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/sched/topology.h>
24#include <linux/sched/signal.h>
25#include <linux/delay.h>
26#include <linux/crash_dump.h>
27#include <linux/prefetch.h>
28
29#include <trace/events/block.h>
30
31#include <linux/blk-mq.h>
32#include "blk.h"
33#include "blk-mq.h"
34#include "blk-mq-debugfs.h"
35#include "blk-mq-tag.h"
36#include "blk-pm.h"
37#include "blk-stat.h"
38#include "blk-mq-sched.h"
39#include "blk-rq-qos.h"
40
41static void blk_mq_poll_stats_start(struct request_queue *q);
42static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
43
44static int blk_mq_poll_stats_bkt(const struct request *rq)
45{
46 int ddir, bytes, bucket;
47
48 ddir = rq_data_dir(rq);
49 bytes = blk_rq_bytes(rq);
50
51 bucket = ddir + 2*(ilog2(bytes) - 9);
52
53 if (bucket < 0)
54 return -1;
55 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
56 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
57
58 return bucket;
59}
60
61/*
62 * Check if any of the ctx's have pending work in this hardware queue
63 */
64static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
65{
66 return !list_empty_careful(&hctx->dispatch) ||
67 sbitmap_any_bit_set(&hctx->ctx_map) ||
68 blk_mq_sched_has_work(hctx);
69}
70
71/*
72 * Mark this ctx as having pending work in this hardware queue
73 */
74static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
75 struct blk_mq_ctx *ctx)
76{
77 const int bit = ctx->index_hw[hctx->type];
78
79 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
80 sbitmap_set_bit(&hctx->ctx_map, bit);
81}
82
83static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
84 struct blk_mq_ctx *ctx)
85{
86 const int bit = ctx->index_hw[hctx->type];
87
88 sbitmap_clear_bit(&hctx->ctx_map, bit);
89}
90
91struct mq_inflight {
92 struct hd_struct *part;
93 unsigned int *inflight;
94};
95
96static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
97 struct request *rq, void *priv,
98 bool reserved)
99{
100 struct mq_inflight *mi = priv;
101
102 /*
103 * index[0] counts the specific partition that was asked for.
104 */
105 if (rq->part == mi->part)
106 mi->inflight[0]++;
107
108 return true;
109}
110
111unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part)
112{
113 unsigned inflight[2];
114 struct mq_inflight mi = { .part = part, .inflight = inflight, };
115
116 inflight[0] = inflight[1] = 0;
117 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
118
119 return inflight[0];
120}
121
122static bool blk_mq_check_inflight_rw(struct blk_mq_hw_ctx *hctx,
123 struct request *rq, void *priv,
124 bool reserved)
125{
126 struct mq_inflight *mi = priv;
127
128 if (rq->part == mi->part)
129 mi->inflight[rq_data_dir(rq)]++;
130
131 return true;
132}
133
134void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
135 unsigned int inflight[2])
136{
137 struct mq_inflight mi = { .part = part, .inflight = inflight, };
138
139 inflight[0] = inflight[1] = 0;
140 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight_rw, &mi);
141}
142
143void blk_freeze_queue_start(struct request_queue *q)
144{
145 int freeze_depth;
146
147 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
148 if (freeze_depth == 1) {
149 percpu_ref_kill(&q->q_usage_counter);
150 if (queue_is_mq(q))
151 blk_mq_run_hw_queues(q, false);
152 }
153}
154EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
155
156void blk_mq_freeze_queue_wait(struct request_queue *q)
157{
158 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
159}
160EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
161
162int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
163 unsigned long timeout)
164{
165 return wait_event_timeout(q->mq_freeze_wq,
166 percpu_ref_is_zero(&q->q_usage_counter),
167 timeout);
168}
169EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
170
171/*
172 * Guarantee no request is in use, so we can change any data structure of
173 * the queue afterward.
174 */
175void blk_freeze_queue(struct request_queue *q)
176{
177 /*
178 * In the !blk_mq case we are only calling this to kill the
179 * q_usage_counter, otherwise this increases the freeze depth
180 * and waits for it to return to zero. For this reason there is
181 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
182 * exported to drivers as the only user for unfreeze is blk_mq.
183 */
184 blk_freeze_queue_start(q);
185 blk_mq_freeze_queue_wait(q);
186}
187
188void blk_mq_freeze_queue(struct request_queue *q)
189{
190 /*
191 * ...just an alias to keep freeze and unfreeze actions balanced
192 * in the blk_mq_* namespace
193 */
194 blk_freeze_queue(q);
195}
196EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
197
198void blk_mq_unfreeze_queue(struct request_queue *q)
199{
200 int freeze_depth;
201
202 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
203 WARN_ON_ONCE(freeze_depth < 0);
204 if (!freeze_depth) {
205 percpu_ref_resurrect(&q->q_usage_counter);
206 wake_up_all(&q->mq_freeze_wq);
207 }
208}
209EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
210
211/*
212 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
213 * mpt3sas driver such that this function can be removed.
214 */
215void blk_mq_quiesce_queue_nowait(struct request_queue *q)
216{
217 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
218}
219EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
220
221/**
222 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
223 * @q: request queue.
224 *
225 * Note: this function does not prevent that the struct request end_io()
226 * callback function is invoked. Once this function is returned, we make
227 * sure no dispatch can happen until the queue is unquiesced via
228 * blk_mq_unquiesce_queue().
229 */
230void blk_mq_quiesce_queue(struct request_queue *q)
231{
232 struct blk_mq_hw_ctx *hctx;
233 unsigned int i;
234 bool rcu = false;
235
236 blk_mq_quiesce_queue_nowait(q);
237
238 queue_for_each_hw_ctx(q, hctx, i) {
239 if (hctx->flags & BLK_MQ_F_BLOCKING)
240 synchronize_srcu(hctx->srcu);
241 else
242 rcu = true;
243 }
244 if (rcu)
245 synchronize_rcu();
246}
247EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
248
249/*
250 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
251 * @q: request queue.
252 *
253 * This function recovers queue into the state before quiescing
254 * which is done by blk_mq_quiesce_queue.
255 */
256void blk_mq_unquiesce_queue(struct request_queue *q)
257{
258 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
259
260 /* dispatch requests which are inserted during quiescing */
261 blk_mq_run_hw_queues(q, true);
262}
263EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
264
265void blk_mq_wake_waiters(struct request_queue *q)
266{
267 struct blk_mq_hw_ctx *hctx;
268 unsigned int i;
269
270 queue_for_each_hw_ctx(q, hctx, i)
271 if (blk_mq_hw_queue_mapped(hctx))
272 blk_mq_tag_wakeup_all(hctx->tags, true);
273}
274
275bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
276{
277 return blk_mq_has_free_tags(hctx->tags);
278}
279EXPORT_SYMBOL(blk_mq_can_queue);
280
281/*
282 * Only need start/end time stamping if we have stats enabled, or using
283 * an IO scheduler.
284 */
285static inline bool blk_mq_need_time_stamp(struct request *rq)
286{
287 return (rq->rq_flags & RQF_IO_STAT) || rq->q->elevator;
288}
289
290static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
291 unsigned int tag, unsigned int op)
292{
293 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
294 struct request *rq = tags->static_rqs[tag];
295 req_flags_t rq_flags = 0;
296
297 if (data->flags & BLK_MQ_REQ_INTERNAL) {
298 rq->tag = -1;
299 rq->internal_tag = tag;
300 } else {
301 if (data->hctx->flags & BLK_MQ_F_TAG_SHARED) {
302 rq_flags = RQF_MQ_INFLIGHT;
303 atomic_inc(&data->hctx->nr_active);
304 }
305 rq->tag = tag;
306 rq->internal_tag = -1;
307 data->hctx->tags->rqs[rq->tag] = rq;
308 }
309
310 /* csd/requeue_work/fifo_time is initialized before use */
311 rq->q = data->q;
312 rq->mq_ctx = data->ctx;
313 rq->mq_hctx = data->hctx;
314 rq->rq_flags = rq_flags;
315 rq->cmd_flags = op;
316 if (data->flags & BLK_MQ_REQ_PREEMPT)
317 rq->rq_flags |= RQF_PREEMPT;
318 if (blk_queue_io_stat(data->q))
319 rq->rq_flags |= RQF_IO_STAT;
320 INIT_LIST_HEAD(&rq->queuelist);
321 INIT_HLIST_NODE(&rq->hash);
322 RB_CLEAR_NODE(&rq->rb_node);
323 rq->rq_disk = NULL;
324 rq->part = NULL;
325 if (blk_mq_need_time_stamp(rq))
326 rq->start_time_ns = ktime_get_ns();
327 else
328 rq->start_time_ns = 0;
329 rq->io_start_time_ns = 0;
330 rq->nr_phys_segments = 0;
331#if defined(CONFIG_BLK_DEV_INTEGRITY)
332 rq->nr_integrity_segments = 0;
333#endif
334 /* tag was already set */
335 rq->extra_len = 0;
336 WRITE_ONCE(rq->deadline, 0);
337
338 rq->timeout = 0;
339
340 rq->end_io = NULL;
341 rq->end_io_data = NULL;
342
343 data->ctx->rq_dispatched[op_is_sync(op)]++;
344 refcount_set(&rq->ref, 1);
345 return rq;
346}
347
348static struct request *blk_mq_get_request(struct request_queue *q,
349 struct bio *bio,
350 struct blk_mq_alloc_data *data)
351{
352 struct elevator_queue *e = q->elevator;
353 struct request *rq;
354 unsigned int tag;
355 bool put_ctx_on_error = false;
356
357 blk_queue_enter_live(q);
358 data->q = q;
359 if (likely(!data->ctx)) {
360 data->ctx = blk_mq_get_ctx(q);
361 put_ctx_on_error = true;
362 }
363 if (likely(!data->hctx))
364 data->hctx = blk_mq_map_queue(q, data->cmd_flags,
365 data->ctx);
366 if (data->cmd_flags & REQ_NOWAIT)
367 data->flags |= BLK_MQ_REQ_NOWAIT;
368
369 if (e) {
370 data->flags |= BLK_MQ_REQ_INTERNAL;
371
372 /*
373 * Flush requests are special and go directly to the
374 * dispatch list. Don't include reserved tags in the
375 * limiting, as it isn't useful.
376 */
377 if (!op_is_flush(data->cmd_flags) &&
378 e->type->ops.limit_depth &&
379 !(data->flags & BLK_MQ_REQ_RESERVED))
380 e->type->ops.limit_depth(data->cmd_flags, data);
381 } else {
382 blk_mq_tag_busy(data->hctx);
383 }
384
385 tag = blk_mq_get_tag(data);
386 if (tag == BLK_MQ_TAG_FAIL) {
387 if (put_ctx_on_error) {
388 blk_mq_put_ctx(data->ctx);
389 data->ctx = NULL;
390 }
391 blk_queue_exit(q);
392 return NULL;
393 }
394
395 rq = blk_mq_rq_ctx_init(data, tag, data->cmd_flags);
396 if (!op_is_flush(data->cmd_flags)) {
397 rq->elv.icq = NULL;
398 if (e && e->type->ops.prepare_request) {
399 if (e->type->icq_cache)
400 blk_mq_sched_assign_ioc(rq);
401
402 e->type->ops.prepare_request(rq, bio);
403 rq->rq_flags |= RQF_ELVPRIV;
404 }
405 }
406 data->hctx->queued++;
407 return rq;
408}
409
410struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
411 blk_mq_req_flags_t flags)
412{
413 struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
414 struct request *rq;
415 int ret;
416
417 ret = blk_queue_enter(q, flags);
418 if (ret)
419 return ERR_PTR(ret);
420
421 rq = blk_mq_get_request(q, NULL, &alloc_data);
422 blk_queue_exit(q);
423
424 if (!rq)
425 return ERR_PTR(-EWOULDBLOCK);
426
427 blk_mq_put_ctx(alloc_data.ctx);
428
429 rq->__data_len = 0;
430 rq->__sector = (sector_t) -1;
431 rq->bio = rq->biotail = NULL;
432 return rq;
433}
434EXPORT_SYMBOL(blk_mq_alloc_request);
435
436struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
437 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
438{
439 struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
440 struct request *rq;
441 unsigned int cpu;
442 int ret;
443
444 /*
445 * If the tag allocator sleeps we could get an allocation for a
446 * different hardware context. No need to complicate the low level
447 * allocator for this for the rare use case of a command tied to
448 * a specific queue.
449 */
450 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
451 return ERR_PTR(-EINVAL);
452
453 if (hctx_idx >= q->nr_hw_queues)
454 return ERR_PTR(-EIO);
455
456 ret = blk_queue_enter(q, flags);
457 if (ret)
458 return ERR_PTR(ret);
459
460 /*
461 * Check if the hardware context is actually mapped to anything.
462 * If not tell the caller that it should skip this queue.
463 */
464 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
465 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
466 blk_queue_exit(q);
467 return ERR_PTR(-EXDEV);
468 }
469 cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
470 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
471
472 rq = blk_mq_get_request(q, NULL, &alloc_data);
473 blk_queue_exit(q);
474
475 if (!rq)
476 return ERR_PTR(-EWOULDBLOCK);
477
478 return rq;
479}
480EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
481
482static void __blk_mq_free_request(struct request *rq)
483{
484 struct request_queue *q = rq->q;
485 struct blk_mq_ctx *ctx = rq->mq_ctx;
486 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
487 const int sched_tag = rq->internal_tag;
488
489 blk_pm_mark_last_busy(rq);
490 rq->mq_hctx = NULL;
491 if (rq->tag != -1)
492 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
493 if (sched_tag != -1)
494 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
495 blk_mq_sched_restart(hctx);
496 blk_queue_exit(q);
497}
498
499void blk_mq_free_request(struct request *rq)
500{
501 struct request_queue *q = rq->q;
502 struct elevator_queue *e = q->elevator;
503 struct blk_mq_ctx *ctx = rq->mq_ctx;
504 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
505
506 if (rq->rq_flags & RQF_ELVPRIV) {
507 if (e && e->type->ops.finish_request)
508 e->type->ops.finish_request(rq);
509 if (rq->elv.icq) {
510 put_io_context(rq->elv.icq->ioc);
511 rq->elv.icq = NULL;
512 }
513 }
514
515 ctx->rq_completed[rq_is_sync(rq)]++;
516 if (rq->rq_flags & RQF_MQ_INFLIGHT)
517 atomic_dec(&hctx->nr_active);
518
519 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
520 laptop_io_completion(q->backing_dev_info);
521
522 rq_qos_done(q, rq);
523
524 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
525 if (refcount_dec_and_test(&rq->ref))
526 __blk_mq_free_request(rq);
527}
528EXPORT_SYMBOL_GPL(blk_mq_free_request);
529
530inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
531{
532 u64 now = 0;
533
534 if (blk_mq_need_time_stamp(rq))
535 now = ktime_get_ns();
536
537 if (rq->rq_flags & RQF_STATS) {
538 blk_mq_poll_stats_start(rq->q);
539 blk_stat_add(rq, now);
540 }
541
542 if (rq->internal_tag != -1)
543 blk_mq_sched_completed_request(rq, now);
544
545 blk_account_io_done(rq, now);
546
547 if (rq->end_io) {
548 rq_qos_done(rq->q, rq);
549 rq->end_io(rq, error);
550 } else {
551 blk_mq_free_request(rq);
552 }
553}
554EXPORT_SYMBOL(__blk_mq_end_request);
555
556void blk_mq_end_request(struct request *rq, blk_status_t error)
557{
558 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
559 BUG();
560 __blk_mq_end_request(rq, error);
561}
562EXPORT_SYMBOL(blk_mq_end_request);
563
564static void __blk_mq_complete_request_remote(void *data)
565{
566 struct request *rq = data;
567 struct request_queue *q = rq->q;
568
569 q->mq_ops->complete(rq);
570}
571
572static void __blk_mq_complete_request(struct request *rq)
573{
574 struct blk_mq_ctx *ctx = rq->mq_ctx;
575 struct request_queue *q = rq->q;
576 bool shared = false;
577 int cpu;
578
579 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
580 /*
581 * Most of single queue controllers, there is only one irq vector
582 * for handling IO completion, and the only irq's affinity is set
583 * as all possible CPUs. On most of ARCHs, this affinity means the
584 * irq is handled on one specific CPU.
585 *
586 * So complete IO reqeust in softirq context in case of single queue
587 * for not degrading IO performance by irqsoff latency.
588 */
589 if (q->nr_hw_queues == 1) {
590 __blk_complete_request(rq);
591 return;
592 }
593
594 /*
595 * For a polled request, always complete locallly, it's pointless
596 * to redirect the completion.
597 */
598 if ((rq->cmd_flags & REQ_HIPRI) ||
599 !test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags)) {
600 q->mq_ops->complete(rq);
601 return;
602 }
603
604 cpu = get_cpu();
605 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &q->queue_flags))
606 shared = cpus_share_cache(cpu, ctx->cpu);
607
608 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
609 rq->csd.func = __blk_mq_complete_request_remote;
610 rq->csd.info = rq;
611 rq->csd.flags = 0;
612 smp_call_function_single_async(ctx->cpu, &rq->csd);
613 } else {
614 q->mq_ops->complete(rq);
615 }
616 put_cpu();
617}
618
619static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
620 __releases(hctx->srcu)
621{
622 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
623 rcu_read_unlock();
624 else
625 srcu_read_unlock(hctx->srcu, srcu_idx);
626}
627
628static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
629 __acquires(hctx->srcu)
630{
631 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
632 /* shut up gcc false positive */
633 *srcu_idx = 0;
634 rcu_read_lock();
635 } else
636 *srcu_idx = srcu_read_lock(hctx->srcu);
637}
638
639/**
640 * blk_mq_complete_request - end I/O on a request
641 * @rq: the request being processed
642 *
643 * Description:
644 * Ends all I/O on a request. It does not handle partial completions.
645 * The actual completion happens out-of-order, through a IPI handler.
646 **/
647bool blk_mq_complete_request(struct request *rq)
648{
649 if (unlikely(blk_should_fake_timeout(rq->q)))
650 return false;
651 __blk_mq_complete_request(rq);
652 return true;
653}
654EXPORT_SYMBOL(blk_mq_complete_request);
655
656int blk_mq_request_started(struct request *rq)
657{
658 return blk_mq_rq_state(rq) != MQ_RQ_IDLE;
659}
660EXPORT_SYMBOL_GPL(blk_mq_request_started);
661
662void blk_mq_start_request(struct request *rq)
663{
664 struct request_queue *q = rq->q;
665
666 blk_mq_sched_started_request(rq);
667
668 trace_block_rq_issue(q, rq);
669
670 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
671 rq->io_start_time_ns = ktime_get_ns();
672#ifdef CONFIG_BLK_DEV_THROTTLING_LOW
673 rq->throtl_size = blk_rq_sectors(rq);
674#endif
675 rq->rq_flags |= RQF_STATS;
676 rq_qos_issue(q, rq);
677 }
678
679 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
680
681 blk_add_timer(rq);
682 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
683
684 if (q->dma_drain_size && blk_rq_bytes(rq)) {
685 /*
686 * Make sure space for the drain appears. We know we can do
687 * this because max_hw_segments has been adjusted to be one
688 * fewer than the device can handle.
689 */
690 rq->nr_phys_segments++;
691 }
692}
693EXPORT_SYMBOL(blk_mq_start_request);
694
695static void __blk_mq_requeue_request(struct request *rq)
696{
697 struct request_queue *q = rq->q;
698
699 blk_mq_put_driver_tag(rq);
700
701 trace_block_rq_requeue(q, rq);
702 rq_qos_requeue(q, rq);
703
704 if (blk_mq_request_started(rq)) {
705 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
706 rq->rq_flags &= ~RQF_TIMED_OUT;
707 if (q->dma_drain_size && blk_rq_bytes(rq))
708 rq->nr_phys_segments--;
709 }
710}
711
712void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
713{
714 __blk_mq_requeue_request(rq);
715
716 /* this request will be re-inserted to io scheduler queue */
717 blk_mq_sched_requeue_request(rq);
718
719 BUG_ON(!list_empty(&rq->queuelist));
720 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
721}
722EXPORT_SYMBOL(blk_mq_requeue_request);
723
724static void blk_mq_requeue_work(struct work_struct *work)
725{
726 struct request_queue *q =
727 container_of(work, struct request_queue, requeue_work.work);
728 LIST_HEAD(rq_list);
729 struct request *rq, *next;
730
731 spin_lock_irq(&q->requeue_lock);
732 list_splice_init(&q->requeue_list, &rq_list);
733 spin_unlock_irq(&q->requeue_lock);
734
735 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
736 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
737 continue;
738
739 rq->rq_flags &= ~RQF_SOFTBARRIER;
740 list_del_init(&rq->queuelist);
741 /*
742 * If RQF_DONTPREP, rq has contained some driver specific
743 * data, so insert it to hctx dispatch list to avoid any
744 * merge.
745 */
746 if (rq->rq_flags & RQF_DONTPREP)
747 blk_mq_request_bypass_insert(rq, false);
748 else
749 blk_mq_sched_insert_request(rq, true, false, false);
750 }
751
752 while (!list_empty(&rq_list)) {
753 rq = list_entry(rq_list.next, struct request, queuelist);
754 list_del_init(&rq->queuelist);
755 blk_mq_sched_insert_request(rq, false, false, false);
756 }
757
758 blk_mq_run_hw_queues(q, false);
759}
760
761void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
762 bool kick_requeue_list)
763{
764 struct request_queue *q = rq->q;
765 unsigned long flags;
766
767 /*
768 * We abuse this flag that is otherwise used by the I/O scheduler to
769 * request head insertion from the workqueue.
770 */
771 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
772
773 spin_lock_irqsave(&q->requeue_lock, flags);
774 if (at_head) {
775 rq->rq_flags |= RQF_SOFTBARRIER;
776 list_add(&rq->queuelist, &q->requeue_list);
777 } else {
778 list_add_tail(&rq->queuelist, &q->requeue_list);
779 }
780 spin_unlock_irqrestore(&q->requeue_lock, flags);
781
782 if (kick_requeue_list)
783 blk_mq_kick_requeue_list(q);
784}
785
786void blk_mq_kick_requeue_list(struct request_queue *q)
787{
788 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
789}
790EXPORT_SYMBOL(blk_mq_kick_requeue_list);
791
792void blk_mq_delay_kick_requeue_list(struct request_queue *q,
793 unsigned long msecs)
794{
795 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
796 msecs_to_jiffies(msecs));
797}
798EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
799
800struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
801{
802 if (tag < tags->nr_tags) {
803 prefetch(tags->rqs[tag]);
804 return tags->rqs[tag];
805 }
806
807 return NULL;
808}
809EXPORT_SYMBOL(blk_mq_tag_to_rq);
810
811static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
812 void *priv, bool reserved)
813{
814 /*
815 * If we find a request that is inflight and the queue matches,
816 * we know the queue is busy. Return false to stop the iteration.
817 */
818 if (rq->state == MQ_RQ_IN_FLIGHT && rq->q == hctx->queue) {
819 bool *busy = priv;
820
821 *busy = true;
822 return false;
823 }
824
825 return true;
826}
827
828bool blk_mq_queue_inflight(struct request_queue *q)
829{
830 bool busy = false;
831
832 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
833 return busy;
834}
835EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
836
837static void blk_mq_rq_timed_out(struct request *req, bool reserved)
838{
839 req->rq_flags |= RQF_TIMED_OUT;
840 if (req->q->mq_ops->timeout) {
841 enum blk_eh_timer_return ret;
842
843 ret = req->q->mq_ops->timeout(req, reserved);
844 if (ret == BLK_EH_DONE)
845 return;
846 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
847 }
848
849 blk_add_timer(req);
850}
851
852static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
853{
854 unsigned long deadline;
855
856 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
857 return false;
858 if (rq->rq_flags & RQF_TIMED_OUT)
859 return false;
860
861 deadline = READ_ONCE(rq->deadline);
862 if (time_after_eq(jiffies, deadline))
863 return true;
864
865 if (*next == 0)
866 *next = deadline;
867 else if (time_after(*next, deadline))
868 *next = deadline;
869 return false;
870}
871
872static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
873 struct request *rq, void *priv, bool reserved)
874{
875 unsigned long *next = priv;
876
877 /*
878 * Just do a quick check if it is expired before locking the request in
879 * so we're not unnecessarilly synchronizing across CPUs.
880 */
881 if (!blk_mq_req_expired(rq, next))
882 return true;
883
884 /*
885 * We have reason to believe the request may be expired. Take a
886 * reference on the request to lock this request lifetime into its
887 * currently allocated context to prevent it from being reallocated in
888 * the event the completion by-passes this timeout handler.
889 *
890 * If the reference was already released, then the driver beat the
891 * timeout handler to posting a natural completion.
892 */
893 if (!refcount_inc_not_zero(&rq->ref))
894 return true;
895
896 /*
897 * The request is now locked and cannot be reallocated underneath the
898 * timeout handler's processing. Re-verify this exact request is truly
899 * expired; if it is not expired, then the request was completed and
900 * reallocated as a new request.
901 */
902 if (blk_mq_req_expired(rq, next))
903 blk_mq_rq_timed_out(rq, reserved);
904 if (refcount_dec_and_test(&rq->ref))
905 __blk_mq_free_request(rq);
906
907 return true;
908}
909
910static void blk_mq_timeout_work(struct work_struct *work)
911{
912 struct request_queue *q =
913 container_of(work, struct request_queue, timeout_work);
914 unsigned long next = 0;
915 struct blk_mq_hw_ctx *hctx;
916 int i;
917
918 /* A deadlock might occur if a request is stuck requiring a
919 * timeout at the same time a queue freeze is waiting
920 * completion, since the timeout code would not be able to
921 * acquire the queue reference here.
922 *
923 * That's why we don't use blk_queue_enter here; instead, we use
924 * percpu_ref_tryget directly, because we need to be able to
925 * obtain a reference even in the short window between the queue
926 * starting to freeze, by dropping the first reference in
927 * blk_freeze_queue_start, and the moment the last request is
928 * consumed, marked by the instant q_usage_counter reaches
929 * zero.
930 */
931 if (!percpu_ref_tryget(&q->q_usage_counter))
932 return;
933
934 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
935
936 if (next != 0) {
937 mod_timer(&q->timeout, next);
938 } else {
939 /*
940 * Request timeouts are handled as a forward rolling timer. If
941 * we end up here it means that no requests are pending and
942 * also that no request has been pending for a while. Mark
943 * each hctx as idle.
944 */
945 queue_for_each_hw_ctx(q, hctx, i) {
946 /* the hctx may be unmapped, so check it here */
947 if (blk_mq_hw_queue_mapped(hctx))
948 blk_mq_tag_idle(hctx);
949 }
950 }
951 blk_queue_exit(q);
952}
953
954struct flush_busy_ctx_data {
955 struct blk_mq_hw_ctx *hctx;
956 struct list_head *list;
957};
958
959static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
960{
961 struct flush_busy_ctx_data *flush_data = data;
962 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
963 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
964 enum hctx_type type = hctx->type;
965
966 spin_lock(&ctx->lock);
967 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
968 sbitmap_clear_bit(sb, bitnr);
969 spin_unlock(&ctx->lock);
970 return true;
971}
972
973/*
974 * Process software queues that have been marked busy, splicing them
975 * to the for-dispatch
976 */
977void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
978{
979 struct flush_busy_ctx_data data = {
980 .hctx = hctx,
981 .list = list,
982 };
983
984 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
985}
986EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
987
988struct dispatch_rq_data {
989 struct blk_mq_hw_ctx *hctx;
990 struct request *rq;
991};
992
993static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
994 void *data)
995{
996 struct dispatch_rq_data *dispatch_data = data;
997 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
998 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
999 enum hctx_type type = hctx->type;
1000
1001 spin_lock(&ctx->lock);
1002 if (!list_empty(&ctx->rq_lists[type])) {
1003 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1004 list_del_init(&dispatch_data->rq->queuelist);
1005 if (list_empty(&ctx->rq_lists[type]))
1006 sbitmap_clear_bit(sb, bitnr);
1007 }
1008 spin_unlock(&ctx->lock);
1009
1010 return !dispatch_data->rq;
1011}
1012
1013struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1014 struct blk_mq_ctx *start)
1015{
1016 unsigned off = start ? start->index_hw[hctx->type] : 0;
1017 struct dispatch_rq_data data = {
1018 .hctx = hctx,
1019 .rq = NULL,
1020 };
1021
1022 __sbitmap_for_each_set(&hctx->ctx_map, off,
1023 dispatch_rq_from_ctx, &data);
1024
1025 return data.rq;
1026}
1027
1028static inline unsigned int queued_to_index(unsigned int queued)
1029{
1030 if (!queued)
1031 return 0;
1032
1033 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1034}
1035
1036bool blk_mq_get_driver_tag(struct request *rq)
1037{
1038 struct blk_mq_alloc_data data = {
1039 .q = rq->q,
1040 .hctx = rq->mq_hctx,
1041 .flags = BLK_MQ_REQ_NOWAIT,
1042 .cmd_flags = rq->cmd_flags,
1043 };
1044 bool shared;
1045
1046 if (rq->tag != -1)
1047 goto done;
1048
1049 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1050 data.flags |= BLK_MQ_REQ_RESERVED;
1051
1052 shared = blk_mq_tag_busy(data.hctx);
1053 rq->tag = blk_mq_get_tag(&data);
1054 if (rq->tag >= 0) {
1055 if (shared) {
1056 rq->rq_flags |= RQF_MQ_INFLIGHT;
1057 atomic_inc(&data.hctx->nr_active);
1058 }
1059 data.hctx->tags->rqs[rq->tag] = rq;
1060 }
1061
1062done:
1063 return rq->tag != -1;
1064}
1065
1066static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1067 int flags, void *key)
1068{
1069 struct blk_mq_hw_ctx *hctx;
1070
1071 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1072
1073 spin_lock(&hctx->dispatch_wait_lock);
1074 list_del_init(&wait->entry);
1075 spin_unlock(&hctx->dispatch_wait_lock);
1076
1077 blk_mq_run_hw_queue(hctx, true);
1078 return 1;
1079}
1080
1081/*
1082 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1083 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1084 * restart. For both cases, take care to check the condition again after
1085 * marking us as waiting.
1086 */
1087static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1088 struct request *rq)
1089{
1090 struct wait_queue_head *wq;
1091 wait_queue_entry_t *wait;
1092 bool ret;
1093
1094 if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1095 blk_mq_sched_mark_restart_hctx(hctx);
1096
1097 /*
1098 * It's possible that a tag was freed in the window between the
1099 * allocation failure and adding the hardware queue to the wait
1100 * queue.
1101 *
1102 * Don't clear RESTART here, someone else could have set it.
1103 * At most this will cost an extra queue run.
1104 */
1105 return blk_mq_get_driver_tag(rq);
1106 }
1107
1108 wait = &hctx->dispatch_wait;
1109 if (!list_empty_careful(&wait->entry))
1110 return false;
1111
1112 wq = &bt_wait_ptr(&hctx->tags->bitmap_tags, hctx)->wait;
1113
1114 spin_lock_irq(&wq->lock);
1115 spin_lock(&hctx->dispatch_wait_lock);
1116 if (!list_empty(&wait->entry)) {
1117 spin_unlock(&hctx->dispatch_wait_lock);
1118 spin_unlock_irq(&wq->lock);
1119 return false;
1120 }
1121
1122 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1123 __add_wait_queue(wq, wait);
1124
1125 /*
1126 * It's possible that a tag was freed in the window between the
1127 * allocation failure and adding the hardware queue to the wait
1128 * queue.
1129 */
1130 ret = blk_mq_get_driver_tag(rq);
1131 if (!ret) {
1132 spin_unlock(&hctx->dispatch_wait_lock);
1133 spin_unlock_irq(&wq->lock);
1134 return false;
1135 }
1136
1137 /*
1138 * We got a tag, remove ourselves from the wait queue to ensure
1139 * someone else gets the wakeup.
1140 */
1141 list_del_init(&wait->entry);
1142 spin_unlock(&hctx->dispatch_wait_lock);
1143 spin_unlock_irq(&wq->lock);
1144
1145 return true;
1146}
1147
1148#define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1149#define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1150/*
1151 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1152 * - EWMA is one simple way to compute running average value
1153 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1154 * - take 4 as factor for avoiding to get too small(0) result, and this
1155 * factor doesn't matter because EWMA decreases exponentially
1156 */
1157static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1158{
1159 unsigned int ewma;
1160
1161 if (hctx->queue->elevator)
1162 return;
1163
1164 ewma = hctx->dispatch_busy;
1165
1166 if (!ewma && !busy)
1167 return;
1168
1169 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1170 if (busy)
1171 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1172 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1173
1174 hctx->dispatch_busy = ewma;
1175}
1176
1177#define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1178
1179/*
1180 * Returns true if we did some work AND can potentially do more.
1181 */
1182bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1183 bool got_budget)
1184{
1185 struct blk_mq_hw_ctx *hctx;
1186 struct request *rq, *nxt;
1187 bool no_tag = false;
1188 int errors, queued;
1189 blk_status_t ret = BLK_STS_OK;
1190
1191 if (list_empty(list))
1192 return false;
1193
1194 WARN_ON(!list_is_singular(list) && got_budget);
1195
1196 /*
1197 * Now process all the entries, sending them to the driver.
1198 */
1199 errors = queued = 0;
1200 do {
1201 struct blk_mq_queue_data bd;
1202
1203 rq = list_first_entry(list, struct request, queuelist);
1204
1205 hctx = rq->mq_hctx;
1206 if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1207 break;
1208
1209 if (!blk_mq_get_driver_tag(rq)) {
1210 /*
1211 * The initial allocation attempt failed, so we need to
1212 * rerun the hardware queue when a tag is freed. The
1213 * waitqueue takes care of that. If the queue is run
1214 * before we add this entry back on the dispatch list,
1215 * we'll re-run it below.
1216 */
1217 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1218 blk_mq_put_dispatch_budget(hctx);
1219 /*
1220 * For non-shared tags, the RESTART check
1221 * will suffice.
1222 */
1223 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1224 no_tag = true;
1225 break;
1226 }
1227 }
1228
1229 list_del_init(&rq->queuelist);
1230
1231 bd.rq = rq;
1232
1233 /*
1234 * Flag last if we have no more requests, or if we have more
1235 * but can't assign a driver tag to it.
1236 */
1237 if (list_empty(list))
1238 bd.last = true;
1239 else {
1240 nxt = list_first_entry(list, struct request, queuelist);
1241 bd.last = !blk_mq_get_driver_tag(nxt);
1242 }
1243
1244 ret = q->mq_ops->queue_rq(hctx, &bd);
1245 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1246 /*
1247 * If an I/O scheduler has been configured and we got a
1248 * driver tag for the next request already, free it
1249 * again.
1250 */
1251 if (!list_empty(list)) {
1252 nxt = list_first_entry(list, struct request, queuelist);
1253 blk_mq_put_driver_tag(nxt);
1254 }
1255 list_add(&rq->queuelist, list);
1256 __blk_mq_requeue_request(rq);
1257 break;
1258 }
1259
1260 if (unlikely(ret != BLK_STS_OK)) {
1261 errors++;
1262 blk_mq_end_request(rq, BLK_STS_IOERR);
1263 continue;
1264 }
1265
1266 queued++;
1267 } while (!list_empty(list));
1268
1269 hctx->dispatched[queued_to_index(queued)]++;
1270
1271 /*
1272 * Any items that need requeuing? Stuff them into hctx->dispatch,
1273 * that is where we will continue on next queue run.
1274 */
1275 if (!list_empty(list)) {
1276 bool needs_restart;
1277
1278 /*
1279 * If we didn't flush the entire list, we could have told
1280 * the driver there was more coming, but that turned out to
1281 * be a lie.
1282 */
1283 if (q->mq_ops->commit_rqs)
1284 q->mq_ops->commit_rqs(hctx);
1285
1286 spin_lock(&hctx->lock);
1287 list_splice_init(list, &hctx->dispatch);
1288 spin_unlock(&hctx->lock);
1289
1290 /*
1291 * If SCHED_RESTART was set by the caller of this function and
1292 * it is no longer set that means that it was cleared by another
1293 * thread and hence that a queue rerun is needed.
1294 *
1295 * If 'no_tag' is set, that means that we failed getting
1296 * a driver tag with an I/O scheduler attached. If our dispatch
1297 * waitqueue is no longer active, ensure that we run the queue
1298 * AFTER adding our entries back to the list.
1299 *
1300 * If no I/O scheduler has been configured it is possible that
1301 * the hardware queue got stopped and restarted before requests
1302 * were pushed back onto the dispatch list. Rerun the queue to
1303 * avoid starvation. Notes:
1304 * - blk_mq_run_hw_queue() checks whether or not a queue has
1305 * been stopped before rerunning a queue.
1306 * - Some but not all block drivers stop a queue before
1307 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1308 * and dm-rq.
1309 *
1310 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1311 * bit is set, run queue after a delay to avoid IO stalls
1312 * that could otherwise occur if the queue is idle.
1313 */
1314 needs_restart = blk_mq_sched_needs_restart(hctx);
1315 if (!needs_restart ||
1316 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1317 blk_mq_run_hw_queue(hctx, true);
1318 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1319 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1320
1321 blk_mq_update_dispatch_busy(hctx, true);
1322 return false;
1323 } else
1324 blk_mq_update_dispatch_busy(hctx, false);
1325
1326 /*
1327 * If the host/device is unable to accept more work, inform the
1328 * caller of that.
1329 */
1330 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1331 return false;
1332
1333 return (queued + errors) != 0;
1334}
1335
1336static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1337{
1338 int srcu_idx;
1339
1340 /*
1341 * We should be running this queue from one of the CPUs that
1342 * are mapped to it.
1343 *
1344 * There are at least two related races now between setting
1345 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1346 * __blk_mq_run_hw_queue():
1347 *
1348 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1349 * but later it becomes online, then this warning is harmless
1350 * at all
1351 *
1352 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1353 * but later it becomes offline, then the warning can't be
1354 * triggered, and we depend on blk-mq timeout handler to
1355 * handle dispatched requests to this hctx
1356 */
1357 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1358 cpu_online(hctx->next_cpu)) {
1359 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1360 raw_smp_processor_id(),
1361 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1362 dump_stack();
1363 }
1364
1365 /*
1366 * We can't run the queue inline with ints disabled. Ensure that
1367 * we catch bad users of this early.
1368 */
1369 WARN_ON_ONCE(in_interrupt());
1370
1371 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1372
1373 hctx_lock(hctx, &srcu_idx);
1374 blk_mq_sched_dispatch_requests(hctx);
1375 hctx_unlock(hctx, srcu_idx);
1376}
1377
1378static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1379{
1380 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1381
1382 if (cpu >= nr_cpu_ids)
1383 cpu = cpumask_first(hctx->cpumask);
1384 return cpu;
1385}
1386
1387/*
1388 * It'd be great if the workqueue API had a way to pass
1389 * in a mask and had some smarts for more clever placement.
1390 * For now we just round-robin here, switching for every
1391 * BLK_MQ_CPU_WORK_BATCH queued items.
1392 */
1393static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1394{
1395 bool tried = false;
1396 int next_cpu = hctx->next_cpu;
1397
1398 if (hctx->queue->nr_hw_queues == 1)
1399 return WORK_CPU_UNBOUND;
1400
1401 if (--hctx->next_cpu_batch <= 0) {
1402select_cpu:
1403 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1404 cpu_online_mask);
1405 if (next_cpu >= nr_cpu_ids)
1406 next_cpu = blk_mq_first_mapped_cpu(hctx);
1407 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1408 }
1409
1410 /*
1411 * Do unbound schedule if we can't find a online CPU for this hctx,
1412 * and it should only happen in the path of handling CPU DEAD.
1413 */
1414 if (!cpu_online(next_cpu)) {
1415 if (!tried) {
1416 tried = true;
1417 goto select_cpu;
1418 }
1419
1420 /*
1421 * Make sure to re-select CPU next time once after CPUs
1422 * in hctx->cpumask become online again.
1423 */
1424 hctx->next_cpu = next_cpu;
1425 hctx->next_cpu_batch = 1;
1426 return WORK_CPU_UNBOUND;
1427 }
1428
1429 hctx->next_cpu = next_cpu;
1430 return next_cpu;
1431}
1432
1433static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1434 unsigned long msecs)
1435{
1436 if (unlikely(blk_mq_hctx_stopped(hctx)))
1437 return;
1438
1439 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1440 int cpu = get_cpu();
1441 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1442 __blk_mq_run_hw_queue(hctx);
1443 put_cpu();
1444 return;
1445 }
1446
1447 put_cpu();
1448 }
1449
1450 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1451 msecs_to_jiffies(msecs));
1452}
1453
1454void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1455{
1456 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1457}
1458EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1459
1460bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1461{
1462 int srcu_idx;
1463 bool need_run;
1464
1465 /*
1466 * When queue is quiesced, we may be switching io scheduler, or
1467 * updating nr_hw_queues, or other things, and we can't run queue
1468 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1469 *
1470 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1471 * quiesced.
1472 */
1473 hctx_lock(hctx, &srcu_idx);
1474 need_run = !blk_queue_quiesced(hctx->queue) &&
1475 blk_mq_hctx_has_pending(hctx);
1476 hctx_unlock(hctx, srcu_idx);
1477
1478 if (need_run) {
1479 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1480 return true;
1481 }
1482
1483 return false;
1484}
1485EXPORT_SYMBOL(blk_mq_run_hw_queue);
1486
1487void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1488{
1489 struct blk_mq_hw_ctx *hctx;
1490 int i;
1491
1492 queue_for_each_hw_ctx(q, hctx, i) {
1493 if (blk_mq_hctx_stopped(hctx))
1494 continue;
1495
1496 blk_mq_run_hw_queue(hctx, async);
1497 }
1498}
1499EXPORT_SYMBOL(blk_mq_run_hw_queues);
1500
1501/**
1502 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1503 * @q: request queue.
1504 *
1505 * The caller is responsible for serializing this function against
1506 * blk_mq_{start,stop}_hw_queue().
1507 */
1508bool blk_mq_queue_stopped(struct request_queue *q)
1509{
1510 struct blk_mq_hw_ctx *hctx;
1511 int i;
1512
1513 queue_for_each_hw_ctx(q, hctx, i)
1514 if (blk_mq_hctx_stopped(hctx))
1515 return true;
1516
1517 return false;
1518}
1519EXPORT_SYMBOL(blk_mq_queue_stopped);
1520
1521/*
1522 * This function is often used for pausing .queue_rq() by driver when
1523 * there isn't enough resource or some conditions aren't satisfied, and
1524 * BLK_STS_RESOURCE is usually returned.
1525 *
1526 * We do not guarantee that dispatch can be drained or blocked
1527 * after blk_mq_stop_hw_queue() returns. Please use
1528 * blk_mq_quiesce_queue() for that requirement.
1529 */
1530void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1531{
1532 cancel_delayed_work(&hctx->run_work);
1533
1534 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1535}
1536EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1537
1538/*
1539 * This function is often used for pausing .queue_rq() by driver when
1540 * there isn't enough resource or some conditions aren't satisfied, and
1541 * BLK_STS_RESOURCE is usually returned.
1542 *
1543 * We do not guarantee that dispatch can be drained or blocked
1544 * after blk_mq_stop_hw_queues() returns. Please use
1545 * blk_mq_quiesce_queue() for that requirement.
1546 */
1547void blk_mq_stop_hw_queues(struct request_queue *q)
1548{
1549 struct blk_mq_hw_ctx *hctx;
1550 int i;
1551
1552 queue_for_each_hw_ctx(q, hctx, i)
1553 blk_mq_stop_hw_queue(hctx);
1554}
1555EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1556
1557void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1558{
1559 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1560
1561 blk_mq_run_hw_queue(hctx, false);
1562}
1563EXPORT_SYMBOL(blk_mq_start_hw_queue);
1564
1565void blk_mq_start_hw_queues(struct request_queue *q)
1566{
1567 struct blk_mq_hw_ctx *hctx;
1568 int i;
1569
1570 queue_for_each_hw_ctx(q, hctx, i)
1571 blk_mq_start_hw_queue(hctx);
1572}
1573EXPORT_SYMBOL(blk_mq_start_hw_queues);
1574
1575void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1576{
1577 if (!blk_mq_hctx_stopped(hctx))
1578 return;
1579
1580 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1581 blk_mq_run_hw_queue(hctx, async);
1582}
1583EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1584
1585void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1586{
1587 struct blk_mq_hw_ctx *hctx;
1588 int i;
1589
1590 queue_for_each_hw_ctx(q, hctx, i)
1591 blk_mq_start_stopped_hw_queue(hctx, async);
1592}
1593EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1594
1595static void blk_mq_run_work_fn(struct work_struct *work)
1596{
1597 struct blk_mq_hw_ctx *hctx;
1598
1599 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1600
1601 /*
1602 * If we are stopped, don't run the queue.
1603 */
1604 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1605 return;
1606
1607 __blk_mq_run_hw_queue(hctx);
1608}
1609
1610static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1611 struct request *rq,
1612 bool at_head)
1613{
1614 struct blk_mq_ctx *ctx = rq->mq_ctx;
1615 enum hctx_type type = hctx->type;
1616
1617 lockdep_assert_held(&ctx->lock);
1618
1619 trace_block_rq_insert(hctx->queue, rq);
1620
1621 if (at_head)
1622 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1623 else
1624 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1625}
1626
1627void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1628 bool at_head)
1629{
1630 struct blk_mq_ctx *ctx = rq->mq_ctx;
1631
1632 lockdep_assert_held(&ctx->lock);
1633
1634 __blk_mq_insert_req_list(hctx, rq, at_head);
1635 blk_mq_hctx_mark_pending(hctx, ctx);
1636}
1637
1638/*
1639 * Should only be used carefully, when the caller knows we want to
1640 * bypass a potential IO scheduler on the target device.
1641 */
1642void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1643{
1644 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1645
1646 spin_lock(&hctx->lock);
1647 list_add_tail(&rq->queuelist, &hctx->dispatch);
1648 spin_unlock(&hctx->lock);
1649
1650 if (run_queue)
1651 blk_mq_run_hw_queue(hctx, false);
1652}
1653
1654void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1655 struct list_head *list)
1656
1657{
1658 struct request *rq;
1659 enum hctx_type type = hctx->type;
1660
1661 /*
1662 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1663 * offline now
1664 */
1665 list_for_each_entry(rq, list, queuelist) {
1666 BUG_ON(rq->mq_ctx != ctx);
1667 trace_block_rq_insert(hctx->queue, rq);
1668 }
1669
1670 spin_lock(&ctx->lock);
1671 list_splice_tail_init(list, &ctx->rq_lists[type]);
1672 blk_mq_hctx_mark_pending(hctx, ctx);
1673 spin_unlock(&ctx->lock);
1674}
1675
1676static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1677{
1678 struct request *rqa = container_of(a, struct request, queuelist);
1679 struct request *rqb = container_of(b, struct request, queuelist);
1680
1681 if (rqa->mq_ctx < rqb->mq_ctx)
1682 return -1;
1683 else if (rqa->mq_ctx > rqb->mq_ctx)
1684 return 1;
1685 else if (rqa->mq_hctx < rqb->mq_hctx)
1686 return -1;
1687 else if (rqa->mq_hctx > rqb->mq_hctx)
1688 return 1;
1689
1690 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1691}
1692
1693void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1694{
1695 struct blk_mq_hw_ctx *this_hctx;
1696 struct blk_mq_ctx *this_ctx;
1697 struct request_queue *this_q;
1698 struct request *rq;
1699 LIST_HEAD(list);
1700 LIST_HEAD(rq_list);
1701 unsigned int depth;
1702
1703 list_splice_init(&plug->mq_list, &list);
1704 plug->rq_count = 0;
1705
1706 if (plug->rq_count > 2 && plug->multiple_queues)
1707 list_sort(NULL, &list, plug_rq_cmp);
1708
1709 this_q = NULL;
1710 this_hctx = NULL;
1711 this_ctx = NULL;
1712 depth = 0;
1713
1714 while (!list_empty(&list)) {
1715 rq = list_entry_rq(list.next);
1716 list_del_init(&rq->queuelist);
1717 BUG_ON(!rq->q);
1718 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) {
1719 if (this_hctx) {
1720 trace_block_unplug(this_q, depth, !from_schedule);
1721 blk_mq_sched_insert_requests(this_hctx, this_ctx,
1722 &rq_list,
1723 from_schedule);
1724 }
1725
1726 this_q = rq->q;
1727 this_ctx = rq->mq_ctx;
1728 this_hctx = rq->mq_hctx;
1729 depth = 0;
1730 }
1731
1732 depth++;
1733 list_add_tail(&rq->queuelist, &rq_list);
1734 }
1735
1736 /*
1737 * If 'this_hctx' is set, we know we have entries to complete
1738 * on 'rq_list'. Do those.
1739 */
1740 if (this_hctx) {
1741 trace_block_unplug(this_q, depth, !from_schedule);
1742 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1743 from_schedule);
1744 }
1745}
1746
1747static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1748{
1749 blk_init_request_from_bio(rq, bio);
1750
1751 blk_account_io_start(rq, true);
1752}
1753
1754static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1755 struct request *rq,
1756 blk_qc_t *cookie, bool last)
1757{
1758 struct request_queue *q = rq->q;
1759 struct blk_mq_queue_data bd = {
1760 .rq = rq,
1761 .last = last,
1762 };
1763 blk_qc_t new_cookie;
1764 blk_status_t ret;
1765
1766 new_cookie = request_to_qc_t(hctx, rq);
1767
1768 /*
1769 * For OK queue, we are done. For error, caller may kill it.
1770 * Any other error (busy), just add it to our list as we
1771 * previously would have done.
1772 */
1773 ret = q->mq_ops->queue_rq(hctx, &bd);
1774 switch (ret) {
1775 case BLK_STS_OK:
1776 blk_mq_update_dispatch_busy(hctx, false);
1777 *cookie = new_cookie;
1778 break;
1779 case BLK_STS_RESOURCE:
1780 case BLK_STS_DEV_RESOURCE:
1781 blk_mq_update_dispatch_busy(hctx, true);
1782 __blk_mq_requeue_request(rq);
1783 break;
1784 default:
1785 blk_mq_update_dispatch_busy(hctx, false);
1786 *cookie = BLK_QC_T_NONE;
1787 break;
1788 }
1789
1790 return ret;
1791}
1792
1793blk_status_t blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1794 struct request *rq,
1795 blk_qc_t *cookie,
1796 bool bypass, bool last)
1797{
1798 struct request_queue *q = rq->q;
1799 bool run_queue = true;
1800 blk_status_t ret = BLK_STS_RESOURCE;
1801 int srcu_idx;
1802 bool force = false;
1803
1804 hctx_lock(hctx, &srcu_idx);
1805 /*
1806 * hctx_lock is needed before checking quiesced flag.
1807 *
1808 * When queue is stopped or quiesced, ignore 'bypass', insert
1809 * and return BLK_STS_OK to caller, and avoid driver to try to
1810 * dispatch again.
1811 */
1812 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q))) {
1813 run_queue = false;
1814 bypass = false;
1815 goto out_unlock;
1816 }
1817
1818 if (unlikely(q->elevator && !bypass))
1819 goto out_unlock;
1820
1821 if (!blk_mq_get_dispatch_budget(hctx))
1822 goto out_unlock;
1823
1824 if (!blk_mq_get_driver_tag(rq)) {
1825 blk_mq_put_dispatch_budget(hctx);
1826 goto out_unlock;
1827 }
1828
1829 /*
1830 * Always add a request that has been through
1831 *.queue_rq() to the hardware dispatch list.
1832 */
1833 force = true;
1834 ret = __blk_mq_issue_directly(hctx, rq, cookie, last);
1835out_unlock:
1836 hctx_unlock(hctx, srcu_idx);
1837 switch (ret) {
1838 case BLK_STS_OK:
1839 break;
1840 case BLK_STS_DEV_RESOURCE:
1841 case BLK_STS_RESOURCE:
1842 if (force) {
1843 blk_mq_request_bypass_insert(rq, run_queue);
1844 /*
1845 * We have to return BLK_STS_OK for the DM
1846 * to avoid livelock. Otherwise, we return
1847 * the real result to indicate whether the
1848 * request is direct-issued successfully.
1849 */
1850 ret = bypass ? BLK_STS_OK : ret;
1851 } else if (!bypass) {
1852 blk_mq_sched_insert_request(rq, false,
1853 run_queue, false);
1854 }
1855 break;
1856 default:
1857 if (!bypass)
1858 blk_mq_end_request(rq, ret);
1859 break;
1860 }
1861
1862 return ret;
1863}
1864
1865void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
1866 struct list_head *list)
1867{
1868 blk_qc_t unused;
1869 blk_status_t ret = BLK_STS_OK;
1870
1871 while (!list_empty(list)) {
1872 struct request *rq = list_first_entry(list, struct request,
1873 queuelist);
1874
1875 list_del_init(&rq->queuelist);
1876 if (ret == BLK_STS_OK)
1877 ret = blk_mq_try_issue_directly(hctx, rq, &unused,
1878 false,
1879 list_empty(list));
1880 else
1881 blk_mq_sched_insert_request(rq, false, true, false);
1882 }
1883
1884 /*
1885 * If we didn't flush the entire list, we could have told
1886 * the driver there was more coming, but that turned out to
1887 * be a lie.
1888 */
1889 if (ret != BLK_STS_OK && hctx->queue->mq_ops->commit_rqs)
1890 hctx->queue->mq_ops->commit_rqs(hctx);
1891}
1892
1893static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1894{
1895 list_add_tail(&rq->queuelist, &plug->mq_list);
1896 plug->rq_count++;
1897 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
1898 struct request *tmp;
1899
1900 tmp = list_first_entry(&plug->mq_list, struct request,
1901 queuelist);
1902 if (tmp->q != rq->q)
1903 plug->multiple_queues = true;
1904 }
1905}
1906
1907static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1908{
1909 const int is_sync = op_is_sync(bio->bi_opf);
1910 const int is_flush_fua = op_is_flush(bio->bi_opf);
1911 struct blk_mq_alloc_data data = { .flags = 0};
1912 struct request *rq;
1913 struct blk_plug *plug;
1914 struct request *same_queue_rq = NULL;
1915 blk_qc_t cookie;
1916
1917 blk_queue_bounce(q, &bio);
1918
1919 blk_queue_split(q, &bio);
1920
1921 if (!bio_integrity_prep(bio))
1922 return BLK_QC_T_NONE;
1923
1924 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1925 blk_attempt_plug_merge(q, bio, &same_queue_rq))
1926 return BLK_QC_T_NONE;
1927
1928 if (blk_mq_sched_bio_merge(q, bio))
1929 return BLK_QC_T_NONE;
1930
1931 rq_qos_throttle(q, bio);
1932
1933 data.cmd_flags = bio->bi_opf;
1934 rq = blk_mq_get_request(q, bio, &data);
1935 if (unlikely(!rq)) {
1936 rq_qos_cleanup(q, bio);
1937 if (bio->bi_opf & REQ_NOWAIT)
1938 bio_wouldblock_error(bio);
1939 return BLK_QC_T_NONE;
1940 }
1941
1942 trace_block_getrq(q, bio, bio->bi_opf);
1943
1944 rq_qos_track(q, rq, bio);
1945
1946 cookie = request_to_qc_t(data.hctx, rq);
1947
1948 plug = current->plug;
1949 if (unlikely(is_flush_fua)) {
1950 blk_mq_put_ctx(data.ctx);
1951 blk_mq_bio_to_request(rq, bio);
1952
1953 /* bypass scheduler for flush rq */
1954 blk_insert_flush(rq);
1955 blk_mq_run_hw_queue(data.hctx, true);
1956 } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs)) {
1957 /*
1958 * Use plugging if we have a ->commit_rqs() hook as well, as
1959 * we know the driver uses bd->last in a smart fashion.
1960 */
1961 unsigned int request_count = plug->rq_count;
1962 struct request *last = NULL;
1963
1964 blk_mq_put_ctx(data.ctx);
1965 blk_mq_bio_to_request(rq, bio);
1966
1967 if (!request_count)
1968 trace_block_plug(q);
1969 else
1970 last = list_entry_rq(plug->mq_list.prev);
1971
1972 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1973 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1974 blk_flush_plug_list(plug, false);
1975 trace_block_plug(q);
1976 }
1977
1978 blk_add_rq_to_plug(plug, rq);
1979 } else if (plug && !blk_queue_nomerges(q)) {
1980 blk_mq_bio_to_request(rq, bio);
1981
1982 /*
1983 * We do limited plugging. If the bio can be merged, do that.
1984 * Otherwise the existing request in the plug list will be
1985 * issued. So the plug list will have one request at most
1986 * The plug list might get flushed before this. If that happens,
1987 * the plug list is empty, and same_queue_rq is invalid.
1988 */
1989 if (list_empty(&plug->mq_list))
1990 same_queue_rq = NULL;
1991 if (same_queue_rq) {
1992 list_del_init(&same_queue_rq->queuelist);
1993 plug->rq_count--;
1994 }
1995 blk_add_rq_to_plug(plug, rq);
1996
1997 blk_mq_put_ctx(data.ctx);
1998
1999 if (same_queue_rq) {
2000 data.hctx = same_queue_rq->mq_hctx;
2001 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2002 &cookie, false, true);
2003 }
2004 } else if ((q->nr_hw_queues > 1 && is_sync) || (!q->elevator &&
2005 !data.hctx->dispatch_busy)) {
2006 blk_mq_put_ctx(data.ctx);
2007 blk_mq_bio_to_request(rq, bio);
2008 blk_mq_try_issue_directly(data.hctx, rq, &cookie, false, true);
2009 } else {
2010 blk_mq_put_ctx(data.ctx);
2011 blk_mq_bio_to_request(rq, bio);
2012 blk_mq_sched_insert_request(rq, false, true, true);
2013 }
2014
2015 return cookie;
2016}
2017
2018void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2019 unsigned int hctx_idx)
2020{
2021 struct page *page;
2022
2023 if (tags->rqs && set->ops->exit_request) {
2024 int i;
2025
2026 for (i = 0; i < tags->nr_tags; i++) {
2027 struct request *rq = tags->static_rqs[i];
2028
2029 if (!rq)
2030 continue;
2031 set->ops->exit_request(set, rq, hctx_idx);
2032 tags->static_rqs[i] = NULL;
2033 }
2034 }
2035
2036 while (!list_empty(&tags->page_list)) {
2037 page = list_first_entry(&tags->page_list, struct page, lru);
2038 list_del_init(&page->lru);
2039 /*
2040 * Remove kmemleak object previously allocated in
2041 * blk_mq_init_rq_map().
2042 */
2043 kmemleak_free(page_address(page));
2044 __free_pages(page, page->private);
2045 }
2046}
2047
2048void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2049{
2050 kfree(tags->rqs);
2051 tags->rqs = NULL;
2052 kfree(tags->static_rqs);
2053 tags->static_rqs = NULL;
2054
2055 blk_mq_free_tags(tags);
2056}
2057
2058struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2059 unsigned int hctx_idx,
2060 unsigned int nr_tags,
2061 unsigned int reserved_tags)
2062{
2063 struct blk_mq_tags *tags;
2064 int node;
2065
2066 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2067 if (node == NUMA_NO_NODE)
2068 node = set->numa_node;
2069
2070 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2071 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2072 if (!tags)
2073 return NULL;
2074
2075 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2076 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2077 node);
2078 if (!tags->rqs) {
2079 blk_mq_free_tags(tags);
2080 return NULL;
2081 }
2082
2083 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2084 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2085 node);
2086 if (!tags->static_rqs) {
2087 kfree(tags->rqs);
2088 blk_mq_free_tags(tags);
2089 return NULL;
2090 }
2091
2092 return tags;
2093}
2094
2095static size_t order_to_size(unsigned int order)
2096{
2097 return (size_t)PAGE_SIZE << order;
2098}
2099
2100static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2101 unsigned int hctx_idx, int node)
2102{
2103 int ret;
2104
2105 if (set->ops->init_request) {
2106 ret = set->ops->init_request(set, rq, hctx_idx, node);
2107 if (ret)
2108 return ret;
2109 }
2110
2111 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2112 return 0;
2113}
2114
2115int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2116 unsigned int hctx_idx, unsigned int depth)
2117{
2118 unsigned int i, j, entries_per_page, max_order = 4;
2119 size_t rq_size, left;
2120 int node;
2121
2122 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2123 if (node == NUMA_NO_NODE)
2124 node = set->numa_node;
2125
2126 INIT_LIST_HEAD(&tags->page_list);
2127
2128 /*
2129 * rq_size is the size of the request plus driver payload, rounded
2130 * to the cacheline size
2131 */
2132 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2133 cache_line_size());
2134 left = rq_size * depth;
2135
2136 for (i = 0; i < depth; ) {
2137 int this_order = max_order;
2138 struct page *page;
2139 int to_do;
2140 void *p;
2141
2142 while (this_order && left < order_to_size(this_order - 1))
2143 this_order--;
2144
2145 do {
2146 page = alloc_pages_node(node,
2147 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2148 this_order);
2149 if (page)
2150 break;
2151 if (!this_order--)
2152 break;
2153 if (order_to_size(this_order) < rq_size)
2154 break;
2155 } while (1);
2156
2157 if (!page)
2158 goto fail;
2159
2160 page->private = this_order;
2161 list_add_tail(&page->lru, &tags->page_list);
2162
2163 p = page_address(page);
2164 /*
2165 * Allow kmemleak to scan these pages as they contain pointers
2166 * to additional allocations like via ops->init_request().
2167 */
2168 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2169 entries_per_page = order_to_size(this_order) / rq_size;
2170 to_do = min(entries_per_page, depth - i);
2171 left -= to_do * rq_size;
2172 for (j = 0; j < to_do; j++) {
2173 struct request *rq = p;
2174
2175 tags->static_rqs[i] = rq;
2176 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2177 tags->static_rqs[i] = NULL;
2178 goto fail;
2179 }
2180
2181 p += rq_size;
2182 i++;
2183 }
2184 }
2185 return 0;
2186
2187fail:
2188 blk_mq_free_rqs(set, tags, hctx_idx);
2189 return -ENOMEM;
2190}
2191
2192/*
2193 * 'cpu' is going away. splice any existing rq_list entries from this
2194 * software queue to the hw queue dispatch list, and ensure that it
2195 * gets run.
2196 */
2197static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2198{
2199 struct blk_mq_hw_ctx *hctx;
2200 struct blk_mq_ctx *ctx;
2201 LIST_HEAD(tmp);
2202 enum hctx_type type;
2203
2204 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2205 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2206 type = hctx->type;
2207
2208 spin_lock(&ctx->lock);
2209 if (!list_empty(&ctx->rq_lists[type])) {
2210 list_splice_init(&ctx->rq_lists[type], &tmp);
2211 blk_mq_hctx_clear_pending(hctx, ctx);
2212 }
2213 spin_unlock(&ctx->lock);
2214
2215 if (list_empty(&tmp))
2216 return 0;
2217
2218 spin_lock(&hctx->lock);
2219 list_splice_tail_init(&tmp, &hctx->dispatch);
2220 spin_unlock(&hctx->lock);
2221
2222 blk_mq_run_hw_queue(hctx, true);
2223 return 0;
2224}
2225
2226static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2227{
2228 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2229 &hctx->cpuhp_dead);
2230}
2231
2232/* hctx->ctxs will be freed in queue's release handler */
2233static void blk_mq_exit_hctx(struct request_queue *q,
2234 struct blk_mq_tag_set *set,
2235 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2236{
2237 if (blk_mq_hw_queue_mapped(hctx))
2238 blk_mq_tag_idle(hctx);
2239
2240 if (set->ops->exit_request)
2241 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2242
2243 if (set->ops->exit_hctx)
2244 set->ops->exit_hctx(hctx, hctx_idx);
2245
2246 if (hctx->flags & BLK_MQ_F_BLOCKING)
2247 cleanup_srcu_struct(hctx->srcu);
2248
2249 blk_mq_remove_cpuhp(hctx);
2250 blk_free_flush_queue(hctx->fq);
2251 sbitmap_free(&hctx->ctx_map);
2252}
2253
2254static void blk_mq_exit_hw_queues(struct request_queue *q,
2255 struct blk_mq_tag_set *set, int nr_queue)
2256{
2257 struct blk_mq_hw_ctx *hctx;
2258 unsigned int i;
2259
2260 queue_for_each_hw_ctx(q, hctx, i) {
2261 if (i == nr_queue)
2262 break;
2263 blk_mq_debugfs_unregister_hctx(hctx);
2264 blk_mq_exit_hctx(q, set, hctx, i);
2265 }
2266}
2267
2268static int blk_mq_init_hctx(struct request_queue *q,
2269 struct blk_mq_tag_set *set,
2270 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2271{
2272 int node;
2273
2274 node = hctx->numa_node;
2275 if (node == NUMA_NO_NODE)
2276 node = hctx->numa_node = set->numa_node;
2277
2278 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2279 spin_lock_init(&hctx->lock);
2280 INIT_LIST_HEAD(&hctx->dispatch);
2281 hctx->queue = q;
2282 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2283
2284 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2285
2286 hctx->tags = set->tags[hctx_idx];
2287
2288 /*
2289 * Allocate space for all possible cpus to avoid allocation at
2290 * runtime
2291 */
2292 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2293 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, node);
2294 if (!hctx->ctxs)
2295 goto unregister_cpu_notifier;
2296
2297 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2298 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, node))
2299 goto free_ctxs;
2300
2301 hctx->nr_ctx = 0;
2302
2303 spin_lock_init(&hctx->dispatch_wait_lock);
2304 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2305 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2306
2307 if (set->ops->init_hctx &&
2308 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2309 goto free_bitmap;
2310
2311 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size,
2312 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
2313 if (!hctx->fq)
2314 goto exit_hctx;
2315
2316 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node))
2317 goto free_fq;
2318
2319 if (hctx->flags & BLK_MQ_F_BLOCKING)
2320 init_srcu_struct(hctx->srcu);
2321
2322 return 0;
2323
2324 free_fq:
2325 kfree(hctx->fq);
2326 exit_hctx:
2327 if (set->ops->exit_hctx)
2328 set->ops->exit_hctx(hctx, hctx_idx);
2329 free_bitmap:
2330 sbitmap_free(&hctx->ctx_map);
2331 free_ctxs:
2332 kfree(hctx->ctxs);
2333 unregister_cpu_notifier:
2334 blk_mq_remove_cpuhp(hctx);
2335 return -1;
2336}
2337
2338static void blk_mq_init_cpu_queues(struct request_queue *q,
2339 unsigned int nr_hw_queues)
2340{
2341 struct blk_mq_tag_set *set = q->tag_set;
2342 unsigned int i, j;
2343
2344 for_each_possible_cpu(i) {
2345 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2346 struct blk_mq_hw_ctx *hctx;
2347 int k;
2348
2349 __ctx->cpu = i;
2350 spin_lock_init(&__ctx->lock);
2351 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2352 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2353
2354 __ctx->queue = q;
2355
2356 /*
2357 * Set local node, IFF we have more than one hw queue. If
2358 * not, we remain on the home node of the device
2359 */
2360 for (j = 0; j < set->nr_maps; j++) {
2361 hctx = blk_mq_map_queue_type(q, j, i);
2362 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2363 hctx->numa_node = local_memory_node(cpu_to_node(i));
2364 }
2365 }
2366}
2367
2368static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2369{
2370 int ret = 0;
2371
2372 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2373 set->queue_depth, set->reserved_tags);
2374 if (!set->tags[hctx_idx])
2375 return false;
2376
2377 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2378 set->queue_depth);
2379 if (!ret)
2380 return true;
2381
2382 blk_mq_free_rq_map(set->tags[hctx_idx]);
2383 set->tags[hctx_idx] = NULL;
2384 return false;
2385}
2386
2387static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2388 unsigned int hctx_idx)
2389{
2390 if (set->tags && set->tags[hctx_idx]) {
2391 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2392 blk_mq_free_rq_map(set->tags[hctx_idx]);
2393 set->tags[hctx_idx] = NULL;
2394 }
2395}
2396
2397static void blk_mq_map_swqueue(struct request_queue *q)
2398{
2399 unsigned int i, j, hctx_idx;
2400 struct blk_mq_hw_ctx *hctx;
2401 struct blk_mq_ctx *ctx;
2402 struct blk_mq_tag_set *set = q->tag_set;
2403
2404 /*
2405 * Avoid others reading imcomplete hctx->cpumask through sysfs
2406 */
2407 mutex_lock(&q->sysfs_lock);
2408
2409 queue_for_each_hw_ctx(q, hctx, i) {
2410 cpumask_clear(hctx->cpumask);
2411 hctx->nr_ctx = 0;
2412 hctx->dispatch_from = NULL;
2413 }
2414
2415 /*
2416 * Map software to hardware queues.
2417 *
2418 * If the cpu isn't present, the cpu is mapped to first hctx.
2419 */
2420 for_each_possible_cpu(i) {
2421 hctx_idx = set->map[HCTX_TYPE_DEFAULT].mq_map[i];
2422 /* unmapped hw queue can be remapped after CPU topo changed */
2423 if (!set->tags[hctx_idx] &&
2424 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2425 /*
2426 * If tags initialization fail for some hctx,
2427 * that hctx won't be brought online. In this
2428 * case, remap the current ctx to hctx[0] which
2429 * is guaranteed to always have tags allocated
2430 */
2431 set->map[HCTX_TYPE_DEFAULT].mq_map[i] = 0;
2432 }
2433
2434 ctx = per_cpu_ptr(q->queue_ctx, i);
2435 for (j = 0; j < set->nr_maps; j++) {
2436 if (!set->map[j].nr_queues) {
2437 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2438 HCTX_TYPE_DEFAULT, i);
2439 continue;
2440 }
2441
2442 hctx = blk_mq_map_queue_type(q, j, i);
2443 ctx->hctxs[j] = hctx;
2444 /*
2445 * If the CPU is already set in the mask, then we've
2446 * mapped this one already. This can happen if
2447 * devices share queues across queue maps.
2448 */
2449 if (cpumask_test_cpu(i, hctx->cpumask))
2450 continue;
2451
2452 cpumask_set_cpu(i, hctx->cpumask);
2453 hctx->type = j;
2454 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2455 hctx->ctxs[hctx->nr_ctx++] = ctx;
2456
2457 /*
2458 * If the nr_ctx type overflows, we have exceeded the
2459 * amount of sw queues we can support.
2460 */
2461 BUG_ON(!hctx->nr_ctx);
2462 }
2463
2464 for (; j < HCTX_MAX_TYPES; j++)
2465 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2466 HCTX_TYPE_DEFAULT, i);
2467 }
2468
2469 mutex_unlock(&q->sysfs_lock);
2470
2471 queue_for_each_hw_ctx(q, hctx, i) {
2472 /*
2473 * If no software queues are mapped to this hardware queue,
2474 * disable it and free the request entries.
2475 */
2476 if (!hctx->nr_ctx) {
2477 /* Never unmap queue 0. We need it as a
2478 * fallback in case of a new remap fails
2479 * allocation
2480 */
2481 if (i && set->tags[i])
2482 blk_mq_free_map_and_requests(set, i);
2483
2484 hctx->tags = NULL;
2485 continue;
2486 }
2487
2488 hctx->tags = set->tags[i];
2489 WARN_ON(!hctx->tags);
2490
2491 /*
2492 * Set the map size to the number of mapped software queues.
2493 * This is more accurate and more efficient than looping
2494 * over all possibly mapped software queues.
2495 */
2496 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2497
2498 /*
2499 * Initialize batch roundrobin counts
2500 */
2501 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2502 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2503 }
2504}
2505
2506/*
2507 * Caller needs to ensure that we're either frozen/quiesced, or that
2508 * the queue isn't live yet.
2509 */
2510static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2511{
2512 struct blk_mq_hw_ctx *hctx;
2513 int i;
2514
2515 queue_for_each_hw_ctx(q, hctx, i) {
2516 if (shared)
2517 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2518 else
2519 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2520 }
2521}
2522
2523static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2524 bool shared)
2525{
2526 struct request_queue *q;
2527
2528 lockdep_assert_held(&set->tag_list_lock);
2529
2530 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2531 blk_mq_freeze_queue(q);
2532 queue_set_hctx_shared(q, shared);
2533 blk_mq_unfreeze_queue(q);
2534 }
2535}
2536
2537static void blk_mq_del_queue_tag_set(struct request_queue *q)
2538{
2539 struct blk_mq_tag_set *set = q->tag_set;
2540
2541 mutex_lock(&set->tag_list_lock);
2542 list_del_rcu(&q->tag_set_list);
2543 if (list_is_singular(&set->tag_list)) {
2544 /* just transitioned to unshared */
2545 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2546 /* update existing queue */
2547 blk_mq_update_tag_set_depth(set, false);
2548 }
2549 mutex_unlock(&set->tag_list_lock);
2550 INIT_LIST_HEAD(&q->tag_set_list);
2551}
2552
2553static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2554 struct request_queue *q)
2555{
2556 mutex_lock(&set->tag_list_lock);
2557
2558 /*
2559 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2560 */
2561 if (!list_empty(&set->tag_list) &&
2562 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2563 set->flags |= BLK_MQ_F_TAG_SHARED;
2564 /* update existing queue */
2565 blk_mq_update_tag_set_depth(set, true);
2566 }
2567 if (set->flags & BLK_MQ_F_TAG_SHARED)
2568 queue_set_hctx_shared(q, true);
2569 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2570
2571 mutex_unlock(&set->tag_list_lock);
2572}
2573
2574/* All allocations will be freed in release handler of q->mq_kobj */
2575static int blk_mq_alloc_ctxs(struct request_queue *q)
2576{
2577 struct blk_mq_ctxs *ctxs;
2578 int cpu;
2579
2580 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2581 if (!ctxs)
2582 return -ENOMEM;
2583
2584 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2585 if (!ctxs->queue_ctx)
2586 goto fail;
2587
2588 for_each_possible_cpu(cpu) {
2589 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2590 ctx->ctxs = ctxs;
2591 }
2592
2593 q->mq_kobj = &ctxs->kobj;
2594 q->queue_ctx = ctxs->queue_ctx;
2595
2596 return 0;
2597 fail:
2598 kfree(ctxs);
2599 return -ENOMEM;
2600}
2601
2602/*
2603 * It is the actual release handler for mq, but we do it from
2604 * request queue's release handler for avoiding use-after-free
2605 * and headache because q->mq_kobj shouldn't have been introduced,
2606 * but we can't group ctx/kctx kobj without it.
2607 */
2608void blk_mq_release(struct request_queue *q)
2609{
2610 struct blk_mq_hw_ctx *hctx;
2611 unsigned int i;
2612
2613 /* hctx kobj stays in hctx */
2614 queue_for_each_hw_ctx(q, hctx, i) {
2615 if (!hctx)
2616 continue;
2617 kobject_put(&hctx->kobj);
2618 }
2619
2620 kfree(q->queue_hw_ctx);
2621
2622 /*
2623 * release .mq_kobj and sw queue's kobject now because
2624 * both share lifetime with request queue.
2625 */
2626 blk_mq_sysfs_deinit(q);
2627}
2628
2629struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2630{
2631 struct request_queue *uninit_q, *q;
2632
2633 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2634 if (!uninit_q)
2635 return ERR_PTR(-ENOMEM);
2636
2637 q = blk_mq_init_allocated_queue(set, uninit_q);
2638 if (IS_ERR(q))
2639 blk_cleanup_queue(uninit_q);
2640
2641 return q;
2642}
2643EXPORT_SYMBOL(blk_mq_init_queue);
2644
2645/*
2646 * Helper for setting up a queue with mq ops, given queue depth, and
2647 * the passed in mq ops flags.
2648 */
2649struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
2650 const struct blk_mq_ops *ops,
2651 unsigned int queue_depth,
2652 unsigned int set_flags)
2653{
2654 struct request_queue *q;
2655 int ret;
2656
2657 memset(set, 0, sizeof(*set));
2658 set->ops = ops;
2659 set->nr_hw_queues = 1;
2660 set->nr_maps = 1;
2661 set->queue_depth = queue_depth;
2662 set->numa_node = NUMA_NO_NODE;
2663 set->flags = set_flags;
2664
2665 ret = blk_mq_alloc_tag_set(set);
2666 if (ret)
2667 return ERR_PTR(ret);
2668
2669 q = blk_mq_init_queue(set);
2670 if (IS_ERR(q)) {
2671 blk_mq_free_tag_set(set);
2672 return q;
2673 }
2674
2675 return q;
2676}
2677EXPORT_SYMBOL(blk_mq_init_sq_queue);
2678
2679static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2680{
2681 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2682
2683 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2684 __alignof__(struct blk_mq_hw_ctx)) !=
2685 sizeof(struct blk_mq_hw_ctx));
2686
2687 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2688 hw_ctx_size += sizeof(struct srcu_struct);
2689
2690 return hw_ctx_size;
2691}
2692
2693static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
2694 struct blk_mq_tag_set *set, struct request_queue *q,
2695 int hctx_idx, int node)
2696{
2697 struct blk_mq_hw_ctx *hctx;
2698
2699 hctx = kzalloc_node(blk_mq_hw_ctx_size(set),
2700 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2701 node);
2702 if (!hctx)
2703 return NULL;
2704
2705 if (!zalloc_cpumask_var_node(&hctx->cpumask,
2706 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2707 node)) {
2708 kfree(hctx);
2709 return NULL;
2710 }
2711
2712 atomic_set(&hctx->nr_active, 0);
2713 hctx->numa_node = node;
2714 hctx->queue_num = hctx_idx;
2715
2716 if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) {
2717 free_cpumask_var(hctx->cpumask);
2718 kfree(hctx);
2719 return NULL;
2720 }
2721 blk_mq_hctx_kobj_init(hctx);
2722
2723 return hctx;
2724}
2725
2726static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2727 struct request_queue *q)
2728{
2729 int i, j, end;
2730 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2731
2732 /* protect against switching io scheduler */
2733 mutex_lock(&q->sysfs_lock);
2734 for (i = 0; i < set->nr_hw_queues; i++) {
2735 int node;
2736 struct blk_mq_hw_ctx *hctx;
2737
2738 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
2739 /*
2740 * If the hw queue has been mapped to another numa node,
2741 * we need to realloc the hctx. If allocation fails, fallback
2742 * to use the previous one.
2743 */
2744 if (hctxs[i] && (hctxs[i]->numa_node == node))
2745 continue;
2746
2747 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
2748 if (hctx) {
2749 if (hctxs[i]) {
2750 blk_mq_exit_hctx(q, set, hctxs[i], i);
2751 kobject_put(&hctxs[i]->kobj);
2752 }
2753 hctxs[i] = hctx;
2754 } else {
2755 if (hctxs[i])
2756 pr_warn("Allocate new hctx on node %d fails,\
2757 fallback to previous one on node %d\n",
2758 node, hctxs[i]->numa_node);
2759 else
2760 break;
2761 }
2762 }
2763 /*
2764 * Increasing nr_hw_queues fails. Free the newly allocated
2765 * hctxs and keep the previous q->nr_hw_queues.
2766 */
2767 if (i != set->nr_hw_queues) {
2768 j = q->nr_hw_queues;
2769 end = i;
2770 } else {
2771 j = i;
2772 end = q->nr_hw_queues;
2773 q->nr_hw_queues = set->nr_hw_queues;
2774 }
2775
2776 for (; j < end; j++) {
2777 struct blk_mq_hw_ctx *hctx = hctxs[j];
2778
2779 if (hctx) {
2780 if (hctx->tags)
2781 blk_mq_free_map_and_requests(set, j);
2782 blk_mq_exit_hctx(q, set, hctx, j);
2783 kobject_put(&hctx->kobj);
2784 hctxs[j] = NULL;
2785
2786 }
2787 }
2788 mutex_unlock(&q->sysfs_lock);
2789}
2790
2791/*
2792 * Maximum number of hardware queues we support. For single sets, we'll never
2793 * have more than the CPUs (software queues). For multiple sets, the tag_set
2794 * user may have set ->nr_hw_queues larger.
2795 */
2796static unsigned int nr_hw_queues(struct blk_mq_tag_set *set)
2797{
2798 if (set->nr_maps == 1)
2799 return nr_cpu_ids;
2800
2801 return max(set->nr_hw_queues, nr_cpu_ids);
2802}
2803
2804struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2805 struct request_queue *q)
2806{
2807 /* mark the queue as mq asap */
2808 q->mq_ops = set->ops;
2809
2810 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2811 blk_mq_poll_stats_bkt,
2812 BLK_MQ_POLL_STATS_BKTS, q);
2813 if (!q->poll_cb)
2814 goto err_exit;
2815
2816 if (blk_mq_alloc_ctxs(q))
2817 goto err_exit;
2818
2819 /* init q->mq_kobj and sw queues' kobjects */
2820 blk_mq_sysfs_init(q);
2821
2822 q->nr_queues = nr_hw_queues(set);
2823 q->queue_hw_ctx = kcalloc_node(q->nr_queues, sizeof(*(q->queue_hw_ctx)),
2824 GFP_KERNEL, set->numa_node);
2825 if (!q->queue_hw_ctx)
2826 goto err_sys_init;
2827
2828 blk_mq_realloc_hw_ctxs(set, q);
2829 if (!q->nr_hw_queues)
2830 goto err_hctxs;
2831
2832 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2833 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2834
2835 q->tag_set = set;
2836
2837 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2838 if (set->nr_maps > HCTX_TYPE_POLL &&
2839 set->map[HCTX_TYPE_POLL].nr_queues)
2840 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
2841
2842 q->sg_reserved_size = INT_MAX;
2843
2844 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2845 INIT_LIST_HEAD(&q->requeue_list);
2846 spin_lock_init(&q->requeue_lock);
2847
2848 blk_queue_make_request(q, blk_mq_make_request);
2849
2850 /*
2851 * Do this after blk_queue_make_request() overrides it...
2852 */
2853 q->nr_requests = set->queue_depth;
2854
2855 /*
2856 * Default to classic polling
2857 */
2858 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
2859
2860 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2861 blk_mq_add_queue_tag_set(set, q);
2862 blk_mq_map_swqueue(q);
2863
2864 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2865 int ret;
2866
2867 ret = elevator_init_mq(q);
2868 if (ret)
2869 return ERR_PTR(ret);
2870 }
2871
2872 return q;
2873
2874err_hctxs:
2875 kfree(q->queue_hw_ctx);
2876err_sys_init:
2877 blk_mq_sysfs_deinit(q);
2878err_exit:
2879 q->mq_ops = NULL;
2880 return ERR_PTR(-ENOMEM);
2881}
2882EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2883
2884void blk_mq_free_queue(struct request_queue *q)
2885{
2886 struct blk_mq_tag_set *set = q->tag_set;
2887
2888 blk_mq_del_queue_tag_set(q);
2889 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2890}
2891
2892static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2893{
2894 int i;
2895
2896 for (i = 0; i < set->nr_hw_queues; i++)
2897 if (!__blk_mq_alloc_rq_map(set, i))
2898 goto out_unwind;
2899
2900 return 0;
2901
2902out_unwind:
2903 while (--i >= 0)
2904 blk_mq_free_rq_map(set->tags[i]);
2905
2906 return -ENOMEM;
2907}
2908
2909/*
2910 * Allocate the request maps associated with this tag_set. Note that this
2911 * may reduce the depth asked for, if memory is tight. set->queue_depth
2912 * will be updated to reflect the allocated depth.
2913 */
2914static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2915{
2916 unsigned int depth;
2917 int err;
2918
2919 depth = set->queue_depth;
2920 do {
2921 err = __blk_mq_alloc_rq_maps(set);
2922 if (!err)
2923 break;
2924
2925 set->queue_depth >>= 1;
2926 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2927 err = -ENOMEM;
2928 break;
2929 }
2930 } while (set->queue_depth);
2931
2932 if (!set->queue_depth || err) {
2933 pr_err("blk-mq: failed to allocate request map\n");
2934 return -ENOMEM;
2935 }
2936
2937 if (depth != set->queue_depth)
2938 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2939 depth, set->queue_depth);
2940
2941 return 0;
2942}
2943
2944static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2945{
2946 if (set->ops->map_queues && !is_kdump_kernel()) {
2947 int i;
2948
2949 /*
2950 * transport .map_queues is usually done in the following
2951 * way:
2952 *
2953 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2954 * mask = get_cpu_mask(queue)
2955 * for_each_cpu(cpu, mask)
2956 * set->map[x].mq_map[cpu] = queue;
2957 * }
2958 *
2959 * When we need to remap, the table has to be cleared for
2960 * killing stale mapping since one CPU may not be mapped
2961 * to any hw queue.
2962 */
2963 for (i = 0; i < set->nr_maps; i++)
2964 blk_mq_clear_mq_map(&set->map[i]);
2965
2966 return set->ops->map_queues(set);
2967 } else {
2968 BUG_ON(set->nr_maps > 1);
2969 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
2970 }
2971}
2972
2973/*
2974 * Alloc a tag set to be associated with one or more request queues.
2975 * May fail with EINVAL for various error conditions. May adjust the
2976 * requested depth down, if it's too large. In that case, the set
2977 * value will be stored in set->queue_depth.
2978 */
2979int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2980{
2981 int i, ret;
2982
2983 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2984
2985 if (!set->nr_hw_queues)
2986 return -EINVAL;
2987 if (!set->queue_depth)
2988 return -EINVAL;
2989 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2990 return -EINVAL;
2991
2992 if (!set->ops->queue_rq)
2993 return -EINVAL;
2994
2995 if (!set->ops->get_budget ^ !set->ops->put_budget)
2996 return -EINVAL;
2997
2998 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2999 pr_info("blk-mq: reduced tag depth to %u\n",
3000 BLK_MQ_MAX_DEPTH);
3001 set->queue_depth = BLK_MQ_MAX_DEPTH;
3002 }
3003
3004 if (!set->nr_maps)
3005 set->nr_maps = 1;
3006 else if (set->nr_maps > HCTX_MAX_TYPES)
3007 return -EINVAL;
3008
3009 /*
3010 * If a crashdump is active, then we are potentially in a very
3011 * memory constrained environment. Limit us to 1 queue and
3012 * 64 tags to prevent using too much memory.
3013 */
3014 if (is_kdump_kernel()) {
3015 set->nr_hw_queues = 1;
3016 set->nr_maps = 1;
3017 set->queue_depth = min(64U, set->queue_depth);
3018 }
3019 /*
3020 * There is no use for more h/w queues than cpus if we just have
3021 * a single map
3022 */
3023 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3024 set->nr_hw_queues = nr_cpu_ids;
3025
3026 set->tags = kcalloc_node(nr_hw_queues(set), sizeof(struct blk_mq_tags *),
3027 GFP_KERNEL, set->numa_node);
3028 if (!set->tags)
3029 return -ENOMEM;
3030
3031 ret = -ENOMEM;
3032 for (i = 0; i < set->nr_maps; i++) {
3033 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3034 sizeof(set->map[i].mq_map[0]),
3035 GFP_KERNEL, set->numa_node);
3036 if (!set->map[i].mq_map)
3037 goto out_free_mq_map;
3038 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3039 }
3040
3041 ret = blk_mq_update_queue_map(set);
3042 if (ret)
3043 goto out_free_mq_map;
3044
3045 ret = blk_mq_alloc_rq_maps(set);
3046 if (ret)
3047 goto out_free_mq_map;
3048
3049 mutex_init(&set->tag_list_lock);
3050 INIT_LIST_HEAD(&set->tag_list);
3051
3052 return 0;
3053
3054out_free_mq_map:
3055 for (i = 0; i < set->nr_maps; i++) {
3056 kfree(set->map[i].mq_map);
3057 set->map[i].mq_map = NULL;
3058 }
3059 kfree(set->tags);
3060 set->tags = NULL;
3061 return ret;
3062}
3063EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3064
3065void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3066{
3067 int i, j;
3068
3069 for (i = 0; i < nr_hw_queues(set); i++)
3070 blk_mq_free_map_and_requests(set, i);
3071
3072 for (j = 0; j < set->nr_maps; j++) {
3073 kfree(set->map[j].mq_map);
3074 set->map[j].mq_map = NULL;
3075 }
3076
3077 kfree(set->tags);
3078 set->tags = NULL;
3079}
3080EXPORT_SYMBOL(blk_mq_free_tag_set);
3081
3082int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3083{
3084 struct blk_mq_tag_set *set = q->tag_set;
3085 struct blk_mq_hw_ctx *hctx;
3086 int i, ret;
3087
3088 if (!set)
3089 return -EINVAL;
3090
3091 if (q->nr_requests == nr)
3092 return 0;
3093
3094 blk_mq_freeze_queue(q);
3095 blk_mq_quiesce_queue(q);
3096
3097 ret = 0;
3098 queue_for_each_hw_ctx(q, hctx, i) {
3099 if (!hctx->tags)
3100 continue;
3101 /*
3102 * If we're using an MQ scheduler, just update the scheduler
3103 * queue depth. This is similar to what the old code would do.
3104 */
3105 if (!hctx->sched_tags) {
3106 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3107 false);
3108 } else {
3109 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3110 nr, true);
3111 }
3112 if (ret)
3113 break;
3114 }
3115
3116 if (!ret)
3117 q->nr_requests = nr;
3118
3119 blk_mq_unquiesce_queue(q);
3120 blk_mq_unfreeze_queue(q);
3121
3122 return ret;
3123}
3124
3125/*
3126 * request_queue and elevator_type pair.
3127 * It is just used by __blk_mq_update_nr_hw_queues to cache
3128 * the elevator_type associated with a request_queue.
3129 */
3130struct blk_mq_qe_pair {
3131 struct list_head node;
3132 struct request_queue *q;
3133 struct elevator_type *type;
3134};
3135
3136/*
3137 * Cache the elevator_type in qe pair list and switch the
3138 * io scheduler to 'none'
3139 */
3140static bool blk_mq_elv_switch_none(struct list_head *head,
3141 struct request_queue *q)
3142{
3143 struct blk_mq_qe_pair *qe;
3144
3145 if (!q->elevator)
3146 return true;
3147
3148 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3149 if (!qe)
3150 return false;
3151
3152 INIT_LIST_HEAD(&qe->node);
3153 qe->q = q;
3154 qe->type = q->elevator->type;
3155 list_add(&qe->node, head);
3156
3157 mutex_lock(&q->sysfs_lock);
3158 /*
3159 * After elevator_switch_mq, the previous elevator_queue will be
3160 * released by elevator_release. The reference of the io scheduler
3161 * module get by elevator_get will also be put. So we need to get
3162 * a reference of the io scheduler module here to prevent it to be
3163 * removed.
3164 */
3165 __module_get(qe->type->elevator_owner);
3166 elevator_switch_mq(q, NULL);
3167 mutex_unlock(&q->sysfs_lock);
3168
3169 return true;
3170}
3171
3172static void blk_mq_elv_switch_back(struct list_head *head,
3173 struct request_queue *q)
3174{
3175 struct blk_mq_qe_pair *qe;
3176 struct elevator_type *t = NULL;
3177
3178 list_for_each_entry(qe, head, node)
3179 if (qe->q == q) {
3180 t = qe->type;
3181 break;
3182 }
3183
3184 if (!t)
3185 return;
3186
3187 list_del(&qe->node);
3188 kfree(qe);
3189
3190 mutex_lock(&q->sysfs_lock);
3191 elevator_switch_mq(q, t);
3192 mutex_unlock(&q->sysfs_lock);
3193}
3194
3195static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3196 int nr_hw_queues)
3197{
3198 struct request_queue *q;
3199 LIST_HEAD(head);
3200 int prev_nr_hw_queues;
3201
3202 lockdep_assert_held(&set->tag_list_lock);
3203
3204 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3205 nr_hw_queues = nr_cpu_ids;
3206 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
3207 return;
3208
3209 list_for_each_entry(q, &set->tag_list, tag_set_list)
3210 blk_mq_freeze_queue(q);
3211 /*
3212 * Sync with blk_mq_queue_tag_busy_iter.
3213 */
3214 synchronize_rcu();
3215 /*
3216 * Switch IO scheduler to 'none', cleaning up the data associated
3217 * with the previous scheduler. We will switch back once we are done
3218 * updating the new sw to hw queue mappings.
3219 */
3220 list_for_each_entry(q, &set->tag_list, tag_set_list)
3221 if (!blk_mq_elv_switch_none(&head, q))
3222 goto switch_back;
3223
3224 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3225 blk_mq_debugfs_unregister_hctxs(q);
3226 blk_mq_sysfs_unregister(q);
3227 }
3228
3229 prev_nr_hw_queues = set->nr_hw_queues;
3230 set->nr_hw_queues = nr_hw_queues;
3231 blk_mq_update_queue_map(set);
3232fallback:
3233 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3234 blk_mq_realloc_hw_ctxs(set, q);
3235 if (q->nr_hw_queues != set->nr_hw_queues) {
3236 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3237 nr_hw_queues, prev_nr_hw_queues);
3238 set->nr_hw_queues = prev_nr_hw_queues;
3239 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3240 goto fallback;
3241 }
3242 blk_mq_map_swqueue(q);
3243 }
3244
3245 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3246 blk_mq_sysfs_register(q);
3247 blk_mq_debugfs_register_hctxs(q);
3248 }
3249
3250switch_back:
3251