mq_cpu_mapped_to_hctx(cpu, hctx))
3725 return 0;
3726
3727 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3728 type = hctx->type;
3729
3730 spin_lock(&ctx->lock);
3731 if (!list_empty(&ctx->rq_lists[type])) {
3732 list_splice_init(&ctx->rq_lists[type], &tmp);
3733 blk_mq_hctx_clear_pending(hctx, ctx);
3734 }
3735 spin_unlock(&ctx->lock);
3736
3737 if (list_empty(&tmp))
3738 return 0;
3739
3740 spin_lock(&hctx->lock);
3741 list_splice_tail_init(&tmp, &hctx->dispatch);
3742 spin_unlock(&hctx->lock);
3743
3744 blk_mq_run_hw_queue(hctx, true);
3745 return 0;
3746}
3747
3748static void __blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3749{
3750 lockdep_assert_held(&blk_mq_cpuhp_lock);
3751
3752 if (!(hctx->flags & BLK_MQ_F_STACKING) &&
3753 !hlist_unhashed(&hctx->cpuhp_online)) {
3754 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3755 &hctx->cpuhp_online);
3756 INIT_HLIST_NODE(&hctx->cpuhp_online);
3757 }
3758
3759 if (!hlist_unhashed(&hctx->cpuhp_dead)) {
3760 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3761 &hctx->cpuhp_dead);
3762 INIT_HLIST_NODE(&hctx->cpuhp_dead);
3763 }
3764}
3765
3766static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3767{
3768 mutex_lock(&blk_mq_cpuhp_lock);
3769 __blk_mq_remove_cpuhp(hctx);
3770 mutex_unlock(&blk_mq_cpuhp_lock);
3771}
3772
3773static void __blk_mq_add_cpuhp(struct blk_mq_hw_ctx *hctx)
3774{
3775 lockdep_assert_held(&blk_mq_cpuhp_lock);
3776
3777 if (!(hctx->flags & BLK_MQ_F_STACKING) &&
3778 hlist_unhashed(&hctx->cpuhp_online))
3779 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3780 &hctx->cpuhp_online);
3781
3782 if (hlist_unhashed(&hctx->cpuhp_dead))
3783 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3784 &hctx->cpuhp_dead);
3785}
3786
3787static void __blk_mq_remove_cpuhp_list(struct list_head *head)
3788{
3789 struct blk_mq_hw_ctx *hctx;
3790
3791 lockdep_assert_held(&blk_mq_cpuhp_lock);
3792
3793 list_for_each_entry(hctx, head, hctx_list)
3794 __blk_mq_remove_cpuhp(hctx);
3795}
3796
3797/*
3798 * Unregister cpuhp callbacks from exited hw queues
3799 *
3800 * Safe to call if this `request_queue` is live
3801 */
3802static void blk_mq_remove_hw_queues_cpuhp(struct request_queue *q)
3803{
3804 LIST_HEAD(hctx_list);
3805
3806 spin_lock(&q->unused_hctx_lock);
3807 list_splice_init(&q->unused_hctx_list, &hctx_list);
3808 spin_unlock(&q->unused_hctx_lock);
3809
3810 mutex_lock(&blk_mq_cpuhp_lock);
3811 __blk_mq_remove_cpuhp_list(&hctx_list);
3812 mutex_unlock(&blk_mq_cpuhp_lock);
3813
3814 spin_lock(&q->unused_hctx_lock);
3815 list_splice(&hctx_list, &q->unused_hctx_list);
3816 spin_unlock(&q->unused_hctx_lock);
3817}
3818
3819/*
3820 * Register cpuhp callbacks from all hw queues
3821 *
3822 * Safe to call if this `request_queue` is live
3823 */
3824static void blk_mq_add_hw_queues_cpuhp(struct request_queue *q)
3825{
3826 struct blk_mq_hw_ctx *hctx;
3827 unsigned long i;
3828
3829 mutex_lock(&blk_mq_cpuhp_lock);
3830 queue_for_each_hw_ctx(q, hctx, i)
3831 __blk_mq_add_cpuhp(hctx);
3832 mutex_unlock(&blk_mq_cpuhp_lock);
3833}
3834
3835/*
3836 * Before freeing hw queue, clearing the flush request reference in
3837 * tags->rqs[] for avoiding potential UAF.
3838 */
3839static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3840 unsigned int queue_depth, struct request *flush_rq)
3841{
3842 int i;
3843 unsigned long flags;
3844
3845 /* The hw queue may not be mapped yet */
3846 if (!tags)
3847 return;
3848
3849 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3850
3851 for (i = 0; i < queue_depth; i++)
3852 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3853
3854 /*
3855 * Wait until all pending iteration is done.
3856 *
3857 * Request reference is cleared and it is guaranteed to be observed
3858 * after the ->lock is released.
3859 */
3860 spin_lock_irqsave(&tags->lock, flags);
3861 spin_unlock_irqrestore(&tags->lock, flags);
3862}
3863
3864/* hctx->ctxs will be freed in queue's release handler */
3865static void blk_mq_exit_hctx(struct request_queue *q,
3866 struct blk_mq_tag_set *set,
3867 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3868{
3869 struct request *flush_rq = hctx->fq->flush_rq;
3870
3871 if (blk_mq_hw_queue_mapped(hctx))
3872 blk_mq_tag_idle(hctx);
3873
3874 if (blk_queue_init_done(q))
3875 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3876 set->queue_depth, flush_rq);
3877 if (set->ops->exit_request)
3878 set->ops->exit_request(set, flush_rq, hctx_idx);
3879
3880 if (set->ops->exit_hctx)
3881 set->ops->exit_hctx(hctx, hctx_idx);
3882
3883 xa_erase(&q->hctx_table, hctx_idx);
3884
3885 spin_lock(&q->unused_hctx_lock);
3886 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3887 spin_unlock(&q->unused_hctx_lock);
3888}
3889
3890static void blk_mq_exit_hw_queues(struct request_queue *q,
3891 struct blk_mq_tag_set *set, int nr_queue)
3892{
3893 struct blk_mq_hw_ctx *hctx;
3894 unsigned long i;
3895
3896 queue_for_each_hw_ctx(q, hctx, i) {
3897 if (i == nr_queue)
3898 break;
3899 blk_mq_remove_cpuhp(hctx);
3900 blk_mq_exit_hctx(q, set, hctx, i);
3901 }
3902}
3903
3904static int blk_mq_init_hctx(struct request_queue *q,
3905 struct blk_mq_tag_set *set,
3906 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3907{
3908 hctx->queue_num = hctx_idx;
3909
3910 hctx->tags = set->tags[hctx_idx];
3911
3912 if (set->ops->init_hctx &&
3913 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3914 goto fail;
3915
3916 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3917 hctx->numa_node))
3918 goto exit_hctx;
3919
3920 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3921 goto exit_flush_rq;
3922
3923 return 0;
3924
3925 exit_flush_rq:
3926 if (set->ops->exit_request)
3927 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3928 exit_hctx:
3929 if (set->ops->exit_hctx)
3930 set->ops->exit_hctx(hctx, hctx_idx);
3931 fail:
3932 return -1;
3933}
3934
3935static struct blk_mq_hw_ctx *
3936blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3937 int node)
3938{
3939 struct blk_mq_hw_ctx *hctx;
3940 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3941
3942 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3943 if (!hctx)
3944 goto fail_alloc_hctx;
3945
3946 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3947 goto free_hctx;
3948
3949 atomic_set(&hctx->nr_active, 0);
3950 if (node == NUMA_NO_NODE)
3951 node = set->numa_node;
3952 hctx->numa_node = node;
3953
3954 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3955 spin_lock_init(&hctx->lock);
3956 INIT_LIST_HEAD(&hctx->dispatch);
3957 INIT_HLIST_NODE(&hctx->cpuhp_dead);
3958 INIT_HLIST_NODE(&hctx->cpuhp_online);
3959 hctx->queue = q;
3960 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3961
3962 INIT_LIST_HEAD(&hctx->hctx_list);
3963
3964 /*
3965 * Allocate space for all possible cpus to avoid allocation at
3966 * runtime
3967 */
3968 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3969 gfp, node);
3970 if (!hctx->ctxs)
3971 goto free_cpumask;
3972
3973 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3974 gfp, node, false, false))
3975 goto free_ctxs;
3976 hctx->nr_ctx = 0;
3977
3978 spin_lock_init(&hctx->dispatch_wait_lock);
3979 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3980 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3981
3982 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3983 if (!hctx->fq)
3984 goto free_bitmap;
3985
3986 blk_mq_hctx_kobj_init(hctx);
3987
3988 return hctx;
3989
3990 free_bitmap:
3991 sbitmap_free(&hctx->ctx_map);
3992 free_ctxs:
3993 kfree(hctx->ctxs);
3994 free_cpumask:
3995 free_cpumask_var(hctx->cpumask);
3996 free_hctx:
3997 kfree(hctx);
3998 fail_alloc_hctx:
3999 return NULL;
4000}
4001
4002static void blk_mq_init_cpu_queues(struct request_queue *q,
4003 unsigned int nr_hw_queues)
4004{
4005 struct blk_mq_tag_set *set = q->tag_set;
4006 unsigned int i, j;
4007
4008 for_each_possible_cpu(i) {
4009 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
4010 struct blk_mq_hw_ctx *hctx;
4011 int k;
4012
4013 __ctx->cpu = i;
4014 spin_lock_init(&__ctx->lock);
4015 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
4016 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
4017
4018 __ctx->queue = q;
4019
4020 /*
4021 * Set local node, IFF we have more than one hw queue. If
4022 * not, we remain on the home node of the device
4023 */
4024 for (j = 0; j < set->nr_maps; j++) {
4025 hctx = blk_mq_map_queue_type(q, j, i);
4026 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
4027 hctx->numa_node = cpu_to_node(i);
4028 }
4029 }
4030}
4031
4032struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
4033 unsigned int hctx_idx,
4034 unsigned int depth)
4035{
4036 struct blk_mq_tags *tags;
4037 int ret;
4038
4039 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
4040 if (!tags)
4041 return NULL;
4042
4043 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
4044 if (ret) {
4045 blk_mq_free_rq_map(tags);
4046 return NULL;
4047 }
4048
4049 return tags;
4050}
4051
4052static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
4053 int hctx_idx)
4054{
4055 if (blk_mq_is_shared_tags(set->flags)) {
4056 set->tags[hctx_idx] = set->shared_tags;
4057
4058 return true;
4059 }
4060
4061 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
4062 set->queue_depth);
4063
4064 return set->tags[hctx_idx];
4065}
4066
4067void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
4068 struct blk_mq_tags *tags,
4069 unsigned int hctx_idx)
4070{
4071 if (tags) {
4072 blk_mq_free_rqs(set, tags, hctx_idx);
4073 blk_mq_free_rq_map(tags);
4074 }
4075}
4076
4077static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
4078 unsigned int hctx_idx)
4079{
4080 if (!blk_mq_is_shared_tags(set->flags))
4081 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
4082
4083 set->tags[hctx_idx] = NULL;
4084}
4085
4086static void blk_mq_map_swqueue(struct request_queue *q)
4087{
4088 unsigned int j, hctx_idx;
4089 unsigned long i;
4090 struct blk_mq_hw_ctx *hctx;
4091 struct blk_mq_ctx *ctx;
4092 struct blk_mq_tag_set *set = q->tag_set;
4093
4094 queue_for_each_hw_ctx(q, hctx, i) {
4095 cpumask_clear(hctx->cpumask);
4096 hctx->nr_ctx = 0;
4097 hctx->dispatch_from = NULL;
4098 }
4099
4100 /*
4101 * Map software to hardware queues.
4102 *
4103 * If the cpu isn't present, the cpu is mapped to first hctx.
4104 */
4105 for_each_possible_cpu(i) {
4106
4107 ctx = per_cpu_ptr(q->queue_ctx, i);
4108 for (j = 0; j < set->nr_maps; j++) {
4109 if (!set->map[j].nr_queues) {
4110 ctx->hctxs[j] = blk_mq_map_queue_type(q,
4111 HCTX_TYPE_DEFAULT, i);
4112 continue;
4113 }
4114 hctx_idx = set->map[j].mq_map[i];
4115 /* unmapped hw queue can be remapped after CPU topo changed */
4116 if (!set->tags[hctx_idx] &&
4117 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
4118 /*
4119 * If tags initialization fail for some hctx,
4120 * that hctx won't be brought online. In this
4121 * case, remap the current ctx to hctx[0] which
4122 * is guaranteed to always have tags allocated
4123 */
4124 set->map[j].mq_map[i] = 0;
4125 }
4126
4127 hctx = blk_mq_map_queue_type(q, j, i);
4128 ctx->hctxs[j] = hctx;
4129 /*
4130 * If the CPU is already set in the mask, then we've
4131 * mapped this one already. This can happen if
4132 * devices share queues across queue maps.
4133 */
4134 if (cpumask_test_cpu(i, hctx->cpumask))
4135 continue;
4136
4137 cpumask_set_cpu(i, hctx->cpumask);
4138 hctx->type = j;
4139 ctx->index_hw[hctx->type] = hctx->nr_ctx;
4140 hctx->ctxs[hctx->nr_ctx++] = ctx;
4141
4142 /*
4143 * If the nr_ctx type overflows, we have exceeded the
4144 * amount of sw queues we can support.
4145 */
4146 BUG_ON(!hctx->nr_ctx);
4147 }
4148
4149 for (; j < HCTX_MAX_TYPES; j++)
4150 ctx->hctxs[j] = blk_mq_map_queue_type(q,
4151 HCTX_TYPE_DEFAULT, i);
4152 }
4153
4154 queue_for_each_hw_ctx(q, hctx, i) {
4155 int cpu;
4156
4157 /*
4158 * If no software queues are mapped to this hardware queue,
4159 * disable it and free the request entries.
4160 */
4161 if (!hctx->nr_ctx) {
4162 /* Never unmap queue 0. We need it as a
4163 * fallback in case of a new remap fails
4164 * allocation
4165 */
4166 if (i)
4167 __blk_mq_free_map_and_rqs(set, i);
4168
4169 hctx->tags = NULL;
4170 continue;
4171 }
4172
4173 hctx->tags = set->tags[i];
4174 WARN_ON(!hctx->tags);
4175
4176 /*
4177 * Set the map size to the number of mapped software queues.
4178 * This is more accurate and more efficient than looping
4179 * over all possibly mapped software queues.
4180 */
4181 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
4182
4183 /*
4184 * Rule out isolated CPUs from hctx->cpumask to avoid
4185 * running block kworker on isolated CPUs
4186 */
4187 for_each_cpu(cpu, hctx->cpumask) {
4188 if (cpu_is_isolated(cpu))
4189 cpumask_clear_cpu(cpu, hctx->cpumask);
4190 }
4191
4192 /*
4193 * Initialize batch roundrobin counts
4194 */
4195 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
4196 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
4197 }
4198}
4199
4200/*
4201 * Caller needs to ensure that we're either frozen/quiesced, or that
4202 * the queue isn't live yet.
4203 */
4204static void queue_set_hctx_shared(struct request_queue *q, bool shared)
4205{
4206 struct blk_mq_hw_ctx *hctx;
4207 unsigned long i;
4208
4209 queue_for_each_hw_ctx(q, hctx, i) {
4210 if (shared) {
4211 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
4212 } else {
4213 blk_mq_tag_idle(hctx);
4214 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
4215 }
4216 }
4217}
4218
4219static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
4220 bool shared)
4221{
4222 struct request_queue *q;
4223
4224 lockdep_assert_held(&set->tag_list_lock);
4225
4226 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4227 blk_mq_freeze_queue(q);
4228 queue_set_hctx_shared(q, shared);
4229 blk_mq_unfreeze_queue(q);
4230 }
4231}
4232
4233static void blk_mq_del_queue_tag_set(struct request_queue *q)
4234{
4235 struct blk_mq_tag_set *set = q->tag_set;
4236
4237 mutex_lock(&set->tag_list_lock);
4238 list_del(&q->tag_set_list);
4239 if (list_is_singular(&set->tag_list)) {
4240 /* just transitioned to unshared */
4241 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
4242 /* update existing queue */
4243 blk_mq_update_tag_set_shared(set, false);
4244 }
4245 mutex_unlock(&set->tag_list_lock);
4246 INIT_LIST_HEAD(&q->tag_set_list);
4247}
4248
4249static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
4250 struct request_queue *q)
4251{
4252 mutex_lock(&set->tag_list_lock);
4253
4254 /*
4255 * Check to see if we're transitioning to shared (from 1 to 2 queues).
4256 */
4257 if (!list_empty(&set->tag_list) &&
4258 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
4259 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
4260 /* update existing queue */
4261 blk_mq_update_tag_set_shared(set, true);
4262 }
4263 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
4264 queue_set_hctx_shared(q, true);
4265 list_add_tail(&q->tag_set_list, &set->tag_list);
4266
4267 mutex_unlock(&set->tag_list_lock);
4268}
4269
4270/* All allocations will be freed in release handler of q->mq_kobj */
4271static int blk_mq_alloc_ctxs(struct request_queue *q)
4272{
4273 struct blk_mq_ctxs *ctxs;
4274 int cpu;
4275
4276 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
4277 if (!ctxs)
4278 return -ENOMEM;
4279
4280 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
4281 if (!ctxs->queue_ctx)
4282 goto fail;
4283
4284 for_each_possible_cpu(cpu) {
4285 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
4286 ctx->ctxs = ctxs;
4287 }
4288
4289 q->mq_kobj = &ctxs->kobj;
4290 q->queue_ctx = ctxs->queue_ctx;
4291
4292 return 0;
4293 fail:
4294 kfree(ctxs);
4295 return -ENOMEM;
4296}
4297
4298/*
4299 * It is the actual release handler for mq, but we do it from
4300 * request queue's release handler for avoiding use-after-free
4301 * and headache because q->mq_kobj shouldn't have been introduced,
4302 * but we can't group ctx/kctx kobj without it.
4303 */
4304void blk_mq_release(struct request_queue *q)
4305{
4306 struct blk_mq_hw_ctx *hctx, *next;
4307 unsigned long i;
4308
4309 queue_for_each_hw_ctx(q, hctx, i)
4310 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4311
4312 /* all hctx are in .unused_hctx_list now */
4313 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4314 list_del_init(&hctx->hctx_list);
4315 kobject_put(&hctx->kobj);
4316 }
4317
4318 xa_destroy(&q->hctx_table);
4319
4320 /*
4321 * release .mq_kobj and sw queue's kobject now because
4322 * both share lifetime with request queue.
4323 */
4324 blk_mq_sysfs_deinit(q);
4325}
4326
4327struct request_queue *blk_mq_alloc_queue(struct blk_mq_tag_set *set,
4328 struct queue_limits *lim, void *queuedata)
4329{
4330 struct queue_limits default_lim = { };
4331 struct request_queue *q;
4332 int ret;
4333
4334 if (!lim)
4335 lim = &default_lim;
4336 lim->features |= BLK_FEAT_IO_STAT | BLK_FEAT_NOWAIT;
4337 if (set->nr_maps > HCTX_TYPE_POLL)
4338 lim->features |= BLK_FEAT_POLL;
4339
4340 q = blk_alloc_queue(lim, set->numa_node);
4341 if (IS_ERR(q))
4342 return q;
4343 q->queuedata = queuedata;
4344 ret = blk_mq_init_allocated_queue(set, q);
4345 if (ret) {
4346 blk_put_queue(q);
4347 return ERR_PTR(ret);
4348 }
4349 return q;
4350}
4351EXPORT_SYMBOL(blk_mq_alloc_queue);
4352
4353/**
4354 * blk_mq_destroy_queue - shutdown a request queue
4355 * @q: request queue to shutdown
4356 *
4357 * This shuts down a request queue allocated by blk_mq_alloc_queue(). All future
4358 * requests will be failed with -ENODEV. The caller is responsible for dropping
4359 * the reference from blk_mq_alloc_queue() by calling blk_put_queue().
4360 *
4361 * Context: can sleep
4362 */
4363void blk_mq_destroy_queue(struct request_queue *q)
4364{
4365 WARN_ON_ONCE(!queue_is_mq(q));
4366 WARN_ON_ONCE(blk_queue_registered(q));
4367
4368 might_sleep();
4369
4370 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4371 blk_queue_start_drain(q);
4372 blk_mq_freeze_queue_wait(q);
4373
4374 blk_sync_queue(q);
4375 blk_mq_cancel_work_sync(q);
4376 blk_mq_exit_queue(q);
4377}
4378EXPORT_SYMBOL(blk_mq_destroy_queue);
4379
4380struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set,
4381 struct queue_limits *lim, void *queuedata,
4382 struct lock_class_key *lkclass)
4383{
4384 struct request_queue *q;
4385 struct gendisk *disk;
4386
4387 q = blk_mq_alloc_queue(set, lim, queuedata);
4388 if (IS_ERR(q))
4389 return ERR_CAST(q);
4390
4391 disk = __alloc_disk_node(q, set->numa_node, lkclass);
4392 if (!disk) {
4393 blk_mq_destroy_queue(q);
4394 blk_put_queue(q);
4395 return ERR_PTR(-ENOMEM);
4396 }
4397 set_bit(GD_OWNS_QUEUE, &disk->state);
4398 return disk;
4399}
4400EXPORT_SYMBOL(__blk_mq_alloc_disk);
4401
4402struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4403 struct lock_class_key *lkclass)
4404{
4405 struct gendisk *disk;
4406
4407 if (!blk_get_queue(q))
4408 return NULL;
4409 disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4410 if (!disk)
4411 blk_put_queue(q);
4412 return disk;
4413}
4414EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4415
4416/*
4417 * Only hctx removed from cpuhp list can be reused
4418 */
4419static bool blk_mq_hctx_is_reusable(struct blk_mq_hw_ctx *hctx)
4420{
4421 return hlist_unhashed(&hctx->cpuhp_online) &&
4422 hlist_unhashed(&hctx->cpuhp_dead);
4423}
4424
4425static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4426 struct blk_mq_tag_set *set, struct request_queue *q,
4427 int hctx_idx, int node)
4428{
4429 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4430
4431 /* reuse dead hctx first */
4432 spin_lock(&q->unused_hctx_lock);
4433 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4434 if (tmp->numa_node == node && blk_mq_hctx_is_reusable(tmp)) {
4435 hctx = tmp;
4436 break;
4437 }
4438 }
4439 if (hctx)
4440 list_del_init(&hctx->hctx_list);
4441 spin_unlock(&q->unused_hctx_lock);
4442
4443 if (!hctx)
4444 hctx = blk_mq_alloc_hctx(q, set, node);
4445 if (!hctx)
4446 goto fail;
4447
4448 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4449 goto free_hctx;
4450
4451 return hctx;
4452
4453 free_hctx:
4454 kobject_put(&hctx->kobj);
4455 fail:
4456 return NULL;
4457}
4458
4459static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4460 struct request_queue *q)
4461{
4462 struct blk_mq_hw_ctx *hctx;
4463 unsigned long i, j;
4464
4465 /* protect against switching io scheduler */
4466 mutex_lock(&q->sysfs_lock);
4467 for (i = 0; i < set->nr_hw_queues; i++) {
4468 int old_node;
4469 int node = blk_mq_get_hctx_node(set, i);
4470 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4471
4472 if (old_hctx) {
4473 old_node = old_hctx->numa_node;
4474 blk_mq_exit_hctx(q, set, old_hctx, i);
4475 }
4476
4477 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4478 if (!old_hctx)
4479 break;
4480 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4481 node, old_node);
4482 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4483 WARN_ON_ONCE(!hctx);
4484 }
4485 }
4486 /*
4487 * Increasing nr_hw_queues fails. Free the newly allocated
4488 * hctxs and keep the previous q->nr_hw_queues.
4489 */
4490 if (i != set->nr_hw_queues) {
4491 j = q->nr_hw_queues;
4492 } else {
4493 j = i;
4494 q->nr_hw_queues = set->nr_hw_queues;
4495 }
4496
4497 xa_for_each_start(&q->hctx_table, j, hctx, j)
4498 blk_mq_exit_hctx(q, set, hctx, j);
4499 mutex_unlock(&q->sysfs_lock);
4500
4501 /* unregister cpuhp callbacks for exited hctxs */
4502 blk_mq_remove_hw_queues_cpuhp(q);
4503
4504 /* register cpuhp for new initialized hctxs */
4505 blk_mq_add_hw_queues_cpuhp(q);
4506}
4507
4508int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4509 struct request_queue *q)
4510{
4511 /* mark the queue as mq asap */
4512 q->mq_ops = set->ops;
4513
4514 /*
4515 * ->tag_set has to be setup before initialize hctx, which cpuphp
4516 * handler needs it for checking queue mapping
4517 */
4518 q->tag_set = set;
4519
4520 if (blk_mq_alloc_ctxs(q))
4521 goto err_exit;
4522
4523 /* init q->mq_kobj and sw queues' kobjects */
4524 blk_mq_sysfs_init(q);
4525
4526 INIT_LIST_HEAD(&q->unused_hctx_list);
4527 spin_lock_init(&q->unused_hctx_lock);
4528
4529 xa_init(&q->hctx_table);
4530
4531 blk_mq_realloc_hw_ctxs(set, q);
4532 if (!q->nr_hw_queues)
4533 goto err_hctxs;
4534
4535 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4536 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4537
4538 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4539
4540 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4541 INIT_LIST_HEAD(&q->flush_list);
4542 INIT_LIST_HEAD(&q->requeue_list);
4543 spin_lock_init(&q->requeue_lock);
4544
4545 q->nr_requests = set->queue_depth;
4546
4547 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4548 blk_mq_add_queue_tag_set(set, q);
4549 blk_mq_map_swqueue(q);
4550 return 0;
4551
4552err_hctxs:
4553 blk_mq_release(q);
4554err_exit:
4555 q->mq_ops = NULL;
4556 return -ENOMEM;
4557}
4558EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4559
4560/* tags can _not_ be used after returning from blk_mq_exit_queue */
4561void blk_mq_exit_queue(struct request_queue *q)
4562{
4563 struct blk_mq_tag_set *set = q->tag_set;
4564
4565 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4566 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4567 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4568 blk_mq_del_queue_tag_set(q);
4569}
4570
4571static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4572{
4573 int i;
4574
4575 if (blk_mq_is_shared_tags(set->flags)) {
4576 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4577 BLK_MQ_NO_HCTX_IDX,
4578 set->queue_depth);
4579 if (!set->shared_tags)
4580 return -ENOMEM;
4581 }
4582
4583 for (i = 0; i < set->nr_hw_queues; i++) {
4584 if (!__blk_mq_alloc_map_and_rqs(set, i))
4585 goto out_unwind;
4586 cond_resched();
4587 }
4588
4589 return 0;
4590
4591out_unwind:
4592 while (--i >= 0)
4593 __blk_mq_free_map_and_rqs(set, i);
4594
4595 if (blk_mq_is_shared_tags(set->flags)) {
4596 blk_mq_free_map_and_rqs(set, set->shared_tags,
4597 BLK_MQ_NO_HCTX_IDX);
4598 }
4599
4600 return -ENOMEM;
4601}
4602
4603/*
4604 * Allocate the request maps associated with this tag_set. Note that this
4605 * may reduce the depth asked for, if memory is tight. set->queue_depth
4606 * will be updated to reflect the allocated depth.
4607 */
4608static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4609{
4610 unsigned int depth;
4611 int err;
4612
4613 depth = set->queue_depth;
4614 do {
4615 err = __blk_mq_alloc_rq_maps(set);
4616 if (!err)
4617 break;
4618
4619 set->queue_depth >>= 1;
4620 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4621 err = -ENOMEM;
4622 break;
4623 }
4624 } while (set->queue_depth);
4625
4626 if (!set->queue_depth || err) {
4627 pr_err("blk-mq: failed to allocate request map\n");
4628 return -ENOMEM;
4629 }
4630
4631 if (depth != set->queue_depth)
4632 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4633 depth, set->queue_depth);
4634
4635 return 0;
4636}
4637
4638static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4639{
4640 /*
4641 * blk_mq_map_queues() and multiple .map_queues() implementations
4642 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4643 * number of hardware queues.
4644 */
4645 if (set->nr_maps == 1)
4646 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4647
4648 if (set->ops->map_queues) {
4649 int i;
4650
4651 /*
4652 * transport .map_queues is usually done in the following
4653 * way:
4654 *
4655 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4656 * mask = get_cpu_mask(queue)
4657 * for_each_cpu(cpu, mask)
4658 * set->map[x].mq_map[cpu] = queue;
4659 * }
4660 *
4661 * When we need to remap, the table has to be cleared for
4662 * killing stale mapping since one CPU may not be mapped
4663 * to any hw queue.
4664 */
4665 for (i = 0; i < set->nr_maps; i++)
4666 blk_mq_clear_mq_map(&set->map[i]);
4667
4668 set->ops->map_queues(set);
4669 } else {
4670 BUG_ON(set->nr_maps > 1);
4671 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4672 }
4673}
4674
4675static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4676 int new_nr_hw_queues)
4677{
4678 struct blk_mq_tags **new_tags;
4679 int i;
4680
4681 if (set->nr_hw_queues >= new_nr_hw_queues)
4682 goto done;
4683
4684 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4685 GFP_KERNEL, set->numa_node);
4686 if (!new_tags)
4687 return -ENOMEM;
4688
4689 if (set->tags)
4690 memcpy(new_tags, set->tags, set->nr_hw_queues *
4691 sizeof(*set->tags));
4692 kfree(set->tags);
4693 set->tags = new_tags;
4694
4695 for (i = set->nr_hw_queues; i < new_nr_hw_queues; i++) {
4696 if (!__blk_mq_alloc_map_and_rqs(set, i)) {
4697 while (--i >= set->nr_hw_queues)
4698 __blk_mq_free_map_and_rqs(set, i);
4699 return -ENOMEM;
4700 }
4701 cond_resched();
4702 }
4703
4704done:
4705 set->nr_hw_queues = new_nr_hw_queues;
4706 return 0;
4707}
4708
4709/*
4710 * Alloc a tag set to be associated with one or more request queues.
4711 * May fail with EINVAL for various error conditions. May adjust the
4712 * requested depth down, if it's too large. In that case, the set
4713 * value will be stored in set->queue_depth.
4714 */
4715int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4716{
4717 int i, ret;
4718
4719 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4720
4721 if (!set->nr_hw_queues)
4722 return -EINVAL;
4723 if (!set->queue_depth)
4724 return -EINVAL;
4725 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4726 return -EINVAL;
4727
4728 if (!set->ops->queue_rq)
4729 return -EINVAL;
4730
4731 if (!set->ops->get_budget ^ !set->ops->put_budget)
4732 return -EINVAL;
4733
4734 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4735 pr_info("blk-mq: reduced tag depth to %u\n",
4736 BLK_MQ_MAX_DEPTH);
4737 set->queue_depth = BLK_MQ_MAX_DEPTH;
4738 }
4739
4740 if (!set->nr_maps)
4741 set->nr_maps = 1;
4742 else if (set->nr_maps > HCTX_MAX_TYPES)
4743 return -EINVAL;
4744
4745 /*
4746 * If a crashdump is active, then we are potentially in a very
4747 * memory constrained environment. Limit us to 64 tags to prevent
4748 * using too much memory.
4749 */
4750 if (is_kdump_kernel())
4751 set->queue_depth = min(64U, set->queue_depth);
4752
4753 /*
4754 * There is no use for more h/w queues than cpus if we just have
4755 * a single map
4756 */
4757 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4758 set->nr_hw_queues = nr_cpu_ids;
4759
4760 if (set->flags & BLK_MQ_F_BLOCKING) {
4761 set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL);
4762 if (!set->srcu)
4763 return -ENOMEM;
4764 ret = init_srcu_struct(set->srcu);
4765 if (ret)
4766 goto out_free_srcu;
4767 }
4768
4769 ret = -ENOMEM;
4770 set->tags = kcalloc_node(set->nr_hw_queues,
4771 sizeof(struct blk_mq_tags *), GFP_KERNEL,
4772 set->numa_node);
4773 if (!set->tags)
4774 goto out_cleanup_srcu;
4775
4776 for (i = 0; i < set->nr_maps; i++) {
4777 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4778 sizeof(set->map[i].mq_map[0]),
4779 GFP_KERNEL, set->numa_node);
4780 if (!set->map[i].mq_map)
4781 goto out_free_mq_map;
4782 set->map[i].nr_queues = set->nr_hw_queues;
4783 }
4784
4785 blk_mq_update_queue_map(set);
4786
4787 ret = blk_mq_alloc_set_map_and_rqs(set);
4788 if (ret)
4789 goto out_free_mq_map;
4790
4791 mutex_init(&set->tag_list_lock);
4792 INIT_LIST_HEAD(&set->tag_list);
4793
4794 return 0;
4795
4796out_free_mq_map:
4797 for (i = 0; i < set->nr_maps; i++) {
4798 kfree(set->map[i].mq_map);
4799 set->map[i].mq_map = NULL;
4800 }
4801 kfree(set->tags);
4802 set->tags = NULL;
4803out_cleanup_srcu:
4804 if (set->flags & BLK_MQ_F_BLOCKING)
4805 cleanup_srcu_struct(set->srcu);
4806out_free_srcu:
4807 if (set->flags & BLK_MQ_F_BLOCKING)
4808 kfree(set->srcu);
4809 return ret;
4810}
4811EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4812
4813/* allocate and initialize a tagset for a simple single-queue device */
4814int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4815 const struct blk_mq_ops *ops, unsigned int queue_depth,
4816 unsigned int set_flags)
4817{
4818 memset(set, 0, sizeof(*set));
4819 set->ops = ops;
4820 set->nr_hw_queues = 1;
4821 set->nr_maps = 1;
4822 set->queue_depth = queue_depth;
4823 set->numa_node = NUMA_NO_NODE;
4824 set->flags = set_flags;
4825 return blk_mq_alloc_tag_set(set);
4826}
4827EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4828
4829void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4830{
4831 int i, j;
4832
4833 for (i = 0; i < set->nr_hw_queues; i++)
4834 __blk_mq_free_map_and_rqs(set, i);
4835
4836 if (blk_mq_is_shared_tags(set->flags)) {
4837 blk_mq_free_map_and_rqs(set, set->shared_tags,
4838 BLK_MQ_NO_HCTX_IDX);
4839 }
4840
4841 for (j = 0; j < set->nr_maps; j++) {
4842 kfree(set->map[j].mq_map);
4843 set->map[j].mq_map = NULL;
4844 }
4845
4846 kfree(set->tags);
4847 set->tags = NULL;
4848 if (set->flags & BLK_MQ_F_BLOCKING) {
4849 cleanup_srcu_struct(set->srcu);
4850 kfree(set->srcu);
4851 }
4852}
4853EXPORT_SYMBOL(blk_mq_free_tag_set);
4854
4855int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4856{
4857 struct blk_mq_tag_set *set = q->tag_set;
4858 struct blk_mq_hw_ctx *hctx;
4859 int ret;
4860 unsigned long i;
4861
4862 if (WARN_ON_ONCE(!q->mq_freeze_depth))
4863 return -EINVAL;
4864
4865 if (!set)
4866 return -EINVAL;
4867
4868 if (q->nr_requests == nr)
4869 return 0;
4870
4871 blk_mq_quiesce_queue(q);
4872
4873 ret = 0;
4874 queue_for_each_hw_ctx(q, hctx, i) {
4875 if (!hctx->tags)
4876 continue;
4877 /*
4878 * If we're using an MQ scheduler, just update the scheduler
4879 * queue depth. This is similar to what the old code would do.
4880 */
4881 if (hctx->sched_tags) {
4882 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4883 nr, true);
4884 } else {
4885 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4886 false);
4887 }
4888 if (ret)
4889 break;
4890 if (q->elevator && q->elevator->type->ops.depth_updated)
4891 q->elevator->type->ops.depth_updated(hctx);
4892 }
4893 if (!ret) {
4894 q->nr_requests = nr;
4895 if (blk_mq_is_shared_tags(set->flags)) {
4896 if (q->elevator)
4897 blk_mq_tag_update_sched_shared_tags(q);
4898 else
4899 blk_mq_tag_resize_shared_tags(set, nr);
4900 }
4901 }
4902
4903 blk_mq_unquiesce_queue(q);
4904
4905 return ret;
4906}
4907
4908/*
4909 * request_queue and elevator_type pair.
4910 * It is just used by __blk_mq_update_nr_hw_queues to cache
4911 * the elevator_type associated with a request_queue.
4912 */
4913struct blk_mq_qe_pair {
4914 struct list_head node;
4915 struct request_queue *q;
4916 struct elevator_type *type;
4917};
4918
4919/*
4920 * Cache the elevator_type in qe pair list and switch the
4921 * io scheduler to 'none'
4922 */
4923static bool blk_mq_elv_switch_none(struct list_head *head,
4924 struct request_queue *q)
4925{
4926 struct blk_mq_qe_pair *qe;
4927
4928 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4929 if (!qe)
4930 return false;
4931
4932 /* q->elevator needs protection from ->sysfs_lock */
4933 mutex_lock(&q->sysfs_lock);
4934
4935 /* the check has to be done with holding sysfs_lock */
4936 if (!q->elevator) {
4937 kfree(qe);
4938 goto unlock;
4939 }
4940
4941 INIT_LIST_HEAD(&qe->node);
4942 qe->q = q;
4943 qe->type = q->elevator->type;
4944 /* keep a reference to the elevator module as we'll switch back */
4945 __elevator_get(qe->type);
4946 list_add(&qe->node, head);
4947 elevator_disable(q);
4948unlock:
4949 mutex_unlock(&q->sysfs_lock);
4950
4951 return true;
4952}
4953
4954static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4955 struct request_queue *q)
4956{
4957 struct blk_mq_qe_pair *qe;
4958
4959 list_for_each_entry(qe, head, node)
4960 if (qe->q == q)
4961 return qe;
4962
4963 return NULL;
4964}
4965
4966static void blk_mq_elv_switch_back(struct list_head *head,
4967 struct request_queue *q)
4968{
4969 struct blk_mq_qe_pair *qe;
4970 struct elevator_type *t;
4971
4972 qe = blk_lookup_qe_pair(head, q);
4973 if (!qe)
4974 return;
4975 t = qe->type;
4976 list_del(&qe->node);
4977 kfree(qe);
4978
4979 mutex_lock(&q->sysfs_lock);
4980 elevator_switch(q, t);
4981 /* drop the reference acquired in blk_mq_elv_switch_none */
4982 elevator_put(t);
4983 mutex_unlock(&q->sysfs_lock);
4984}
4985
4986static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4987 int nr_hw_queues)
4988{
4989 struct request_queue *q;
4990 LIST_HEAD(head);
4991 int prev_nr_hw_queues = set->nr_hw_queues;
4992 int i;
4993
4994 lockdep_assert_held(&set->tag_list_lock);
4995
4996 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4997 nr_hw_queues = nr_cpu_ids;
4998 if (nr_hw_queues < 1)
4999 return;
5000 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
5001 return;
5002
5003 list_for_each_entry(q, &set->tag_list, tag_set_list)
5004 blk_mq_freeze_queue(q);
5005 /*
5006 * Switch IO scheduler to 'none', cleaning up the data associated
5007 * with the previous scheduler. We will switch back once we are done
5008 * updating the new sw to hw queue mappings.
5009 */
5010 list_for_each_entry(q, &set->tag_list, tag_set_list)
5011 if (!blk_mq_elv_switch_none(&head, q))
5012 goto switch_back;
5013
5014 list_for_each_entry(q, &set->tag_list, tag_set_list) {
5015 blk_mq_debugfs_unregister_hctxs(q);
5016 blk_mq_sysfs_unregister_hctxs(q);
5017 }
5018
5019 if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0)
5020 goto reregister;
5021
5022fallback:
5023 blk_mq_update_queue_map(set);
5024 list_for_each_entry(q, &set->tag_list, tag_set_list) {
5025 blk_mq_realloc_hw_ctxs(set, q);
5026
5027 if (q->nr_hw_queues != set->nr_hw_queues) {
5028 int i = prev_nr_hw_queues;
5029
5030 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
5031 nr_hw_queues, prev_nr_hw_queues);
5032 for (; i < set->nr_hw_queues; i++)
5033 __blk_mq_free_map_and_rqs(set, i);
5034
5035 set->nr_hw_queues = prev_nr_hw_queues;
5036 goto fallback;
5037 }
5038 blk_mq_map_swqueue(q);
5039 }
5040
5041reregister:
5042 list_for_each_entry(q, &set->tag_list, tag_set_list) {
5043 blk_mq_sysfs_register_hctxs(q);
5044 blk_mq_debugfs_register_hctxs(q);
5045 }
5046
5047switch_back:
5048 list_for_each_entry(q, &set->tag_list, tag_set_list)
5049 blk_mq_elv_switch_back(&head, q);
5050
5051 list_for_each_entry(q, &set->tag_list, tag_set_list)
5052 blk_mq_unfreeze_queue(q);
5053
5054 /* Free the excess tags when nr_hw_queues shrink. */
5055 for (i = set->nr_hw_queues; i < prev_nr_hw_queues; i++)
5056 __blk_mq_free_map_and_rqs(set, i);
5057}
5058
5059void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
5060{
5061 mutex_lock(&set->tag_list_lock);
5062 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
5063 mutex_unlock(&set->tag_list_lock);
5064}
5065EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
5066
5067static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
5068 struct io_comp_batch *iob, unsigned int flags)
5069{
5070 long state = get_current_state();
5071 int ret;
5072
5073 do {
5074 ret = q->mq_ops->poll(hctx, iob);
5075 if (ret > 0) {
5076 __set_current_state(TASK_RUNNING);
5077 return ret;
5078 }
5079
5080 if (signal_pending_state(state, current))
5081 __set_current_state(TASK_RUNNING);
5082 if (task_is_running(current))
5083 return 1;
5084
5085 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
5086 break;
5087 cpu_relax();
5088 } while (!need_resched());
5089
5090 __set_current_state(TASK_RUNNING);
5091 return 0;
5092}
5093
5094int blk_mq_poll(struct request_queue *q, blk_qc_t cookie,
5095 struct io_comp_batch *iob, unsigned int flags)
5096{
5097 if (!blk_mq_can_poll(q))
5098 return 0;
5099 return blk_hctx_poll(q, xa_load(&q->hctx_table, cookie), iob, flags);
5100}
5101
5102int blk_rq_poll(struct request *rq, struct io_comp_batch *iob,
5103 unsigned int poll_flags)
5104{
5105 struct request_queue *q = rq->q;
5106 int ret;
5107
5108 if (!blk_rq_is_poll(rq))
5109 return 0;
5110 if (!percpu_ref_tryget(&q->q_usage_counter))
5111 return 0;
5112
5113 ret = blk_hctx_poll(q, rq->mq_hctx, iob, poll_flags);
5114 blk_queue_exit(q);
5115
5116 return ret;
5117}
5118EXPORT_SYMBOL_GPL(blk_rq_poll);
5119
5120unsigned int blk_mq_rq_cpu(struct request *rq)
5121{
5122 return rq->mq_ctx->cpu;
5123}
5124EXPORT_SYMBOL(blk_mq_rq_cpu);
5125
5126void blk_mq_cancel_work_sync(struct request_queue *q)
5127{
5128 struct blk_mq_hw_ctx *hctx;
5129 unsigned long i;
5130
5131 cancel_delayed_work_sync(&q->requeue_work);
5132
5133 queue_for_each_hw_ctx(q, hctx, i)
5134 cancel_delayed_work_sync(&hctx->run_work);
5135}
5136
5137static int __init blk_mq_init(void)
5138{
5139 int i;
5140
5141 for_each_possible_cpu(i)
5142 init_llist_head(&per_cpu(blk_cpu_done, i));
5143 for_each_possible_cpu(i)
5144 INIT_CSD(&per_cpu(blk_cpu_csd, i),
5145 __blk_mq_complete_request_remote, NULL);
5146 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
5147
5148 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
5149 "block/softirq:dead", NULL,
5150 blk_softirq_cpu_dead);
5151 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
5152 blk_mq_hctx_notify_dead);
5153 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
5154 blk_mq_hctx_notify_online,
5155 blk_mq_hctx_notify_offline);
5156 return 0;
5157}
5158subsys_initcall(blk_mq_init);