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/*
2 * Block multiqueue core code
3 *
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
6 */
7#include <linux/kernel.h>
8#include <linux/module.h>
9#include <linux/backing-dev.h>
10#include <linux/bio.h>
11#include <linux/blkdev.h>
12#include <linux/kmemleak.h>
13#include <linux/mm.h>
14#include <linux/init.h>
15#include <linux/slab.h>
16#include <linux/workqueue.h>
17#include <linux/smp.h>
18#include <linux/llist.h>
19#include <linux/list_sort.h>
20#include <linux/cpu.h>
21#include <linux/cache.h>
22#include <linux/sched/sysctl.h>
23#include <linux/delay.h>
24#include <linux/crash_dump.h>
25
26#include <trace/events/block.h>
27
28#include <linux/blk-mq.h>
29#include "blk.h"
30#include "blk-mq.h"
31#include "blk-mq-tag.h"
32
33static DEFINE_MUTEX(all_q_mutex);
34static LIST_HEAD(all_q_list);
35
36static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
37
38/*
39 * Check if any of the ctx's have pending work in this hardware queue
40 */
41static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
42{
43 unsigned int i;
44
45 for (i = 0; i < hctx->ctx_map.size; i++)
46 if (hctx->ctx_map.map[i].word)
47 return true;
48
49 return false;
50}
51
52static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx,
53 struct blk_mq_ctx *ctx)
54{
55 return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word];
56}
57
58#define CTX_TO_BIT(hctx, ctx) \
59 ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
60
61/*
62 * Mark this ctx as having pending work in this hardware queue
63 */
64static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
65 struct blk_mq_ctx *ctx)
66{
67 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
68
69 if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word))
70 set_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
71}
72
73static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
74 struct blk_mq_ctx *ctx)
75{
76 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
77
78 clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
79}
80
81void blk_mq_freeze_queue_start(struct request_queue *q)
82{
83 int freeze_depth;
84
85 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
86 if (freeze_depth == 1) {
87 percpu_ref_kill(&q->q_usage_counter);
88 blk_mq_run_hw_queues(q, false);
89 }
90}
91EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
92
93static void blk_mq_freeze_queue_wait(struct request_queue *q)
94{
95 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
96}
97
98/*
99 * Guarantee no request is in use, so we can change any data structure of
100 * the queue afterward.
101 */
102void blk_freeze_queue(struct request_queue *q)
103{
104 /*
105 * In the !blk_mq case we are only calling this to kill the
106 * q_usage_counter, otherwise this increases the freeze depth
107 * and waits for it to return to zero. For this reason there is
108 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
109 * exported to drivers as the only user for unfreeze is blk_mq.
110 */
111 blk_mq_freeze_queue_start(q);
112 blk_mq_freeze_queue_wait(q);
113}
114
115void blk_mq_freeze_queue(struct request_queue *q)
116{
117 /*
118 * ...just an alias to keep freeze and unfreeze actions balanced
119 * in the blk_mq_* namespace
120 */
121 blk_freeze_queue(q);
122}
123EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
124
125void blk_mq_unfreeze_queue(struct request_queue *q)
126{
127 int freeze_depth;
128
129 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
130 WARN_ON_ONCE(freeze_depth < 0);
131 if (!freeze_depth) {
132 percpu_ref_reinit(&q->q_usage_counter);
133 wake_up_all(&q->mq_freeze_wq);
134 }
135}
136EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
137
138void blk_mq_wake_waiters(struct request_queue *q)
139{
140 struct blk_mq_hw_ctx *hctx;
141 unsigned int i;
142
143 queue_for_each_hw_ctx(q, hctx, i)
144 if (blk_mq_hw_queue_mapped(hctx))
145 blk_mq_tag_wakeup_all(hctx->tags, true);
146
147 /*
148 * If we are called because the queue has now been marked as
149 * dying, we need to ensure that processes currently waiting on
150 * the queue are notified as well.
151 */
152 wake_up_all(&q->mq_freeze_wq);
153}
154
155bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
156{
157 return blk_mq_has_free_tags(hctx->tags);
158}
159EXPORT_SYMBOL(blk_mq_can_queue);
160
161static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
162 struct request *rq, unsigned int rw_flags)
163{
164 if (blk_queue_io_stat(q))
165 rw_flags |= REQ_IO_STAT;
166
167 INIT_LIST_HEAD(&rq->queuelist);
168 /* csd/requeue_work/fifo_time is initialized before use */
169 rq->q = q;
170 rq->mq_ctx = ctx;
171 rq->cmd_flags |= rw_flags;
172 /* do not touch atomic flags, it needs atomic ops against the timer */
173 rq->cpu = -1;
174 INIT_HLIST_NODE(&rq->hash);
175 RB_CLEAR_NODE(&rq->rb_node);
176 rq->rq_disk = NULL;
177 rq->part = NULL;
178 rq->start_time = jiffies;
179#ifdef CONFIG_BLK_CGROUP
180 rq->rl = NULL;
181 set_start_time_ns(rq);
182 rq->io_start_time_ns = 0;
183#endif
184 rq->nr_phys_segments = 0;
185#if defined(CONFIG_BLK_DEV_INTEGRITY)
186 rq->nr_integrity_segments = 0;
187#endif
188 rq->special = NULL;
189 /* tag was already set */
190 rq->errors = 0;
191
192 rq->cmd = rq->__cmd;
193
194 rq->extra_len = 0;
195 rq->sense_len = 0;
196 rq->resid_len = 0;
197 rq->sense = NULL;
198
199 INIT_LIST_HEAD(&rq->timeout_list);
200 rq->timeout = 0;
201
202 rq->end_io = NULL;
203 rq->end_io_data = NULL;
204 rq->next_rq = NULL;
205
206 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
207}
208
209static struct request *
210__blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
211{
212 struct request *rq;
213 unsigned int tag;
214
215 tag = blk_mq_get_tag(data);
216 if (tag != BLK_MQ_TAG_FAIL) {
217 rq = data->hctx->tags->rqs[tag];
218
219 if (blk_mq_tag_busy(data->hctx)) {
220 rq->cmd_flags = REQ_MQ_INFLIGHT;
221 atomic_inc(&data->hctx->nr_active);
222 }
223
224 rq->tag = tag;
225 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
226 return rq;
227 }
228
229 return NULL;
230}
231
232struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
233 unsigned int flags)
234{
235 struct blk_mq_ctx *ctx;
236 struct blk_mq_hw_ctx *hctx;
237 struct request *rq;
238 struct blk_mq_alloc_data alloc_data;
239 int ret;
240
241 ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
242 if (ret)
243 return ERR_PTR(ret);
244
245 ctx = blk_mq_get_ctx(q);
246 hctx = q->mq_ops->map_queue(q, ctx->cpu);
247 blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
248
249 rq = __blk_mq_alloc_request(&alloc_data, rw);
250 if (!rq && !(flags & BLK_MQ_REQ_NOWAIT)) {
251 __blk_mq_run_hw_queue(hctx);
252 blk_mq_put_ctx(ctx);
253
254 ctx = blk_mq_get_ctx(q);
255 hctx = q->mq_ops->map_queue(q, ctx->cpu);
256 blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
257 rq = __blk_mq_alloc_request(&alloc_data, rw);
258 ctx = alloc_data.ctx;
259 }
260 blk_mq_put_ctx(ctx);
261 if (!rq) {
262 blk_queue_exit(q);
263 return ERR_PTR(-EWOULDBLOCK);
264 }
265 return rq;
266}
267EXPORT_SYMBOL(blk_mq_alloc_request);
268
269static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
270 struct blk_mq_ctx *ctx, struct request *rq)
271{
272 const int tag = rq->tag;
273 struct request_queue *q = rq->q;
274
275 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
276 atomic_dec(&hctx->nr_active);
277 rq->cmd_flags = 0;
278
279 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
280 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
281 blk_queue_exit(q);
282}
283
284void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
285{
286 struct blk_mq_ctx *ctx = rq->mq_ctx;
287
288 ctx->rq_completed[rq_is_sync(rq)]++;
289 __blk_mq_free_request(hctx, ctx, rq);
290
291}
292EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
293
294void blk_mq_free_request(struct request *rq)
295{
296 struct blk_mq_hw_ctx *hctx;
297 struct request_queue *q = rq->q;
298
299 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
300 blk_mq_free_hctx_request(hctx, rq);
301}
302EXPORT_SYMBOL_GPL(blk_mq_free_request);
303
304inline void __blk_mq_end_request(struct request *rq, int error)
305{
306 blk_account_io_done(rq);
307
308 if (rq->end_io) {
309 rq->end_io(rq, error);
310 } else {
311 if (unlikely(blk_bidi_rq(rq)))
312 blk_mq_free_request(rq->next_rq);
313 blk_mq_free_request(rq);
314 }
315}
316EXPORT_SYMBOL(__blk_mq_end_request);
317
318void blk_mq_end_request(struct request *rq, int error)
319{
320 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
321 BUG();
322 __blk_mq_end_request(rq, error);
323}
324EXPORT_SYMBOL(blk_mq_end_request);
325
326static void __blk_mq_complete_request_remote(void *data)
327{
328 struct request *rq = data;
329
330 rq->q->softirq_done_fn(rq);
331}
332
333static void blk_mq_ipi_complete_request(struct request *rq)
334{
335 struct blk_mq_ctx *ctx = rq->mq_ctx;
336 bool shared = false;
337 int cpu;
338
339 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
340 rq->q->softirq_done_fn(rq);
341 return;
342 }
343
344 cpu = get_cpu();
345 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
346 shared = cpus_share_cache(cpu, ctx->cpu);
347
348 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
349 rq->csd.func = __blk_mq_complete_request_remote;
350 rq->csd.info = rq;
351 rq->csd.flags = 0;
352 smp_call_function_single_async(ctx->cpu, &rq->csd);
353 } else {
354 rq->q->softirq_done_fn(rq);
355 }
356 put_cpu();
357}
358
359static void __blk_mq_complete_request(struct request *rq)
360{
361 struct request_queue *q = rq->q;
362
363 if (!q->softirq_done_fn)
364 blk_mq_end_request(rq, rq->errors);
365 else
366 blk_mq_ipi_complete_request(rq);
367}
368
369/**
370 * blk_mq_complete_request - end I/O on a request
371 * @rq: the request being processed
372 *
373 * Description:
374 * Ends all I/O on a request. It does not handle partial completions.
375 * The actual completion happens out-of-order, through a IPI handler.
376 **/
377void blk_mq_complete_request(struct request *rq, int error)
378{
379 struct request_queue *q = rq->q;
380
381 if (unlikely(blk_should_fake_timeout(q)))
382 return;
383 if (!blk_mark_rq_complete(rq)) {
384 rq->errors = error;
385 __blk_mq_complete_request(rq);
386 }
387}
388EXPORT_SYMBOL(blk_mq_complete_request);
389
390int blk_mq_request_started(struct request *rq)
391{
392 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
393}
394EXPORT_SYMBOL_GPL(blk_mq_request_started);
395
396void blk_mq_start_request(struct request *rq)
397{
398 struct request_queue *q = rq->q;
399
400 trace_block_rq_issue(q, rq);
401
402 rq->resid_len = blk_rq_bytes(rq);
403 if (unlikely(blk_bidi_rq(rq)))
404 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
405
406 blk_add_timer(rq);
407
408 /*
409 * Ensure that ->deadline is visible before set the started
410 * flag and clear the completed flag.
411 */
412 smp_mb__before_atomic();
413
414 /*
415 * Mark us as started and clear complete. Complete might have been
416 * set if requeue raced with timeout, which then marked it as
417 * complete. So be sure to clear complete again when we start
418 * the request, otherwise we'll ignore the completion event.
419 */
420 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
421 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
422 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
423 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
424
425 if (q->dma_drain_size && blk_rq_bytes(rq)) {
426 /*
427 * Make sure space for the drain appears. We know we can do
428 * this because max_hw_segments has been adjusted to be one
429 * fewer than the device can handle.
430 */
431 rq->nr_phys_segments++;
432 }
433}
434EXPORT_SYMBOL(blk_mq_start_request);
435
436static void __blk_mq_requeue_request(struct request *rq)
437{
438 struct request_queue *q = rq->q;
439
440 trace_block_rq_requeue(q, rq);
441
442 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
443 if (q->dma_drain_size && blk_rq_bytes(rq))
444 rq->nr_phys_segments--;
445 }
446}
447
448void blk_mq_requeue_request(struct request *rq)
449{
450 __blk_mq_requeue_request(rq);
451
452 BUG_ON(blk_queued_rq(rq));
453 blk_mq_add_to_requeue_list(rq, true);
454}
455EXPORT_SYMBOL(blk_mq_requeue_request);
456
457static void blk_mq_requeue_work(struct work_struct *work)
458{
459 struct request_queue *q =
460 container_of(work, struct request_queue, requeue_work);
461 LIST_HEAD(rq_list);
462 struct request *rq, *next;
463 unsigned long flags;
464
465 spin_lock_irqsave(&q->requeue_lock, flags);
466 list_splice_init(&q->requeue_list, &rq_list);
467 spin_unlock_irqrestore(&q->requeue_lock, flags);
468
469 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
470 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
471 continue;
472
473 rq->cmd_flags &= ~REQ_SOFTBARRIER;
474 list_del_init(&rq->queuelist);
475 blk_mq_insert_request(rq, true, false, false);
476 }
477
478 while (!list_empty(&rq_list)) {
479 rq = list_entry(rq_list.next, struct request, queuelist);
480 list_del_init(&rq->queuelist);
481 blk_mq_insert_request(rq, false, false, false);
482 }
483
484 /*
485 * Use the start variant of queue running here, so that running
486 * the requeue work will kick stopped queues.
487 */
488 blk_mq_start_hw_queues(q);
489}
490
491void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
492{
493 struct request_queue *q = rq->q;
494 unsigned long flags;
495
496 /*
497 * We abuse this flag that is otherwise used by the I/O scheduler to
498 * request head insertation from the workqueue.
499 */
500 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
501
502 spin_lock_irqsave(&q->requeue_lock, flags);
503 if (at_head) {
504 rq->cmd_flags |= REQ_SOFTBARRIER;
505 list_add(&rq->queuelist, &q->requeue_list);
506 } else {
507 list_add_tail(&rq->queuelist, &q->requeue_list);
508 }
509 spin_unlock_irqrestore(&q->requeue_lock, flags);
510}
511EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
512
513void blk_mq_cancel_requeue_work(struct request_queue *q)
514{
515 cancel_work_sync(&q->requeue_work);
516}
517EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
518
519void blk_mq_kick_requeue_list(struct request_queue *q)
520{
521 kblockd_schedule_work(&q->requeue_work);
522}
523EXPORT_SYMBOL(blk_mq_kick_requeue_list);
524
525void blk_mq_abort_requeue_list(struct request_queue *q)
526{
527 unsigned long flags;
528 LIST_HEAD(rq_list);
529
530 spin_lock_irqsave(&q->requeue_lock, flags);
531 list_splice_init(&q->requeue_list, &rq_list);
532 spin_unlock_irqrestore(&q->requeue_lock, flags);
533
534 while (!list_empty(&rq_list)) {
535 struct request *rq;
536
537 rq = list_first_entry(&rq_list, struct request, queuelist);
538 list_del_init(&rq->queuelist);
539 rq->errors = -EIO;
540 blk_mq_end_request(rq, rq->errors);
541 }
542}
543EXPORT_SYMBOL(blk_mq_abort_requeue_list);
544
545struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
546{
547 if (tag < tags->nr_tags)
548 return tags->rqs[tag];
549
550 return NULL;
551}
552EXPORT_SYMBOL(blk_mq_tag_to_rq);
553
554struct blk_mq_timeout_data {
555 unsigned long next;
556 unsigned int next_set;
557};
558
559void blk_mq_rq_timed_out(struct request *req, bool reserved)
560{
561 struct blk_mq_ops *ops = req->q->mq_ops;
562 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
563
564 /*
565 * We know that complete is set at this point. If STARTED isn't set
566 * anymore, then the request isn't active and the "timeout" should
567 * just be ignored. This can happen due to the bitflag ordering.
568 * Timeout first checks if STARTED is set, and if it is, assumes
569 * the request is active. But if we race with completion, then
570 * we both flags will get cleared. So check here again, and ignore
571 * a timeout event with a request that isn't active.
572 */
573 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
574 return;
575
576 if (ops->timeout)
577 ret = ops->timeout(req, reserved);
578
579 switch (ret) {
580 case BLK_EH_HANDLED:
581 __blk_mq_complete_request(req);
582 break;
583 case BLK_EH_RESET_TIMER:
584 blk_add_timer(req);
585 blk_clear_rq_complete(req);
586 break;
587 case BLK_EH_NOT_HANDLED:
588 break;
589 default:
590 printk(KERN_ERR "block: bad eh return: %d\n", ret);
591 break;
592 }
593}
594
595static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
596 struct request *rq, void *priv, bool reserved)
597{
598 struct blk_mq_timeout_data *data = priv;
599
600 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
601 /*
602 * If a request wasn't started before the queue was
603 * marked dying, kill it here or it'll go unnoticed.
604 */
605 if (unlikely(blk_queue_dying(rq->q))) {
606 rq->errors = -EIO;
607 blk_mq_end_request(rq, rq->errors);
608 }
609 return;
610 }
611
612 if (time_after_eq(jiffies, rq->deadline)) {
613 if (!blk_mark_rq_complete(rq))
614 blk_mq_rq_timed_out(rq, reserved);
615 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
616 data->next = rq->deadline;
617 data->next_set = 1;
618 }
619}
620
621static void blk_mq_timeout_work(struct work_struct *work)
622{
623 struct request_queue *q =
624 container_of(work, struct request_queue, timeout_work);
625 struct blk_mq_timeout_data data = {
626 .next = 0,
627 .next_set = 0,
628 };
629 int i;
630
631 if (blk_queue_enter(q, true))
632 return;
633
634 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
635
636 if (data.next_set) {
637 data.next = blk_rq_timeout(round_jiffies_up(data.next));
638 mod_timer(&q->timeout, data.next);
639 } else {
640 struct blk_mq_hw_ctx *hctx;
641
642 queue_for_each_hw_ctx(q, hctx, i) {
643 /* the hctx may be unmapped, so check it here */
644 if (blk_mq_hw_queue_mapped(hctx))
645 blk_mq_tag_idle(hctx);
646 }
647 }
648 blk_queue_exit(q);
649}
650
651/*
652 * Reverse check our software queue for entries that we could potentially
653 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
654 * too much time checking for merges.
655 */
656static bool blk_mq_attempt_merge(struct request_queue *q,
657 struct blk_mq_ctx *ctx, struct bio *bio)
658{
659 struct request *rq;
660 int checked = 8;
661
662 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
663 int el_ret;
664
665 if (!checked--)
666 break;
667
668 if (!blk_rq_merge_ok(rq, bio))
669 continue;
670
671 el_ret = blk_try_merge(rq, bio);
672 if (el_ret == ELEVATOR_BACK_MERGE) {
673 if (bio_attempt_back_merge(q, rq, bio)) {
674 ctx->rq_merged++;
675 return true;
676 }
677 break;
678 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
679 if (bio_attempt_front_merge(q, rq, bio)) {
680 ctx->rq_merged++;
681 return true;
682 }
683 break;
684 }
685 }
686
687 return false;
688}
689
690/*
691 * Process software queues that have been marked busy, splicing them
692 * to the for-dispatch
693 */
694static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
695{
696 struct blk_mq_ctx *ctx;
697 int i;
698
699 for (i = 0; i < hctx->ctx_map.size; i++) {
700 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
701 unsigned int off, bit;
702
703 if (!bm->word)
704 continue;
705
706 bit = 0;
707 off = i * hctx->ctx_map.bits_per_word;
708 do {
709 bit = find_next_bit(&bm->word, bm->depth, bit);
710 if (bit >= bm->depth)
711 break;
712
713 ctx = hctx->ctxs[bit + off];
714 clear_bit(bit, &bm->word);
715 spin_lock(&ctx->lock);
716 list_splice_tail_init(&ctx->rq_list, list);
717 spin_unlock(&ctx->lock);
718
719 bit++;
720 } while (1);
721 }
722}
723
724/*
725 * Run this hardware queue, pulling any software queues mapped to it in.
726 * Note that this function currently has various problems around ordering
727 * of IO. In particular, we'd like FIFO behaviour on handling existing
728 * items on the hctx->dispatch list. Ignore that for now.
729 */
730static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
731{
732 struct request_queue *q = hctx->queue;
733 struct request *rq;
734 LIST_HEAD(rq_list);
735 LIST_HEAD(driver_list);
736 struct list_head *dptr;
737 int queued;
738
739 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
740
741 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
742 return;
743
744 hctx->run++;
745
746 /*
747 * Touch any software queue that has pending entries.
748 */
749 flush_busy_ctxs(hctx, &rq_list);
750
751 /*
752 * If we have previous entries on our dispatch list, grab them
753 * and stuff them at the front for more fair dispatch.
754 */
755 if (!list_empty_careful(&hctx->dispatch)) {
756 spin_lock(&hctx->lock);
757 if (!list_empty(&hctx->dispatch))
758 list_splice_init(&hctx->dispatch, &rq_list);
759 spin_unlock(&hctx->lock);
760 }
761
762 /*
763 * Start off with dptr being NULL, so we start the first request
764 * immediately, even if we have more pending.
765 */
766 dptr = NULL;
767
768 /*
769 * Now process all the entries, sending them to the driver.
770 */
771 queued = 0;
772 while (!list_empty(&rq_list)) {
773 struct blk_mq_queue_data bd;
774 int ret;
775
776 rq = list_first_entry(&rq_list, struct request, queuelist);
777 list_del_init(&rq->queuelist);
778
779 bd.rq = rq;
780 bd.list = dptr;
781 bd.last = list_empty(&rq_list);
782
783 ret = q->mq_ops->queue_rq(hctx, &bd);
784 switch (ret) {
785 case BLK_MQ_RQ_QUEUE_OK:
786 queued++;
787 continue;
788 case BLK_MQ_RQ_QUEUE_BUSY:
789 list_add(&rq->queuelist, &rq_list);
790 __blk_mq_requeue_request(rq);
791 break;
792 default:
793 pr_err("blk-mq: bad return on queue: %d\n", ret);
794 case BLK_MQ_RQ_QUEUE_ERROR:
795 rq->errors = -EIO;
796 blk_mq_end_request(rq, rq->errors);
797 break;
798 }
799
800 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
801 break;
802
803 /*
804 * We've done the first request. If we have more than 1
805 * left in the list, set dptr to defer issue.
806 */
807 if (!dptr && rq_list.next != rq_list.prev)
808 dptr = &driver_list;
809 }
810
811 if (!queued)
812 hctx->dispatched[0]++;
813 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
814 hctx->dispatched[ilog2(queued) + 1]++;
815
816 /*
817 * Any items that need requeuing? Stuff them into hctx->dispatch,
818 * that is where we will continue on next queue run.
819 */
820 if (!list_empty(&rq_list)) {
821 spin_lock(&hctx->lock);
822 list_splice(&rq_list, &hctx->dispatch);
823 spin_unlock(&hctx->lock);
824 /*
825 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
826 * it's possible the queue is stopped and restarted again
827 * before this. Queue restart will dispatch requests. And since
828 * requests in rq_list aren't added into hctx->dispatch yet,
829 * the requests in rq_list might get lost.
830 *
831 * blk_mq_run_hw_queue() already checks the STOPPED bit
832 **/
833 blk_mq_run_hw_queue(hctx, true);
834 }
835}
836
837/*
838 * It'd be great if the workqueue API had a way to pass
839 * in a mask and had some smarts for more clever placement.
840 * For now we just round-robin here, switching for every
841 * BLK_MQ_CPU_WORK_BATCH queued items.
842 */
843static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
844{
845 if (hctx->queue->nr_hw_queues == 1)
846 return WORK_CPU_UNBOUND;
847
848 if (--hctx->next_cpu_batch <= 0) {
849 int cpu = hctx->next_cpu, next_cpu;
850
851 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
852 if (next_cpu >= nr_cpu_ids)
853 next_cpu = cpumask_first(hctx->cpumask);
854
855 hctx->next_cpu = next_cpu;
856 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
857
858 return cpu;
859 }
860
861 return hctx->next_cpu;
862}
863
864void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
865{
866 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
867 !blk_mq_hw_queue_mapped(hctx)))
868 return;
869
870 if (!async) {
871 int cpu = get_cpu();
872 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
873 __blk_mq_run_hw_queue(hctx);
874 put_cpu();
875 return;
876 }
877
878 put_cpu();
879 }
880
881 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
882 &hctx->run_work, 0);
883}
884
885void blk_mq_run_hw_queues(struct request_queue *q, bool async)
886{
887 struct blk_mq_hw_ctx *hctx;
888 int i;
889
890 queue_for_each_hw_ctx(q, hctx, i) {
891 if ((!blk_mq_hctx_has_pending(hctx) &&
892 list_empty_careful(&hctx->dispatch)) ||
893 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
894 continue;
895
896 blk_mq_run_hw_queue(hctx, async);
897 }
898}
899EXPORT_SYMBOL(blk_mq_run_hw_queues);
900
901void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
902{
903 cancel_delayed_work(&hctx->run_work);
904 cancel_delayed_work(&hctx->delay_work);
905 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
906}
907EXPORT_SYMBOL(blk_mq_stop_hw_queue);
908
909void blk_mq_stop_hw_queues(struct request_queue *q)
910{
911 struct blk_mq_hw_ctx *hctx;
912 int i;
913
914 queue_for_each_hw_ctx(q, hctx, i)
915 blk_mq_stop_hw_queue(hctx);
916}
917EXPORT_SYMBOL(blk_mq_stop_hw_queues);
918
919void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
920{
921 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
922
923 blk_mq_run_hw_queue(hctx, false);
924}
925EXPORT_SYMBOL(blk_mq_start_hw_queue);
926
927void blk_mq_start_hw_queues(struct request_queue *q)
928{
929 struct blk_mq_hw_ctx *hctx;
930 int i;
931
932 queue_for_each_hw_ctx(q, hctx, i)
933 blk_mq_start_hw_queue(hctx);
934}
935EXPORT_SYMBOL(blk_mq_start_hw_queues);
936
937void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
938{
939 struct blk_mq_hw_ctx *hctx;
940 int i;
941
942 queue_for_each_hw_ctx(q, hctx, i) {
943 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
944 continue;
945
946 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
947 blk_mq_run_hw_queue(hctx, async);
948 }
949}
950EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
951
952static void blk_mq_run_work_fn(struct work_struct *work)
953{
954 struct blk_mq_hw_ctx *hctx;
955
956 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
957
958 __blk_mq_run_hw_queue(hctx);
959}
960
961static void blk_mq_delay_work_fn(struct work_struct *work)
962{
963 struct blk_mq_hw_ctx *hctx;
964
965 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
966
967 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
968 __blk_mq_run_hw_queue(hctx);
969}
970
971void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
972{
973 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
974 return;
975
976 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
977 &hctx->delay_work, msecs_to_jiffies(msecs));
978}
979EXPORT_SYMBOL(blk_mq_delay_queue);
980
981static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
982 struct blk_mq_ctx *ctx,
983 struct request *rq,
984 bool at_head)
985{
986 trace_block_rq_insert(hctx->queue, rq);
987
988 if (at_head)
989 list_add(&rq->queuelist, &ctx->rq_list);
990 else
991 list_add_tail(&rq->queuelist, &ctx->rq_list);
992}
993
994static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
995 struct request *rq, bool at_head)
996{
997 struct blk_mq_ctx *ctx = rq->mq_ctx;
998
999 __blk_mq_insert_req_list(hctx, ctx, rq, at_head);
1000 blk_mq_hctx_mark_pending(hctx, ctx);
1001}
1002
1003void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1004 bool async)
1005{
1006 struct request_queue *q = rq->q;
1007 struct blk_mq_hw_ctx *hctx;
1008 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
1009
1010 current_ctx = blk_mq_get_ctx(q);
1011 if (!cpu_online(ctx->cpu))
1012 rq->mq_ctx = ctx = current_ctx;
1013
1014 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1015
1016 spin_lock(&ctx->lock);
1017 __blk_mq_insert_request(hctx, rq, at_head);
1018 spin_unlock(&ctx->lock);
1019
1020 if (run_queue)
1021 blk_mq_run_hw_queue(hctx, async);
1022
1023 blk_mq_put_ctx(current_ctx);
1024}
1025
1026static void blk_mq_insert_requests(struct request_queue *q,
1027 struct blk_mq_ctx *ctx,
1028 struct list_head *list,
1029 int depth,
1030 bool from_schedule)
1031
1032{
1033 struct blk_mq_hw_ctx *hctx;
1034 struct blk_mq_ctx *current_ctx;
1035
1036 trace_block_unplug(q, depth, !from_schedule);
1037
1038 current_ctx = blk_mq_get_ctx(q);
1039
1040 if (!cpu_online(ctx->cpu))
1041 ctx = current_ctx;
1042 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1043
1044 /*
1045 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1046 * offline now
1047 */
1048 spin_lock(&ctx->lock);
1049 while (!list_empty(list)) {
1050 struct request *rq;
1051
1052 rq = list_first_entry(list, struct request, queuelist);
1053 list_del_init(&rq->queuelist);
1054 rq->mq_ctx = ctx;
1055 __blk_mq_insert_req_list(hctx, ctx, rq, false);
1056 }
1057 blk_mq_hctx_mark_pending(hctx, ctx);
1058 spin_unlock(&ctx->lock);
1059
1060 blk_mq_run_hw_queue(hctx, from_schedule);
1061 blk_mq_put_ctx(current_ctx);
1062}
1063
1064static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1065{
1066 struct request *rqa = container_of(a, struct request, queuelist);
1067 struct request *rqb = container_of(b, struct request, queuelist);
1068
1069 return !(rqa->mq_ctx < rqb->mq_ctx ||
1070 (rqa->mq_ctx == rqb->mq_ctx &&
1071 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1072}
1073
1074void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1075{
1076 struct blk_mq_ctx *this_ctx;
1077 struct request_queue *this_q;
1078 struct request *rq;
1079 LIST_HEAD(list);
1080 LIST_HEAD(ctx_list);
1081 unsigned int depth;
1082
1083 list_splice_init(&plug->mq_list, &list);
1084
1085 list_sort(NULL, &list, plug_ctx_cmp);
1086
1087 this_q = NULL;
1088 this_ctx = NULL;
1089 depth = 0;
1090
1091 while (!list_empty(&list)) {
1092 rq = list_entry_rq(list.next);
1093 list_del_init(&rq->queuelist);
1094 BUG_ON(!rq->q);
1095 if (rq->mq_ctx != this_ctx) {
1096 if (this_ctx) {
1097 blk_mq_insert_requests(this_q, this_ctx,
1098 &ctx_list, depth,
1099 from_schedule);
1100 }
1101
1102 this_ctx = rq->mq_ctx;
1103 this_q = rq->q;
1104 depth = 0;
1105 }
1106
1107 depth++;
1108 list_add_tail(&rq->queuelist, &ctx_list);
1109 }
1110
1111 /*
1112 * If 'this_ctx' is set, we know we have entries to complete
1113 * on 'ctx_list'. Do those.
1114 */
1115 if (this_ctx) {
1116 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1117 from_schedule);
1118 }
1119}
1120
1121static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1122{
1123 init_request_from_bio(rq, bio);
1124
1125 if (blk_do_io_stat(rq))
1126 blk_account_io_start(rq, 1);
1127}
1128
1129static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1130{
1131 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1132 !blk_queue_nomerges(hctx->queue);
1133}
1134
1135static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1136 struct blk_mq_ctx *ctx,
1137 struct request *rq, struct bio *bio)
1138{
1139 if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
1140 blk_mq_bio_to_request(rq, bio);
1141 spin_lock(&ctx->lock);
1142insert_rq:
1143 __blk_mq_insert_request(hctx, rq, false);
1144 spin_unlock(&ctx->lock);
1145 return false;
1146 } else {
1147 struct request_queue *q = hctx->queue;
1148
1149 spin_lock(&ctx->lock);
1150 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1151 blk_mq_bio_to_request(rq, bio);
1152 goto insert_rq;
1153 }
1154
1155 spin_unlock(&ctx->lock);
1156 __blk_mq_free_request(hctx, ctx, rq);
1157 return true;
1158 }
1159}
1160
1161struct blk_map_ctx {
1162 struct blk_mq_hw_ctx *hctx;
1163 struct blk_mq_ctx *ctx;
1164};
1165
1166static struct request *blk_mq_map_request(struct request_queue *q,
1167 struct bio *bio,
1168 struct blk_map_ctx *data)
1169{
1170 struct blk_mq_hw_ctx *hctx;
1171 struct blk_mq_ctx *ctx;
1172 struct request *rq;
1173 int rw = bio_data_dir(bio);
1174 struct blk_mq_alloc_data alloc_data;
1175
1176 blk_queue_enter_live(q);
1177 ctx = blk_mq_get_ctx(q);
1178 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1179
1180 if (rw_is_sync(bio->bi_rw))
1181 rw |= REQ_SYNC;
1182
1183 trace_block_getrq(q, bio, rw);
1184 blk_mq_set_alloc_data(&alloc_data, q, BLK_MQ_REQ_NOWAIT, ctx, hctx);
1185 rq = __blk_mq_alloc_request(&alloc_data, rw);
1186 if (unlikely(!rq)) {
1187 __blk_mq_run_hw_queue(hctx);
1188 blk_mq_put_ctx(ctx);
1189 trace_block_sleeprq(q, bio, rw);
1190
1191 ctx = blk_mq_get_ctx(q);
1192 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1193 blk_mq_set_alloc_data(&alloc_data, q, 0, ctx, hctx);
1194 rq = __blk_mq_alloc_request(&alloc_data, rw);
1195 ctx = alloc_data.ctx;
1196 hctx = alloc_data.hctx;
1197 }
1198
1199 hctx->queued++;
1200 data->hctx = hctx;
1201 data->ctx = ctx;
1202 return rq;
1203}
1204
1205static int blk_mq_direct_issue_request(struct request *rq, blk_qc_t *cookie)
1206{
1207 int ret;
1208 struct request_queue *q = rq->q;
1209 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q,
1210 rq->mq_ctx->cpu);
1211 struct blk_mq_queue_data bd = {
1212 .rq = rq,
1213 .list = NULL,
1214 .last = 1
1215 };
1216 blk_qc_t new_cookie = blk_tag_to_qc_t(rq->tag, hctx->queue_num);
1217
1218 /*
1219 * For OK queue, we are done. For error, kill it. Any other
1220 * error (busy), just add it to our list as we previously
1221 * would have done
1222 */
1223 ret = q->mq_ops->queue_rq(hctx, &bd);
1224 if (ret == BLK_MQ_RQ_QUEUE_OK) {
1225 *cookie = new_cookie;
1226 return 0;
1227 }
1228
1229 __blk_mq_requeue_request(rq);
1230
1231 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1232 *cookie = BLK_QC_T_NONE;
1233 rq->errors = -EIO;
1234 blk_mq_end_request(rq, rq->errors);
1235 return 0;
1236 }
1237
1238 return -1;
1239}
1240
1241/*
1242 * Multiple hardware queue variant. This will not use per-process plugs,
1243 * but will attempt to bypass the hctx queueing if we can go straight to
1244 * hardware for SYNC IO.
1245 */
1246static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1247{
1248 const int is_sync = rw_is_sync(bio->bi_rw);
1249 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1250 struct blk_map_ctx data;
1251 struct request *rq;
1252 unsigned int request_count = 0;
1253 struct blk_plug *plug;
1254 struct request *same_queue_rq = NULL;
1255 blk_qc_t cookie;
1256
1257 blk_queue_bounce(q, &bio);
1258
1259 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1260 bio_io_error(bio);
1261 return BLK_QC_T_NONE;
1262 }
1263
1264 blk_queue_split(q, &bio, q->bio_split);
1265
1266 if (!is_flush_fua && !blk_queue_nomerges(q)) {
1267 if (blk_attempt_plug_merge(q, bio, &request_count,
1268 &same_queue_rq))
1269 return BLK_QC_T_NONE;
1270 } else
1271 request_count = blk_plug_queued_count(q);
1272
1273 rq = blk_mq_map_request(q, bio, &data);
1274 if (unlikely(!rq))
1275 return BLK_QC_T_NONE;
1276
1277 cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
1278
1279 if (unlikely(is_flush_fua)) {
1280 blk_mq_bio_to_request(rq, bio);
1281 blk_insert_flush(rq);
1282 goto run_queue;
1283 }
1284
1285 plug = current->plug;
1286 /*
1287 * If the driver supports defer issued based on 'last', then
1288 * queue it up like normal since we can potentially save some
1289 * CPU this way.
1290 */
1291 if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1292 !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1293 struct request *old_rq = NULL;
1294
1295 blk_mq_bio_to_request(rq, bio);
1296
1297 /*
1298 * We do limited pluging. If the bio can be merged, do that.
1299 * Otherwise the existing request in the plug list will be
1300 * issued. So the plug list will have one request at most
1301 */
1302 if (plug) {
1303 /*
1304 * The plug list might get flushed before this. If that
1305 * happens, same_queue_rq is invalid and plug list is
1306 * empty
1307 */
1308 if (same_queue_rq && !list_empty(&plug->mq_list)) {
1309 old_rq = same_queue_rq;
1310 list_del_init(&old_rq->queuelist);
1311 }
1312 list_add_tail(&rq->queuelist, &plug->mq_list);
1313 } else /* is_sync */
1314 old_rq = rq;
1315 blk_mq_put_ctx(data.ctx);
1316 if (!old_rq)
1317 goto done;
1318 if (!blk_mq_direct_issue_request(old_rq, &cookie))
1319 goto done;
1320 blk_mq_insert_request(old_rq, false, true, true);
1321 goto done;
1322 }
1323
1324 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1325 /*
1326 * For a SYNC request, send it to the hardware immediately. For
1327 * an ASYNC request, just ensure that we run it later on. The
1328 * latter allows for merging opportunities and more efficient
1329 * dispatching.
1330 */
1331run_queue:
1332 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1333 }
1334 blk_mq_put_ctx(data.ctx);
1335done:
1336 return cookie;
1337}
1338
1339/*
1340 * Single hardware queue variant. This will attempt to use any per-process
1341 * plug for merging and IO deferral.
1342 */
1343static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio)
1344{
1345 const int is_sync = rw_is_sync(bio->bi_rw);
1346 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1347 struct blk_plug *plug;
1348 unsigned int request_count = 0;
1349 struct blk_map_ctx data;
1350 struct request *rq;
1351 blk_qc_t cookie;
1352
1353 blk_queue_bounce(q, &bio);
1354
1355 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1356 bio_io_error(bio);
1357 return BLK_QC_T_NONE;
1358 }
1359
1360 blk_queue_split(q, &bio, q->bio_split);
1361
1362 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1363 blk_attempt_plug_merge(q, bio, &request_count, NULL))
1364 return BLK_QC_T_NONE;
1365
1366 rq = blk_mq_map_request(q, bio, &data);
1367 if (unlikely(!rq))
1368 return BLK_QC_T_NONE;
1369
1370 cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
1371
1372 if (unlikely(is_flush_fua)) {
1373 blk_mq_bio_to_request(rq, bio);
1374 blk_insert_flush(rq);
1375 goto run_queue;
1376 }
1377
1378 /*
1379 * A task plug currently exists. Since this is completely lockless,
1380 * utilize that to temporarily store requests until the task is
1381 * either done or scheduled away.
1382 */
1383 plug = current->plug;
1384 if (plug) {
1385 blk_mq_bio_to_request(rq, bio);
1386 if (!request_count)
1387 trace_block_plug(q);
1388
1389 blk_mq_put_ctx(data.ctx);
1390
1391 if (request_count >= BLK_MAX_REQUEST_COUNT) {
1392 blk_flush_plug_list(plug, false);
1393 trace_block_plug(q);
1394 }
1395
1396 list_add_tail(&rq->queuelist, &plug->mq_list);
1397 return cookie;
1398 }
1399
1400 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1401 /*
1402 * For a SYNC request, send it to the hardware immediately. For
1403 * an ASYNC request, just ensure that we run it later on. The
1404 * latter allows for merging opportunities and more efficient
1405 * dispatching.
1406 */
1407run_queue:
1408 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1409 }
1410
1411 blk_mq_put_ctx(data.ctx);
1412 return cookie;
1413}
1414
1415/*
1416 * Default mapping to a software queue, since we use one per CPU.
1417 */
1418struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1419{
1420 return q->queue_hw_ctx[q->mq_map[cpu]];
1421}
1422EXPORT_SYMBOL(blk_mq_map_queue);
1423
1424static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1425 struct blk_mq_tags *tags, unsigned int hctx_idx)
1426{
1427 struct page *page;
1428
1429 if (tags->rqs && set->ops->exit_request) {
1430 int i;
1431
1432 for (i = 0; i < tags->nr_tags; i++) {
1433 if (!tags->rqs[i])
1434 continue;
1435 set->ops->exit_request(set->driver_data, tags->rqs[i],
1436 hctx_idx, i);
1437 tags->rqs[i] = NULL;
1438 }
1439 }
1440
1441 while (!list_empty(&tags->page_list)) {
1442 page = list_first_entry(&tags->page_list, struct page, lru);
1443 list_del_init(&page->lru);
1444 /*
1445 * Remove kmemleak object previously allocated in
1446 * blk_mq_init_rq_map().
1447 */
1448 kmemleak_free(page_address(page));
1449 __free_pages(page, page->private);
1450 }
1451
1452 kfree(tags->rqs);
1453
1454 blk_mq_free_tags(tags);
1455}
1456
1457static size_t order_to_size(unsigned int order)
1458{
1459 return (size_t)PAGE_SIZE << order;
1460}
1461
1462static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1463 unsigned int hctx_idx)
1464{
1465 struct blk_mq_tags *tags;
1466 unsigned int i, j, entries_per_page, max_order = 4;
1467 size_t rq_size, left;
1468
1469 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1470 set->numa_node,
1471 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1472 if (!tags)
1473 return NULL;
1474
1475 INIT_LIST_HEAD(&tags->page_list);
1476
1477 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1478 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1479 set->numa_node);
1480 if (!tags->rqs) {
1481 blk_mq_free_tags(tags);
1482 return NULL;
1483 }
1484
1485 /*
1486 * rq_size is the size of the request plus driver payload, rounded
1487 * to the cacheline size
1488 */
1489 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1490 cache_line_size());
1491 left = rq_size * set->queue_depth;
1492
1493 for (i = 0; i < set->queue_depth; ) {
1494 int this_order = max_order;
1495 struct page *page;
1496 int to_do;
1497 void *p;
1498
1499 while (left < order_to_size(this_order - 1) && this_order)
1500 this_order--;
1501
1502 do {
1503 page = alloc_pages_node(set->numa_node,
1504 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1505 this_order);
1506 if (page)
1507 break;
1508 if (!this_order--)
1509 break;
1510 if (order_to_size(this_order) < rq_size)
1511 break;
1512 } while (1);
1513
1514 if (!page)
1515 goto fail;
1516
1517 page->private = this_order;
1518 list_add_tail(&page->lru, &tags->page_list);
1519
1520 p = page_address(page);
1521 /*
1522 * Allow kmemleak to scan these pages as they contain pointers
1523 * to additional allocations like via ops->init_request().
1524 */
1525 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_KERNEL);
1526 entries_per_page = order_to_size(this_order) / rq_size;
1527 to_do = min(entries_per_page, set->queue_depth - i);
1528 left -= to_do * rq_size;
1529 for (j = 0; j < to_do; j++) {
1530 tags->rqs[i] = p;
1531 if (set->ops->init_request) {
1532 if (set->ops->init_request(set->driver_data,
1533 tags->rqs[i], hctx_idx, i,
1534 set->numa_node)) {
1535 tags->rqs[i] = NULL;
1536 goto fail;
1537 }
1538 }
1539
1540 p += rq_size;
1541 i++;
1542 }
1543 }
1544 return tags;
1545
1546fail:
1547 blk_mq_free_rq_map(set, tags, hctx_idx);
1548 return NULL;
1549}
1550
1551static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1552{
1553 kfree(bitmap->map);
1554}
1555
1556static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1557{
1558 unsigned int bpw = 8, total, num_maps, i;
1559
1560 bitmap->bits_per_word = bpw;
1561
1562 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1563 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1564 GFP_KERNEL, node);
1565 if (!bitmap->map)
1566 return -ENOMEM;
1567
1568 total = nr_cpu_ids;
1569 for (i = 0; i < num_maps; i++) {
1570 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1571 total -= bitmap->map[i].depth;
1572 }
1573
1574 return 0;
1575}
1576
1577static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1578{
1579 struct request_queue *q = hctx->queue;
1580 struct blk_mq_ctx *ctx;
1581 LIST_HEAD(tmp);
1582
1583 /*
1584 * Move ctx entries to new CPU, if this one is going away.
1585 */
1586 ctx = __blk_mq_get_ctx(q, cpu);
1587
1588 spin_lock(&ctx->lock);
1589 if (!list_empty(&ctx->rq_list)) {
1590 list_splice_init(&ctx->rq_list, &tmp);
1591 blk_mq_hctx_clear_pending(hctx, ctx);
1592 }
1593 spin_unlock(&ctx->lock);
1594
1595 if (list_empty(&tmp))
1596 return NOTIFY_OK;
1597
1598 ctx = blk_mq_get_ctx(q);
1599 spin_lock(&ctx->lock);
1600
1601 while (!list_empty(&tmp)) {
1602 struct request *rq;
1603
1604 rq = list_first_entry(&tmp, struct request, queuelist);
1605 rq->mq_ctx = ctx;
1606 list_move_tail(&rq->queuelist, &ctx->rq_list);
1607 }
1608
1609 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1610 blk_mq_hctx_mark_pending(hctx, ctx);
1611
1612 spin_unlock(&ctx->lock);
1613
1614 blk_mq_run_hw_queue(hctx, true);
1615 blk_mq_put_ctx(ctx);
1616 return NOTIFY_OK;
1617}
1618
1619static int blk_mq_hctx_notify(void *data, unsigned long action,
1620 unsigned int cpu)
1621{
1622 struct blk_mq_hw_ctx *hctx = data;
1623
1624 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1625 return blk_mq_hctx_cpu_offline(hctx, cpu);
1626
1627 /*
1628 * In case of CPU online, tags may be reallocated
1629 * in blk_mq_map_swqueue() after mapping is updated.
1630 */
1631
1632 return NOTIFY_OK;
1633}
1634
1635/* hctx->ctxs will be freed in queue's release handler */
1636static void blk_mq_exit_hctx(struct request_queue *q,
1637 struct blk_mq_tag_set *set,
1638 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1639{
1640 unsigned flush_start_tag = set->queue_depth;
1641
1642 blk_mq_tag_idle(hctx);
1643
1644 if (set->ops->exit_request)
1645 set->ops->exit_request(set->driver_data,
1646 hctx->fq->flush_rq, hctx_idx,
1647 flush_start_tag + hctx_idx);
1648
1649 if (set->ops->exit_hctx)
1650 set->ops->exit_hctx(hctx, hctx_idx);
1651
1652 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1653 blk_free_flush_queue(hctx->fq);
1654 blk_mq_free_bitmap(&hctx->ctx_map);
1655}
1656
1657static void blk_mq_exit_hw_queues(struct request_queue *q,
1658 struct blk_mq_tag_set *set, int nr_queue)
1659{
1660 struct blk_mq_hw_ctx *hctx;
1661 unsigned int i;
1662
1663 queue_for_each_hw_ctx(q, hctx, i) {
1664 if (i == nr_queue)
1665 break;
1666 blk_mq_exit_hctx(q, set, hctx, i);
1667 }
1668}
1669
1670static void blk_mq_free_hw_queues(struct request_queue *q,
1671 struct blk_mq_tag_set *set)
1672{
1673 struct blk_mq_hw_ctx *hctx;
1674 unsigned int i;
1675
1676 queue_for_each_hw_ctx(q, hctx, i)
1677 free_cpumask_var(hctx->cpumask);
1678}
1679
1680static int blk_mq_init_hctx(struct request_queue *q,
1681 struct blk_mq_tag_set *set,
1682 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1683{
1684 int node;
1685 unsigned flush_start_tag = set->queue_depth;
1686
1687 node = hctx->numa_node;
1688 if (node == NUMA_NO_NODE)
1689 node = hctx->numa_node = set->numa_node;
1690
1691 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1692 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1693 spin_lock_init(&hctx->lock);
1694 INIT_LIST_HEAD(&hctx->dispatch);
1695 hctx->queue = q;
1696 hctx->queue_num = hctx_idx;
1697 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1698
1699 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1700 blk_mq_hctx_notify, hctx);
1701 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1702
1703 hctx->tags = set->tags[hctx_idx];
1704
1705 /*
1706 * Allocate space for all possible cpus to avoid allocation at
1707 * runtime
1708 */
1709 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1710 GFP_KERNEL, node);
1711 if (!hctx->ctxs)
1712 goto unregister_cpu_notifier;
1713
1714 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1715 goto free_ctxs;
1716
1717 hctx->nr_ctx = 0;
1718
1719 if (set->ops->init_hctx &&
1720 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1721 goto free_bitmap;
1722
1723 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1724 if (!hctx->fq)
1725 goto exit_hctx;
1726
1727 if (set->ops->init_request &&
1728 set->ops->init_request(set->driver_data,
1729 hctx->fq->flush_rq, hctx_idx,
1730 flush_start_tag + hctx_idx, node))
1731 goto free_fq;
1732
1733 return 0;
1734
1735 free_fq:
1736 kfree(hctx->fq);
1737 exit_hctx:
1738 if (set->ops->exit_hctx)
1739 set->ops->exit_hctx(hctx, hctx_idx);
1740 free_bitmap:
1741 blk_mq_free_bitmap(&hctx->ctx_map);
1742 free_ctxs:
1743 kfree(hctx->ctxs);
1744 unregister_cpu_notifier:
1745 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1746
1747 return -1;
1748}
1749
1750static void blk_mq_init_cpu_queues(struct request_queue *q,
1751 unsigned int nr_hw_queues)
1752{
1753 unsigned int i;
1754
1755 for_each_possible_cpu(i) {
1756 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1757 struct blk_mq_hw_ctx *hctx;
1758
1759 memset(__ctx, 0, sizeof(*__ctx));
1760 __ctx->cpu = i;
1761 spin_lock_init(&__ctx->lock);
1762 INIT_LIST_HEAD(&__ctx->rq_list);
1763 __ctx->queue = q;
1764
1765 /* If the cpu isn't online, the cpu is mapped to first hctx */
1766 if (!cpu_online(i))
1767 continue;
1768
1769 hctx = q->mq_ops->map_queue(q, i);
1770
1771 /*
1772 * Set local node, IFF we have more than one hw queue. If
1773 * not, we remain on the home node of the device
1774 */
1775 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1776 hctx->numa_node = local_memory_node(cpu_to_node(i));
1777 }
1778}
1779
1780static void blk_mq_map_swqueue(struct request_queue *q,
1781 const struct cpumask *online_mask)
1782{
1783 unsigned int i;
1784 struct blk_mq_hw_ctx *hctx;
1785 struct blk_mq_ctx *ctx;
1786 struct blk_mq_tag_set *set = q->tag_set;
1787
1788 /*
1789 * Avoid others reading imcomplete hctx->cpumask through sysfs
1790 */
1791 mutex_lock(&q->sysfs_lock);
1792
1793 queue_for_each_hw_ctx(q, hctx, i) {
1794 cpumask_clear(hctx->cpumask);
1795 hctx->nr_ctx = 0;
1796 }
1797
1798 /*
1799 * Map software to hardware queues
1800 */
1801 for_each_possible_cpu(i) {
1802 /* If the cpu isn't online, the cpu is mapped to first hctx */
1803 if (!cpumask_test_cpu(i, online_mask))
1804 continue;
1805
1806 ctx = per_cpu_ptr(q->queue_ctx, i);
1807 hctx = q->mq_ops->map_queue(q, i);
1808
1809 cpumask_set_cpu(i, hctx->cpumask);
1810 ctx->index_hw = hctx->nr_ctx;
1811 hctx->ctxs[hctx->nr_ctx++] = ctx;
1812 }
1813
1814 mutex_unlock(&q->sysfs_lock);
1815
1816 queue_for_each_hw_ctx(q, hctx, i) {
1817 struct blk_mq_ctxmap *map = &hctx->ctx_map;
1818
1819 /*
1820 * If no software queues are mapped to this hardware queue,
1821 * disable it and free the request entries.
1822 */
1823 if (!hctx->nr_ctx) {
1824 if (set->tags[i]) {
1825 blk_mq_free_rq_map(set, set->tags[i], i);
1826 set->tags[i] = NULL;
1827 }
1828 hctx->tags = NULL;
1829 continue;
1830 }
1831
1832 /* unmapped hw queue can be remapped after CPU topo changed */
1833 if (!set->tags[i])
1834 set->tags[i] = blk_mq_init_rq_map(set, i);
1835 hctx->tags = set->tags[i];
1836 WARN_ON(!hctx->tags);
1837
1838 cpumask_copy(hctx->tags->cpumask, hctx->cpumask);
1839 /*
1840 * Set the map size to the number of mapped software queues.
1841 * This is more accurate and more efficient than looping
1842 * over all possibly mapped software queues.
1843 */
1844 map->size = DIV_ROUND_UP(hctx->nr_ctx, map->bits_per_word);
1845
1846 /*
1847 * Initialize batch roundrobin counts
1848 */
1849 hctx->next_cpu = cpumask_first(hctx->cpumask);
1850 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1851 }
1852}
1853
1854static void queue_set_hctx_shared(struct request_queue *q, bool shared)
1855{
1856 struct blk_mq_hw_ctx *hctx;
1857 int i;
1858
1859 queue_for_each_hw_ctx(q, hctx, i) {
1860 if (shared)
1861 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1862 else
1863 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1864 }
1865}
1866
1867static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
1868{
1869 struct request_queue *q;
1870
1871 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1872 blk_mq_freeze_queue(q);
1873 queue_set_hctx_shared(q, shared);
1874 blk_mq_unfreeze_queue(q);
1875 }
1876}
1877
1878static void blk_mq_del_queue_tag_set(struct request_queue *q)
1879{
1880 struct blk_mq_tag_set *set = q->tag_set;
1881
1882 mutex_lock(&set->tag_list_lock);
1883 list_del_init(&q->tag_set_list);
1884 if (list_is_singular(&set->tag_list)) {
1885 /* just transitioned to unshared */
1886 set->flags &= ~BLK_MQ_F_TAG_SHARED;
1887 /* update existing queue */
1888 blk_mq_update_tag_set_depth(set, false);
1889 }
1890 mutex_unlock(&set->tag_list_lock);
1891}
1892
1893static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1894 struct request_queue *q)
1895{
1896 q->tag_set = set;
1897
1898 mutex_lock(&set->tag_list_lock);
1899
1900 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
1901 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
1902 set->flags |= BLK_MQ_F_TAG_SHARED;
1903 /* update existing queue */
1904 blk_mq_update_tag_set_depth(set, true);
1905 }
1906 if (set->flags & BLK_MQ_F_TAG_SHARED)
1907 queue_set_hctx_shared(q, true);
1908 list_add_tail(&q->tag_set_list, &set->tag_list);
1909
1910 mutex_unlock(&set->tag_list_lock);
1911}
1912
1913/*
1914 * It is the actual release handler for mq, but we do it from
1915 * request queue's release handler for avoiding use-after-free
1916 * and headache because q->mq_kobj shouldn't have been introduced,
1917 * but we can't group ctx/kctx kobj without it.
1918 */
1919void blk_mq_release(struct request_queue *q)
1920{
1921 struct blk_mq_hw_ctx *hctx;
1922 unsigned int i;
1923
1924 /* hctx kobj stays in hctx */
1925 queue_for_each_hw_ctx(q, hctx, i) {
1926 if (!hctx)
1927 continue;
1928 kfree(hctx->ctxs);
1929 kfree(hctx);
1930 }
1931
1932 kfree(q->mq_map);
1933 q->mq_map = NULL;
1934
1935 kfree(q->queue_hw_ctx);
1936
1937 /* ctx kobj stays in queue_ctx */
1938 free_percpu(q->queue_ctx);
1939}
1940
1941struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1942{
1943 struct request_queue *uninit_q, *q;
1944
1945 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1946 if (!uninit_q)
1947 return ERR_PTR(-ENOMEM);
1948
1949 q = blk_mq_init_allocated_queue(set, uninit_q);
1950 if (IS_ERR(q))
1951 blk_cleanup_queue(uninit_q);
1952
1953 return q;
1954}
1955EXPORT_SYMBOL(blk_mq_init_queue);
1956
1957static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
1958 struct request_queue *q)
1959{
1960 int i, j;
1961 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
1962
1963 blk_mq_sysfs_unregister(q);
1964 for (i = 0; i < set->nr_hw_queues; i++) {
1965 int node;
1966
1967 if (hctxs[i])
1968 continue;
1969
1970 node = blk_mq_hw_queue_to_node(q->mq_map, i);
1971 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1972 GFP_KERNEL, node);
1973 if (!hctxs[i])
1974 break;
1975
1976 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1977 node)) {
1978 kfree(hctxs[i]);
1979 hctxs[i] = NULL;
1980 break;
1981 }
1982
1983 atomic_set(&hctxs[i]->nr_active, 0);
1984 hctxs[i]->numa_node = node;
1985 hctxs[i]->queue_num = i;
1986
1987 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
1988 free_cpumask_var(hctxs[i]->cpumask);
1989 kfree(hctxs[i]);
1990 hctxs[i] = NULL;
1991 break;
1992 }
1993 blk_mq_hctx_kobj_init(hctxs[i]);
1994 }
1995 for (j = i; j < q->nr_hw_queues; j++) {
1996 struct blk_mq_hw_ctx *hctx = hctxs[j];
1997
1998 if (hctx) {
1999 if (hctx->tags) {
2000 blk_mq_free_rq_map(set, hctx->tags, j);
2001 set->tags[j] = NULL;
2002 }
2003 blk_mq_exit_hctx(q, set, hctx, j);
2004 free_cpumask_var(hctx->cpumask);
2005 kobject_put(&hctx->kobj);
2006 kfree(hctx->ctxs);
2007 kfree(hctx);
2008 hctxs[j] = NULL;
2009
2010 }
2011 }
2012 q->nr_hw_queues = i;
2013 blk_mq_sysfs_register(q);
2014}
2015
2016struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2017 struct request_queue *q)
2018{
2019 /* mark the queue as mq asap */
2020 q->mq_ops = set->ops;
2021
2022 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2023 if (!q->queue_ctx)
2024 return ERR_PTR(-ENOMEM);
2025
2026 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2027 GFP_KERNEL, set->numa_node);
2028 if (!q->queue_hw_ctx)
2029 goto err_percpu;
2030
2031 q->mq_map = blk_mq_make_queue_map(set);
2032 if (!q->mq_map)
2033 goto err_map;
2034
2035 blk_mq_realloc_hw_ctxs(set, q);
2036 if (!q->nr_hw_queues)
2037 goto err_hctxs;
2038
2039 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2040 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2041
2042 q->nr_queues = nr_cpu_ids;
2043
2044 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2045
2046 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2047 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2048
2049 q->sg_reserved_size = INT_MAX;
2050
2051 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
2052 INIT_LIST_HEAD(&q->requeue_list);
2053 spin_lock_init(&q->requeue_lock);
2054
2055 if (q->nr_hw_queues > 1)
2056 blk_queue_make_request(q, blk_mq_make_request);
2057 else
2058 blk_queue_make_request(q, blk_sq_make_request);
2059
2060 /*
2061 * Do this after blk_queue_make_request() overrides it...
2062 */
2063 q->nr_requests = set->queue_depth;
2064
2065 if (set->ops->complete)
2066 blk_queue_softirq_done(q, set->ops->complete);
2067
2068 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2069
2070 get_online_cpus();
2071 mutex_lock(&all_q_mutex);
2072
2073 list_add_tail(&q->all_q_node, &all_q_list);
2074 blk_mq_add_queue_tag_set(set, q);
2075 blk_mq_map_swqueue(q, cpu_online_mask);
2076
2077 mutex_unlock(&all_q_mutex);
2078 put_online_cpus();
2079
2080 return q;
2081
2082err_hctxs:
2083 kfree(q->mq_map);
2084err_map:
2085 kfree(q->queue_hw_ctx);
2086err_percpu:
2087 free_percpu(q->queue_ctx);
2088 return ERR_PTR(-ENOMEM);
2089}
2090EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2091
2092void blk_mq_free_queue(struct request_queue *q)
2093{
2094 struct blk_mq_tag_set *set = q->tag_set;
2095
2096 mutex_lock(&all_q_mutex);
2097 list_del_init(&q->all_q_node);
2098 mutex_unlock(&all_q_mutex);
2099
2100 blk_mq_del_queue_tag_set(q);
2101
2102 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2103 blk_mq_free_hw_queues(q, set);
2104}
2105
2106/* Basically redo blk_mq_init_queue with queue frozen */
2107static void blk_mq_queue_reinit(struct request_queue *q,
2108 const struct cpumask *online_mask)
2109{
2110 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2111
2112 blk_mq_sysfs_unregister(q);
2113
2114 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues, online_mask);
2115
2116 /*
2117 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2118 * we should change hctx numa_node according to new topology (this
2119 * involves free and re-allocate memory, worthy doing?)
2120 */
2121
2122 blk_mq_map_swqueue(q, online_mask);
2123
2124 blk_mq_sysfs_register(q);
2125}
2126
2127static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2128 unsigned long action, void *hcpu)
2129{
2130 struct request_queue *q;
2131 int cpu = (unsigned long)hcpu;
2132 /*
2133 * New online cpumask which is going to be set in this hotplug event.
2134 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2135 * one-by-one and dynamically allocating this could result in a failure.
2136 */
2137 static struct cpumask online_new;
2138
2139 /*
2140 * Before hotadded cpu starts handling requests, new mappings must
2141 * be established. Otherwise, these requests in hw queue might
2142 * never be dispatched.
2143 *
2144 * For example, there is a single hw queue (hctx) and two CPU queues
2145 * (ctx0 for CPU0, and ctx1 for CPU1).
2146 *
2147 * Now CPU1 is just onlined and a request is inserted into
2148 * ctx1->rq_list and set bit0 in pending bitmap as ctx1->index_hw is
2149 * still zero.
2150 *
2151 * And then while running hw queue, flush_busy_ctxs() finds bit0 is
2152 * set in pending bitmap and tries to retrieve requests in
2153 * hctx->ctxs[0]->rq_list. But htx->ctxs[0] is a pointer to ctx0,
2154 * so the request in ctx1->rq_list is ignored.
2155 */
2156 switch (action & ~CPU_TASKS_FROZEN) {
2157 case CPU_DEAD:
2158 case CPU_UP_CANCELED:
2159 cpumask_copy(&online_new, cpu_online_mask);
2160 break;
2161 case CPU_UP_PREPARE:
2162 cpumask_copy(&online_new, cpu_online_mask);
2163 cpumask_set_cpu(cpu, &online_new);
2164 break;
2165 default:
2166 return NOTIFY_OK;
2167 }
2168
2169 mutex_lock(&all_q_mutex);
2170
2171 /*
2172 * We need to freeze and reinit all existing queues. Freezing
2173 * involves synchronous wait for an RCU grace period and doing it
2174 * one by one may take a long time. Start freezing all queues in
2175 * one swoop and then wait for the completions so that freezing can
2176 * take place in parallel.
2177 */
2178 list_for_each_entry(q, &all_q_list, all_q_node)
2179 blk_mq_freeze_queue_start(q);
2180 list_for_each_entry(q, &all_q_list, all_q_node) {
2181 blk_mq_freeze_queue_wait(q);
2182
2183 /*
2184 * timeout handler can't touch hw queue during the
2185 * reinitialization
2186 */
2187 del_timer_sync(&q->timeout);
2188 }
2189
2190 list_for_each_entry(q, &all_q_list, all_q_node)
2191 blk_mq_queue_reinit(q, &online_new);
2192
2193 list_for_each_entry(q, &all_q_list, all_q_node)
2194 blk_mq_unfreeze_queue(q);
2195
2196 mutex_unlock(&all_q_mutex);
2197 return NOTIFY_OK;
2198}
2199
2200static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2201{
2202 int i;
2203
2204 for (i = 0; i < set->nr_hw_queues; i++) {
2205 set->tags[i] = blk_mq_init_rq_map(set, i);
2206 if (!set->tags[i])
2207 goto out_unwind;
2208 }
2209
2210 return 0;
2211
2212out_unwind:
2213 while (--i >= 0)
2214 blk_mq_free_rq_map(set, set->tags[i], i);
2215
2216 return -ENOMEM;
2217}
2218
2219/*
2220 * Allocate the request maps associated with this tag_set. Note that this
2221 * may reduce the depth asked for, if memory is tight. set->queue_depth
2222 * will be updated to reflect the allocated depth.
2223 */
2224static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2225{
2226 unsigned int depth;
2227 int err;
2228
2229 depth = set->queue_depth;
2230 do {
2231 err = __blk_mq_alloc_rq_maps(set);
2232 if (!err)
2233 break;
2234
2235 set->queue_depth >>= 1;
2236 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2237 err = -ENOMEM;
2238 break;
2239 }
2240 } while (set->queue_depth);
2241
2242 if (!set->queue_depth || err) {
2243 pr_err("blk-mq: failed to allocate request map\n");
2244 return -ENOMEM;
2245 }
2246
2247 if (depth != set->queue_depth)
2248 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2249 depth, set->queue_depth);
2250
2251 return 0;
2252}
2253
2254struct cpumask *blk_mq_tags_cpumask(struct blk_mq_tags *tags)
2255{
2256 return tags->cpumask;
2257}
2258EXPORT_SYMBOL_GPL(blk_mq_tags_cpumask);
2259
2260/*
2261 * Alloc a tag set to be associated with one or more request queues.
2262 * May fail with EINVAL for various error conditions. May adjust the
2263 * requested depth down, if if it too large. In that case, the set
2264 * value will be stored in set->queue_depth.
2265 */
2266int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2267{
2268 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2269
2270 if (!set->nr_hw_queues)
2271 return -EINVAL;
2272 if (!set->queue_depth)
2273 return -EINVAL;
2274 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2275 return -EINVAL;
2276
2277 if (!set->ops->queue_rq || !set->ops->map_queue)
2278 return -EINVAL;
2279
2280 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2281 pr_info("blk-mq: reduced tag depth to %u\n",
2282 BLK_MQ_MAX_DEPTH);
2283 set->queue_depth = BLK_MQ_MAX_DEPTH;
2284 }
2285
2286 /*
2287 * If a crashdump is active, then we are potentially in a very
2288 * memory constrained environment. Limit us to 1 queue and
2289 * 64 tags to prevent using too much memory.
2290 */
2291 if (is_kdump_kernel()) {
2292 set->nr_hw_queues = 1;
2293 set->queue_depth = min(64U, set->queue_depth);
2294 }
2295 /*
2296 * There is no use for more h/w queues than cpus.
2297 */
2298 if (set->nr_hw_queues > nr_cpu_ids)
2299 set->nr_hw_queues = nr_cpu_ids;
2300
2301 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2302 GFP_KERNEL, set->numa_node);
2303 if (!set->tags)
2304 return -ENOMEM;
2305
2306 if (blk_mq_alloc_rq_maps(set))
2307 goto enomem;
2308
2309 mutex_init(&set->tag_list_lock);
2310 INIT_LIST_HEAD(&set->tag_list);
2311
2312 return 0;
2313enomem:
2314 kfree(set->tags);
2315 set->tags = NULL;
2316 return -ENOMEM;
2317}
2318EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2319
2320void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2321{
2322 int i;
2323
2324 for (i = 0; i < nr_cpu_ids; i++) {
2325 if (set->tags[i])
2326 blk_mq_free_rq_map(set, set->tags[i], i);
2327 }
2328
2329 kfree(set->tags);
2330 set->tags = NULL;
2331}
2332EXPORT_SYMBOL(blk_mq_free_tag_set);
2333
2334int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2335{
2336 struct blk_mq_tag_set *set = q->tag_set;
2337 struct blk_mq_hw_ctx *hctx;
2338 int i, ret;
2339
2340 if (!set || nr > set->queue_depth)
2341 return -EINVAL;
2342
2343 ret = 0;
2344 queue_for_each_hw_ctx(q, hctx, i) {
2345 if (!hctx->tags)
2346 continue;
2347 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2348 if (ret)
2349 break;
2350 }
2351
2352 if (!ret)
2353 q->nr_requests = nr;
2354
2355 return ret;
2356}
2357
2358void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2359{
2360 struct request_queue *q;
2361
2362 if (nr_hw_queues > nr_cpu_ids)
2363 nr_hw_queues = nr_cpu_ids;
2364 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2365 return;
2366
2367 list_for_each_entry(q, &set->tag_list, tag_set_list)
2368 blk_mq_freeze_queue(q);
2369
2370 set->nr_hw_queues = nr_hw_queues;
2371 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2372 blk_mq_realloc_hw_ctxs(set, q);
2373
2374 if (q->nr_hw_queues > 1)
2375 blk_queue_make_request(q, blk_mq_make_request);
2376 else
2377 blk_queue_make_request(q, blk_sq_make_request);
2378
2379 blk_mq_queue_reinit(q, cpu_online_mask);
2380 }
2381
2382 list_for_each_entry(q, &set->tag_list, tag_set_list)
2383 blk_mq_unfreeze_queue(q);
2384}
2385EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2386
2387void blk_mq_disable_hotplug(void)
2388{
2389 mutex_lock(&all_q_mutex);
2390}
2391
2392void blk_mq_enable_hotplug(void)
2393{
2394 mutex_unlock(&all_q_mutex);
2395}
2396
2397static int __init blk_mq_init(void)
2398{
2399 blk_mq_cpu_init();
2400
2401 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2402
2403 return 0;
2404}
2405subsys_initcall(blk_mq_init);