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/kmemleak.h>
14#include <linux/mm.h>
15#include <linux/init.h>
16#include <linux/slab.h>
17#include <linux/workqueue.h>
18#include <linux/smp.h>
19#include <linux/llist.h>
20#include <linux/list_sort.h>
21#include <linux/cpu.h>
22#include <linux/cache.h>
23#include <linux/sched/sysctl.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
30#include <trace/events/block.h>
31
32#include <linux/blk-mq.h>
33#include <linux/t10-pi.h>
34#include "blk.h"
35#include "blk-mq.h"
36#include "blk-mq-debugfs.h"
37#include "blk-mq-tag.h"
38#include "blk-pm.h"
39#include "blk-stat.h"
40#include "blk-mq-sched.h"
41#include "blk-rq-qos.h"
42
43static void blk_mq_poll_stats_start(struct request_queue *q);
44static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
45
46static int blk_mq_poll_stats_bkt(const struct request *rq)
47{
48 int ddir, sectors, bucket;
49
50 ddir = rq_data_dir(rq);
51 sectors = blk_rq_stats_sectors(rq);
52
53 bucket = ddir + 2 * ilog2(sectors);
54
55 if (bucket < 0)
56 return -1;
57 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
58 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
59
60 return bucket;
61}
62
63/*
64 * Check if any of the ctx, dispatch list or elevator
65 * have pending work in this hardware queue.
66 */
67static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
68{
69 return !list_empty_careful(&hctx->dispatch) ||
70 sbitmap_any_bit_set(&hctx->ctx_map) ||
71 blk_mq_sched_has_work(hctx);
72}
73
74/*
75 * Mark this ctx as having pending work in this hardware queue
76 */
77static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
78 struct blk_mq_ctx *ctx)
79{
80 const int bit = ctx->index_hw[hctx->type];
81
82 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
83 sbitmap_set_bit(&hctx->ctx_map, bit);
84}
85
86static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
87 struct blk_mq_ctx *ctx)
88{
89 const int bit = ctx->index_hw[hctx->type];
90
91 sbitmap_clear_bit(&hctx->ctx_map, bit);
92}
93
94struct mq_inflight {
95 struct hd_struct *part;
96 unsigned int *inflight;
97};
98
99static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
100 struct request *rq, void *priv,
101 bool reserved)
102{
103 struct mq_inflight *mi = priv;
104
105 /*
106 * index[0] counts the specific partition that was asked for.
107 */
108 if (rq->part == mi->part)
109 mi->inflight[0]++;
110
111 return true;
112}
113
114unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part)
115{
116 unsigned inflight[2];
117 struct mq_inflight mi = { .part = part, .inflight = inflight, };
118
119 inflight[0] = inflight[1] = 0;
120 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
121
122 return inflight[0];
123}
124
125static bool blk_mq_check_inflight_rw(struct blk_mq_hw_ctx *hctx,
126 struct request *rq, void *priv,
127 bool reserved)
128{
129 struct mq_inflight *mi = priv;
130
131 if (rq->part == mi->part)
132 mi->inflight[rq_data_dir(rq)]++;
133
134 return true;
135}
136
137void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
138 unsigned int inflight[2])
139{
140 struct mq_inflight mi = { .part = part, .inflight = inflight, };
141
142 inflight[0] = inflight[1] = 0;
143 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight_rw, &mi);
144}
145
146void blk_freeze_queue_start(struct request_queue *q)
147{
148 mutex_lock(&q->mq_freeze_lock);
149 if (++q->mq_freeze_depth == 1) {
150 percpu_ref_kill(&q->q_usage_counter);
151 mutex_unlock(&q->mq_freeze_lock);
152 if (queue_is_mq(q))
153 blk_mq_run_hw_queues(q, false);
154 } else {
155 mutex_unlock(&q->mq_freeze_lock);
156 }
157}
158EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
159
160void blk_mq_freeze_queue_wait(struct request_queue *q)
161{
162 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
163}
164EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
165
166int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
167 unsigned long timeout)
168{
169 return wait_event_timeout(q->mq_freeze_wq,
170 percpu_ref_is_zero(&q->q_usage_counter),
171 timeout);
172}
173EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
174
175/*
176 * Guarantee no request is in use, so we can change any data structure of
177 * the queue afterward.
178 */
179void blk_freeze_queue(struct request_queue *q)
180{
181 /*
182 * In the !blk_mq case we are only calling this to kill the
183 * q_usage_counter, otherwise this increases the freeze depth
184 * and waits for it to return to zero. For this reason there is
185 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
186 * exported to drivers as the only user for unfreeze is blk_mq.
187 */
188 blk_freeze_queue_start(q);
189 blk_mq_freeze_queue_wait(q);
190}
191
192void blk_mq_freeze_queue(struct request_queue *q)
193{
194 /*
195 * ...just an alias to keep freeze and unfreeze actions balanced
196 * in the blk_mq_* namespace
197 */
198 blk_freeze_queue(q);
199}
200EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
201
202void blk_mq_unfreeze_queue(struct request_queue *q)
203{
204 mutex_lock(&q->mq_freeze_lock);
205 q->mq_freeze_depth--;
206 WARN_ON_ONCE(q->mq_freeze_depth < 0);
207 if (!q->mq_freeze_depth) {
208 percpu_ref_resurrect(&q->q_usage_counter);
209 wake_up_all(&q->mq_freeze_wq);
210 }
211 mutex_unlock(&q->mq_freeze_lock);
212}
213EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
214
215/*
216 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
217 * mpt3sas driver such that this function can be removed.
218 */
219void blk_mq_quiesce_queue_nowait(struct request_queue *q)
220{
221 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
222}
223EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
224
225/**
226 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
227 * @q: request queue.
228 *
229 * Note: this function does not prevent that the struct request end_io()
230 * callback function is invoked. Once this function is returned, we make
231 * sure no dispatch can happen until the queue is unquiesced via
232 * blk_mq_unquiesce_queue().
233 */
234void blk_mq_quiesce_queue(struct request_queue *q)
235{
236 struct blk_mq_hw_ctx *hctx;
237 unsigned int i;
238 bool rcu = false;
239
240 blk_mq_quiesce_queue_nowait(q);
241
242 queue_for_each_hw_ctx(q, hctx, i) {
243 if (hctx->flags & BLK_MQ_F_BLOCKING)
244 synchronize_srcu(hctx->srcu);
245 else
246 rcu = true;
247 }
248 if (rcu)
249 synchronize_rcu();
250}
251EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
252
253/*
254 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
255 * @q: request queue.
256 *
257 * This function recovers queue into the state before quiescing
258 * which is done by blk_mq_quiesce_queue.
259 */
260void blk_mq_unquiesce_queue(struct request_queue *q)
261{
262 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
263
264 /* dispatch requests which are inserted during quiescing */
265 blk_mq_run_hw_queues(q, true);
266}
267EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
268
269void blk_mq_wake_waiters(struct request_queue *q)
270{
271 struct blk_mq_hw_ctx *hctx;
272 unsigned int i;
273
274 queue_for_each_hw_ctx(q, hctx, i)
275 if (blk_mq_hw_queue_mapped(hctx))
276 blk_mq_tag_wakeup_all(hctx->tags, true);
277}
278
279bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
280{
281 return blk_mq_has_free_tags(hctx->tags);
282}
283EXPORT_SYMBOL(blk_mq_can_queue);
284
285/*
286 * Only need start/end time stamping if we have iostat or
287 * blk stats enabled, or using an IO scheduler.
288 */
289static inline bool blk_mq_need_time_stamp(struct request *rq)
290{
291 return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator;
292}
293
294static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
295 unsigned int tag, unsigned int op, u64 alloc_time_ns)
296{
297 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
298 struct request *rq = tags->static_rqs[tag];
299 req_flags_t rq_flags = 0;
300
301 if (data->flags & BLK_MQ_REQ_INTERNAL) {
302 rq->tag = -1;
303 rq->internal_tag = tag;
304 } else {
305 if (data->hctx->flags & BLK_MQ_F_TAG_SHARED) {
306 rq_flags = RQF_MQ_INFLIGHT;
307 atomic_inc(&data->hctx->nr_active);
308 }
309 rq->tag = tag;
310 rq->internal_tag = -1;
311 data->hctx->tags->rqs[rq->tag] = rq;
312 }
313
314 /* csd/requeue_work/fifo_time is initialized before use */
315 rq->q = data->q;
316 rq->mq_ctx = data->ctx;
317 rq->mq_hctx = data->hctx;
318 rq->rq_flags = rq_flags;
319 rq->cmd_flags = op;
320 if (data->flags & BLK_MQ_REQ_PREEMPT)
321 rq->rq_flags |= RQF_PREEMPT;
322 if (blk_queue_io_stat(data->q))
323 rq->rq_flags |= RQF_IO_STAT;
324 INIT_LIST_HEAD(&rq->queuelist);
325 INIT_HLIST_NODE(&rq->hash);
326 RB_CLEAR_NODE(&rq->rb_node);
327 rq->rq_disk = NULL;
328 rq->part = NULL;
329#ifdef CONFIG_BLK_RQ_ALLOC_TIME
330 rq->alloc_time_ns = alloc_time_ns;
331#endif
332 if (blk_mq_need_time_stamp(rq))
333 rq->start_time_ns = ktime_get_ns();
334 else
335 rq->start_time_ns = 0;
336 rq->io_start_time_ns = 0;
337 rq->stats_sectors = 0;
338 rq->nr_phys_segments = 0;
339#if defined(CONFIG_BLK_DEV_INTEGRITY)
340 rq->nr_integrity_segments = 0;
341#endif
342 /* tag was already set */
343 rq->extra_len = 0;
344 WRITE_ONCE(rq->deadline, 0);
345
346 rq->timeout = 0;
347
348 rq->end_io = NULL;
349 rq->end_io_data = NULL;
350
351 data->ctx->rq_dispatched[op_is_sync(op)]++;
352 refcount_set(&rq->ref, 1);
353 return rq;
354}
355
356static struct request *blk_mq_get_request(struct request_queue *q,
357 struct bio *bio,
358 struct blk_mq_alloc_data *data)
359{
360 struct elevator_queue *e = q->elevator;
361 struct request *rq;
362 unsigned int tag;
363 bool clear_ctx_on_error = false;
364 u64 alloc_time_ns = 0;
365
366 blk_queue_enter_live(q);
367
368 /* alloc_time includes depth and tag waits */
369 if (blk_queue_rq_alloc_time(q))
370 alloc_time_ns = ktime_get_ns();
371
372 data->q = q;
373 if (likely(!data->ctx)) {
374 data->ctx = blk_mq_get_ctx(q);
375 clear_ctx_on_error = true;
376 }
377 if (likely(!data->hctx))
378 data->hctx = blk_mq_map_queue(q, data->cmd_flags,
379 data->ctx);
380 if (data->cmd_flags & REQ_NOWAIT)
381 data->flags |= BLK_MQ_REQ_NOWAIT;
382
383 if (e) {
384 data->flags |= BLK_MQ_REQ_INTERNAL;
385
386 /*
387 * Flush requests are special and go directly to the
388 * dispatch list. Don't include reserved tags in the
389 * limiting, as it isn't useful.
390 */
391 if (!op_is_flush(data->cmd_flags) &&
392 e->type->ops.limit_depth &&
393 !(data->flags & BLK_MQ_REQ_RESERVED))
394 e->type->ops.limit_depth(data->cmd_flags, data);
395 } else {
396 blk_mq_tag_busy(data->hctx);
397 }
398
399 tag = blk_mq_get_tag(data);
400 if (tag == BLK_MQ_TAG_FAIL) {
401 if (clear_ctx_on_error)
402 data->ctx = NULL;
403 blk_queue_exit(q);
404 return NULL;
405 }
406
407 rq = blk_mq_rq_ctx_init(data, tag, data->cmd_flags, alloc_time_ns);
408 if (!op_is_flush(data->cmd_flags)) {
409 rq->elv.icq = NULL;
410 if (e && e->type->ops.prepare_request) {
411 if (e->type->icq_cache)
412 blk_mq_sched_assign_ioc(rq);
413
414 e->type->ops.prepare_request(rq, bio);
415 rq->rq_flags |= RQF_ELVPRIV;
416 }
417 }
418 data->hctx->queued++;
419 return rq;
420}
421
422struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
423 blk_mq_req_flags_t flags)
424{
425 struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
426 struct request *rq;
427 int ret;
428
429 ret = blk_queue_enter(q, flags);
430 if (ret)
431 return ERR_PTR(ret);
432
433 rq = blk_mq_get_request(q, NULL, &alloc_data);
434 blk_queue_exit(q);
435
436 if (!rq)
437 return ERR_PTR(-EWOULDBLOCK);
438
439 rq->__data_len = 0;
440 rq->__sector = (sector_t) -1;
441 rq->bio = rq->biotail = NULL;
442 return rq;
443}
444EXPORT_SYMBOL(blk_mq_alloc_request);
445
446struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
447 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
448{
449 struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
450 struct request *rq;
451 unsigned int cpu;
452 int ret;
453
454 /*
455 * If the tag allocator sleeps we could get an allocation for a
456 * different hardware context. No need to complicate the low level
457 * allocator for this for the rare use case of a command tied to
458 * a specific queue.
459 */
460 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
461 return ERR_PTR(-EINVAL);
462
463 if (hctx_idx >= q->nr_hw_queues)
464 return ERR_PTR(-EIO);
465
466 ret = blk_queue_enter(q, flags);
467 if (ret)
468 return ERR_PTR(ret);
469
470 /*
471 * Check if the hardware context is actually mapped to anything.
472 * If not tell the caller that it should skip this queue.
473 */
474 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
475 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
476 blk_queue_exit(q);
477 return ERR_PTR(-EXDEV);
478 }
479 cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
480 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
481
482 rq = blk_mq_get_request(q, NULL, &alloc_data);
483 blk_queue_exit(q);
484
485 if (!rq)
486 return ERR_PTR(-EWOULDBLOCK);
487
488 return rq;
489}
490EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
491
492static void __blk_mq_free_request(struct request *rq)
493{
494 struct request_queue *q = rq->q;
495 struct blk_mq_ctx *ctx = rq->mq_ctx;
496 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
497 const int sched_tag = rq->internal_tag;
498
499 blk_pm_mark_last_busy(rq);
500 rq->mq_hctx = NULL;
501 if (rq->tag != -1)
502 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
503 if (sched_tag != -1)
504 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
505 blk_mq_sched_restart(hctx);
506 blk_queue_exit(q);
507}
508
509void blk_mq_free_request(struct request *rq)
510{
511 struct request_queue *q = rq->q;
512 struct elevator_queue *e = q->elevator;
513 struct blk_mq_ctx *ctx = rq->mq_ctx;
514 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
515
516 if (rq->rq_flags & RQF_ELVPRIV) {
517 if (e && e->type->ops.finish_request)
518 e->type->ops.finish_request(rq);
519 if (rq->elv.icq) {
520 put_io_context(rq->elv.icq->ioc);
521 rq->elv.icq = NULL;
522 }
523 }
524
525 ctx->rq_completed[rq_is_sync(rq)]++;
526 if (rq->rq_flags & RQF_MQ_INFLIGHT)
527 atomic_dec(&hctx->nr_active);
528
529 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
530 laptop_io_completion(q->backing_dev_info);
531
532 rq_qos_done(q, rq);
533
534 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
535 if (refcount_dec_and_test(&rq->ref))
536 __blk_mq_free_request(rq);
537}
538EXPORT_SYMBOL_GPL(blk_mq_free_request);
539
540inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
541{
542 u64 now = 0;
543
544 if (blk_mq_need_time_stamp(rq))
545 now = ktime_get_ns();
546
547 if (rq->rq_flags & RQF_STATS) {
548 blk_mq_poll_stats_start(rq->q);
549 blk_stat_add(rq, now);
550 }
551
552 if (rq->internal_tag != -1)
553 blk_mq_sched_completed_request(rq, now);
554
555 blk_account_io_done(rq, now);
556
557 if (rq->end_io) {
558 rq_qos_done(rq->q, rq);
559 rq->end_io(rq, error);
560 } else {
561 blk_mq_free_request(rq);
562 }
563}
564EXPORT_SYMBOL(__blk_mq_end_request);
565
566void blk_mq_end_request(struct request *rq, blk_status_t error)
567{
568 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
569 BUG();
570 __blk_mq_end_request(rq, error);
571}
572EXPORT_SYMBOL(blk_mq_end_request);
573
574static void __blk_mq_complete_request_remote(void *data)
575{
576 struct request *rq = data;
577 struct request_queue *q = rq->q;
578
579 q->mq_ops->complete(rq);
580}
581
582static void __blk_mq_complete_request(struct request *rq)
583{
584 struct blk_mq_ctx *ctx = rq->mq_ctx;
585 struct request_queue *q = rq->q;
586 bool shared = false;
587 int cpu;
588
589 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
590 /*
591 * Most of single queue controllers, there is only one irq vector
592 * for handling IO completion, and the only irq's affinity is set
593 * as all possible CPUs. On most of ARCHs, this affinity means the
594 * irq is handled on one specific CPU.
595 *
596 * So complete IO reqeust in softirq context in case of single queue
597 * for not degrading IO performance by irqsoff latency.
598 */
599 if (q->nr_hw_queues == 1) {
600 __blk_complete_request(rq);
601 return;
602 }
603
604 /*
605 * For a polled request, always complete locallly, it's pointless
606 * to redirect the completion.
607 */
608 if ((rq->cmd_flags & REQ_HIPRI) ||
609 !test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags)) {
610 q->mq_ops->complete(rq);
611 return;
612 }
613
614 cpu = get_cpu();
615 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &q->queue_flags))
616 shared = cpus_share_cache(cpu, ctx->cpu);
617
618 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
619 rq->csd.func = __blk_mq_complete_request_remote;
620 rq->csd.info = rq;
621 rq->csd.flags = 0;
622 smp_call_function_single_async(ctx->cpu, &rq->csd);
623 } else {
624 q->mq_ops->complete(rq);
625 }
626 put_cpu();
627}
628
629static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
630 __releases(hctx->srcu)
631{
632 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
633 rcu_read_unlock();
634 else
635 srcu_read_unlock(hctx->srcu, srcu_idx);
636}
637
638static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
639 __acquires(hctx->srcu)
640{
641 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
642 /* shut up gcc false positive */
643 *srcu_idx = 0;
644 rcu_read_lock();
645 } else
646 *srcu_idx = srcu_read_lock(hctx->srcu);
647}
648
649/**
650 * blk_mq_complete_request - end I/O on a request
651 * @rq: the request being processed
652 *
653 * Description:
654 * Ends all I/O on a request. It does not handle partial completions.
655 * The actual completion happens out-of-order, through a IPI handler.
656 **/
657bool blk_mq_complete_request(struct request *rq)
658{
659 if (unlikely(blk_should_fake_timeout(rq->q)))
660 return false;
661 __blk_mq_complete_request(rq);
662 return true;
663}
664EXPORT_SYMBOL(blk_mq_complete_request);
665
666int blk_mq_request_started(struct request *rq)
667{
668 return blk_mq_rq_state(rq) != MQ_RQ_IDLE;
669}
670EXPORT_SYMBOL_GPL(blk_mq_request_started);
671
672int blk_mq_request_completed(struct request *rq)
673{
674 return blk_mq_rq_state(rq) == MQ_RQ_COMPLETE;
675}
676EXPORT_SYMBOL_GPL(blk_mq_request_completed);
677
678void blk_mq_start_request(struct request *rq)
679{
680 struct request_queue *q = rq->q;
681
682 trace_block_rq_issue(q, rq);
683
684 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
685 rq->io_start_time_ns = ktime_get_ns();
686 rq->stats_sectors = blk_rq_sectors(rq);
687 rq->rq_flags |= RQF_STATS;
688 rq_qos_issue(q, rq);
689 }
690
691 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
692
693 blk_add_timer(rq);
694 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
695
696 if (q->dma_drain_size && blk_rq_bytes(rq)) {
697 /*
698 * Make sure space for the drain appears. We know we can do
699 * this because max_hw_segments has been adjusted to be one
700 * fewer than the device can handle.
701 */
702 rq->nr_phys_segments++;
703 }
704
705#ifdef CONFIG_BLK_DEV_INTEGRITY
706 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
707 q->integrity.profile->prepare_fn(rq);
708#endif
709}
710EXPORT_SYMBOL(blk_mq_start_request);
711
712static void __blk_mq_requeue_request(struct request *rq)
713{
714 struct request_queue *q = rq->q;
715
716 blk_mq_put_driver_tag(rq);
717
718 trace_block_rq_requeue(q, rq);
719 rq_qos_requeue(q, rq);
720
721 if (blk_mq_request_started(rq)) {
722 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
723 rq->rq_flags &= ~RQF_TIMED_OUT;
724 if (q->dma_drain_size && blk_rq_bytes(rq))
725 rq->nr_phys_segments--;
726 }
727}
728
729void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
730{
731 __blk_mq_requeue_request(rq);
732
733 /* this request will be re-inserted to io scheduler queue */
734 blk_mq_sched_requeue_request(rq);
735
736 BUG_ON(!list_empty(&rq->queuelist));
737 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
738}
739EXPORT_SYMBOL(blk_mq_requeue_request);
740
741static void blk_mq_requeue_work(struct work_struct *work)
742{
743 struct request_queue *q =
744 container_of(work, struct request_queue, requeue_work.work);
745 LIST_HEAD(rq_list);
746 struct request *rq, *next;
747
748 spin_lock_irq(&q->requeue_lock);
749 list_splice_init(&q->requeue_list, &rq_list);
750 spin_unlock_irq(&q->requeue_lock);
751
752 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
753 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
754 continue;
755
756 rq->rq_flags &= ~RQF_SOFTBARRIER;
757 list_del_init(&rq->queuelist);
758 /*
759 * If RQF_DONTPREP, rq has contained some driver specific
760 * data, so insert it to hctx dispatch list to avoid any
761 * merge.
762 */
763 if (rq->rq_flags & RQF_DONTPREP)
764 blk_mq_request_bypass_insert(rq, false);
765 else
766 blk_mq_sched_insert_request(rq, true, false, false);
767 }
768
769 while (!list_empty(&rq_list)) {
770 rq = list_entry(rq_list.next, struct request, queuelist);
771 list_del_init(&rq->queuelist);
772 blk_mq_sched_insert_request(rq, false, false, false);
773 }
774
775 blk_mq_run_hw_queues(q, false);
776}
777
778void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
779 bool kick_requeue_list)
780{
781 struct request_queue *q = rq->q;
782 unsigned long flags;
783
784 /*
785 * We abuse this flag that is otherwise used by the I/O scheduler to
786 * request head insertion from the workqueue.
787 */
788 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
789
790 spin_lock_irqsave(&q->requeue_lock, flags);
791 if (at_head) {
792 rq->rq_flags |= RQF_SOFTBARRIER;
793 list_add(&rq->queuelist, &q->requeue_list);
794 } else {
795 list_add_tail(&rq->queuelist, &q->requeue_list);
796 }
797 spin_unlock_irqrestore(&q->requeue_lock, flags);
798
799 if (kick_requeue_list)
800 blk_mq_kick_requeue_list(q);
801}
802
803void blk_mq_kick_requeue_list(struct request_queue *q)
804{
805 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
806}
807EXPORT_SYMBOL(blk_mq_kick_requeue_list);
808
809void blk_mq_delay_kick_requeue_list(struct request_queue *q,
810 unsigned long msecs)
811{
812 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
813 msecs_to_jiffies(msecs));
814}
815EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
816
817struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
818{
819 if (tag < tags->nr_tags) {
820 prefetch(tags->rqs[tag]);
821 return tags->rqs[tag];
822 }
823
824 return NULL;
825}
826EXPORT_SYMBOL(blk_mq_tag_to_rq);
827
828static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
829 void *priv, bool reserved)
830{
831 /*
832 * If we find a request that is inflight and the queue matches,
833 * we know the queue is busy. Return false to stop the iteration.
834 */
835 if (rq->state == MQ_RQ_IN_FLIGHT && rq->q == hctx->queue) {
836 bool *busy = priv;
837
838 *busy = true;
839 return false;
840 }
841
842 return true;
843}
844
845bool blk_mq_queue_inflight(struct request_queue *q)
846{
847 bool busy = false;
848
849 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
850 return busy;
851}
852EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
853
854static void blk_mq_rq_timed_out(struct request *req, bool reserved)
855{
856 req->rq_flags |= RQF_TIMED_OUT;
857 if (req->q->mq_ops->timeout) {
858 enum blk_eh_timer_return ret;
859
860 ret = req->q->mq_ops->timeout(req, reserved);
861 if (ret == BLK_EH_DONE)
862 return;
863 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
864 }
865
866 blk_add_timer(req);
867}
868
869static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
870{
871 unsigned long deadline;
872
873 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
874 return false;
875 if (rq->rq_flags & RQF_TIMED_OUT)
876 return false;
877
878 deadline = READ_ONCE(rq->deadline);
879 if (time_after_eq(jiffies, deadline))
880 return true;
881
882 if (*next == 0)
883 *next = deadline;
884 else if (time_after(*next, deadline))
885 *next = deadline;
886 return false;
887}
888
889static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
890 struct request *rq, void *priv, bool reserved)
891{
892 unsigned long *next = priv;
893
894 /*
895 * Just do a quick check if it is expired before locking the request in
896 * so we're not unnecessarilly synchronizing across CPUs.
897 */
898 if (!blk_mq_req_expired(rq, next))
899 return true;
900
901 /*
902 * We have reason to believe the request may be expired. Take a
903 * reference on the request to lock this request lifetime into its
904 * currently allocated context to prevent it from being reallocated in
905 * the event the completion by-passes this timeout handler.
906 *
907 * If the reference was already released, then the driver beat the
908 * timeout handler to posting a natural completion.
909 */
910 if (!refcount_inc_not_zero(&rq->ref))
911 return true;
912
913 /*
914 * The request is now locked and cannot be reallocated underneath the
915 * timeout handler's processing. Re-verify this exact request is truly
916 * expired; if it is not expired, then the request was completed and
917 * reallocated as a new request.
918 */
919 if (blk_mq_req_expired(rq, next))
920 blk_mq_rq_timed_out(rq, reserved);
921
922 if (is_flush_rq(rq, hctx))
923 rq->end_io(rq, 0);
924 else if (refcount_dec_and_test(&rq->ref))
925 __blk_mq_free_request(rq);
926
927 return true;
928}
929
930static void blk_mq_timeout_work(struct work_struct *work)
931{
932 struct request_queue *q =
933 container_of(work, struct request_queue, timeout_work);
934 unsigned long next = 0;
935 struct blk_mq_hw_ctx *hctx;
936 int i;
937
938 /* A deadlock might occur if a request is stuck requiring a
939 * timeout at the same time a queue freeze is waiting
940 * completion, since the timeout code would not be able to
941 * acquire the queue reference here.
942 *
943 * That's why we don't use blk_queue_enter here; instead, we use
944 * percpu_ref_tryget directly, because we need to be able to
945 * obtain a reference even in the short window between the queue
946 * starting to freeze, by dropping the first reference in
947 * blk_freeze_queue_start, and the moment the last request is
948 * consumed, marked by the instant q_usage_counter reaches
949 * zero.
950 */
951 if (!percpu_ref_tryget(&q->q_usage_counter))
952 return;
953
954 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
955
956 if (next != 0) {
957 mod_timer(&q->timeout, next);
958 } else {
959 /*
960 * Request timeouts are handled as a forward rolling timer. If
961 * we end up here it means that no requests are pending and
962 * also that no request has been pending for a while. Mark
963 * each hctx as idle.
964 */
965 queue_for_each_hw_ctx(q, hctx, i) {
966 /* the hctx may be unmapped, so check it here */
967 if (blk_mq_hw_queue_mapped(hctx))
968 blk_mq_tag_idle(hctx);
969 }
970 }
971 blk_queue_exit(q);
972}
973
974struct flush_busy_ctx_data {
975 struct blk_mq_hw_ctx *hctx;
976 struct list_head *list;
977};
978
979static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
980{
981 struct flush_busy_ctx_data *flush_data = data;
982 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
983 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
984 enum hctx_type type = hctx->type;
985
986 spin_lock(&ctx->lock);
987 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
988 sbitmap_clear_bit(sb, bitnr);
989 spin_unlock(&ctx->lock);
990 return true;
991}
992
993/*
994 * Process software queues that have been marked busy, splicing them
995 * to the for-dispatch
996 */
997void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
998{
999 struct flush_busy_ctx_data data = {
1000 .hctx = hctx,
1001 .list = list,
1002 };
1003
1004 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1005}
1006EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1007
1008struct dispatch_rq_data {
1009 struct blk_mq_hw_ctx *hctx;
1010 struct request *rq;
1011};
1012
1013static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1014 void *data)
1015{
1016 struct dispatch_rq_data *dispatch_data = data;
1017 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1018 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1019 enum hctx_type type = hctx->type;
1020
1021 spin_lock(&ctx->lock);
1022 if (!list_empty(&ctx->rq_lists[type])) {
1023 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1024 list_del_init(&dispatch_data->rq->queuelist);
1025 if (list_empty(&ctx->rq_lists[type]))
1026 sbitmap_clear_bit(sb, bitnr);
1027 }
1028 spin_unlock(&ctx->lock);
1029
1030 return !dispatch_data->rq;
1031}
1032
1033struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1034 struct blk_mq_ctx *start)
1035{
1036 unsigned off = start ? start->index_hw[hctx->type] : 0;
1037 struct dispatch_rq_data data = {
1038 .hctx = hctx,
1039 .rq = NULL,
1040 };
1041
1042 __sbitmap_for_each_set(&hctx->ctx_map, off,
1043 dispatch_rq_from_ctx, &data);
1044
1045 return data.rq;
1046}
1047
1048static inline unsigned int queued_to_index(unsigned int queued)
1049{
1050 if (!queued)
1051 return 0;
1052
1053 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1054}
1055
1056bool blk_mq_get_driver_tag(struct request *rq)
1057{
1058 struct blk_mq_alloc_data data = {
1059 .q = rq->q,
1060 .hctx = rq->mq_hctx,
1061 .flags = BLK_MQ_REQ_NOWAIT,
1062 .cmd_flags = rq->cmd_flags,
1063 };
1064 bool shared;
1065
1066 if (rq->tag != -1)
1067 goto done;
1068
1069 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1070 data.flags |= BLK_MQ_REQ_RESERVED;
1071
1072 shared = blk_mq_tag_busy(data.hctx);
1073 rq->tag = blk_mq_get_tag(&data);
1074 if (rq->tag >= 0) {
1075 if (shared) {
1076 rq->rq_flags |= RQF_MQ_INFLIGHT;
1077 atomic_inc(&data.hctx->nr_active);
1078 }
1079 data.hctx->tags->rqs[rq->tag] = rq;
1080 }
1081
1082done:
1083 return rq->tag != -1;
1084}
1085
1086static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1087 int flags, void *key)
1088{
1089 struct blk_mq_hw_ctx *hctx;
1090
1091 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1092
1093 spin_lock(&hctx->dispatch_wait_lock);
1094 if (!list_empty(&wait->entry)) {
1095 struct sbitmap_queue *sbq;
1096
1097 list_del_init(&wait->entry);
1098 sbq = &hctx->tags->bitmap_tags;
1099 atomic_dec(&sbq->ws_active);
1100 }
1101 spin_unlock(&hctx->dispatch_wait_lock);
1102
1103 blk_mq_run_hw_queue(hctx, true);
1104 return 1;
1105}
1106
1107/*
1108 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1109 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1110 * restart. For both cases, take care to check the condition again after
1111 * marking us as waiting.
1112 */
1113static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1114 struct request *rq)
1115{
1116 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1117 struct wait_queue_head *wq;
1118 wait_queue_entry_t *wait;
1119 bool ret;
1120
1121 if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1122 blk_mq_sched_mark_restart_hctx(hctx);
1123
1124 /*
1125 * It's possible that a tag was freed in the window between the
1126 * allocation failure and adding the hardware queue to the wait
1127 * queue.
1128 *
1129 * Don't clear RESTART here, someone else could have set it.
1130 * At most this will cost an extra queue run.
1131 */
1132 return blk_mq_get_driver_tag(rq);
1133 }
1134
1135 wait = &hctx->dispatch_wait;
1136 if (!list_empty_careful(&wait->entry))
1137 return false;
1138
1139 wq = &bt_wait_ptr(sbq, hctx)->wait;
1140
1141 spin_lock_irq(&wq->lock);
1142 spin_lock(&hctx->dispatch_wait_lock);
1143 if (!list_empty(&wait->entry)) {
1144 spin_unlock(&hctx->dispatch_wait_lock);
1145 spin_unlock_irq(&wq->lock);
1146 return false;
1147 }
1148
1149 atomic_inc(&sbq->ws_active);
1150 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1151 __add_wait_queue(wq, wait);
1152
1153 /*
1154 * It's possible that a tag was freed in the window between the
1155 * allocation failure and adding the hardware queue to the wait
1156 * queue.
1157 */
1158 ret = blk_mq_get_driver_tag(rq);
1159 if (!ret) {
1160 spin_unlock(&hctx->dispatch_wait_lock);
1161 spin_unlock_irq(&wq->lock);
1162 return false;
1163 }
1164
1165 /*
1166 * We got a tag, remove ourselves from the wait queue to ensure
1167 * someone else gets the wakeup.
1168 */
1169 list_del_init(&wait->entry);
1170 atomic_dec(&sbq->ws_active);
1171 spin_unlock(&hctx->dispatch_wait_lock);
1172 spin_unlock_irq(&wq->lock);
1173
1174 return true;
1175}
1176
1177#define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1178#define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1179/*
1180 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1181 * - EWMA is one simple way to compute running average value
1182 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1183 * - take 4 as factor for avoiding to get too small(0) result, and this
1184 * factor doesn't matter because EWMA decreases exponentially
1185 */
1186static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1187{
1188 unsigned int ewma;
1189
1190 if (hctx->queue->elevator)
1191 return;
1192
1193 ewma = hctx->dispatch_busy;
1194
1195 if (!ewma && !busy)
1196 return;
1197
1198 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1199 if (busy)
1200 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1201 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1202
1203 hctx->dispatch_busy = ewma;
1204}
1205
1206#define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1207
1208/*
1209 * Returns true if we did some work AND can potentially do more.
1210 */
1211bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1212 bool got_budget)
1213{
1214 struct blk_mq_hw_ctx *hctx;
1215 struct request *rq, *nxt;
1216 bool no_tag = false;
1217 int errors, queued;
1218 blk_status_t ret = BLK_STS_OK;
1219
1220 if (list_empty(list))
1221 return false;
1222
1223 WARN_ON(!list_is_singular(list) && got_budget);
1224
1225 /*
1226 * Now process all the entries, sending them to the driver.
1227 */
1228 errors = queued = 0;
1229 do {
1230 struct blk_mq_queue_data bd;
1231
1232 rq = list_first_entry(list, struct request, queuelist);
1233
1234 hctx = rq->mq_hctx;
1235 if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1236 break;
1237
1238 if (!blk_mq_get_driver_tag(rq)) {
1239 /*
1240 * The initial allocation attempt failed, so we need to
1241 * rerun the hardware queue when a tag is freed. The
1242 * waitqueue takes care of that. If the queue is run
1243 * before we add this entry back on the dispatch list,
1244 * we'll re-run it below.
1245 */
1246 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1247 blk_mq_put_dispatch_budget(hctx);
1248 /*
1249 * For non-shared tags, the RESTART check
1250 * will suffice.
1251 */
1252 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1253 no_tag = true;
1254 break;
1255 }
1256 }
1257
1258 list_del_init(&rq->queuelist);
1259
1260 bd.rq = rq;
1261
1262 /*
1263 * Flag last if we have no more requests, or if we have more
1264 * but can't assign a driver tag to it.
1265 */
1266 if (list_empty(list))
1267 bd.last = true;
1268 else {
1269 nxt = list_first_entry(list, struct request, queuelist);
1270 bd.last = !blk_mq_get_driver_tag(nxt);
1271 }
1272
1273 ret = q->mq_ops->queue_rq(hctx, &bd);
1274 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1275 /*
1276 * If an I/O scheduler has been configured and we got a
1277 * driver tag for the next request already, free it
1278 * again.
1279 */
1280 if (!list_empty(list)) {
1281 nxt = list_first_entry(list, struct request, queuelist);
1282 blk_mq_put_driver_tag(nxt);
1283 }
1284 list_add(&rq->queuelist, list);
1285 __blk_mq_requeue_request(rq);
1286 break;
1287 }
1288
1289 if (unlikely(ret != BLK_STS_OK)) {
1290 errors++;
1291 blk_mq_end_request(rq, BLK_STS_IOERR);
1292 continue;
1293 }
1294
1295 queued++;
1296 } while (!list_empty(list));
1297
1298 hctx->dispatched[queued_to_index(queued)]++;
1299
1300 /*
1301 * Any items that need requeuing? Stuff them into hctx->dispatch,
1302 * that is where we will continue on next queue run.
1303 */
1304 if (!list_empty(list)) {
1305 bool needs_restart;
1306
1307 /*
1308 * If we didn't flush the entire list, we could have told
1309 * the driver there was more coming, but that turned out to
1310 * be a lie.
1311 */
1312 if (q->mq_ops->commit_rqs)
1313 q->mq_ops->commit_rqs(hctx);
1314
1315 spin_lock(&hctx->lock);
1316 list_splice_init(list, &hctx->dispatch);
1317 spin_unlock(&hctx->lock);
1318
1319 /*
1320 * If SCHED_RESTART was set by the caller of this function and
1321 * it is no longer set that means that it was cleared by another
1322 * thread and hence that a queue rerun is needed.
1323 *
1324 * If 'no_tag' is set, that means that we failed getting
1325 * a driver tag with an I/O scheduler attached. If our dispatch
1326 * waitqueue is no longer active, ensure that we run the queue
1327 * AFTER adding our entries back to the list.
1328 *
1329 * If no I/O scheduler has been configured it is possible that
1330 * the hardware queue got stopped and restarted before requests
1331 * were pushed back onto the dispatch list. Rerun the queue to
1332 * avoid starvation. Notes:
1333 * - blk_mq_run_hw_queue() checks whether or not a queue has
1334 * been stopped before rerunning a queue.
1335 * - Some but not all block drivers stop a queue before
1336 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1337 * and dm-rq.
1338 *
1339 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1340 * bit is set, run queue after a delay to avoid IO stalls
1341 * that could otherwise occur if the queue is idle.
1342 */
1343 needs_restart = blk_mq_sched_needs_restart(hctx);
1344 if (!needs_restart ||
1345 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1346 blk_mq_run_hw_queue(hctx, true);
1347 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1348 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1349
1350 blk_mq_update_dispatch_busy(hctx, true);
1351 return false;
1352 } else
1353 blk_mq_update_dispatch_busy(hctx, false);
1354
1355 /*
1356 * If the host/device is unable to accept more work, inform the
1357 * caller of that.
1358 */
1359 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1360 return false;
1361
1362 return (queued + errors) != 0;
1363}
1364
1365static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1366{
1367 int srcu_idx;
1368
1369 /*
1370 * We should be running this queue from one of the CPUs that
1371 * are mapped to it.
1372 *
1373 * There are at least two related races now between setting
1374 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1375 * __blk_mq_run_hw_queue():
1376 *
1377 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1378 * but later it becomes online, then this warning is harmless
1379 * at all
1380 *
1381 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1382 * but later it becomes offline, then the warning can't be
1383 * triggered, and we depend on blk-mq timeout handler to
1384 * handle dispatched requests to this hctx
1385 */
1386 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1387 cpu_online(hctx->next_cpu)) {
1388 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1389 raw_smp_processor_id(),
1390 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1391 dump_stack();
1392 }
1393
1394 /*
1395 * We can't run the queue inline with ints disabled. Ensure that
1396 * we catch bad users of this early.
1397 */
1398 WARN_ON_ONCE(in_interrupt());
1399
1400 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1401
1402 hctx_lock(hctx, &srcu_idx);
1403 blk_mq_sched_dispatch_requests(hctx);
1404 hctx_unlock(hctx, srcu_idx);
1405}
1406
1407static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1408{
1409 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1410
1411 if (cpu >= nr_cpu_ids)
1412 cpu = cpumask_first(hctx->cpumask);
1413 return cpu;
1414}
1415
1416/*
1417 * It'd be great if the workqueue API had a way to pass
1418 * in a mask and had some smarts for more clever placement.
1419 * For now we just round-robin here, switching for every
1420 * BLK_MQ_CPU_WORK_BATCH queued items.
1421 */
1422static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1423{
1424 bool tried = false;
1425 int next_cpu = hctx->next_cpu;
1426
1427 if (hctx->queue->nr_hw_queues == 1)
1428 return WORK_CPU_UNBOUND;
1429
1430 if (--hctx->next_cpu_batch <= 0) {
1431select_cpu:
1432 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1433 cpu_online_mask);
1434 if (next_cpu >= nr_cpu_ids)
1435 next_cpu = blk_mq_first_mapped_cpu(hctx);
1436 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1437 }
1438
1439 /*
1440 * Do unbound schedule if we can't find a online CPU for this hctx,
1441 * and it should only happen in the path of handling CPU DEAD.
1442 */
1443 if (!cpu_online(next_cpu)) {
1444 if (!tried) {
1445 tried = true;
1446 goto select_cpu;
1447 }
1448
1449 /*
1450 * Make sure to re-select CPU next time once after CPUs
1451 * in hctx->cpumask become online again.
1452 */
1453 hctx->next_cpu = next_cpu;
1454 hctx->next_cpu_batch = 1;
1455 return WORK_CPU_UNBOUND;
1456 }
1457
1458 hctx->next_cpu = next_cpu;
1459 return next_cpu;
1460}
1461
1462static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1463 unsigned long msecs)
1464{
1465 if (unlikely(blk_mq_hctx_stopped(hctx)))
1466 return;
1467
1468 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1469 int cpu = get_cpu();
1470 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1471 __blk_mq_run_hw_queue(hctx);
1472 put_cpu();
1473 return;
1474 }
1475
1476 put_cpu();
1477 }
1478
1479 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1480 msecs_to_jiffies(msecs));
1481}
1482
1483void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1484{
1485 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1486}
1487EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1488
1489bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1490{
1491 int srcu_idx;
1492 bool need_run;
1493
1494 /*
1495 * When queue is quiesced, we may be switching io scheduler, or
1496 * updating nr_hw_queues, or other things, and we can't run queue
1497 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1498 *
1499 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1500 * quiesced.
1501 */
1502 hctx_lock(hctx, &srcu_idx);
1503 need_run = !blk_queue_quiesced(hctx->queue) &&
1504 blk_mq_hctx_has_pending(hctx);
1505 hctx_unlock(hctx, srcu_idx);
1506
1507 if (need_run) {
1508 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1509 return true;
1510 }
1511
1512 return false;
1513}
1514EXPORT_SYMBOL(blk_mq_run_hw_queue);
1515
1516void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1517{
1518 struct blk_mq_hw_ctx *hctx;
1519 int i;
1520
1521 queue_for_each_hw_ctx(q, hctx, i) {
1522 if (blk_mq_hctx_stopped(hctx))
1523 continue;
1524
1525 blk_mq_run_hw_queue(hctx, async);
1526 }
1527}
1528EXPORT_SYMBOL(blk_mq_run_hw_queues);
1529
1530/**
1531 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1532 * @q: request queue.
1533 *
1534 * The caller is responsible for serializing this function against
1535 * blk_mq_{start,stop}_hw_queue().
1536 */
1537bool blk_mq_queue_stopped(struct request_queue *q)
1538{
1539 struct blk_mq_hw_ctx *hctx;
1540 int i;
1541
1542 queue_for_each_hw_ctx(q, hctx, i)
1543 if (blk_mq_hctx_stopped(hctx))
1544 return true;
1545
1546 return false;
1547}
1548EXPORT_SYMBOL(blk_mq_queue_stopped);
1549
1550/*
1551 * This function is often used for pausing .queue_rq() by driver when
1552 * there isn't enough resource or some conditions aren't satisfied, and
1553 * BLK_STS_RESOURCE is usually returned.
1554 *
1555 * We do not guarantee that dispatch can be drained or blocked
1556 * after blk_mq_stop_hw_queue() returns. Please use
1557 * blk_mq_quiesce_queue() for that requirement.
1558 */
1559void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1560{
1561 cancel_delayed_work(&hctx->run_work);
1562
1563 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1564}
1565EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1566
1567/*
1568 * This function is often used for pausing .queue_rq() by driver when
1569 * there isn't enough resource or some conditions aren't satisfied, and
1570 * BLK_STS_RESOURCE is usually returned.
1571 *
1572 * We do not guarantee that dispatch can be drained or blocked
1573 * after blk_mq_stop_hw_queues() returns. Please use
1574 * blk_mq_quiesce_queue() for that requirement.
1575 */
1576void blk_mq_stop_hw_queues(struct request_queue *q)
1577{
1578 struct blk_mq_hw_ctx *hctx;
1579 int i;
1580
1581 queue_for_each_hw_ctx(q, hctx, i)
1582 blk_mq_stop_hw_queue(hctx);
1583}
1584EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1585
1586void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1587{
1588 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1589
1590 blk_mq_run_hw_queue(hctx, false);
1591}
1592EXPORT_SYMBOL(blk_mq_start_hw_queue);
1593
1594void blk_mq_start_hw_queues(struct request_queue *q)
1595{
1596 struct blk_mq_hw_ctx *hctx;
1597 int i;
1598
1599 queue_for_each_hw_ctx(q, hctx, i)
1600 blk_mq_start_hw_queue(hctx);
1601}
1602EXPORT_SYMBOL(blk_mq_start_hw_queues);
1603
1604void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1605{
1606 if (!blk_mq_hctx_stopped(hctx))
1607 return;
1608
1609 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1610 blk_mq_run_hw_queue(hctx, async);
1611}
1612EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1613
1614void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1615{
1616 struct blk_mq_hw_ctx *hctx;
1617 int i;
1618
1619 queue_for_each_hw_ctx(q, hctx, i)
1620 blk_mq_start_stopped_hw_queue(hctx, async);
1621}
1622EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1623
1624static void blk_mq_run_work_fn(struct work_struct *work)
1625{
1626 struct blk_mq_hw_ctx *hctx;
1627
1628 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1629
1630 /*
1631 * If we are stopped, don't run the queue.
1632 */
1633 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1634 return;
1635
1636 __blk_mq_run_hw_queue(hctx);
1637}
1638
1639static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1640 struct request *rq,
1641 bool at_head)
1642{
1643 struct blk_mq_ctx *ctx = rq->mq_ctx;
1644 enum hctx_type type = hctx->type;
1645
1646 lockdep_assert_held(&ctx->lock);
1647
1648 trace_block_rq_insert(hctx->queue, rq);
1649
1650 if (at_head)
1651 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1652 else
1653 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1654}
1655
1656void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1657 bool at_head)
1658{
1659 struct blk_mq_ctx *ctx = rq->mq_ctx;
1660
1661 lockdep_assert_held(&ctx->lock);
1662
1663 __blk_mq_insert_req_list(hctx, rq, at_head);
1664 blk_mq_hctx_mark_pending(hctx, ctx);
1665}
1666
1667/*
1668 * Should only be used carefully, when the caller knows we want to
1669 * bypass a potential IO scheduler on the target device.
1670 */
1671void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1672{
1673 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1674
1675 spin_lock(&hctx->lock);
1676 list_add_tail(&rq->queuelist, &hctx->dispatch);
1677 spin_unlock(&hctx->lock);
1678
1679 if (run_queue)
1680 blk_mq_run_hw_queue(hctx, false);
1681}
1682
1683void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1684 struct list_head *list)
1685
1686{
1687 struct request *rq;
1688 enum hctx_type type = hctx->type;
1689
1690 /*
1691 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1692 * offline now
1693 */
1694 list_for_each_entry(rq, list, queuelist) {
1695 BUG_ON(rq->mq_ctx != ctx);
1696 trace_block_rq_insert(hctx->queue, rq);
1697 }
1698
1699 spin_lock(&ctx->lock);
1700 list_splice_tail_init(list, &ctx->rq_lists[type]);
1701 blk_mq_hctx_mark_pending(hctx, ctx);
1702 spin_unlock(&ctx->lock);
1703}
1704
1705static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1706{
1707 struct request *rqa = container_of(a, struct request, queuelist);
1708 struct request *rqb = container_of(b, struct request, queuelist);
1709
1710 if (rqa->mq_ctx < rqb->mq_ctx)
1711 return -1;
1712 else if (rqa->mq_ctx > rqb->mq_ctx)
1713 return 1;
1714 else if (rqa->mq_hctx < rqb->mq_hctx)
1715 return -1;
1716 else if (rqa->mq_hctx > rqb->mq_hctx)
1717 return 1;
1718
1719 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1720}
1721
1722void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1723{
1724 struct blk_mq_hw_ctx *this_hctx;
1725 struct blk_mq_ctx *this_ctx;
1726 struct request_queue *this_q;
1727 struct request *rq;
1728 LIST_HEAD(list);
1729 LIST_HEAD(rq_list);
1730 unsigned int depth;
1731
1732 list_splice_init(&plug->mq_list, &list);
1733
1734 if (plug->rq_count > 2 && plug->multiple_queues)
1735 list_sort(NULL, &list, plug_rq_cmp);
1736
1737 plug->rq_count = 0;
1738
1739 this_q = NULL;
1740 this_hctx = NULL;
1741 this_ctx = NULL;
1742 depth = 0;
1743
1744 while (!list_empty(&list)) {
1745 rq = list_entry_rq(list.next);
1746 list_del_init(&rq->queuelist);
1747 BUG_ON(!rq->q);
1748 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) {
1749 if (this_hctx) {
1750 trace_block_unplug(this_q, depth, !from_schedule);
1751 blk_mq_sched_insert_requests(this_hctx, this_ctx,
1752 &rq_list,
1753 from_schedule);
1754 }
1755
1756 this_q = rq->q;
1757 this_ctx = rq->mq_ctx;
1758 this_hctx = rq->mq_hctx;
1759 depth = 0;
1760 }
1761
1762 depth++;
1763 list_add_tail(&rq->queuelist, &rq_list);
1764 }
1765
1766 /*
1767 * If 'this_hctx' is set, we know we have entries to complete
1768 * on 'rq_list'. Do those.
1769 */
1770 if (this_hctx) {
1771 trace_block_unplug(this_q, depth, !from_schedule);
1772 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1773 from_schedule);
1774 }
1775}
1776
1777static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1778 unsigned int nr_segs)
1779{
1780 if (bio->bi_opf & REQ_RAHEAD)
1781 rq->cmd_flags |= REQ_FAILFAST_MASK;
1782
1783 rq->__sector = bio->bi_iter.bi_sector;
1784 rq->write_hint = bio->bi_write_hint;
1785 blk_rq_bio_prep(rq, bio, nr_segs);
1786
1787 blk_account_io_start(rq, true);
1788}
1789
1790static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1791 struct request *rq,
1792 blk_qc_t *cookie, bool last)
1793{
1794 struct request_queue *q = rq->q;
1795 struct blk_mq_queue_data bd = {
1796 .rq = rq,
1797 .last = last,
1798 };
1799 blk_qc_t new_cookie;
1800 blk_status_t ret;
1801
1802 new_cookie = request_to_qc_t(hctx, rq);
1803
1804 /*
1805 * For OK queue, we are done. For error, caller may kill it.
1806 * Any other error (busy), just add it to our list as we
1807 * previously would have done.
1808 */
1809 ret = q->mq_ops->queue_rq(hctx, &bd);
1810 switch (ret) {
1811 case BLK_STS_OK:
1812 blk_mq_update_dispatch_busy(hctx, false);
1813 *cookie = new_cookie;
1814 break;
1815 case BLK_STS_RESOURCE:
1816 case BLK_STS_DEV_RESOURCE:
1817 blk_mq_update_dispatch_busy(hctx, true);
1818 __blk_mq_requeue_request(rq);
1819 break;
1820 default:
1821 blk_mq_update_dispatch_busy(hctx, false);
1822 *cookie = BLK_QC_T_NONE;
1823 break;
1824 }
1825
1826 return ret;
1827}
1828
1829static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1830 struct request *rq,
1831 blk_qc_t *cookie,
1832 bool bypass_insert, bool last)
1833{
1834 struct request_queue *q = rq->q;
1835 bool run_queue = true;
1836
1837 /*
1838 * RCU or SRCU read lock is needed before checking quiesced flag.
1839 *
1840 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1841 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1842 * and avoid driver to try to dispatch again.
1843 */
1844 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1845 run_queue = false;
1846 bypass_insert = false;
1847 goto insert;
1848 }
1849
1850 if (q->elevator && !bypass_insert)
1851 goto insert;
1852
1853 if (!blk_mq_get_dispatch_budget(hctx))
1854 goto insert;
1855
1856 if (!blk_mq_get_driver_tag(rq)) {
1857 blk_mq_put_dispatch_budget(hctx);
1858 goto insert;
1859 }
1860
1861 return __blk_mq_issue_directly(hctx, rq, cookie, last);
1862insert:
1863 if (bypass_insert)
1864 return BLK_STS_RESOURCE;
1865
1866 blk_mq_request_bypass_insert(rq, run_queue);
1867 return BLK_STS_OK;
1868}
1869
1870static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1871 struct request *rq, blk_qc_t *cookie)
1872{
1873 blk_status_t ret;
1874 int srcu_idx;
1875
1876 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1877
1878 hctx_lock(hctx, &srcu_idx);
1879
1880 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
1881 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1882 blk_mq_request_bypass_insert(rq, true);
1883 else if (ret != BLK_STS_OK)
1884 blk_mq_end_request(rq, ret);
1885
1886 hctx_unlock(hctx, srcu_idx);
1887}
1888
1889blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
1890{
1891 blk_status_t ret;
1892 int srcu_idx;
1893 blk_qc_t unused_cookie;
1894 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1895
1896 hctx_lock(hctx, &srcu_idx);
1897 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
1898 hctx_unlock(hctx, srcu_idx);
1899
1900 return ret;
1901}
1902
1903void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
1904 struct list_head *list)
1905{
1906 while (!list_empty(list)) {
1907 blk_status_t ret;
1908 struct request *rq = list_first_entry(list, struct request,
1909 queuelist);
1910
1911 list_del_init(&rq->queuelist);
1912 ret = blk_mq_request_issue_directly(rq, list_empty(list));
1913 if (ret != BLK_STS_OK) {
1914 if (ret == BLK_STS_RESOURCE ||
1915 ret == BLK_STS_DEV_RESOURCE) {
1916 blk_mq_request_bypass_insert(rq,
1917 list_empty(list));
1918 break;
1919 }
1920 blk_mq_end_request(rq, ret);
1921 }
1922 }
1923
1924 /*
1925 * If we didn't flush the entire list, we could have told
1926 * the driver there was more coming, but that turned out to
1927 * be a lie.
1928 */
1929 if (!list_empty(list) && hctx->queue->mq_ops->commit_rqs)
1930 hctx->queue->mq_ops->commit_rqs(hctx);
1931}
1932
1933static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1934{
1935 list_add_tail(&rq->queuelist, &plug->mq_list);
1936 plug->rq_count++;
1937 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
1938 struct request *tmp;
1939
1940 tmp = list_first_entry(&plug->mq_list, struct request,
1941 queuelist);
1942 if (tmp->q != rq->q)
1943 plug->multiple_queues = true;
1944 }
1945}
1946
1947static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1948{
1949 const int is_sync = op_is_sync(bio->bi_opf);
1950 const int is_flush_fua = op_is_flush(bio->bi_opf);
1951 struct blk_mq_alloc_data data = { .flags = 0};
1952 struct request *rq;
1953 struct blk_plug *plug;
1954 struct request *same_queue_rq = NULL;
1955 unsigned int nr_segs;
1956 blk_qc_t cookie;
1957
1958 blk_queue_bounce(q, &bio);
1959 __blk_queue_split(q, &bio, &nr_segs);
1960
1961 if (!bio_integrity_prep(bio))
1962 return BLK_QC_T_NONE;
1963
1964 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1965 blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
1966 return BLK_QC_T_NONE;
1967
1968 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
1969 return BLK_QC_T_NONE;
1970
1971 rq_qos_throttle(q, bio);
1972
1973 data.cmd_flags = bio->bi_opf;
1974 rq = blk_mq_get_request(q, bio, &data);
1975 if (unlikely(!rq)) {
1976 rq_qos_cleanup(q, bio);
1977 if (bio->bi_opf & REQ_NOWAIT)
1978 bio_wouldblock_error(bio);
1979 return BLK_QC_T_NONE;
1980 }
1981
1982 trace_block_getrq(q, bio, bio->bi_opf);
1983
1984 rq_qos_track(q, rq, bio);
1985
1986 cookie = request_to_qc_t(data.hctx, rq);
1987
1988 blk_mq_bio_to_request(rq, bio, nr_segs);
1989
1990 plug = blk_mq_plug(q, bio);
1991 if (unlikely(is_flush_fua)) {
1992 /* bypass scheduler for flush rq */
1993 blk_insert_flush(rq);
1994 blk_mq_run_hw_queue(data.hctx, true);
1995 } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs ||
1996 !blk_queue_nonrot(q))) {
1997 /*
1998 * Use plugging if we have a ->commit_rqs() hook as well, as
1999 * we know the driver uses bd->last in a smart fashion.
2000 *
2001 * Use normal plugging if this disk is slow HDD, as sequential
2002 * IO may benefit a lot from plug merging.
2003 */
2004 unsigned int request_count = plug->rq_count;
2005 struct request *last = NULL;
2006
2007 if (!request_count)
2008 trace_block_plug(q);
2009 else
2010 last = list_entry_rq(plug->mq_list.prev);
2011
2012 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
2013 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2014 blk_flush_plug_list(plug, false);
2015 trace_block_plug(q);
2016 }
2017
2018 blk_add_rq_to_plug(plug, rq);
2019 } else if (q->elevator) {
2020 blk_mq_sched_insert_request(rq, false, true, true);
2021 } else if (plug && !blk_queue_nomerges(q)) {
2022 /*
2023 * We do limited plugging. If the bio can be merged, do that.
2024 * Otherwise the existing request in the plug list will be
2025 * issued. So the plug list will have one request at most
2026 * The plug list might get flushed before this. If that happens,
2027 * the plug list is empty, and same_queue_rq is invalid.
2028 */
2029 if (list_empty(&plug->mq_list))
2030 same_queue_rq = NULL;
2031 if (same_queue_rq) {
2032 list_del_init(&same_queue_rq->queuelist);
2033 plug->rq_count--;
2034 }
2035 blk_add_rq_to_plug(plug, rq);
2036 trace_block_plug(q);
2037
2038 if (same_queue_rq) {
2039 data.hctx = same_queue_rq->mq_hctx;
2040 trace_block_unplug(q, 1, true);
2041 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2042 &cookie);
2043 }
2044 } else if ((q->nr_hw_queues > 1 && is_sync) ||
2045 !data.hctx->dispatch_busy) {
2046 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2047 } else {
2048 blk_mq_sched_insert_request(rq, false, true, true);
2049 }
2050
2051 return cookie;
2052}
2053
2054void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2055 unsigned int hctx_idx)
2056{
2057 struct page *page;
2058
2059 if (tags->rqs && set->ops->exit_request) {
2060 int i;
2061
2062 for (i = 0; i < tags->nr_tags; i++) {
2063 struct request *rq = tags->static_rqs[i];
2064
2065 if (!rq)
2066 continue;
2067 set->ops->exit_request(set, rq, hctx_idx);
2068 tags->static_rqs[i] = NULL;
2069 }
2070 }
2071
2072 while (!list_empty(&tags->page_list)) {
2073 page = list_first_entry(&tags->page_list, struct page, lru);
2074 list_del_init(&page->lru);
2075 /*
2076 * Remove kmemleak object previously allocated in
2077 * blk_mq_alloc_rqs().
2078 */
2079 kmemleak_free(page_address(page));
2080 __free_pages(page, page->private);
2081 }
2082}
2083
2084void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2085{
2086 kfree(tags->rqs);
2087 tags->rqs = NULL;
2088 kfree(tags->static_rqs);
2089 tags->static_rqs = NULL;
2090
2091 blk_mq_free_tags(tags);
2092}
2093
2094struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2095 unsigned int hctx_idx,
2096 unsigned int nr_tags,
2097 unsigned int reserved_tags)
2098{
2099 struct blk_mq_tags *tags;
2100 int node;
2101
2102 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2103 if (node == NUMA_NO_NODE)
2104 node = set->numa_node;
2105
2106 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2107 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2108 if (!tags)
2109 return NULL;
2110
2111 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2112 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2113 node);
2114 if (!tags->rqs) {
2115 blk_mq_free_tags(tags);
2116 return NULL;
2117 }
2118
2119 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2120 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2121 node);
2122 if (!tags->static_rqs) {
2123 kfree(tags->rqs);
2124 blk_mq_free_tags(tags);
2125 return NULL;
2126 }
2127
2128 return tags;
2129}
2130
2131static size_t order_to_size(unsigned int order)
2132{
2133 return (size_t)PAGE_SIZE << order;
2134}
2135
2136static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2137 unsigned int hctx_idx, int node)
2138{
2139 int ret;
2140
2141 if (set->ops->init_request) {
2142 ret = set->ops->init_request(set, rq, hctx_idx, node);
2143 if (ret)
2144 return ret;
2145 }
2146
2147 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2148 return 0;
2149}
2150
2151int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2152 unsigned int hctx_idx, unsigned int depth)
2153{
2154 unsigned int i, j, entries_per_page, max_order = 4;
2155 size_t rq_size, left;
2156 int node;
2157
2158 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2159 if (node == NUMA_NO_NODE)
2160 node = set->numa_node;
2161
2162 INIT_LIST_HEAD(&tags->page_list);
2163
2164 /*
2165 * rq_size is the size of the request plus driver payload, rounded
2166 * to the cacheline size
2167 */
2168 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2169 cache_line_size());
2170 left = rq_size * depth;
2171
2172 for (i = 0; i < depth; ) {
2173 int this_order = max_order;
2174 struct page *page;
2175 int to_do;
2176 void *p;
2177
2178 while (this_order && left < order_to_size(this_order - 1))
2179 this_order--;
2180
2181 do {
2182 page = alloc_pages_node(node,
2183 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2184 this_order);
2185 if (page)
2186 break;
2187 if (!this_order--)
2188 break;
2189 if (order_to_size(this_order) < rq_size)
2190 break;
2191 } while (1);
2192
2193 if (!page)
2194 goto fail;
2195
2196 page->private = this_order;
2197 list_add_tail(&page->lru, &tags->page_list);
2198
2199 p = page_address(page);
2200 /*
2201 * Allow kmemleak to scan these pages as they contain pointers
2202 * to additional allocations like via ops->init_request().
2203 */
2204 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2205 entries_per_page = order_to_size(this_order) / rq_size;
2206 to_do = min(entries_per_page, depth - i);
2207 left -= to_do * rq_size;
2208 for (j = 0; j < to_do; j++) {
2209 struct request *rq = p;
2210
2211 tags->static_rqs[i] = rq;
2212 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2213 tags->static_rqs[i] = NULL;
2214 goto fail;
2215 }
2216
2217 p += rq_size;
2218 i++;
2219 }
2220 }
2221 return 0;
2222
2223fail:
2224 blk_mq_free_rqs(set, tags, hctx_idx);
2225 return -ENOMEM;
2226}
2227
2228/*
2229 * 'cpu' is going away. splice any existing rq_list entries from this
2230 * software queue to the hw queue dispatch list, and ensure that it
2231 * gets run.
2232 */
2233static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2234{
2235 struct blk_mq_hw_ctx *hctx;
2236 struct blk_mq_ctx *ctx;
2237 LIST_HEAD(tmp);
2238 enum hctx_type type;
2239
2240 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2241 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2242 type = hctx->type;
2243
2244 spin_lock(&ctx->lock);
2245 if (!list_empty(&ctx->rq_lists[type])) {
2246 list_splice_init(&ctx->rq_lists[type], &tmp);
2247 blk_mq_hctx_clear_pending(hctx, ctx);
2248 }
2249 spin_unlock(&ctx->lock);
2250
2251 if (list_empty(&tmp))
2252 return 0;
2253
2254 spin_lock(&hctx->lock);
2255 list_splice_tail_init(&tmp, &hctx->dispatch);
2256 spin_unlock(&hctx->lock);
2257
2258 blk_mq_run_hw_queue(hctx, true);
2259 return 0;
2260}
2261
2262static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2263{
2264 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2265 &hctx->cpuhp_dead);
2266}
2267
2268/* hctx->ctxs will be freed in queue's release handler */
2269static void blk_mq_exit_hctx(struct request_queue *q,
2270 struct blk_mq_tag_set *set,
2271 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2272{
2273 if (blk_mq_hw_queue_mapped(hctx))
2274 blk_mq_tag_idle(hctx);
2275
2276 if (set->ops->exit_request)
2277 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2278
2279 if (set->ops->exit_hctx)
2280 set->ops->exit_hctx(hctx, hctx_idx);
2281
2282 blk_mq_remove_cpuhp(hctx);
2283
2284 spin_lock(&q->unused_hctx_lock);
2285 list_add(&hctx->hctx_list, &q->unused_hctx_list);
2286 spin_unlock(&q->unused_hctx_lock);
2287}
2288
2289static void blk_mq_exit_hw_queues(struct request_queue *q,
2290 struct blk_mq_tag_set *set, int nr_queue)
2291{
2292 struct blk_mq_hw_ctx *hctx;
2293 unsigned int i;
2294
2295 queue_for_each_hw_ctx(q, hctx, i) {
2296 if (i == nr_queue)
2297 break;
2298 blk_mq_debugfs_unregister_hctx(hctx);
2299 blk_mq_exit_hctx(q, set, hctx, i);
2300 }
2301}
2302
2303static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2304{
2305 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2306
2307 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2308 __alignof__(struct blk_mq_hw_ctx)) !=
2309 sizeof(struct blk_mq_hw_ctx));
2310
2311 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2312 hw_ctx_size += sizeof(struct srcu_struct);
2313
2314 return hw_ctx_size;
2315}
2316
2317static int blk_mq_init_hctx(struct request_queue *q,
2318 struct blk_mq_tag_set *set,
2319 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2320{
2321 hctx->queue_num = hctx_idx;
2322
2323 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2324
2325 hctx->tags = set->tags[hctx_idx];
2326
2327 if (set->ops->init_hctx &&
2328 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2329 goto unregister_cpu_notifier;
2330
2331 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2332 hctx->numa_node))
2333 goto exit_hctx;
2334 return 0;
2335
2336 exit_hctx:
2337 if (set->ops->exit_hctx)
2338 set->ops->exit_hctx(hctx, hctx_idx);
2339 unregister_cpu_notifier:
2340 blk_mq_remove_cpuhp(hctx);
2341 return -1;
2342}
2343
2344static struct blk_mq_hw_ctx *
2345blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2346 int node)
2347{
2348 struct blk_mq_hw_ctx *hctx;
2349 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2350
2351 hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2352 if (!hctx)
2353 goto fail_alloc_hctx;
2354
2355 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2356 goto free_hctx;
2357
2358 atomic_set(&hctx->nr_active, 0);
2359 if (node == NUMA_NO_NODE)
2360 node = set->numa_node;
2361 hctx->numa_node = node;
2362
2363 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2364 spin_lock_init(&hctx->lock);
2365 INIT_LIST_HEAD(&hctx->dispatch);
2366 hctx->queue = q;
2367 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2368
2369 INIT_LIST_HEAD(&hctx->hctx_list);
2370
2371 /*
2372 * Allocate space for all possible cpus to avoid allocation at
2373 * runtime
2374 */
2375 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2376 gfp, node);
2377 if (!hctx->ctxs)
2378 goto free_cpumask;
2379
2380 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2381 gfp, node))
2382 goto free_ctxs;
2383 hctx->nr_ctx = 0;
2384
2385 spin_lock_init(&hctx->dispatch_wait_lock);
2386 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2387 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2388
2389 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size,
2390 gfp);
2391 if (!hctx->fq)
2392 goto free_bitmap;
2393
2394 if (hctx->flags & BLK_MQ_F_BLOCKING)
2395 init_srcu_struct(hctx->srcu);
2396 blk_mq_hctx_kobj_init(hctx);
2397
2398 return hctx;
2399
2400 free_bitmap:
2401 sbitmap_free(&hctx->ctx_map);
2402 free_ctxs:
2403 kfree(hctx->ctxs);
2404 free_cpumask:
2405 free_cpumask_var(hctx->cpumask);
2406 free_hctx:
2407 kfree(hctx);
2408 fail_alloc_hctx:
2409 return NULL;
2410}
2411
2412static void blk_mq_init_cpu_queues(struct request_queue *q,
2413 unsigned int nr_hw_queues)
2414{
2415 struct blk_mq_tag_set *set = q->tag_set;
2416 unsigned int i, j;
2417
2418 for_each_possible_cpu(i) {
2419 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2420 struct blk_mq_hw_ctx *hctx;
2421 int k;
2422
2423 __ctx->cpu = i;
2424 spin_lock_init(&__ctx->lock);
2425 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2426 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2427
2428 __ctx->queue = q;
2429
2430 /*
2431 * Set local node, IFF we have more than one hw queue. If
2432 * not, we remain on the home node of the device
2433 */
2434 for (j = 0; j < set->nr_maps; j++) {
2435 hctx = blk_mq_map_queue_type(q, j, i);
2436 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2437 hctx->numa_node = local_memory_node(cpu_to_node(i));
2438 }
2439 }
2440}
2441
2442static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2443{
2444 int ret = 0;
2445
2446 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2447 set->queue_depth, set->reserved_tags);
2448 if (!set->tags[hctx_idx])
2449 return false;
2450
2451 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2452 set->queue_depth);
2453 if (!ret)
2454 return true;
2455
2456 blk_mq_free_rq_map(set->tags[hctx_idx]);
2457 set->tags[hctx_idx] = NULL;
2458 return false;
2459}
2460
2461static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2462 unsigned int hctx_idx)
2463{
2464 if (set->tags && set->tags[hctx_idx]) {
2465 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2466 blk_mq_free_rq_map(set->tags[hctx_idx]);
2467 set->tags[hctx_idx] = NULL;
2468 }
2469}
2470
2471static void blk_mq_map_swqueue(struct request_queue *q)
2472{
2473 unsigned int i, j, hctx_idx;
2474 struct blk_mq_hw_ctx *hctx;
2475 struct blk_mq_ctx *ctx;
2476 struct blk_mq_tag_set *set = q->tag_set;
2477
2478 queue_for_each_hw_ctx(q, hctx, i) {
2479 cpumask_clear(hctx->cpumask);
2480 hctx->nr_ctx = 0;
2481 hctx->dispatch_from = NULL;
2482 }
2483
2484 /*
2485 * Map software to hardware queues.
2486 *
2487 * If the cpu isn't present, the cpu is mapped to first hctx.
2488 */
2489 for_each_possible_cpu(i) {
2490 hctx_idx = set->map[HCTX_TYPE_DEFAULT].mq_map[i];
2491 /* unmapped hw queue can be remapped after CPU topo changed */
2492 if (!set->tags[hctx_idx] &&
2493 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2494 /*
2495 * If tags initialization fail for some hctx,
2496 * that hctx won't be brought online. In this
2497 * case, remap the current ctx to hctx[0] which
2498 * is guaranteed to always have tags allocated
2499 */
2500 set->map[HCTX_TYPE_DEFAULT].mq_map[i] = 0;
2501 }
2502
2503 ctx = per_cpu_ptr(q->queue_ctx, i);
2504 for (j = 0; j < set->nr_maps; j++) {
2505 if (!set->map[j].nr_queues) {
2506 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2507 HCTX_TYPE_DEFAULT, i);
2508 continue;
2509 }
2510
2511 hctx = blk_mq_map_queue_type(q, j, i);
2512 ctx->hctxs[j] = hctx;
2513 /*
2514 * If the CPU is already set in the mask, then we've
2515 * mapped this one already. This can happen if
2516 * devices share queues across queue maps.
2517 */
2518 if (cpumask_test_cpu(i, hctx->cpumask))
2519 continue;
2520
2521 cpumask_set_cpu(i, hctx->cpumask);
2522 hctx->type = j;
2523 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2524 hctx->ctxs[hctx->nr_ctx++] = ctx;
2525
2526 /*
2527 * If the nr_ctx type overflows, we have exceeded the
2528 * amount of sw queues we can support.
2529 */
2530 BUG_ON(!hctx->nr_ctx);
2531 }
2532
2533 for (; j < HCTX_MAX_TYPES; j++)
2534 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2535 HCTX_TYPE_DEFAULT, i);
2536 }
2537
2538 queue_for_each_hw_ctx(q, hctx, i) {
2539 /*
2540 * If no software queues are mapped to this hardware queue,
2541 * disable it and free the request entries.
2542 */
2543 if (!hctx->nr_ctx) {
2544 /* Never unmap queue 0. We need it as a
2545 * fallback in case of a new remap fails
2546 * allocation
2547 */
2548 if (i && set->tags[i])
2549 blk_mq_free_map_and_requests(set, i);
2550
2551 hctx->tags = NULL;
2552 continue;
2553 }
2554
2555 hctx->tags = set->tags[i];
2556 WARN_ON(!hctx->tags);
2557
2558 /*
2559 * Set the map size to the number of mapped software queues.
2560 * This is more accurate and more efficient than looping
2561 * over all possibly mapped software queues.
2562 */
2563 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2564
2565 /*
2566 * Initialize batch roundrobin counts
2567 */
2568 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2569 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2570 }
2571}
2572
2573/*
2574 * Caller needs to ensure that we're either frozen/quiesced, or that
2575 * the queue isn't live yet.
2576 */
2577static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2578{
2579 struct blk_mq_hw_ctx *hctx;
2580 int i;
2581
2582 queue_for_each_hw_ctx(q, hctx, i) {
2583 if (shared)
2584 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2585 else
2586 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2587 }
2588}
2589
2590static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2591 bool shared)
2592{
2593 struct request_queue *q;
2594
2595 lockdep_assert_held(&set->tag_list_lock);
2596
2597 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2598 blk_mq_freeze_queue(q);
2599 queue_set_hctx_shared(q, shared);
2600 blk_mq_unfreeze_queue(q);
2601 }
2602}
2603
2604static void blk_mq_del_queue_tag_set(struct request_queue *q)
2605{
2606 struct blk_mq_tag_set *set = q->tag_set;
2607
2608 mutex_lock(&set->tag_list_lock);
2609 list_del_rcu(&q->tag_set_list);
2610 if (list_is_singular(&set->tag_list)) {
2611 /* just transitioned to unshared */
2612 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2613 /* update existing queue */
2614 blk_mq_update_tag_set_depth(set, false);
2615 }
2616 mutex_unlock(&set->tag_list_lock);
2617 INIT_LIST_HEAD(&q->tag_set_list);
2618}
2619
2620static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2621 struct request_queue *q)
2622{
2623 mutex_lock(&set->tag_list_lock);
2624
2625 /*
2626 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2627 */
2628 if (!list_empty(&set->tag_list) &&
2629 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2630 set->flags |= BLK_MQ_F_TAG_SHARED;
2631 /* update existing queue */
2632 blk_mq_update_tag_set_depth(set, true);
2633 }
2634 if (set->flags & BLK_MQ_F_TAG_SHARED)
2635 queue_set_hctx_shared(q, true);
2636 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2637
2638 mutex_unlock(&set->tag_list_lock);
2639}
2640
2641/* All allocations will be freed in release handler of q->mq_kobj */
2642static int blk_mq_alloc_ctxs(struct request_queue *q)
2643{
2644 struct blk_mq_ctxs *ctxs;
2645 int cpu;
2646
2647 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2648 if (!ctxs)
2649 return -ENOMEM;
2650
2651 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2652 if (!ctxs->queue_ctx)
2653 goto fail;
2654
2655 for_each_possible_cpu(cpu) {
2656 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2657 ctx->ctxs = ctxs;
2658 }
2659
2660 q->mq_kobj = &ctxs->kobj;
2661 q->queue_ctx = ctxs->queue_ctx;
2662
2663 return 0;
2664 fail:
2665 kfree(ctxs);
2666 return -ENOMEM;
2667}
2668
2669/*
2670 * It is the actual release handler for mq, but we do it from
2671 * request queue's release handler for avoiding use-after-free
2672 * and headache because q->mq_kobj shouldn't have been introduced,
2673 * but we can't group ctx/kctx kobj without it.
2674 */
2675void blk_mq_release(struct request_queue *q)
2676{
2677 struct blk_mq_hw_ctx *hctx, *next;
2678 int i;
2679
2680 queue_for_each_hw_ctx(q, hctx, i)
2681 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
2682
2683 /* all hctx are in .unused_hctx_list now */
2684 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
2685 list_del_init(&hctx->hctx_list);
2686 kobject_put(&hctx->kobj);
2687 }
2688
2689 kfree(q->queue_hw_ctx);
2690
2691 /*
2692 * release .mq_kobj and sw queue's kobject now because
2693 * both share lifetime with request queue.
2694 */
2695 blk_mq_sysfs_deinit(q);
2696}
2697
2698struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2699{
2700 struct request_queue *uninit_q, *q;
2701
2702 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2703 if (!uninit_q)
2704 return ERR_PTR(-ENOMEM);
2705
2706 /*
2707 * Initialize the queue without an elevator. device_add_disk() will do
2708 * the initialization.
2709 */
2710 q = blk_mq_init_allocated_queue(set, uninit_q, false);
2711 if (IS_ERR(q))
2712 blk_cleanup_queue(uninit_q);
2713
2714 return q;
2715}
2716EXPORT_SYMBOL(blk_mq_init_queue);
2717
2718/*
2719 * Helper for setting up a queue with mq ops, given queue depth, and
2720 * the passed in mq ops flags.
2721 */
2722struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
2723 const struct blk_mq_ops *ops,
2724 unsigned int queue_depth,
2725 unsigned int set_flags)
2726{
2727 struct request_queue *q;
2728 int ret;
2729
2730 memset(set, 0, sizeof(*set));
2731 set->ops = ops;
2732 set->nr_hw_queues = 1;
2733 set->nr_maps = 1;
2734 set->queue_depth = queue_depth;
2735 set->numa_node = NUMA_NO_NODE;
2736 set->flags = set_flags;
2737
2738 ret = blk_mq_alloc_tag_set(set);
2739 if (ret)
2740 return ERR_PTR(ret);
2741
2742 q = blk_mq_init_queue(set);
2743 if (IS_ERR(q)) {
2744 blk_mq_free_tag_set(set);
2745 return q;
2746 }
2747
2748 return q;
2749}
2750EXPORT_SYMBOL(blk_mq_init_sq_queue);
2751
2752static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
2753 struct blk_mq_tag_set *set, struct request_queue *q,
2754 int hctx_idx, int node)
2755{
2756 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
2757
2758 /* reuse dead hctx first */
2759 spin_lock(&q->unused_hctx_lock);
2760 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
2761 if (tmp->numa_node == node) {
2762 hctx = tmp;
2763 break;
2764 }
2765 }
2766 if (hctx)
2767 list_del_init(&hctx->hctx_list);
2768 spin_unlock(&q->unused_hctx_lock);
2769
2770 if (!hctx)
2771 hctx = blk_mq_alloc_hctx(q, set, node);
2772 if (!hctx)
2773 goto fail;
2774
2775 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
2776 goto free_hctx;
2777
2778 return hctx;
2779
2780 free_hctx:
2781 kobject_put(&hctx->kobj);
2782 fail:
2783 return NULL;
2784}
2785
2786static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2787 struct request_queue *q)
2788{
2789 int i, j, end;
2790 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2791
2792 /* protect against switching io scheduler */
2793 mutex_lock(&q->sysfs_lock);
2794 for (i = 0; i < set->nr_hw_queues; i++) {
2795 int node;
2796 struct blk_mq_hw_ctx *hctx;
2797
2798 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
2799 /*
2800 * If the hw queue has been mapped to another numa node,
2801 * we need to realloc the hctx. If allocation fails, fallback
2802 * to use the previous one.
2803 */
2804 if (hctxs[i] && (hctxs[i]->numa_node == node))
2805 continue;
2806
2807 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
2808 if (hctx) {
2809 if (hctxs[i])
2810 blk_mq_exit_hctx(q, set, hctxs[i], i);
2811 hctxs[i] = hctx;
2812 } else {
2813 if (hctxs[i])
2814 pr_warn("Allocate new hctx on node %d fails,\
2815 fallback to previous one on node %d\n",
2816 node, hctxs[i]->numa_node);
2817 else
2818 break;
2819 }
2820 }
2821 /*
2822 * Increasing nr_hw_queues fails. Free the newly allocated
2823 * hctxs and keep the previous q->nr_hw_queues.
2824 */
2825 if (i != set->nr_hw_queues) {
2826 j = q->nr_hw_queues;
2827 end = i;
2828 } else {
2829 j = i;
2830 end = q->nr_hw_queues;
2831 q->nr_hw_queues = set->nr_hw_queues;
2832 }
2833
2834 for (; j < end; j++) {
2835 struct blk_mq_hw_ctx *hctx = hctxs[j];
2836
2837 if (hctx) {
2838 if (hctx->tags)
2839 blk_mq_free_map_and_requests(set, j);
2840 blk_mq_exit_hctx(q, set, hctx, j);
2841 hctxs[j] = NULL;
2842 }
2843 }
2844 mutex_unlock(&q->sysfs_lock);
2845}
2846
2847/*
2848 * Maximum number of hardware queues we support. For single sets, we'll never
2849 * have more than the CPUs (software queues). For multiple sets, the tag_set
2850 * user may have set ->nr_hw_queues larger.
2851 */
2852static unsigned int nr_hw_queues(struct blk_mq_tag_set *set)
2853{
2854 if (set->nr_maps == 1)
2855 return nr_cpu_ids;
2856
2857 return max(set->nr_hw_queues, nr_cpu_ids);
2858}
2859
2860struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2861 struct request_queue *q,
2862 bool elevator_init)
2863{
2864 /* mark the queue as mq asap */
2865 q->mq_ops = set->ops;
2866
2867 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2868 blk_mq_poll_stats_bkt,
2869 BLK_MQ_POLL_STATS_BKTS, q);
2870 if (!q->poll_cb)
2871 goto err_exit;
2872
2873 if (blk_mq_alloc_ctxs(q))
2874 goto err_poll;
2875
2876 /* init q->mq_kobj and sw queues' kobjects */
2877 blk_mq_sysfs_init(q);
2878
2879 q->nr_queues = nr_hw_queues(set);
2880 q->queue_hw_ctx = kcalloc_node(q->nr_queues, sizeof(*(q->queue_hw_ctx)),
2881 GFP_KERNEL, set->numa_node);
2882 if (!q->queue_hw_ctx)
2883 goto err_sys_init;
2884
2885 INIT_LIST_HEAD(&q->unused_hctx_list);
2886 spin_lock_init(&q->unused_hctx_lock);
2887
2888 blk_mq_realloc_hw_ctxs(set, q);
2889 if (!q->nr_hw_queues)
2890 goto err_hctxs;
2891
2892 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2893 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2894
2895 q->tag_set = set;
2896
2897 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2898 if (set->nr_maps > HCTX_TYPE_POLL &&
2899 set->map[HCTX_TYPE_POLL].nr_queues)
2900 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
2901
2902 q->sg_reserved_size = INT_MAX;
2903
2904 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2905 INIT_LIST_HEAD(&q->requeue_list);
2906 spin_lock_init(&q->requeue_lock);
2907
2908 blk_queue_make_request(q, blk_mq_make_request);
2909
2910 /*
2911 * Do this after blk_queue_make_request() overrides it...
2912 */
2913 q->nr_requests = set->queue_depth;
2914
2915 /*
2916 * Default to classic polling
2917 */
2918 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
2919
2920 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2921 blk_mq_add_queue_tag_set(set, q);
2922 blk_mq_map_swqueue(q);
2923
2924 if (elevator_init)
2925 elevator_init_mq(q);
2926
2927 return q;
2928
2929err_hctxs:
2930 kfree(q->queue_hw_ctx);
2931 q->nr_hw_queues = 0;
2932err_sys_init:
2933 blk_mq_sysfs_deinit(q);
2934err_poll:
2935 blk_stat_free_callback(q->poll_cb);
2936 q->poll_cb = NULL;
2937err_exit:
2938 q->mq_ops = NULL;
2939 return ERR_PTR(-ENOMEM);
2940}
2941EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2942
2943/* tags can _not_ be used after returning from blk_mq_exit_queue */
2944void blk_mq_exit_queue(struct request_queue *q)
2945{
2946 struct blk_mq_tag_set *set = q->tag_set;
2947
2948 blk_mq_del_queue_tag_set(q);
2949 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2950}
2951
2952static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2953{
2954 int i;
2955
2956 for (i = 0; i < set->nr_hw_queues; i++)
2957 if (!__blk_mq_alloc_rq_map(set, i))
2958 goto out_unwind;
2959
2960 return 0;
2961
2962out_unwind:
2963 while (--i >= 0)
2964 blk_mq_free_rq_map(set->tags[i]);
2965
2966 return -ENOMEM;
2967}
2968
2969/*
2970 * Allocate the request maps associated with this tag_set. Note that this
2971 * may reduce the depth asked for, if memory is tight. set->queue_depth
2972 * will be updated to reflect the allocated depth.
2973 */
2974static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2975{
2976 unsigned int depth;
2977 int err;
2978
2979 depth = set->queue_depth;
2980 do {
2981 err = __blk_mq_alloc_rq_maps(set);
2982 if (!err)
2983 break;
2984
2985 set->queue_depth >>= 1;
2986 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2987 err = -ENOMEM;
2988 break;
2989 }
2990 } while (set->queue_depth);
2991
2992 if (!set->queue_depth || err) {
2993 pr_err("blk-mq: failed to allocate request map\n");
2994 return -ENOMEM;
2995 }
2996
2997 if (depth != set->queue_depth)
2998 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2999 depth, set->queue_depth);
3000
3001 return 0;
3002}
3003
3004static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3005{
3006 if (set->ops->map_queues && !is_kdump_kernel()) {
3007 int i;
3008
3009 /*
3010 * transport .map_queues is usually done in the following
3011 * way:
3012 *
3013 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3014 * mask = get_cpu_mask(queue)
3015 * for_each_cpu(cpu, mask)
3016 * set->map[x].mq_map[cpu] = queue;
3017 * }
3018 *
3019 * When we need to remap, the table has to be cleared for
3020 * killing stale mapping since one CPU may not be mapped
3021 * to any hw queue.
3022 */
3023 for (i = 0; i < set->nr_maps; i++)
3024 blk_mq_clear_mq_map(&set->map[i]);
3025
3026 return set->ops->map_queues(set);
3027 } else {
3028 BUG_ON(set->nr_maps > 1);
3029 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3030 }
3031}
3032
3033/*
3034 * Alloc a tag set to be associated with one or more request queues.
3035 * May fail with EINVAL for various error conditions. May adjust the
3036 * requested depth down, if it's too large. In that case, the set
3037 * value will be stored in set->queue_depth.
3038 */
3039int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3040{
3041 int i, ret;
3042
3043 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3044
3045 if (!set->nr_hw_queues)
3046 return -EINVAL;
3047 if (!set->queue_depth)
3048 return -EINVAL;
3049 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3050 return -EINVAL;
3051
3052 if (!set->ops->queue_rq)
3053 return -EINVAL;
3054
3055 if (!set->ops->get_budget ^ !set->ops->put_budget)
3056 return -EINVAL;
3057
3058 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3059 pr_info("blk-mq: reduced tag depth to %u\n",
3060 BLK_MQ_MAX_DEPTH);
3061 set->queue_depth = BLK_MQ_MAX_DEPTH;
3062 }
3063
3064 if (!set->nr_maps)
3065 set->nr_maps = 1;
3066 else if (set->nr_maps > HCTX_MAX_TYPES)
3067 return -EINVAL;
3068
3069 /*
3070 * If a crashdump is active, then we are potentially in a very
3071 * memory constrained environment. Limit us to 1 queue and
3072 * 64 tags to prevent using too much memory.
3073 */
3074 if (is_kdump_kernel()) {
3075 set->nr_hw_queues = 1;
3076 set->nr_maps = 1;
3077 set->queue_depth = min(64U, set->queue_depth);
3078 }
3079 /*
3080 * There is no use for more h/w queues than cpus if we just have
3081 * a single map
3082 */
3083 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3084 set->nr_hw_queues = nr_cpu_ids;
3085
3086 set->tags = kcalloc_node(nr_hw_queues(set), sizeof(struct blk_mq_tags *),
3087 GFP_KERNEL, set->numa_node);
3088 if (!set->tags)
3089 return -ENOMEM;
3090
3091 ret = -ENOMEM;
3092 for (i = 0; i < set->nr_maps; i++) {
3093 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3094 sizeof(set->map[i].mq_map[0]),
3095 GFP_KERNEL, set->numa_node);
3096 if (!set->map[i].mq_map)
3097 goto out_free_mq_map;
3098 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3099 }
3100
3101 ret = blk_mq_update_queue_map(set);
3102 if (ret)
3103 goto out_free_mq_map;
3104
3105 ret = blk_mq_alloc_rq_maps(set);
3106 if (ret)
3107 goto out_free_mq_map;
3108
3109 mutex_init(&set->tag_list_lock);
3110 INIT_LIST_HEAD(&set->tag_list);
3111
3112 return 0;
3113
3114out_free_mq_map:
3115 for (i = 0; i < set->nr_maps; i++) {
3116 kfree(set->map[i].mq_map);
3117 set->map[i].mq_map = NULL;
3118 }
3119 kfree(set->tags);
3120 set->tags = NULL;
3121 return ret;
3122}
3123EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3124
3125void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3126{
3127 int i, j;
3128
3129 for (i = 0; i < nr_hw_queues(set); i++)
3130 blk_mq_free_map_and_requests(set, i);
3131
3132 for (j = 0; j < set->nr_maps; j++) {
3133 kfree(set->map[j].mq_map);
3134 set->map[j].mq_map = NULL;
3135 }
3136
3137 kfree(set->tags);
3138 set->tags = NULL;
3139}
3140EXPORT_SYMBOL(blk_mq_free_tag_set);
3141
3142int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3143{
3144 struct blk_mq_tag_set *set = q->tag_set;
3145 struct blk_mq_hw_ctx *hctx;
3146 int i, ret;
3147
3148 if (!set)
3149 return -EINVAL;
3150
3151 if (q->nr_requests == nr)
3152 return 0;
3153
3154 blk_mq_freeze_queue(q);
3155 blk_mq_quiesce_queue(q);
3156
3157 ret = 0;
3158 queue_for_each_hw_ctx(q, hctx, i) {
3159 if (!hctx->tags)
3160 continue;
3161 /*
3162 * If we're using an MQ scheduler, just update the scheduler
3163 * queue depth. This is similar to what the old code would do.
3164 */
3165 if (!hctx->sched_tags) {
3166 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3167 false);
3168 } else {
3169 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3170 nr, true);
3171 }
3172 if (ret)
3173 break;
3174 if (q->elevator && q->elevator->type->ops.depth_updated)
3175 q->elevator->type->ops.depth_updated(hctx);
3176 }
3177
3178 if (!ret)
3179 q->nr_requests = nr;
3180
3181 blk_mq_unquiesce_queue(q);
3182 blk_mq_unfreeze_queue(q);
3183
3184 return ret;
3185}
3186
3187/*
3188 * request_queue and elevator_type pair.
3189 * It is just used by __blk_mq_update_nr_hw_queues to cache
3190 * the elevator_type associated with a request_queue.
3191 */
3192struct blk_mq_qe_pair {
3193 struct list_head node;
3194 struct request_queue *q;
3195 struct elevator_type *type;
3196};
3197
3198/*
3199 * Cache the elevator_type in qe pair list and switch the
3200 * io scheduler to 'none'
3201 */
3202static bool blk_mq_elv_switch_none(struct list_head *head,
3203 struct request_queue *q)
3204{
3205 struct blk_mq_qe_pair *qe;
3206
3207 if (!q->elevator)
3208 return true;
3209
3210 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3211 if (!qe)
3212 return false;
3213
3214 INIT_LIST_HEAD(&qe->node);
3215 qe->q = q;
3216 qe->type = q->elevator->type;
3217 list_add(&qe->node, head);
3218
3219 mutex_lock(&q->sysfs_lock);
3220 /*
3221 * After elevator_switch_mq, the previous elevator_queue will be
3222 * released by elevator_release. The reference of the io scheduler
3223 * module get by elevator_get will also be put. So we need to get
3224 * a reference of the io scheduler module here to prevent it to be
3225 * removed.
3226 */
3227 __module_get(qe->type->elevator_owner);
3228 elevator_switch_mq(q, NULL);
3229 mutex_unlock(&q->sysfs_lock);
3230
3231 return true;
3232}
3233
3234static void blk_mq_elv_switch_back(struct list_head *head,
3235 struct request_queue *q)
3236{
3237 struct blk_mq_qe_pair *qe;
3238 struct elevator_type *t = NULL;
3239
3240 list_for_each_entry(qe, head, node)
3241 if (qe->q == q) {
3242 t = qe->type;
3243 break;
3244 }
3245
3246 if (!t)
3247 return;
3248
3249 list_del(&qe->node);
3250 kfree(qe);
3251
3252 mutex_lock(&q->sysfs_lock);
3253 elevator_switch_mq(q, t);
3254 mutex_unlock(&q->sysfs_lock);
3255}
3256
3257static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3258 int nr_hw_queues)
3259{
3260 struct request_queue *q;
3261 LIST_HEAD(head);
3262 int prev_nr_hw_queues;
3263
3264 lockdep_assert_held(&set->tag_list_lock);
3265
3266 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3267 nr_hw_queues = nr_cpu_ids;
3268 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
3269 return;
3270
3271 list_for_each_entry(q, &set->tag_list, tag_set_list)
3272 blk_mq_freeze_queue(q);
3273 /*
3274 * Sync with blk_mq_queue_tag_busy_iter.
3275 */
3276 synchronize_rcu();
3277 /*
3278 * Switch IO scheduler to 'none', cleaning up the data associated
3279 * with the previous scheduler. We will switch back once we are done
3280 * updating the new sw to hw queue mappings.
3281 */
3282 list_for_each_entry(q, &set->tag_list, tag_set_list)
3283 if (!blk_mq_elv_switch_none(&head, q))
3284 goto switch_back;
3285
3286 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3287 blk_mq_debugfs_unregister_hctxs(q);
3288 blk_mq_sysfs_unregister(q);
3289 }
3290
3291 prev_nr_hw_queues = set->nr_hw_queues;
3292 set->nr_hw_queues = nr_hw_queues;
3293 blk_mq_update_queue_map(set);
3294fallback:
3295 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3296 blk_mq_realloc_hw_ctxs(set, q);
3297 if (q->nr_hw_queues != set->nr_hw_queues) {
3298 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3299 nr_hw_queues, prev_nr_hw_queues);
3300 set->nr_hw_queues = prev_nr_hw_queues;
3301 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3302 goto fallback;
3303 }
3304 blk_mq_map_swqueue(q);
3305 }
3306
3307 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3308 blk_mq_sysfs_register(q);
3309 blk_mq_debugfs_register_hctxs(q);
3310 }
3311
3312switch_back:
3313 list_for_each_entry(q, &set->tag_list, tag_set_list)
3314 blk_mq_elv_switch_back(&head, q);
3315
3316 list_for_each_entry(q, &set->tag_list, tag_set_list)
3317 blk_mq_unfreeze_queue(q);
3318}
3319
3320void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3321{
3322 mutex_lock(&set->tag_list_lock);
3323 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3324 mutex_unlock(&set->tag_list_lock);
3325}
3326EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3327
3328/* Enable polling stats and return whether they were already enabled. */
3329static bool blk_poll_stats_enable(struct request_queue *q)
3330{
3331 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3332 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3333 return true;
3334 blk_stat_add_callback(q, q->poll_cb);
3335 return false;
3336}
3337
3338static void blk_mq_poll_stats_start(struct request_queue *q)
3339{
3340 /*
3341 * We don't arm the callback if polling stats are not enabled or the
3342 * callback is already active.
3343 */
3344 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3345 blk_stat_is_active(q->poll_cb))
3346 return;
3347
3348 blk_stat_activate_msecs(q->poll_cb, 100);
3349}
3350
3351static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3352{
3353 struct request_queue *q = cb->data;
3354 int bucket;
3355
3356 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3357 if (cb->stat[bucket].nr_samples)
3358 q->poll_stat[bucket] = cb->stat[bucket];
3359 }
3360}
3361
3362static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3363 struct blk_mq_hw_ctx *hctx,
3364 struct request *rq)
3365{
3366 unsigned long ret = 0;
3367 int bucket;
3368
3369 /*
3370 * If stats collection isn't on, don't sleep but turn it on for
3371 * future users
3372 */
3373 if (!blk_poll_stats_enable(q))
3374 return 0;
3375
3376 /*
3377 * As an optimistic guess, use half of the mean service time
3378 * for this type of request. We can (and should) make this smarter.
3379 * For instance, if the completion latencies are tight, we can
3380 * get closer than just half the mean. This is especially
3381 * important on devices where the completion latencies are longer
3382 * than ~10 usec. We do use the stats for the relevant IO size
3383 * if available which does lead to better estimates.
3384 */
3385 bucket = blk_mq_poll_stats_bkt(rq);
3386 if (bucket < 0)
3387 return ret;
3388
3389 if (q->poll_stat[bucket].nr_samples)
3390 ret = (q->poll_stat[bucket].mean + 1) / 2;
3391
3392 return ret;
3393}
3394
3395static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3396 struct blk_mq_hw_ctx *hctx,
3397 struct request *rq)
3398{
3399 struct hrtimer_sleeper hs;
3400 enum hrtimer_mode mode;
3401 unsigned int nsecs;
3402 ktime_t kt;
3403
3404 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3405 return false;
3406
3407 /*
3408 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3409 *
3410 * 0: use half of prev avg
3411 * >0: use this specific value
3412 */
3413 if (q->poll_nsec > 0)
3414 nsecs = q->poll_nsec;
3415 else
3416 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
3417
3418 if (!nsecs)
3419 return false;
3420
3421 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3422
3423 /*
3424 * This will be replaced with the stats tracking code, using
3425 * 'avg_completion_time / 2' as the pre-sleep target.
3426 */
3427 kt = nsecs;
3428
3429 mode = HRTIMER_MODE_REL;
3430 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3431 hrtimer_set_expires(&hs.timer, kt);
3432
3433 do {
3434 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3435 break;
3436 set_current_state(TASK_UNINTERRUPTIBLE);
3437 hrtimer_sleeper_start_expires(&hs, mode);
3438 if (hs.task)
3439 io_schedule();
3440 hrtimer_cancel(&hs.timer);
3441 mode = HRTIMER_MODE_ABS;
3442 } while (hs.task && !signal_pending(current));
3443
3444 __set_current_state(TASK_RUNNING);
3445 destroy_hrtimer_on_stack(&hs.timer);
3446 return true;
3447}
3448
3449static bool blk_mq_poll_hybrid(struct request_queue *q,
3450 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3451{
3452 struct request *rq;
3453
3454 if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3455 return false;
3456
3457 if (!blk_qc_t_is_internal(cookie))
3458 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3459 else {
3460 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3461 /*
3462 * With scheduling, if the request has completed, we'll
3463 * get a NULL return here, as we clear the sched tag when
3464 * that happens. The request still remains valid, like always,
3465 * so we should be safe with just the NULL check.
3466 */
3467 if (!rq)
3468 return false;
3469 }
3470
3471 return blk_mq_poll_hybrid_sleep(q, hctx, rq);
3472}
3473
3474/**
3475 * blk_poll - poll for IO completions
3476 * @q: the queue
3477 * @cookie: cookie passed back at IO submission time
3478 * @spin: whether to spin for completions
3479 *
3480 * Description:
3481 * Poll for completions on the passed in queue. Returns number of
3482 * completed entries found. If @spin is true, then blk_poll will continue
3483 * looping until at least one completion is found, unless the task is
3484 * otherwise marked running (or we need to reschedule).
3485 */
3486int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3487{
3488 struct blk_mq_hw_ctx *hctx;
3489 long state;
3490
3491 if (!blk_qc_t_valid(cookie) ||
3492 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3493 return 0;
3494
3495 if (current->plug)
3496 blk_flush_plug_list(current->plug, false);
3497
3498 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3499
3500 /*
3501 * If we sleep, have the caller restart the poll loop to reset
3502 * the state. Like for the other success return cases, the
3503 * caller is responsible for checking if the IO completed. If
3504 * the IO isn't complete, we'll get called again and will go
3505 * straight to the busy poll loop.
3506 */
3507 if (blk_mq_poll_hybrid(q, hctx, cookie))
3508 return 1;
3509
3510 hctx->poll_considered++;
3511
3512 state = current->state;
3513 do {
3514 int ret;
3515
3516 hctx->poll_invoked++;
3517
3518 ret = q->mq_ops->poll(hctx);
3519 if (ret > 0) {
3520 hctx->poll_success++;
3521 __set_current_state(TASK_RUNNING);
3522 return ret;
3523 }
3524
3525 if (signal_pending_state(state, current))
3526 __set_current_state(TASK_RUNNING);
3527
3528 if (current->state == TASK_RUNNING)
3529 return 1;
3530 if (ret < 0 || !spin)
3531 break;
3532 cpu_relax();
3533 } while (!need_resched());
3534
3535 __set_current_state(TASK_RUNNING);
3536 return 0;
3537}
3538EXPORT_SYMBOL_GPL(blk_poll);
3539
3540unsigned int blk_mq_rq_cpu(struct request *rq)
3541{
3542 return rq->mq_ctx->cpu;
3543}
3544EXPORT_SYMBOL(blk_mq_rq_cpu);
3545
3546static int __init blk_mq_init(void)
3547{
3548 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3549 blk_mq_hctx_notify_dead);
3550 return 0;
3551}
3552subsys_initcall(blk_mq_init);