Loading...
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-stat.h"
37#include "blk-wbt.h"
38#include "blk-mq-sched.h"
39
40static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie);
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 if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
78 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
79}
80
81static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
82 struct blk_mq_ctx *ctx)
83{
84 sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
85}
86
87struct mq_inflight {
88 struct hd_struct *part;
89 unsigned int *inflight;
90};
91
92static void blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
93 struct request *rq, void *priv,
94 bool reserved)
95{
96 struct mq_inflight *mi = priv;
97
98 /*
99 * index[0] counts the specific partition that was asked for. index[1]
100 * counts the ones that are active on the whole device, so increment
101 * that if mi->part is indeed a partition, and not a whole device.
102 */
103 if (rq->part == mi->part)
104 mi->inflight[0]++;
105 if (mi->part->partno)
106 mi->inflight[1]++;
107}
108
109void blk_mq_in_flight(struct request_queue *q, struct hd_struct *part,
110 unsigned int inflight[2])
111{
112 struct mq_inflight mi = { .part = part, .inflight = inflight, };
113
114 inflight[0] = inflight[1] = 0;
115 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
116}
117
118static void blk_mq_check_inflight_rw(struct blk_mq_hw_ctx *hctx,
119 struct request *rq, void *priv,
120 bool reserved)
121{
122 struct mq_inflight *mi = priv;
123
124 if (rq->part == mi->part)
125 mi->inflight[rq_data_dir(rq)]++;
126}
127
128void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
129 unsigned int inflight[2])
130{
131 struct mq_inflight mi = { .part = part, .inflight = inflight, };
132
133 inflight[0] = inflight[1] = 0;
134 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight_rw, &mi);
135}
136
137void blk_freeze_queue_start(struct request_queue *q)
138{
139 int freeze_depth;
140
141 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
142 if (freeze_depth == 1) {
143 percpu_ref_kill(&q->q_usage_counter);
144 if (q->mq_ops)
145 blk_mq_run_hw_queues(q, false);
146 }
147}
148EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
149
150void blk_mq_freeze_queue_wait(struct request_queue *q)
151{
152 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
153}
154EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
155
156int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
157 unsigned long timeout)
158{
159 return wait_event_timeout(q->mq_freeze_wq,
160 percpu_ref_is_zero(&q->q_usage_counter),
161 timeout);
162}
163EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
164
165/*
166 * Guarantee no request is in use, so we can change any data structure of
167 * the queue afterward.
168 */
169void blk_freeze_queue(struct request_queue *q)
170{
171 /*
172 * In the !blk_mq case we are only calling this to kill the
173 * q_usage_counter, otherwise this increases the freeze depth
174 * and waits for it to return to zero. For this reason there is
175 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
176 * exported to drivers as the only user for unfreeze is blk_mq.
177 */
178 blk_freeze_queue_start(q);
179 if (!q->mq_ops)
180 blk_drain_queue(q);
181 blk_mq_freeze_queue_wait(q);
182}
183
184void blk_mq_freeze_queue(struct request_queue *q)
185{
186 /*
187 * ...just an alias to keep freeze and unfreeze actions balanced
188 * in the blk_mq_* namespace
189 */
190 blk_freeze_queue(q);
191}
192EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
193
194void blk_mq_unfreeze_queue(struct request_queue *q)
195{
196 int freeze_depth;
197
198 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
199 WARN_ON_ONCE(freeze_depth < 0);
200 if (!freeze_depth) {
201 percpu_ref_reinit(&q->q_usage_counter);
202 wake_up_all(&q->mq_freeze_wq);
203 }
204}
205EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
206
207/*
208 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
209 * mpt3sas driver such that this function can be removed.
210 */
211void blk_mq_quiesce_queue_nowait(struct request_queue *q)
212{
213 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
214}
215EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
216
217/**
218 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
219 * @q: request queue.
220 *
221 * Note: this function does not prevent that the struct request end_io()
222 * callback function is invoked. Once this function is returned, we make
223 * sure no dispatch can happen until the queue is unquiesced via
224 * blk_mq_unquiesce_queue().
225 */
226void blk_mq_quiesce_queue(struct request_queue *q)
227{
228 struct blk_mq_hw_ctx *hctx;
229 unsigned int i;
230 bool rcu = false;
231
232 blk_mq_quiesce_queue_nowait(q);
233
234 queue_for_each_hw_ctx(q, hctx, i) {
235 if (hctx->flags & BLK_MQ_F_BLOCKING)
236 synchronize_srcu(hctx->srcu);
237 else
238 rcu = true;
239 }
240 if (rcu)
241 synchronize_rcu();
242}
243EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
244
245/*
246 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
247 * @q: request queue.
248 *
249 * This function recovers queue into the state before quiescing
250 * which is done by blk_mq_quiesce_queue.
251 */
252void blk_mq_unquiesce_queue(struct request_queue *q)
253{
254 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
255
256 /* dispatch requests which are inserted during quiescing */
257 blk_mq_run_hw_queues(q, true);
258}
259EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
260
261void blk_mq_wake_waiters(struct request_queue *q)
262{
263 struct blk_mq_hw_ctx *hctx;
264 unsigned int i;
265
266 queue_for_each_hw_ctx(q, hctx, i)
267 if (blk_mq_hw_queue_mapped(hctx))
268 blk_mq_tag_wakeup_all(hctx->tags, true);
269}
270
271bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
272{
273 return blk_mq_has_free_tags(hctx->tags);
274}
275EXPORT_SYMBOL(blk_mq_can_queue);
276
277static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
278 unsigned int tag, unsigned int op)
279{
280 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
281 struct request *rq = tags->static_rqs[tag];
282 req_flags_t rq_flags = 0;
283
284 if (data->flags & BLK_MQ_REQ_INTERNAL) {
285 rq->tag = -1;
286 rq->internal_tag = tag;
287 } else {
288 if (blk_mq_tag_busy(data->hctx)) {
289 rq_flags = RQF_MQ_INFLIGHT;
290 atomic_inc(&data->hctx->nr_active);
291 }
292 rq->tag = tag;
293 rq->internal_tag = -1;
294 data->hctx->tags->rqs[rq->tag] = rq;
295 }
296
297 /* csd/requeue_work/fifo_time is initialized before use */
298 rq->q = data->q;
299 rq->mq_ctx = data->ctx;
300 rq->rq_flags = rq_flags;
301 rq->cpu = -1;
302 rq->cmd_flags = op;
303 if (data->flags & BLK_MQ_REQ_PREEMPT)
304 rq->rq_flags |= RQF_PREEMPT;
305 if (blk_queue_io_stat(data->q))
306 rq->rq_flags |= RQF_IO_STAT;
307 INIT_LIST_HEAD(&rq->queuelist);
308 INIT_HLIST_NODE(&rq->hash);
309 RB_CLEAR_NODE(&rq->rb_node);
310 rq->rq_disk = NULL;
311 rq->part = NULL;
312 rq->start_time = jiffies;
313 rq->nr_phys_segments = 0;
314#if defined(CONFIG_BLK_DEV_INTEGRITY)
315 rq->nr_integrity_segments = 0;
316#endif
317 rq->special = NULL;
318 /* tag was already set */
319 rq->extra_len = 0;
320 rq->__deadline = 0;
321
322 INIT_LIST_HEAD(&rq->timeout_list);
323 rq->timeout = 0;
324
325 rq->end_io = NULL;
326 rq->end_io_data = NULL;
327 rq->next_rq = NULL;
328
329#ifdef CONFIG_BLK_CGROUP
330 rq->rl = NULL;
331 set_start_time_ns(rq);
332 rq->io_start_time_ns = 0;
333#endif
334
335 data->ctx->rq_dispatched[op_is_sync(op)]++;
336 return rq;
337}
338
339static struct request *blk_mq_get_request(struct request_queue *q,
340 struct bio *bio, unsigned int op,
341 struct blk_mq_alloc_data *data)
342{
343 struct elevator_queue *e = q->elevator;
344 struct request *rq;
345 unsigned int tag;
346 bool put_ctx_on_error = false;
347
348 blk_queue_enter_live(q);
349 data->q = q;
350 if (likely(!data->ctx)) {
351 data->ctx = blk_mq_get_ctx(q);
352 put_ctx_on_error = true;
353 }
354 if (likely(!data->hctx))
355 data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
356 if (op & REQ_NOWAIT)
357 data->flags |= BLK_MQ_REQ_NOWAIT;
358
359 if (e) {
360 data->flags |= BLK_MQ_REQ_INTERNAL;
361
362 /*
363 * Flush requests are special and go directly to the
364 * dispatch list.
365 */
366 if (!op_is_flush(op) && e->type->ops.mq.limit_depth)
367 e->type->ops.mq.limit_depth(op, data);
368 }
369
370 tag = blk_mq_get_tag(data);
371 if (tag == BLK_MQ_TAG_FAIL) {
372 if (put_ctx_on_error) {
373 blk_mq_put_ctx(data->ctx);
374 data->ctx = NULL;
375 }
376 blk_queue_exit(q);
377 return NULL;
378 }
379
380 rq = blk_mq_rq_ctx_init(data, tag, op);
381 if (!op_is_flush(op)) {
382 rq->elv.icq = NULL;
383 if (e && e->type->ops.mq.prepare_request) {
384 if (e->type->icq_cache && rq_ioc(bio))
385 blk_mq_sched_assign_ioc(rq, bio);
386
387 e->type->ops.mq.prepare_request(rq, bio);
388 rq->rq_flags |= RQF_ELVPRIV;
389 }
390 }
391 data->hctx->queued++;
392 return rq;
393}
394
395struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
396 blk_mq_req_flags_t flags)
397{
398 struct blk_mq_alloc_data alloc_data = { .flags = flags };
399 struct request *rq;
400 int ret;
401
402 ret = blk_queue_enter(q, flags);
403 if (ret)
404 return ERR_PTR(ret);
405
406 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
407 blk_queue_exit(q);
408
409 if (!rq)
410 return ERR_PTR(-EWOULDBLOCK);
411
412 blk_mq_put_ctx(alloc_data.ctx);
413
414 rq->__data_len = 0;
415 rq->__sector = (sector_t) -1;
416 rq->bio = rq->biotail = NULL;
417 return rq;
418}
419EXPORT_SYMBOL(blk_mq_alloc_request);
420
421struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
422 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
423{
424 struct blk_mq_alloc_data alloc_data = { .flags = flags };
425 struct request *rq;
426 unsigned int cpu;
427 int ret;
428
429 /*
430 * If the tag allocator sleeps we could get an allocation for a
431 * different hardware context. No need to complicate the low level
432 * allocator for this for the rare use case of a command tied to
433 * a specific queue.
434 */
435 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
436 return ERR_PTR(-EINVAL);
437
438 if (hctx_idx >= q->nr_hw_queues)
439 return ERR_PTR(-EIO);
440
441 ret = blk_queue_enter(q, flags);
442 if (ret)
443 return ERR_PTR(ret);
444
445 /*
446 * Check if the hardware context is actually mapped to anything.
447 * If not tell the caller that it should skip this queue.
448 */
449 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
450 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
451 blk_queue_exit(q);
452 return ERR_PTR(-EXDEV);
453 }
454 cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
455 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
456
457 rq = blk_mq_get_request(q, NULL, op, &alloc_data);
458 blk_queue_exit(q);
459
460 if (!rq)
461 return ERR_PTR(-EWOULDBLOCK);
462
463 return rq;
464}
465EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
466
467void blk_mq_free_request(struct request *rq)
468{
469 struct request_queue *q = rq->q;
470 struct elevator_queue *e = q->elevator;
471 struct blk_mq_ctx *ctx = rq->mq_ctx;
472 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
473 const int sched_tag = rq->internal_tag;
474
475 if (rq->rq_flags & RQF_ELVPRIV) {
476 if (e && e->type->ops.mq.finish_request)
477 e->type->ops.mq.finish_request(rq);
478 if (rq->elv.icq) {
479 put_io_context(rq->elv.icq->ioc);
480 rq->elv.icq = NULL;
481 }
482 }
483
484 ctx->rq_completed[rq_is_sync(rq)]++;
485 if (rq->rq_flags & RQF_MQ_INFLIGHT)
486 atomic_dec(&hctx->nr_active);
487
488 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
489 laptop_io_completion(q->backing_dev_info);
490
491 wbt_done(q->rq_wb, &rq->issue_stat);
492
493 if (blk_rq_rl(rq))
494 blk_put_rl(blk_rq_rl(rq));
495
496 blk_mq_rq_update_state(rq, MQ_RQ_IDLE);
497 if (rq->tag != -1)
498 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
499 if (sched_tag != -1)
500 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
501 blk_mq_sched_restart(hctx);
502 blk_queue_exit(q);
503}
504EXPORT_SYMBOL_GPL(blk_mq_free_request);
505
506inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
507{
508 blk_account_io_done(rq);
509
510 if (rq->end_io) {
511 wbt_done(rq->q->rq_wb, &rq->issue_stat);
512 rq->end_io(rq, error);
513 } else {
514 if (unlikely(blk_bidi_rq(rq)))
515 blk_mq_free_request(rq->next_rq);
516 blk_mq_free_request(rq);
517 }
518}
519EXPORT_SYMBOL(__blk_mq_end_request);
520
521void blk_mq_end_request(struct request *rq, blk_status_t error)
522{
523 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
524 BUG();
525 __blk_mq_end_request(rq, error);
526}
527EXPORT_SYMBOL(blk_mq_end_request);
528
529static void __blk_mq_complete_request_remote(void *data)
530{
531 struct request *rq = data;
532
533 rq->q->softirq_done_fn(rq);
534}
535
536static void __blk_mq_complete_request(struct request *rq)
537{
538 struct blk_mq_ctx *ctx = rq->mq_ctx;
539 bool shared = false;
540 int cpu;
541
542 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT);
543 blk_mq_rq_update_state(rq, MQ_RQ_COMPLETE);
544
545 if (rq->internal_tag != -1)
546 blk_mq_sched_completed_request(rq);
547 if (rq->rq_flags & RQF_STATS) {
548 blk_mq_poll_stats_start(rq->q);
549 blk_stat_add(rq);
550 }
551
552 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
553 rq->q->softirq_done_fn(rq);
554 return;
555 }
556
557 cpu = get_cpu();
558 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
559 shared = cpus_share_cache(cpu, ctx->cpu);
560
561 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
562 rq->csd.func = __blk_mq_complete_request_remote;
563 rq->csd.info = rq;
564 rq->csd.flags = 0;
565 smp_call_function_single_async(ctx->cpu, &rq->csd);
566 } else {
567 rq->q->softirq_done_fn(rq);
568 }
569 put_cpu();
570}
571
572static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
573 __releases(hctx->srcu)
574{
575 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
576 rcu_read_unlock();
577 else
578 srcu_read_unlock(hctx->srcu, srcu_idx);
579}
580
581static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
582 __acquires(hctx->srcu)
583{
584 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
585 /* shut up gcc false positive */
586 *srcu_idx = 0;
587 rcu_read_lock();
588 } else
589 *srcu_idx = srcu_read_lock(hctx->srcu);
590}
591
592static void blk_mq_rq_update_aborted_gstate(struct request *rq, u64 gstate)
593{
594 unsigned long flags;
595
596 /*
597 * blk_mq_rq_aborted_gstate() is used from the completion path and
598 * can thus be called from irq context. u64_stats_fetch in the
599 * middle of update on the same CPU leads to lockup. Disable irq
600 * while updating.
601 */
602 local_irq_save(flags);
603 u64_stats_update_begin(&rq->aborted_gstate_sync);
604 rq->aborted_gstate = gstate;
605 u64_stats_update_end(&rq->aborted_gstate_sync);
606 local_irq_restore(flags);
607}
608
609static u64 blk_mq_rq_aborted_gstate(struct request *rq)
610{
611 unsigned int start;
612 u64 aborted_gstate;
613
614 do {
615 start = u64_stats_fetch_begin(&rq->aborted_gstate_sync);
616 aborted_gstate = rq->aborted_gstate;
617 } while (u64_stats_fetch_retry(&rq->aborted_gstate_sync, start));
618
619 return aborted_gstate;
620}
621
622/**
623 * blk_mq_complete_request - end I/O on a request
624 * @rq: the request being processed
625 *
626 * Description:
627 * Ends all I/O on a request. It does not handle partial completions.
628 * The actual completion happens out-of-order, through a IPI handler.
629 **/
630void blk_mq_complete_request(struct request *rq)
631{
632 struct request_queue *q = rq->q;
633 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu);
634 int srcu_idx;
635
636 if (unlikely(blk_should_fake_timeout(q)))
637 return;
638
639 /*
640 * If @rq->aborted_gstate equals the current instance, timeout is
641 * claiming @rq and we lost. This is synchronized through
642 * hctx_lock(). See blk_mq_timeout_work() for details.
643 *
644 * Completion path never blocks and we can directly use RCU here
645 * instead of hctx_lock() which can be either RCU or SRCU.
646 * However, that would complicate paths which want to synchronize
647 * against us. Let stay in sync with the issue path so that
648 * hctx_lock() covers both issue and completion paths.
649 */
650 hctx_lock(hctx, &srcu_idx);
651 if (blk_mq_rq_aborted_gstate(rq) != rq->gstate)
652 __blk_mq_complete_request(rq);
653 hctx_unlock(hctx, srcu_idx);
654}
655EXPORT_SYMBOL(blk_mq_complete_request);
656
657int blk_mq_request_started(struct request *rq)
658{
659 return blk_mq_rq_state(rq) != MQ_RQ_IDLE;
660}
661EXPORT_SYMBOL_GPL(blk_mq_request_started);
662
663void blk_mq_start_request(struct request *rq)
664{
665 struct request_queue *q = rq->q;
666
667 blk_mq_sched_started_request(rq);
668
669 trace_block_rq_issue(q, rq);
670
671 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
672 blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq));
673 rq->rq_flags |= RQF_STATS;
674 wbt_issue(q->rq_wb, &rq->issue_stat);
675 }
676
677 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
678
679 /*
680 * Mark @rq in-flight which also advances the generation number,
681 * and register for timeout. Protect with a seqcount to allow the
682 * timeout path to read both @rq->gstate and @rq->deadline
683 * coherently.
684 *
685 * This is the only place where a request is marked in-flight. If
686 * the timeout path reads an in-flight @rq->gstate, the
687 * @rq->deadline it reads together under @rq->gstate_seq is
688 * guaranteed to be the matching one.
689 */
690 preempt_disable();
691 write_seqcount_begin(&rq->gstate_seq);
692
693 blk_mq_rq_update_state(rq, MQ_RQ_IN_FLIGHT);
694 blk_add_timer(rq);
695
696 write_seqcount_end(&rq->gstate_seq);
697 preempt_enable();
698
699 if (q->dma_drain_size && blk_rq_bytes(rq)) {
700 /*
701 * Make sure space for the drain appears. We know we can do
702 * this because max_hw_segments has been adjusted to be one
703 * fewer than the device can handle.
704 */
705 rq->nr_phys_segments++;
706 }
707}
708EXPORT_SYMBOL(blk_mq_start_request);
709
710/*
711 * When we reach here because queue is busy, it's safe to change the state
712 * to IDLE without checking @rq->aborted_gstate because we should still be
713 * holding the RCU read lock and thus protected against timeout.
714 */
715static void __blk_mq_requeue_request(struct request *rq)
716{
717 struct request_queue *q = rq->q;
718
719 blk_mq_put_driver_tag(rq);
720
721 trace_block_rq_requeue(q, rq);
722 wbt_requeue(q->rq_wb, &rq->issue_stat);
723
724 if (blk_mq_rq_state(rq) != MQ_RQ_IDLE) {
725 blk_mq_rq_update_state(rq, MQ_RQ_IDLE);
726 if (q->dma_drain_size && blk_rq_bytes(rq))
727 rq->nr_phys_segments--;
728 }
729}
730
731void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
732{
733 __blk_mq_requeue_request(rq);
734
735 /* this request will be re-inserted to io scheduler queue */
736 blk_mq_sched_requeue_request(rq);
737
738 BUG_ON(blk_queued_rq(rq));
739 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
740}
741EXPORT_SYMBOL(blk_mq_requeue_request);
742
743static void blk_mq_requeue_work(struct work_struct *work)
744{
745 struct request_queue *q =
746 container_of(work, struct request_queue, requeue_work.work);
747 LIST_HEAD(rq_list);
748 struct request *rq, *next;
749
750 spin_lock_irq(&q->requeue_lock);
751 list_splice_init(&q->requeue_list, &rq_list);
752 spin_unlock_irq(&q->requeue_lock);
753
754 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
755 if (!(rq->rq_flags & RQF_SOFTBARRIER))
756 continue;
757
758 rq->rq_flags &= ~RQF_SOFTBARRIER;
759 list_del_init(&rq->queuelist);
760 blk_mq_sched_insert_request(rq, true, false, false);
761 }
762
763 while (!list_empty(&rq_list)) {
764 rq = list_entry(rq_list.next, struct request, queuelist);
765 list_del_init(&rq->queuelist);
766 blk_mq_sched_insert_request(rq, false, false, false);
767 }
768
769 blk_mq_run_hw_queues(q, false);
770}
771
772void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
773 bool kick_requeue_list)
774{
775 struct request_queue *q = rq->q;
776 unsigned long flags;
777
778 /*
779 * We abuse this flag that is otherwise used by the I/O scheduler to
780 * request head insertion from the workqueue.
781 */
782 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
783
784 spin_lock_irqsave(&q->requeue_lock, flags);
785 if (at_head) {
786 rq->rq_flags |= RQF_SOFTBARRIER;
787 list_add(&rq->queuelist, &q->requeue_list);
788 } else {
789 list_add_tail(&rq->queuelist, &q->requeue_list);
790 }
791 spin_unlock_irqrestore(&q->requeue_lock, flags);
792
793 if (kick_requeue_list)
794 blk_mq_kick_requeue_list(q);
795}
796EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
797
798void blk_mq_kick_requeue_list(struct request_queue *q)
799{
800 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
801}
802EXPORT_SYMBOL(blk_mq_kick_requeue_list);
803
804void blk_mq_delay_kick_requeue_list(struct request_queue *q,
805 unsigned long msecs)
806{
807 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
808 msecs_to_jiffies(msecs));
809}
810EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
811
812struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
813{
814 if (tag < tags->nr_tags) {
815 prefetch(tags->rqs[tag]);
816 return tags->rqs[tag];
817 }
818
819 return NULL;
820}
821EXPORT_SYMBOL(blk_mq_tag_to_rq);
822
823struct blk_mq_timeout_data {
824 unsigned long next;
825 unsigned int next_set;
826 unsigned int nr_expired;
827};
828
829static void blk_mq_rq_timed_out(struct request *req, bool reserved)
830{
831 const struct blk_mq_ops *ops = req->q->mq_ops;
832 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
833
834 req->rq_flags |= RQF_MQ_TIMEOUT_EXPIRED;
835
836 if (ops->timeout)
837 ret = ops->timeout(req, reserved);
838
839 switch (ret) {
840 case BLK_EH_HANDLED:
841 __blk_mq_complete_request(req);
842 break;
843 case BLK_EH_RESET_TIMER:
844 /*
845 * As nothing prevents from completion happening while
846 * ->aborted_gstate is set, this may lead to ignored
847 * completions and further spurious timeouts.
848 */
849 blk_mq_rq_update_aborted_gstate(req, 0);
850 blk_add_timer(req);
851 break;
852 case BLK_EH_NOT_HANDLED:
853 break;
854 default:
855 printk(KERN_ERR "block: bad eh return: %d\n", ret);
856 break;
857 }
858}
859
860static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
861 struct request *rq, void *priv, bool reserved)
862{
863 struct blk_mq_timeout_data *data = priv;
864 unsigned long gstate, deadline;
865 int start;
866
867 might_sleep();
868
869 if (rq->rq_flags & RQF_MQ_TIMEOUT_EXPIRED)
870 return;
871
872 /* read coherent snapshots of @rq->state_gen and @rq->deadline */
873 while (true) {
874 start = read_seqcount_begin(&rq->gstate_seq);
875 gstate = READ_ONCE(rq->gstate);
876 deadline = blk_rq_deadline(rq);
877 if (!read_seqcount_retry(&rq->gstate_seq, start))
878 break;
879 cond_resched();
880 }
881
882 /* if in-flight && overdue, mark for abortion */
883 if ((gstate & MQ_RQ_STATE_MASK) == MQ_RQ_IN_FLIGHT &&
884 time_after_eq(jiffies, deadline)) {
885 blk_mq_rq_update_aborted_gstate(rq, gstate);
886 data->nr_expired++;
887 hctx->nr_expired++;
888 } else if (!data->next_set || time_after(data->next, deadline)) {
889 data->next = deadline;
890 data->next_set = 1;
891 }
892}
893
894static void blk_mq_terminate_expired(struct blk_mq_hw_ctx *hctx,
895 struct request *rq, void *priv, bool reserved)
896{
897 /*
898 * We marked @rq->aborted_gstate and waited for RCU. If there were
899 * completions that we lost to, they would have finished and
900 * updated @rq->gstate by now; otherwise, the completion path is
901 * now guaranteed to see @rq->aborted_gstate and yield. If
902 * @rq->aborted_gstate still matches @rq->gstate, @rq is ours.
903 */
904 if (!(rq->rq_flags & RQF_MQ_TIMEOUT_EXPIRED) &&
905 READ_ONCE(rq->gstate) == rq->aborted_gstate)
906 blk_mq_rq_timed_out(rq, reserved);
907}
908
909static void blk_mq_timeout_work(struct work_struct *work)
910{
911 struct request_queue *q =
912 container_of(work, struct request_queue, timeout_work);
913 struct blk_mq_timeout_data data = {
914 .next = 0,
915 .next_set = 0,
916 .nr_expired = 0,
917 };
918 struct blk_mq_hw_ctx *hctx;
919 int i;
920
921 /* A deadlock might occur if a request is stuck requiring a
922 * timeout at the same time a queue freeze is waiting
923 * completion, since the timeout code would not be able to
924 * acquire the queue reference here.
925 *
926 * That's why we don't use blk_queue_enter here; instead, we use
927 * percpu_ref_tryget directly, because we need to be able to
928 * obtain a reference even in the short window between the queue
929 * starting to freeze, by dropping the first reference in
930 * blk_freeze_queue_start, and the moment the last request is
931 * consumed, marked by the instant q_usage_counter reaches
932 * zero.
933 */
934 if (!percpu_ref_tryget(&q->q_usage_counter))
935 return;
936
937 /* scan for the expired ones and set their ->aborted_gstate */
938 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
939
940 if (data.nr_expired) {
941 bool has_rcu = false;
942
943 /*
944 * Wait till everyone sees ->aborted_gstate. The
945 * sequential waits for SRCUs aren't ideal. If this ever
946 * becomes a problem, we can add per-hw_ctx rcu_head and
947 * wait in parallel.
948 */
949 queue_for_each_hw_ctx(q, hctx, i) {
950 if (!hctx->nr_expired)
951 continue;
952
953 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
954 has_rcu = true;
955 else
956 synchronize_srcu(hctx->srcu);
957
958 hctx->nr_expired = 0;
959 }
960 if (has_rcu)
961 synchronize_rcu();
962
963 /* terminate the ones we won */
964 blk_mq_queue_tag_busy_iter(q, blk_mq_terminate_expired, NULL);
965 }
966
967 if (data.next_set) {
968 data.next = blk_rq_timeout(round_jiffies_up(data.next));
969 mod_timer(&q->timeout, data.next);
970 } else {
971 /*
972 * Request timeouts are handled as a forward rolling timer. If
973 * we end up here it means that no requests are pending and
974 * also that no request has been pending for a while. Mark
975 * each hctx as idle.
976 */
977 queue_for_each_hw_ctx(q, hctx, i) {
978 /* the hctx may be unmapped, so check it here */
979 if (blk_mq_hw_queue_mapped(hctx))
980 blk_mq_tag_idle(hctx);
981 }
982 }
983 blk_queue_exit(q);
984}
985
986struct flush_busy_ctx_data {
987 struct blk_mq_hw_ctx *hctx;
988 struct list_head *list;
989};
990
991static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
992{
993 struct flush_busy_ctx_data *flush_data = data;
994 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
995 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
996
997 spin_lock(&ctx->lock);
998 list_splice_tail_init(&ctx->rq_list, flush_data->list);
999 sbitmap_clear_bit(sb, bitnr);
1000 spin_unlock(&ctx->lock);
1001 return true;
1002}
1003
1004/*
1005 * Process software queues that have been marked busy, splicing them
1006 * to the for-dispatch
1007 */
1008void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1009{
1010 struct flush_busy_ctx_data data = {
1011 .hctx = hctx,
1012 .list = list,
1013 };
1014
1015 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1016}
1017EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1018
1019struct dispatch_rq_data {
1020 struct blk_mq_hw_ctx *hctx;
1021 struct request *rq;
1022};
1023
1024static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1025 void *data)
1026{
1027 struct dispatch_rq_data *dispatch_data = data;
1028 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1029 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1030
1031 spin_lock(&ctx->lock);
1032 if (unlikely(!list_empty(&ctx->rq_list))) {
1033 dispatch_data->rq = list_entry_rq(ctx->rq_list.next);
1034 list_del_init(&dispatch_data->rq->queuelist);
1035 if (list_empty(&ctx->rq_list))
1036 sbitmap_clear_bit(sb, bitnr);
1037 }
1038 spin_unlock(&ctx->lock);
1039
1040 return !dispatch_data->rq;
1041}
1042
1043struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1044 struct blk_mq_ctx *start)
1045{
1046 unsigned off = start ? start->index_hw : 0;
1047 struct dispatch_rq_data data = {
1048 .hctx = hctx,
1049 .rq = NULL,
1050 };
1051
1052 __sbitmap_for_each_set(&hctx->ctx_map, off,
1053 dispatch_rq_from_ctx, &data);
1054
1055 return data.rq;
1056}
1057
1058static inline unsigned int queued_to_index(unsigned int queued)
1059{
1060 if (!queued)
1061 return 0;
1062
1063 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1064}
1065
1066bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
1067 bool wait)
1068{
1069 struct blk_mq_alloc_data data = {
1070 .q = rq->q,
1071 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
1072 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
1073 };
1074
1075 might_sleep_if(wait);
1076
1077 if (rq->tag != -1)
1078 goto done;
1079
1080 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1081 data.flags |= BLK_MQ_REQ_RESERVED;
1082
1083 rq->tag = blk_mq_get_tag(&data);
1084 if (rq->tag >= 0) {
1085 if (blk_mq_tag_busy(data.hctx)) {
1086 rq->rq_flags |= RQF_MQ_INFLIGHT;
1087 atomic_inc(&data.hctx->nr_active);
1088 }
1089 data.hctx->tags->rqs[rq->tag] = rq;
1090 }
1091
1092done:
1093 if (hctx)
1094 *hctx = data.hctx;
1095 return rq->tag != -1;
1096}
1097
1098static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1099 int flags, void *key)
1100{
1101 struct blk_mq_hw_ctx *hctx;
1102
1103 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1104
1105 list_del_init(&wait->entry);
1106 blk_mq_run_hw_queue(hctx, true);
1107 return 1;
1108}
1109
1110/*
1111 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1112 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1113 * restart. For both cases, take care to check the condition again after
1114 * marking us as waiting.
1115 */
1116static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx **hctx,
1117 struct request *rq)
1118{
1119 struct blk_mq_hw_ctx *this_hctx = *hctx;
1120 struct sbq_wait_state *ws;
1121 wait_queue_entry_t *wait;
1122 bool ret;
1123
1124 if (!(this_hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1125 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state))
1126 set_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state);
1127
1128 /*
1129 * It's possible that a tag was freed in the window between the
1130 * allocation failure and adding the hardware queue to the wait
1131 * queue.
1132 *
1133 * Don't clear RESTART here, someone else could have set it.
1134 * At most this will cost an extra queue run.
1135 */
1136 return blk_mq_get_driver_tag(rq, hctx, false);
1137 }
1138
1139 wait = &this_hctx->dispatch_wait;
1140 if (!list_empty_careful(&wait->entry))
1141 return false;
1142
1143 spin_lock(&this_hctx->lock);
1144 if (!list_empty(&wait->entry)) {
1145 spin_unlock(&this_hctx->lock);
1146 return false;
1147 }
1148
1149 ws = bt_wait_ptr(&this_hctx->tags->bitmap_tags, this_hctx);
1150 add_wait_queue(&ws->wait, wait);
1151
1152 /*
1153 * It's possible that a tag was freed in the window between the
1154 * allocation failure and adding the hardware queue to the wait
1155 * queue.
1156 */
1157 ret = blk_mq_get_driver_tag(rq, hctx, false);
1158 if (!ret) {
1159 spin_unlock(&this_hctx->lock);
1160 return false;
1161 }
1162
1163 /*
1164 * We got a tag, remove ourselves from the wait queue to ensure
1165 * someone else gets the wakeup.
1166 */
1167 spin_lock_irq(&ws->wait.lock);
1168 list_del_init(&wait->entry);
1169 spin_unlock_irq(&ws->wait.lock);
1170 spin_unlock(&this_hctx->lock);
1171
1172 return true;
1173}
1174
1175#define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1176
1177bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1178 bool got_budget)
1179{
1180 struct blk_mq_hw_ctx *hctx;
1181 struct request *rq, *nxt;
1182 bool no_tag = false;
1183 int errors, queued;
1184 blk_status_t ret = BLK_STS_OK;
1185
1186 if (list_empty(list))
1187 return false;
1188
1189 WARN_ON(!list_is_singular(list) && got_budget);
1190
1191 /*
1192 * Now process all the entries, sending them to the driver.
1193 */
1194 errors = queued = 0;
1195 do {
1196 struct blk_mq_queue_data bd;
1197
1198 rq = list_first_entry(list, struct request, queuelist);
1199
1200 hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
1201 if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1202 break;
1203
1204 if (!blk_mq_get_driver_tag(rq, NULL, false)) {
1205 /*
1206 * The initial allocation attempt failed, so we need to
1207 * rerun the hardware queue when a tag is freed. The
1208 * waitqueue takes care of that. If the queue is run
1209 * before we add this entry back on the dispatch list,
1210 * we'll re-run it below.
1211 */
1212 if (!blk_mq_mark_tag_wait(&hctx, rq)) {
1213 blk_mq_put_dispatch_budget(hctx);
1214 /*
1215 * For non-shared tags, the RESTART check
1216 * will suffice.
1217 */
1218 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1219 no_tag = true;
1220 break;
1221 }
1222 }
1223
1224 list_del_init(&rq->queuelist);
1225
1226 bd.rq = rq;
1227
1228 /*
1229 * Flag last if we have no more requests, or if we have more
1230 * but can't assign a driver tag to it.
1231 */
1232 if (list_empty(list))
1233 bd.last = true;
1234 else {
1235 nxt = list_first_entry(list, struct request, queuelist);
1236 bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1237 }
1238
1239 ret = q->mq_ops->queue_rq(hctx, &bd);
1240 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1241 /*
1242 * If an I/O scheduler has been configured and we got a
1243 * driver tag for the next request already, free it
1244 * again.
1245 */
1246 if (!list_empty(list)) {
1247 nxt = list_first_entry(list, struct request, queuelist);
1248 blk_mq_put_driver_tag(nxt);
1249 }
1250 list_add(&rq->queuelist, list);
1251 __blk_mq_requeue_request(rq);
1252 break;
1253 }
1254
1255 if (unlikely(ret != BLK_STS_OK)) {
1256 errors++;
1257 blk_mq_end_request(rq, BLK_STS_IOERR);
1258 continue;
1259 }
1260
1261 queued++;
1262 } while (!list_empty(list));
1263
1264 hctx->dispatched[queued_to_index(queued)]++;
1265
1266 /*
1267 * Any items that need requeuing? Stuff them into hctx->dispatch,
1268 * that is where we will continue on next queue run.
1269 */
1270 if (!list_empty(list)) {
1271 bool needs_restart;
1272
1273 spin_lock(&hctx->lock);
1274 list_splice_init(list, &hctx->dispatch);
1275 spin_unlock(&hctx->lock);
1276
1277 /*
1278 * If SCHED_RESTART was set by the caller of this function and
1279 * it is no longer set that means that it was cleared by another
1280 * thread and hence that a queue rerun is needed.
1281 *
1282 * If 'no_tag' is set, that means that we failed getting
1283 * a driver tag with an I/O scheduler attached. If our dispatch
1284 * waitqueue is no longer active, ensure that we run the queue
1285 * AFTER adding our entries back to the list.
1286 *
1287 * If no I/O scheduler has been configured it is possible that
1288 * the hardware queue got stopped and restarted before requests
1289 * were pushed back onto the dispatch list. Rerun the queue to
1290 * avoid starvation. Notes:
1291 * - blk_mq_run_hw_queue() checks whether or not a queue has
1292 * been stopped before rerunning a queue.
1293 * - Some but not all block drivers stop a queue before
1294 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1295 * and dm-rq.
1296 *
1297 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1298 * bit is set, run queue after a delay to avoid IO stalls
1299 * that could otherwise occur if the queue is idle.
1300 */
1301 needs_restart = blk_mq_sched_needs_restart(hctx);
1302 if (!needs_restart ||
1303 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1304 blk_mq_run_hw_queue(hctx, true);
1305 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1306 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1307 }
1308
1309 return (queued + errors) != 0;
1310}
1311
1312static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1313{
1314 int srcu_idx;
1315
1316 /*
1317 * We should be running this queue from one of the CPUs that
1318 * are mapped to it.
1319 *
1320 * There are at least two related races now between setting
1321 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1322 * __blk_mq_run_hw_queue():
1323 *
1324 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1325 * but later it becomes online, then this warning is harmless
1326 * at all
1327 *
1328 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1329 * but later it becomes offline, then the warning can't be
1330 * triggered, and we depend on blk-mq timeout handler to
1331 * handle dispatched requests to this hctx
1332 */
1333 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1334 cpu_online(hctx->next_cpu)) {
1335 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1336 raw_smp_processor_id(),
1337 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1338 dump_stack();
1339 }
1340
1341 /*
1342 * We can't run the queue inline with ints disabled. Ensure that
1343 * we catch bad users of this early.
1344 */
1345 WARN_ON_ONCE(in_interrupt());
1346
1347 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1348
1349 hctx_lock(hctx, &srcu_idx);
1350 blk_mq_sched_dispatch_requests(hctx);
1351 hctx_unlock(hctx, srcu_idx);
1352}
1353
1354static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1355{
1356 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1357
1358 if (cpu >= nr_cpu_ids)
1359 cpu = cpumask_first(hctx->cpumask);
1360 return cpu;
1361}
1362
1363/*
1364 * It'd be great if the workqueue API had a way to pass
1365 * in a mask and had some smarts for more clever placement.
1366 * For now we just round-robin here, switching for every
1367 * BLK_MQ_CPU_WORK_BATCH queued items.
1368 */
1369static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1370{
1371 bool tried = false;
1372 int next_cpu = hctx->next_cpu;
1373
1374 if (hctx->queue->nr_hw_queues == 1)
1375 return WORK_CPU_UNBOUND;
1376
1377 if (--hctx->next_cpu_batch <= 0) {
1378select_cpu:
1379 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1380 cpu_online_mask);
1381 if (next_cpu >= nr_cpu_ids)
1382 next_cpu = blk_mq_first_mapped_cpu(hctx);
1383 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1384 }
1385
1386 /*
1387 * Do unbound schedule if we can't find a online CPU for this hctx,
1388 * and it should only happen in the path of handling CPU DEAD.
1389 */
1390 if (!cpu_online(next_cpu)) {
1391 if (!tried) {
1392 tried = true;
1393 goto select_cpu;
1394 }
1395
1396 /*
1397 * Make sure to re-select CPU next time once after CPUs
1398 * in hctx->cpumask become online again.
1399 */
1400 hctx->next_cpu = next_cpu;
1401 hctx->next_cpu_batch = 1;
1402 return WORK_CPU_UNBOUND;
1403 }
1404
1405 hctx->next_cpu = next_cpu;
1406 return next_cpu;
1407}
1408
1409static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1410 unsigned long msecs)
1411{
1412 if (unlikely(blk_mq_hctx_stopped(hctx)))
1413 return;
1414
1415 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1416 int cpu = get_cpu();
1417 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1418 __blk_mq_run_hw_queue(hctx);
1419 put_cpu();
1420 return;
1421 }
1422
1423 put_cpu();
1424 }
1425
1426 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1427 msecs_to_jiffies(msecs));
1428}
1429
1430void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1431{
1432 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1433}
1434EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1435
1436bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1437{
1438 int srcu_idx;
1439 bool need_run;
1440
1441 /*
1442 * When queue is quiesced, we may be switching io scheduler, or
1443 * updating nr_hw_queues, or other things, and we can't run queue
1444 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1445 *
1446 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1447 * quiesced.
1448 */
1449 hctx_lock(hctx, &srcu_idx);
1450 need_run = !blk_queue_quiesced(hctx->queue) &&
1451 blk_mq_hctx_has_pending(hctx);
1452 hctx_unlock(hctx, srcu_idx);
1453
1454 if (need_run) {
1455 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1456 return true;
1457 }
1458
1459 return false;
1460}
1461EXPORT_SYMBOL(blk_mq_run_hw_queue);
1462
1463void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1464{
1465 struct blk_mq_hw_ctx *hctx;
1466 int i;
1467
1468 queue_for_each_hw_ctx(q, hctx, i) {
1469 if (blk_mq_hctx_stopped(hctx))
1470 continue;
1471
1472 blk_mq_run_hw_queue(hctx, async);
1473 }
1474}
1475EXPORT_SYMBOL(blk_mq_run_hw_queues);
1476
1477/**
1478 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1479 * @q: request queue.
1480 *
1481 * The caller is responsible for serializing this function against
1482 * blk_mq_{start,stop}_hw_queue().
1483 */
1484bool blk_mq_queue_stopped(struct request_queue *q)
1485{
1486 struct blk_mq_hw_ctx *hctx;
1487 int i;
1488
1489 queue_for_each_hw_ctx(q, hctx, i)
1490 if (blk_mq_hctx_stopped(hctx))
1491 return true;
1492
1493 return false;
1494}
1495EXPORT_SYMBOL(blk_mq_queue_stopped);
1496
1497/*
1498 * This function is often used for pausing .queue_rq() by driver when
1499 * there isn't enough resource or some conditions aren't satisfied, and
1500 * BLK_STS_RESOURCE is usually returned.
1501 *
1502 * We do not guarantee that dispatch can be drained or blocked
1503 * after blk_mq_stop_hw_queue() returns. Please use
1504 * blk_mq_quiesce_queue() for that requirement.
1505 */
1506void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1507{
1508 cancel_delayed_work(&hctx->run_work);
1509
1510 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1511}
1512EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1513
1514/*
1515 * This function is often used for pausing .queue_rq() by driver when
1516 * there isn't enough resource or some conditions aren't satisfied, and
1517 * BLK_STS_RESOURCE is usually returned.
1518 *
1519 * We do not guarantee that dispatch can be drained or blocked
1520 * after blk_mq_stop_hw_queues() returns. Please use
1521 * blk_mq_quiesce_queue() for that requirement.
1522 */
1523void blk_mq_stop_hw_queues(struct request_queue *q)
1524{
1525 struct blk_mq_hw_ctx *hctx;
1526 int i;
1527
1528 queue_for_each_hw_ctx(q, hctx, i)
1529 blk_mq_stop_hw_queue(hctx);
1530}
1531EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1532
1533void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1534{
1535 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1536
1537 blk_mq_run_hw_queue(hctx, false);
1538}
1539EXPORT_SYMBOL(blk_mq_start_hw_queue);
1540
1541void blk_mq_start_hw_queues(struct request_queue *q)
1542{
1543 struct blk_mq_hw_ctx *hctx;
1544 int i;
1545
1546 queue_for_each_hw_ctx(q, hctx, i)
1547 blk_mq_start_hw_queue(hctx);
1548}
1549EXPORT_SYMBOL(blk_mq_start_hw_queues);
1550
1551void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1552{
1553 if (!blk_mq_hctx_stopped(hctx))
1554 return;
1555
1556 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1557 blk_mq_run_hw_queue(hctx, async);
1558}
1559EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1560
1561void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1562{
1563 struct blk_mq_hw_ctx *hctx;
1564 int i;
1565
1566 queue_for_each_hw_ctx(q, hctx, i)
1567 blk_mq_start_stopped_hw_queue(hctx, async);
1568}
1569EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1570
1571static void blk_mq_run_work_fn(struct work_struct *work)
1572{
1573 struct blk_mq_hw_ctx *hctx;
1574
1575 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1576
1577 /*
1578 * If we are stopped, don't run the queue.
1579 */
1580 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1581 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1582
1583 __blk_mq_run_hw_queue(hctx);
1584}
1585
1586static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1587 struct request *rq,
1588 bool at_head)
1589{
1590 struct blk_mq_ctx *ctx = rq->mq_ctx;
1591
1592 lockdep_assert_held(&ctx->lock);
1593
1594 trace_block_rq_insert(hctx->queue, rq);
1595
1596 if (at_head)
1597 list_add(&rq->queuelist, &ctx->rq_list);
1598 else
1599 list_add_tail(&rq->queuelist, &ctx->rq_list);
1600}
1601
1602void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1603 bool at_head)
1604{
1605 struct blk_mq_ctx *ctx = rq->mq_ctx;
1606
1607 lockdep_assert_held(&ctx->lock);
1608
1609 __blk_mq_insert_req_list(hctx, rq, at_head);
1610 blk_mq_hctx_mark_pending(hctx, ctx);
1611}
1612
1613/*
1614 * Should only be used carefully, when the caller knows we want to
1615 * bypass a potential IO scheduler on the target device.
1616 */
1617void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1618{
1619 struct blk_mq_ctx *ctx = rq->mq_ctx;
1620 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1621
1622 spin_lock(&hctx->lock);
1623 list_add_tail(&rq->queuelist, &hctx->dispatch);
1624 spin_unlock(&hctx->lock);
1625
1626 if (run_queue)
1627 blk_mq_run_hw_queue(hctx, false);
1628}
1629
1630void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1631 struct list_head *list)
1632
1633{
1634 /*
1635 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1636 * offline now
1637 */
1638 spin_lock(&ctx->lock);
1639 while (!list_empty(list)) {
1640 struct request *rq;
1641
1642 rq = list_first_entry(list, struct request, queuelist);
1643 BUG_ON(rq->mq_ctx != ctx);
1644 list_del_init(&rq->queuelist);
1645 __blk_mq_insert_req_list(hctx, rq, false);
1646 }
1647 blk_mq_hctx_mark_pending(hctx, ctx);
1648 spin_unlock(&ctx->lock);
1649}
1650
1651static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1652{
1653 struct request *rqa = container_of(a, struct request, queuelist);
1654 struct request *rqb = container_of(b, struct request, queuelist);
1655
1656 return !(rqa->mq_ctx < rqb->mq_ctx ||
1657 (rqa->mq_ctx == rqb->mq_ctx &&
1658 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1659}
1660
1661void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1662{
1663 struct blk_mq_ctx *this_ctx;
1664 struct request_queue *this_q;
1665 struct request *rq;
1666 LIST_HEAD(list);
1667 LIST_HEAD(ctx_list);
1668 unsigned int depth;
1669
1670 list_splice_init(&plug->mq_list, &list);
1671
1672 list_sort(NULL, &list, plug_ctx_cmp);
1673
1674 this_q = NULL;
1675 this_ctx = NULL;
1676 depth = 0;
1677
1678 while (!list_empty(&list)) {
1679 rq = list_entry_rq(list.next);
1680 list_del_init(&rq->queuelist);
1681 BUG_ON(!rq->q);
1682 if (rq->mq_ctx != this_ctx) {
1683 if (this_ctx) {
1684 trace_block_unplug(this_q, depth, from_schedule);
1685 blk_mq_sched_insert_requests(this_q, this_ctx,
1686 &ctx_list,
1687 from_schedule);
1688 }
1689
1690 this_ctx = rq->mq_ctx;
1691 this_q = rq->q;
1692 depth = 0;
1693 }
1694
1695 depth++;
1696 list_add_tail(&rq->queuelist, &ctx_list);
1697 }
1698
1699 /*
1700 * If 'this_ctx' is set, we know we have entries to complete
1701 * on 'ctx_list'. Do those.
1702 */
1703 if (this_ctx) {
1704 trace_block_unplug(this_q, depth, from_schedule);
1705 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1706 from_schedule);
1707 }
1708}
1709
1710static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1711{
1712 blk_init_request_from_bio(rq, bio);
1713
1714 blk_rq_set_rl(rq, blk_get_rl(rq->q, bio));
1715
1716 blk_account_io_start(rq, true);
1717}
1718
1719static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx,
1720 struct blk_mq_ctx *ctx,
1721 struct request *rq)
1722{
1723 spin_lock(&ctx->lock);
1724 __blk_mq_insert_request(hctx, rq, false);
1725 spin_unlock(&ctx->lock);
1726}
1727
1728static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1729{
1730 if (rq->tag != -1)
1731 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1732
1733 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1734}
1735
1736static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1737 struct request *rq,
1738 blk_qc_t *cookie)
1739{
1740 struct request_queue *q = rq->q;
1741 struct blk_mq_queue_data bd = {
1742 .rq = rq,
1743 .last = true,
1744 };
1745 blk_qc_t new_cookie;
1746 blk_status_t ret;
1747
1748 new_cookie = request_to_qc_t(hctx, rq);
1749
1750 /*
1751 * For OK queue, we are done. For error, caller may kill it.
1752 * Any other error (busy), just add it to our list as we
1753 * previously would have done.
1754 */
1755 ret = q->mq_ops->queue_rq(hctx, &bd);
1756 switch (ret) {
1757 case BLK_STS_OK:
1758 *cookie = new_cookie;
1759 break;
1760 case BLK_STS_RESOURCE:
1761 case BLK_STS_DEV_RESOURCE:
1762 __blk_mq_requeue_request(rq);
1763 break;
1764 default:
1765 *cookie = BLK_QC_T_NONE;
1766 break;
1767 }
1768
1769 return ret;
1770}
1771
1772static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1773 struct request *rq,
1774 blk_qc_t *cookie,
1775 bool bypass_insert)
1776{
1777 struct request_queue *q = rq->q;
1778 bool run_queue = true;
1779
1780 /*
1781 * RCU or SRCU read lock is needed before checking quiesced flag.
1782 *
1783 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1784 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1785 * and avoid driver to try to dispatch again.
1786 */
1787 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1788 run_queue = false;
1789 bypass_insert = false;
1790 goto insert;
1791 }
1792
1793 if (q->elevator && !bypass_insert)
1794 goto insert;
1795
1796 if (!blk_mq_get_dispatch_budget(hctx))
1797 goto insert;
1798
1799 if (!blk_mq_get_driver_tag(rq, NULL, false)) {
1800 blk_mq_put_dispatch_budget(hctx);
1801 goto insert;
1802 }
1803
1804 return __blk_mq_issue_directly(hctx, rq, cookie);
1805insert:
1806 if (bypass_insert)
1807 return BLK_STS_RESOURCE;
1808
1809 blk_mq_sched_insert_request(rq, false, run_queue, false);
1810 return BLK_STS_OK;
1811}
1812
1813static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1814 struct request *rq, blk_qc_t *cookie)
1815{
1816 blk_status_t ret;
1817 int srcu_idx;
1818
1819 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1820
1821 hctx_lock(hctx, &srcu_idx);
1822
1823 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1824 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1825 blk_mq_sched_insert_request(rq, false, true, false);
1826 else if (ret != BLK_STS_OK)
1827 blk_mq_end_request(rq, ret);
1828
1829 hctx_unlock(hctx, srcu_idx);
1830}
1831
1832blk_status_t blk_mq_request_issue_directly(struct request *rq)
1833{
1834 blk_status_t ret;
1835 int srcu_idx;
1836 blk_qc_t unused_cookie;
1837 struct blk_mq_ctx *ctx = rq->mq_ctx;
1838 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1839
1840 hctx_lock(hctx, &srcu_idx);
1841 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true);
1842 hctx_unlock(hctx, srcu_idx);
1843
1844 return ret;
1845}
1846
1847static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1848{
1849 const int is_sync = op_is_sync(bio->bi_opf);
1850 const int is_flush_fua = op_is_flush(bio->bi_opf);
1851 struct blk_mq_alloc_data data = { .flags = 0 };
1852 struct request *rq;
1853 unsigned int request_count = 0;
1854 struct blk_plug *plug;
1855 struct request *same_queue_rq = NULL;
1856 blk_qc_t cookie;
1857 unsigned int wb_acct;
1858
1859 blk_queue_bounce(q, &bio);
1860
1861 blk_queue_split(q, &bio);
1862
1863 if (!bio_integrity_prep(bio))
1864 return BLK_QC_T_NONE;
1865
1866 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1867 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1868 return BLK_QC_T_NONE;
1869
1870 if (blk_mq_sched_bio_merge(q, bio))
1871 return BLK_QC_T_NONE;
1872
1873 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1874
1875 trace_block_getrq(q, bio, bio->bi_opf);
1876
1877 rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1878 if (unlikely(!rq)) {
1879 __wbt_done(q->rq_wb, wb_acct);
1880 if (bio->bi_opf & REQ_NOWAIT)
1881 bio_wouldblock_error(bio);
1882 return BLK_QC_T_NONE;
1883 }
1884
1885 wbt_track(&rq->issue_stat, wb_acct);
1886
1887 cookie = request_to_qc_t(data.hctx, rq);
1888
1889 plug = current->plug;
1890 if (unlikely(is_flush_fua)) {
1891 blk_mq_put_ctx(data.ctx);
1892 blk_mq_bio_to_request(rq, bio);
1893
1894 /* bypass scheduler for flush rq */
1895 blk_insert_flush(rq);
1896 blk_mq_run_hw_queue(data.hctx, true);
1897 } else if (plug && q->nr_hw_queues == 1) {
1898 struct request *last = NULL;
1899
1900 blk_mq_put_ctx(data.ctx);
1901 blk_mq_bio_to_request(rq, bio);
1902
1903 /*
1904 * @request_count may become stale because of schedule
1905 * out, so check the list again.
1906 */
1907 if (list_empty(&plug->mq_list))
1908 request_count = 0;
1909 else if (blk_queue_nomerges(q))
1910 request_count = blk_plug_queued_count(q);
1911
1912 if (!request_count)
1913 trace_block_plug(q);
1914 else
1915 last = list_entry_rq(plug->mq_list.prev);
1916
1917 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1918 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1919 blk_flush_plug_list(plug, false);
1920 trace_block_plug(q);
1921 }
1922
1923 list_add_tail(&rq->queuelist, &plug->mq_list);
1924 } else if (plug && !blk_queue_nomerges(q)) {
1925 blk_mq_bio_to_request(rq, bio);
1926
1927 /*
1928 * We do limited plugging. If the bio can be merged, do that.
1929 * Otherwise the existing request in the plug list will be
1930 * issued. So the plug list will have one request at most
1931 * The plug list might get flushed before this. If that happens,
1932 * the plug list is empty, and same_queue_rq is invalid.
1933 */
1934 if (list_empty(&plug->mq_list))
1935 same_queue_rq = NULL;
1936 if (same_queue_rq)
1937 list_del_init(&same_queue_rq->queuelist);
1938 list_add_tail(&rq->queuelist, &plug->mq_list);
1939
1940 blk_mq_put_ctx(data.ctx);
1941
1942 if (same_queue_rq) {
1943 data.hctx = blk_mq_map_queue(q,
1944 same_queue_rq->mq_ctx->cpu);
1945 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1946 &cookie);
1947 }
1948 } else if (q->nr_hw_queues > 1 && is_sync) {
1949 blk_mq_put_ctx(data.ctx);
1950 blk_mq_bio_to_request(rq, bio);
1951 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1952 } else if (q->elevator) {
1953 blk_mq_put_ctx(data.ctx);
1954 blk_mq_bio_to_request(rq, bio);
1955 blk_mq_sched_insert_request(rq, false, true, true);
1956 } else {
1957 blk_mq_put_ctx(data.ctx);
1958 blk_mq_bio_to_request(rq, bio);
1959 blk_mq_queue_io(data.hctx, data.ctx, rq);
1960 blk_mq_run_hw_queue(data.hctx, true);
1961 }
1962
1963 return cookie;
1964}
1965
1966void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1967 unsigned int hctx_idx)
1968{
1969 struct page *page;
1970
1971 if (tags->rqs && set->ops->exit_request) {
1972 int i;
1973
1974 for (i = 0; i < tags->nr_tags; i++) {
1975 struct request *rq = tags->static_rqs[i];
1976
1977 if (!rq)
1978 continue;
1979 set->ops->exit_request(set, rq, hctx_idx);
1980 tags->static_rqs[i] = NULL;
1981 }
1982 }
1983
1984 while (!list_empty(&tags->page_list)) {
1985 page = list_first_entry(&tags->page_list, struct page, lru);
1986 list_del_init(&page->lru);
1987 /*
1988 * Remove kmemleak object previously allocated in
1989 * blk_mq_init_rq_map().
1990 */
1991 kmemleak_free(page_address(page));
1992 __free_pages(page, page->private);
1993 }
1994}
1995
1996void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1997{
1998 kfree(tags->rqs);
1999 tags->rqs = NULL;
2000 kfree(tags->static_rqs);
2001 tags->static_rqs = NULL;
2002
2003 blk_mq_free_tags(tags);
2004}
2005
2006struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2007 unsigned int hctx_idx,
2008 unsigned int nr_tags,
2009 unsigned int reserved_tags)
2010{
2011 struct blk_mq_tags *tags;
2012 int node;
2013
2014 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
2015 if (node == NUMA_NO_NODE)
2016 node = set->numa_node;
2017
2018 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2019 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2020 if (!tags)
2021 return NULL;
2022
2023 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
2024 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2025 node);
2026 if (!tags->rqs) {
2027 blk_mq_free_tags(tags);
2028 return NULL;
2029 }
2030
2031 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
2032 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2033 node);
2034 if (!tags->static_rqs) {
2035 kfree(tags->rqs);
2036 blk_mq_free_tags(tags);
2037 return NULL;
2038 }
2039
2040 return tags;
2041}
2042
2043static size_t order_to_size(unsigned int order)
2044{
2045 return (size_t)PAGE_SIZE << order;
2046}
2047
2048static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2049 unsigned int hctx_idx, int node)
2050{
2051 int ret;
2052
2053 if (set->ops->init_request) {
2054 ret = set->ops->init_request(set, rq, hctx_idx, node);
2055 if (ret)
2056 return ret;
2057 }
2058
2059 seqcount_init(&rq->gstate_seq);
2060 u64_stats_init(&rq->aborted_gstate_sync);
2061 /*
2062 * start gstate with gen 1 instead of 0, otherwise it will be equal
2063 * to aborted_gstate, and be identified timed out by
2064 * blk_mq_terminate_expired.
2065 */
2066 WRITE_ONCE(rq->gstate, MQ_RQ_GEN_INC);
2067
2068 return 0;
2069}
2070
2071int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2072 unsigned int hctx_idx, unsigned int depth)
2073{
2074 unsigned int i, j, entries_per_page, max_order = 4;
2075 size_t rq_size, left;
2076 int node;
2077
2078 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
2079 if (node == NUMA_NO_NODE)
2080 node = set->numa_node;
2081
2082 INIT_LIST_HEAD(&tags->page_list);
2083
2084 /*
2085 * rq_size is the size of the request plus driver payload, rounded
2086 * to the cacheline size
2087 */
2088 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2089 cache_line_size());
2090 left = rq_size * depth;
2091
2092 for (i = 0; i < depth; ) {
2093 int this_order = max_order;
2094 struct page *page;
2095 int to_do;
2096 void *p;
2097
2098 while (this_order && left < order_to_size(this_order - 1))
2099 this_order--;
2100
2101 do {
2102 page = alloc_pages_node(node,
2103 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2104 this_order);
2105 if (page)
2106 break;
2107 if (!this_order--)
2108 break;
2109 if (order_to_size(this_order) < rq_size)
2110 break;
2111 } while (1);
2112
2113 if (!page)
2114 goto fail;
2115
2116 page->private = this_order;
2117 list_add_tail(&page->lru, &tags->page_list);
2118
2119 p = page_address(page);
2120 /*
2121 * Allow kmemleak to scan these pages as they contain pointers
2122 * to additional allocations like via ops->init_request().
2123 */
2124 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2125 entries_per_page = order_to_size(this_order) / rq_size;
2126 to_do = min(entries_per_page, depth - i);
2127 left -= to_do * rq_size;
2128 for (j = 0; j < to_do; j++) {
2129 struct request *rq = p;
2130
2131 tags->static_rqs[i] = rq;
2132 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2133 tags->static_rqs[i] = NULL;
2134 goto fail;
2135 }
2136
2137 p += rq_size;
2138 i++;
2139 }
2140 }
2141 return 0;
2142
2143fail:
2144 blk_mq_free_rqs(set, tags, hctx_idx);
2145 return -ENOMEM;
2146}
2147
2148/*
2149 * 'cpu' is going away. splice any existing rq_list entries from this
2150 * software queue to the hw queue dispatch list, and ensure that it
2151 * gets run.
2152 */
2153static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2154{
2155 struct blk_mq_hw_ctx *hctx;
2156 struct blk_mq_ctx *ctx;
2157 LIST_HEAD(tmp);
2158
2159 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2160 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2161
2162 spin_lock(&ctx->lock);
2163 if (!list_empty(&ctx->rq_list)) {
2164 list_splice_init(&ctx->rq_list, &tmp);
2165 blk_mq_hctx_clear_pending(hctx, ctx);
2166 }
2167 spin_unlock(&ctx->lock);
2168
2169 if (list_empty(&tmp))
2170 return 0;
2171
2172 spin_lock(&hctx->lock);
2173 list_splice_tail_init(&tmp, &hctx->dispatch);
2174 spin_unlock(&hctx->lock);
2175
2176 blk_mq_run_hw_queue(hctx, true);
2177 return 0;
2178}
2179
2180static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2181{
2182 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2183 &hctx->cpuhp_dead);
2184}
2185
2186/* hctx->ctxs will be freed in queue's release handler */
2187static void blk_mq_exit_hctx(struct request_queue *q,
2188 struct blk_mq_tag_set *set,
2189 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2190{
2191 blk_mq_debugfs_unregister_hctx(hctx);
2192
2193 if (blk_mq_hw_queue_mapped(hctx))
2194 blk_mq_tag_idle(hctx);
2195
2196 if (set->ops->exit_request)
2197 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2198
2199 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2200
2201 if (set->ops->exit_hctx)
2202 set->ops->exit_hctx(hctx, hctx_idx);
2203
2204 if (hctx->flags & BLK_MQ_F_BLOCKING)
2205 cleanup_srcu_struct(hctx->srcu);
2206
2207 blk_mq_remove_cpuhp(hctx);
2208 blk_free_flush_queue(hctx->fq);
2209 sbitmap_free(&hctx->ctx_map);
2210}
2211
2212static void blk_mq_exit_hw_queues(struct request_queue *q,
2213 struct blk_mq_tag_set *set, int nr_queue)
2214{
2215 struct blk_mq_hw_ctx *hctx;
2216 unsigned int i;
2217
2218 queue_for_each_hw_ctx(q, hctx, i) {
2219 if (i == nr_queue)
2220 break;
2221 blk_mq_exit_hctx(q, set, hctx, i);
2222 }
2223}
2224
2225static int blk_mq_init_hctx(struct request_queue *q,
2226 struct blk_mq_tag_set *set,
2227 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2228{
2229 int node;
2230
2231 node = hctx->numa_node;
2232 if (node == NUMA_NO_NODE)
2233 node = hctx->numa_node = set->numa_node;
2234
2235 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2236 spin_lock_init(&hctx->lock);
2237 INIT_LIST_HEAD(&hctx->dispatch);
2238 hctx->queue = q;
2239 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2240
2241 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2242
2243 hctx->tags = set->tags[hctx_idx];
2244
2245 /*
2246 * Allocate space for all possible cpus to avoid allocation at
2247 * runtime
2248 */
2249 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2250 GFP_KERNEL, node);
2251 if (!hctx->ctxs)
2252 goto unregister_cpu_notifier;
2253
2254 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
2255 node))
2256 goto free_ctxs;
2257
2258 hctx->nr_ctx = 0;
2259
2260 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2261 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2262
2263 if (set->ops->init_hctx &&
2264 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2265 goto free_bitmap;
2266
2267 if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
2268 goto exit_hctx;
2269
2270 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
2271 if (!hctx->fq)
2272 goto sched_exit_hctx;
2273
2274 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node))
2275 goto free_fq;
2276
2277 if (hctx->flags & BLK_MQ_F_BLOCKING)
2278 init_srcu_struct(hctx->srcu);
2279
2280 blk_mq_debugfs_register_hctx(q, hctx);
2281
2282 return 0;
2283
2284 free_fq:
2285 kfree(hctx->fq);
2286 sched_exit_hctx:
2287 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2288 exit_hctx:
2289 if (set->ops->exit_hctx)
2290 set->ops->exit_hctx(hctx, hctx_idx);
2291 free_bitmap:
2292 sbitmap_free(&hctx->ctx_map);
2293 free_ctxs:
2294 kfree(hctx->ctxs);
2295 unregister_cpu_notifier:
2296 blk_mq_remove_cpuhp(hctx);
2297 return -1;
2298}
2299
2300static void blk_mq_init_cpu_queues(struct request_queue *q,
2301 unsigned int nr_hw_queues)
2302{
2303 unsigned int i;
2304
2305 for_each_possible_cpu(i) {
2306 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2307 struct blk_mq_hw_ctx *hctx;
2308
2309 __ctx->cpu = i;
2310 spin_lock_init(&__ctx->lock);
2311 INIT_LIST_HEAD(&__ctx->rq_list);
2312 __ctx->queue = q;
2313
2314 /*
2315 * Set local node, IFF we have more than one hw queue. If
2316 * not, we remain on the home node of the device
2317 */
2318 hctx = blk_mq_map_queue(q, i);
2319 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2320 hctx->numa_node = local_memory_node(cpu_to_node(i));
2321 }
2322}
2323
2324static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2325{
2326 int ret = 0;
2327
2328 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2329 set->queue_depth, set->reserved_tags);
2330 if (!set->tags[hctx_idx])
2331 return false;
2332
2333 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2334 set->queue_depth);
2335 if (!ret)
2336 return true;
2337
2338 blk_mq_free_rq_map(set->tags[hctx_idx]);
2339 set->tags[hctx_idx] = NULL;
2340 return false;
2341}
2342
2343static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2344 unsigned int hctx_idx)
2345{
2346 if (set->tags[hctx_idx]) {
2347 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2348 blk_mq_free_rq_map(set->tags[hctx_idx]);
2349 set->tags[hctx_idx] = NULL;
2350 }
2351}
2352
2353static void blk_mq_map_swqueue(struct request_queue *q)
2354{
2355 unsigned int i, hctx_idx;
2356 struct blk_mq_hw_ctx *hctx;
2357 struct blk_mq_ctx *ctx;
2358 struct blk_mq_tag_set *set = q->tag_set;
2359
2360 /*
2361 * Avoid others reading imcomplete hctx->cpumask through sysfs
2362 */
2363 mutex_lock(&q->sysfs_lock);
2364
2365 queue_for_each_hw_ctx(q, hctx, i) {
2366 cpumask_clear(hctx->cpumask);
2367 hctx->nr_ctx = 0;
2368 }
2369
2370 /*
2371 * Map software to hardware queues.
2372 *
2373 * If the cpu isn't present, the cpu is mapped to first hctx.
2374 */
2375 for_each_possible_cpu(i) {
2376 hctx_idx = q->mq_map[i];
2377 /* unmapped hw queue can be remapped after CPU topo changed */
2378 if (!set->tags[hctx_idx] &&
2379 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2380 /*
2381 * If tags initialization fail for some hctx,
2382 * that hctx won't be brought online. In this
2383 * case, remap the current ctx to hctx[0] which
2384 * is guaranteed to always have tags allocated
2385 */
2386 q->mq_map[i] = 0;
2387 }
2388
2389 ctx = per_cpu_ptr(q->queue_ctx, i);
2390 hctx = blk_mq_map_queue(q, i);
2391
2392 cpumask_set_cpu(i, hctx->cpumask);
2393 ctx->index_hw = hctx->nr_ctx;
2394 hctx->ctxs[hctx->nr_ctx++] = ctx;
2395 }
2396
2397 mutex_unlock(&q->sysfs_lock);
2398
2399 queue_for_each_hw_ctx(q, hctx, i) {
2400 /*
2401 * If no software queues are mapped to this hardware queue,
2402 * disable it and free the request entries.
2403 */
2404 if (!hctx->nr_ctx) {
2405 /* Never unmap queue 0. We need it as a
2406 * fallback in case of a new remap fails
2407 * allocation
2408 */
2409 if (i && set->tags[i])
2410 blk_mq_free_map_and_requests(set, i);
2411
2412 hctx->tags = NULL;
2413 continue;
2414 }
2415
2416 hctx->tags = set->tags[i];
2417 WARN_ON(!hctx->tags);
2418
2419 /*
2420 * Set the map size to the number of mapped software queues.
2421 * This is more accurate and more efficient than looping
2422 * over all possibly mapped software queues.
2423 */
2424 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2425
2426 /*
2427 * Initialize batch roundrobin counts
2428 */
2429 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2430 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2431 }
2432}
2433
2434/*
2435 * Caller needs to ensure that we're either frozen/quiesced, or that
2436 * the queue isn't live yet.
2437 */
2438static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2439{
2440 struct blk_mq_hw_ctx *hctx;
2441 int i;
2442
2443 queue_for_each_hw_ctx(q, hctx, i) {
2444 if (shared) {
2445 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2446 atomic_inc(&q->shared_hctx_restart);
2447 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2448 } else {
2449 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2450 atomic_dec(&q->shared_hctx_restart);
2451 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2452 }
2453 }
2454}
2455
2456static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2457 bool shared)
2458{
2459 struct request_queue *q;
2460
2461 lockdep_assert_held(&set->tag_list_lock);
2462
2463 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2464 blk_mq_freeze_queue(q);
2465 queue_set_hctx_shared(q, shared);
2466 blk_mq_unfreeze_queue(q);
2467 }
2468}
2469
2470static void blk_mq_del_queue_tag_set(struct request_queue *q)
2471{
2472 struct blk_mq_tag_set *set = q->tag_set;
2473
2474 mutex_lock(&set->tag_list_lock);
2475 list_del_rcu(&q->tag_set_list);
2476 INIT_LIST_HEAD(&q->tag_set_list);
2477 if (list_is_singular(&set->tag_list)) {
2478 /* just transitioned to unshared */
2479 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2480 /* update existing queue */
2481 blk_mq_update_tag_set_depth(set, false);
2482 }
2483 mutex_unlock(&set->tag_list_lock);
2484
2485 synchronize_rcu();
2486}
2487
2488static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2489 struct request_queue *q)
2490{
2491 q->tag_set = set;
2492
2493 mutex_lock(&set->tag_list_lock);
2494
2495 /*
2496 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2497 */
2498 if (!list_empty(&set->tag_list) &&
2499 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2500 set->flags |= BLK_MQ_F_TAG_SHARED;
2501 /* update existing queue */
2502 blk_mq_update_tag_set_depth(set, true);
2503 }
2504 if (set->flags & BLK_MQ_F_TAG_SHARED)
2505 queue_set_hctx_shared(q, true);
2506 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2507
2508 mutex_unlock(&set->tag_list_lock);
2509}
2510
2511/*
2512 * It is the actual release handler for mq, but we do it from
2513 * request queue's release handler for avoiding use-after-free
2514 * and headache because q->mq_kobj shouldn't have been introduced,
2515 * but we can't group ctx/kctx kobj without it.
2516 */
2517void blk_mq_release(struct request_queue *q)
2518{
2519 struct blk_mq_hw_ctx *hctx;
2520 unsigned int i;
2521
2522 /* hctx kobj stays in hctx */
2523 queue_for_each_hw_ctx(q, hctx, i) {
2524 if (!hctx)
2525 continue;
2526 kobject_put(&hctx->kobj);
2527 }
2528
2529 q->mq_map = NULL;
2530
2531 kfree(q->queue_hw_ctx);
2532
2533 /*
2534 * release .mq_kobj and sw queue's kobject now because
2535 * both share lifetime with request queue.
2536 */
2537 blk_mq_sysfs_deinit(q);
2538
2539 free_percpu(q->queue_ctx);
2540}
2541
2542struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2543{
2544 struct request_queue *uninit_q, *q;
2545
2546 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node, NULL);
2547 if (!uninit_q)
2548 return ERR_PTR(-ENOMEM);
2549
2550 q = blk_mq_init_allocated_queue(set, uninit_q);
2551 if (IS_ERR(q))
2552 blk_cleanup_queue(uninit_q);
2553
2554 return q;
2555}
2556EXPORT_SYMBOL(blk_mq_init_queue);
2557
2558static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2559{
2560 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2561
2562 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2563 __alignof__(struct blk_mq_hw_ctx)) !=
2564 sizeof(struct blk_mq_hw_ctx));
2565
2566 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2567 hw_ctx_size += sizeof(struct srcu_struct);
2568
2569 return hw_ctx_size;
2570}
2571
2572static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2573 struct request_queue *q)
2574{
2575 int i, j;
2576 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2577
2578 blk_mq_sysfs_unregister(q);
2579
2580 /* protect against switching io scheduler */
2581 mutex_lock(&q->sysfs_lock);
2582 for (i = 0; i < set->nr_hw_queues; i++) {
2583 int node;
2584
2585 if (hctxs[i])
2586 continue;
2587
2588 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2589 hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set),
2590 GFP_KERNEL, node);
2591 if (!hctxs[i])
2592 break;
2593
2594 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2595 node)) {
2596 kfree(hctxs[i]);
2597 hctxs[i] = NULL;
2598 break;
2599 }
2600
2601 atomic_set(&hctxs[i]->nr_active, 0);
2602 hctxs[i]->numa_node = node;
2603 hctxs[i]->queue_num = i;
2604
2605 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2606 free_cpumask_var(hctxs[i]->cpumask);
2607 kfree(hctxs[i]);
2608 hctxs[i] = NULL;
2609 break;
2610 }
2611 blk_mq_hctx_kobj_init(hctxs[i]);
2612 }
2613 for (j = i; j < q->nr_hw_queues; j++) {
2614 struct blk_mq_hw_ctx *hctx = hctxs[j];
2615
2616 if (hctx) {
2617 if (hctx->tags)
2618 blk_mq_free_map_and_requests(set, j);
2619 blk_mq_exit_hctx(q, set, hctx, j);
2620 kobject_put(&hctx->kobj);
2621 hctxs[j] = NULL;
2622
2623 }
2624 }
2625 q->nr_hw_queues = i;
2626 mutex_unlock(&q->sysfs_lock);
2627 blk_mq_sysfs_register(q);
2628}
2629
2630struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2631 struct request_queue *q)
2632{
2633 /* mark the queue as mq asap */
2634 q->mq_ops = set->ops;
2635
2636 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2637 blk_mq_poll_stats_bkt,
2638 BLK_MQ_POLL_STATS_BKTS, q);
2639 if (!q->poll_cb)
2640 goto err_exit;
2641
2642 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2643 if (!q->queue_ctx)
2644 goto err_exit;
2645
2646 /* init q->mq_kobj and sw queues' kobjects */
2647 blk_mq_sysfs_init(q);
2648
2649 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2650 GFP_KERNEL, set->numa_node);
2651 if (!q->queue_hw_ctx)
2652 goto err_percpu;
2653
2654 q->mq_map = set->mq_map;
2655
2656 blk_mq_realloc_hw_ctxs(set, q);
2657 if (!q->nr_hw_queues)
2658 goto err_hctxs;
2659
2660 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2661 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2662
2663 q->nr_queues = nr_cpu_ids;
2664
2665 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2666
2667 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2668 queue_flag_set_unlocked(QUEUE_FLAG_NO_SG_MERGE, q);
2669
2670 q->sg_reserved_size = INT_MAX;
2671
2672 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2673 INIT_LIST_HEAD(&q->requeue_list);
2674 spin_lock_init(&q->requeue_lock);
2675
2676 blk_queue_make_request(q, blk_mq_make_request);
2677 if (q->mq_ops->poll)
2678 q->poll_fn = blk_mq_poll;
2679
2680 /*
2681 * Do this after blk_queue_make_request() overrides it...
2682 */
2683 q->nr_requests = set->queue_depth;
2684
2685 /*
2686 * Default to classic polling
2687 */
2688 q->poll_nsec = -1;
2689
2690 if (set->ops->complete)
2691 blk_queue_softirq_done(q, set->ops->complete);
2692
2693 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2694 blk_mq_add_queue_tag_set(set, q);
2695 blk_mq_map_swqueue(q);
2696
2697 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2698 int ret;
2699
2700 ret = blk_mq_sched_init(q);
2701 if (ret)
2702 return ERR_PTR(ret);
2703 }
2704
2705 return q;
2706
2707err_hctxs:
2708 kfree(q->queue_hw_ctx);
2709err_percpu:
2710 free_percpu(q->queue_ctx);
2711err_exit:
2712 q->mq_ops = NULL;
2713 return ERR_PTR(-ENOMEM);
2714}
2715EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2716
2717void blk_mq_free_queue(struct request_queue *q)
2718{
2719 struct blk_mq_tag_set *set = q->tag_set;
2720
2721 blk_mq_del_queue_tag_set(q);
2722 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2723}
2724
2725/* Basically redo blk_mq_init_queue with queue frozen */
2726static void blk_mq_queue_reinit(struct request_queue *q)
2727{
2728 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2729
2730 blk_mq_debugfs_unregister_hctxs(q);
2731 blk_mq_sysfs_unregister(q);
2732
2733 /*
2734 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2735 * we should change hctx numa_node according to the new topology (this
2736 * involves freeing and re-allocating memory, worth doing?)
2737 */
2738 blk_mq_map_swqueue(q);
2739
2740 blk_mq_sysfs_register(q);
2741 blk_mq_debugfs_register_hctxs(q);
2742}
2743
2744static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2745{
2746 int i;
2747
2748 for (i = 0; i < set->nr_hw_queues; i++)
2749 if (!__blk_mq_alloc_rq_map(set, i))
2750 goto out_unwind;
2751
2752 return 0;
2753
2754out_unwind:
2755 while (--i >= 0)
2756 blk_mq_free_rq_map(set->tags[i]);
2757
2758 return -ENOMEM;
2759}
2760
2761/*
2762 * Allocate the request maps associated with this tag_set. Note that this
2763 * may reduce the depth asked for, if memory is tight. set->queue_depth
2764 * will be updated to reflect the allocated depth.
2765 */
2766static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2767{
2768 unsigned int depth;
2769 int err;
2770
2771 depth = set->queue_depth;
2772 do {
2773 err = __blk_mq_alloc_rq_maps(set);
2774 if (!err)
2775 break;
2776
2777 set->queue_depth >>= 1;
2778 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2779 err = -ENOMEM;
2780 break;
2781 }
2782 } while (set->queue_depth);
2783
2784 if (!set->queue_depth || err) {
2785 pr_err("blk-mq: failed to allocate request map\n");
2786 return -ENOMEM;
2787 }
2788
2789 if (depth != set->queue_depth)
2790 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2791 depth, set->queue_depth);
2792
2793 return 0;
2794}
2795
2796static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2797{
2798 if (set->ops->map_queues) {
2799 int cpu;
2800 /*
2801 * transport .map_queues is usually done in the following
2802 * way:
2803 *
2804 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2805 * mask = get_cpu_mask(queue)
2806 * for_each_cpu(cpu, mask)
2807 * set->mq_map[cpu] = queue;
2808 * }
2809 *
2810 * When we need to remap, the table has to be cleared for
2811 * killing stale mapping since one CPU may not be mapped
2812 * to any hw queue.
2813 */
2814 for_each_possible_cpu(cpu)
2815 set->mq_map[cpu] = 0;
2816
2817 return set->ops->map_queues(set);
2818 } else
2819 return blk_mq_map_queues(set);
2820}
2821
2822/*
2823 * Alloc a tag set to be associated with one or more request queues.
2824 * May fail with EINVAL for various error conditions. May adjust the
2825 * requested depth down, if if it too large. In that case, the set
2826 * value will be stored in set->queue_depth.
2827 */
2828int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2829{
2830 int ret;
2831
2832 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2833
2834 if (!set->nr_hw_queues)
2835 return -EINVAL;
2836 if (!set->queue_depth)
2837 return -EINVAL;
2838 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2839 return -EINVAL;
2840
2841 if (!set->ops->queue_rq)
2842 return -EINVAL;
2843
2844 if (!set->ops->get_budget ^ !set->ops->put_budget)
2845 return -EINVAL;
2846
2847 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2848 pr_info("blk-mq: reduced tag depth to %u\n",
2849 BLK_MQ_MAX_DEPTH);
2850 set->queue_depth = BLK_MQ_MAX_DEPTH;
2851 }
2852
2853 /*
2854 * If a crashdump is active, then we are potentially in a very
2855 * memory constrained environment. Limit us to 1 queue and
2856 * 64 tags to prevent using too much memory.
2857 */
2858 if (is_kdump_kernel()) {
2859 set->nr_hw_queues = 1;
2860 set->queue_depth = min(64U, set->queue_depth);
2861 }
2862 /*
2863 * There is no use for more h/w queues than cpus.
2864 */
2865 if (set->nr_hw_queues > nr_cpu_ids)
2866 set->nr_hw_queues = nr_cpu_ids;
2867
2868 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2869 GFP_KERNEL, set->numa_node);
2870 if (!set->tags)
2871 return -ENOMEM;
2872
2873 ret = -ENOMEM;
2874 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2875 GFP_KERNEL, set->numa_node);
2876 if (!set->mq_map)
2877 goto out_free_tags;
2878
2879 ret = blk_mq_update_queue_map(set);
2880 if (ret)
2881 goto out_free_mq_map;
2882
2883 ret = blk_mq_alloc_rq_maps(set);
2884 if (ret)
2885 goto out_free_mq_map;
2886
2887 mutex_init(&set->tag_list_lock);
2888 INIT_LIST_HEAD(&set->tag_list);
2889
2890 return 0;
2891
2892out_free_mq_map:
2893 kfree(set->mq_map);
2894 set->mq_map = NULL;
2895out_free_tags:
2896 kfree(set->tags);
2897 set->tags = NULL;
2898 return ret;
2899}
2900EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2901
2902void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2903{
2904 int i;
2905
2906 for (i = 0; i < nr_cpu_ids; i++)
2907 blk_mq_free_map_and_requests(set, i);
2908
2909 kfree(set->mq_map);
2910 set->mq_map = NULL;
2911
2912 kfree(set->tags);
2913 set->tags = NULL;
2914}
2915EXPORT_SYMBOL(blk_mq_free_tag_set);
2916
2917int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2918{
2919 struct blk_mq_tag_set *set = q->tag_set;
2920 struct blk_mq_hw_ctx *hctx;
2921 int i, ret;
2922
2923 if (!set)
2924 return -EINVAL;
2925
2926 blk_mq_freeze_queue(q);
2927 blk_mq_quiesce_queue(q);
2928
2929 ret = 0;
2930 queue_for_each_hw_ctx(q, hctx, i) {
2931 if (!hctx->tags)
2932 continue;
2933 /*
2934 * If we're using an MQ scheduler, just update the scheduler
2935 * queue depth. This is similar to what the old code would do.
2936 */
2937 if (!hctx->sched_tags) {
2938 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
2939 false);
2940 } else {
2941 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2942 nr, true);
2943 }
2944 if (ret)
2945 break;
2946 }
2947
2948 if (!ret)
2949 q->nr_requests = nr;
2950
2951 blk_mq_unquiesce_queue(q);
2952 blk_mq_unfreeze_queue(q);
2953
2954 return ret;
2955}
2956
2957static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2958 int nr_hw_queues)
2959{
2960 struct request_queue *q;
2961
2962 lockdep_assert_held(&set->tag_list_lock);
2963
2964 if (nr_hw_queues > nr_cpu_ids)
2965 nr_hw_queues = nr_cpu_ids;
2966 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2967 return;
2968
2969 list_for_each_entry(q, &set->tag_list, tag_set_list)
2970 blk_mq_freeze_queue(q);
2971
2972 set->nr_hw_queues = nr_hw_queues;
2973 blk_mq_update_queue_map(set);
2974 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2975 blk_mq_realloc_hw_ctxs(set, q);
2976 blk_mq_queue_reinit(q);
2977 }
2978
2979 list_for_each_entry(q, &set->tag_list, tag_set_list)
2980 blk_mq_unfreeze_queue(q);
2981}
2982
2983void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2984{
2985 mutex_lock(&set->tag_list_lock);
2986 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2987 mutex_unlock(&set->tag_list_lock);
2988}
2989EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2990
2991/* Enable polling stats and return whether they were already enabled. */
2992static bool blk_poll_stats_enable(struct request_queue *q)
2993{
2994 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2995 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
2996 return true;
2997 blk_stat_add_callback(q, q->poll_cb);
2998 return false;
2999}
3000
3001static void blk_mq_poll_stats_start(struct request_queue *q)
3002{
3003 /*
3004 * We don't arm the callback if polling stats are not enabled or the
3005 * callback is already active.
3006 */
3007 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3008 blk_stat_is_active(q->poll_cb))
3009 return;
3010
3011 blk_stat_activate_msecs(q->poll_cb, 100);
3012}
3013
3014static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3015{
3016 struct request_queue *q = cb->data;
3017 int bucket;
3018
3019 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3020 if (cb->stat[bucket].nr_samples)
3021 q->poll_stat[bucket] = cb->stat[bucket];
3022 }
3023}
3024
3025static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3026 struct blk_mq_hw_ctx *hctx,
3027 struct request *rq)
3028{
3029 unsigned long ret = 0;
3030 int bucket;
3031
3032 /*
3033 * If stats collection isn't on, don't sleep but turn it on for
3034 * future users
3035 */
3036 if (!blk_poll_stats_enable(q))
3037 return 0;
3038
3039 /*
3040 * As an optimistic guess, use half of the mean service time
3041 * for this type of request. We can (and should) make this smarter.
3042 * For instance, if the completion latencies are tight, we can
3043 * get closer than just half the mean. This is especially
3044 * important on devices where the completion latencies are longer
3045 * than ~10 usec. We do use the stats for the relevant IO size
3046 * if available which does lead to better estimates.
3047 */
3048 bucket = blk_mq_poll_stats_bkt(rq);
3049 if (bucket < 0)
3050 return ret;
3051
3052 if (q->poll_stat[bucket].nr_samples)
3053 ret = (q->poll_stat[bucket].mean + 1) / 2;
3054
3055 return ret;
3056}
3057
3058static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3059 struct blk_mq_hw_ctx *hctx,
3060 struct request *rq)
3061{
3062 struct hrtimer_sleeper hs;
3063 enum hrtimer_mode mode;
3064 unsigned int nsecs;
3065 ktime_t kt;
3066
3067 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3068 return false;
3069
3070 /*
3071 * poll_nsec can be:
3072 *
3073 * -1: don't ever hybrid sleep
3074 * 0: use half of prev avg
3075 * >0: use this specific value
3076 */
3077 if (q->poll_nsec == -1)
3078 return false;
3079 else if (q->poll_nsec > 0)
3080 nsecs = q->poll_nsec;
3081 else
3082 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
3083
3084 if (!nsecs)
3085 return false;
3086
3087 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3088
3089 /*
3090 * This will be replaced with the stats tracking code, using
3091 * 'avg_completion_time / 2' as the pre-sleep target.
3092 */
3093 kt = nsecs;
3094
3095 mode = HRTIMER_MODE_REL;
3096 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
3097 hrtimer_set_expires(&hs.timer, kt);
3098
3099 hrtimer_init_sleeper(&hs, current);
3100 do {
3101 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3102 break;
3103 set_current_state(TASK_UNINTERRUPTIBLE);
3104 hrtimer_start_expires(&hs.timer, mode);
3105 if (hs.task)
3106 io_schedule();
3107 hrtimer_cancel(&hs.timer);
3108 mode = HRTIMER_MODE_ABS;
3109 } while (hs.task && !signal_pending(current));
3110
3111 __set_current_state(TASK_RUNNING);
3112 destroy_hrtimer_on_stack(&hs.timer);
3113 return true;
3114}
3115
3116static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
3117{
3118 struct request_queue *q = hctx->queue;
3119 long state;
3120
3121 /*
3122 * If we sleep, have the caller restart the poll loop to reset
3123 * the state. Like for the other success return cases, the
3124 * caller is responsible for checking if the IO completed. If
3125 * the IO isn't complete, we'll get called again and will go
3126 * straight to the busy poll loop.
3127 */
3128 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
3129 return true;
3130
3131 hctx->poll_considered++;
3132
3133 state = current->state;
3134 while (!need_resched()) {
3135 int ret;
3136
3137 hctx->poll_invoked++;
3138
3139 ret = q->mq_ops->poll(hctx, rq->tag);
3140 if (ret > 0) {
3141 hctx->poll_success++;
3142 set_current_state(TASK_RUNNING);
3143 return true;
3144 }
3145
3146 if (signal_pending_state(state, current))
3147 set_current_state(TASK_RUNNING);
3148
3149 if (current->state == TASK_RUNNING)
3150 return true;
3151 if (ret < 0)
3152 break;
3153 cpu_relax();
3154 }
3155
3156 __set_current_state(TASK_RUNNING);
3157 return false;
3158}
3159
3160static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
3161{
3162 struct blk_mq_hw_ctx *hctx;
3163 struct request *rq;
3164
3165 if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3166 return false;
3167
3168 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3169 if (!blk_qc_t_is_internal(cookie))
3170 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3171 else {
3172 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3173 /*
3174 * With scheduling, if the request has completed, we'll
3175 * get a NULL return here, as we clear the sched tag when
3176 * that happens. The request still remains valid, like always,
3177 * so we should be safe with just the NULL check.
3178 */
3179 if (!rq)
3180 return false;
3181 }
3182
3183 return __blk_mq_poll(hctx, rq);
3184}
3185
3186static int __init blk_mq_init(void)
3187{
3188 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3189 blk_mq_hctx_notify_dead);
3190 return 0;
3191}
3192subsys_initcall(blk_mq_init);
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Block multiqueue core code
4 *
5 * Copyright (C) 2013-2014 Jens Axboe
6 * Copyright (C) 2013-2014 Christoph Hellwig
7 */
8#include <linux/kernel.h>
9#include <linux/module.h>
10#include <linux/backing-dev.h>
11#include <linux/bio.h>
12#include <linux/blkdev.h>
13#include <linux/kmemleak.h>
14#include <linux/mm.h>
15#include <linux/init.h>
16#include <linux/slab.h>
17#include <linux/workqueue.h>
18#include <linux/smp.h>
19#include <linux/llist.h>
20#include <linux/list_sort.h>
21#include <linux/cpu.h>
22#include <linux/cache.h>
23#include <linux/sched/sysctl.h>
24#include <linux/sched/topology.h>
25#include <linux/sched/signal.h>
26#include <linux/delay.h>
27#include <linux/crash_dump.h>
28#include <linux/prefetch.h>
29#include <linux/blk-crypto.h>
30
31#include <trace/events/block.h>
32
33#include <linux/blk-mq.h>
34#include <linux/t10-pi.h>
35#include "blk.h"
36#include "blk-mq.h"
37#include "blk-mq-debugfs.h"
38#include "blk-mq-tag.h"
39#include "blk-pm.h"
40#include "blk-stat.h"
41#include "blk-mq-sched.h"
42#include "blk-rq-qos.h"
43
44static DEFINE_PER_CPU(struct list_head, blk_cpu_done);
45
46static void blk_mq_poll_stats_start(struct request_queue *q);
47static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
48
49static int blk_mq_poll_stats_bkt(const struct request *rq)
50{
51 int ddir, sectors, bucket;
52
53 ddir = rq_data_dir(rq);
54 sectors = blk_rq_stats_sectors(rq);
55
56 bucket = ddir + 2 * ilog2(sectors);
57
58 if (bucket < 0)
59 return -1;
60 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
61 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
62
63 return bucket;
64}
65
66/*
67 * Check if any of the ctx, dispatch list or elevator
68 * have pending work in this hardware queue.
69 */
70static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
71{
72 return !list_empty_careful(&hctx->dispatch) ||
73 sbitmap_any_bit_set(&hctx->ctx_map) ||
74 blk_mq_sched_has_work(hctx);
75}
76
77/*
78 * Mark this ctx as having pending work in this hardware queue
79 */
80static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
81 struct blk_mq_ctx *ctx)
82{
83 const int bit = ctx->index_hw[hctx->type];
84
85 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
86 sbitmap_set_bit(&hctx->ctx_map, bit);
87}
88
89static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
90 struct blk_mq_ctx *ctx)
91{
92 const int bit = ctx->index_hw[hctx->type];
93
94 sbitmap_clear_bit(&hctx->ctx_map, bit);
95}
96
97struct mq_inflight {
98 struct hd_struct *part;
99 unsigned int inflight[2];
100};
101
102static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
103 struct request *rq, void *priv,
104 bool reserved)
105{
106 struct mq_inflight *mi = priv;
107
108 if (rq->part == mi->part)
109 mi->inflight[rq_data_dir(rq)]++;
110
111 return true;
112}
113
114unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part)
115{
116 struct mq_inflight mi = { .part = part };
117
118 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
119
120 return mi.inflight[0] + mi.inflight[1];
121}
122
123void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
124 unsigned int inflight[2])
125{
126 struct mq_inflight mi = { .part = part };
127
128 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
129 inflight[0] = mi.inflight[0];
130 inflight[1] = mi.inflight[1];
131}
132
133void blk_freeze_queue_start(struct request_queue *q)
134{
135 mutex_lock(&q->mq_freeze_lock);
136 if (++q->mq_freeze_depth == 1) {
137 percpu_ref_kill(&q->q_usage_counter);
138 mutex_unlock(&q->mq_freeze_lock);
139 if (queue_is_mq(q))
140 blk_mq_run_hw_queues(q, false);
141 } else {
142 mutex_unlock(&q->mq_freeze_lock);
143 }
144}
145EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
146
147void blk_mq_freeze_queue_wait(struct request_queue *q)
148{
149 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
150}
151EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
152
153int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
154 unsigned long timeout)
155{
156 return wait_event_timeout(q->mq_freeze_wq,
157 percpu_ref_is_zero(&q->q_usage_counter),
158 timeout);
159}
160EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
161
162/*
163 * Guarantee no request is in use, so we can change any data structure of
164 * the queue afterward.
165 */
166void blk_freeze_queue(struct request_queue *q)
167{
168 /*
169 * In the !blk_mq case we are only calling this to kill the
170 * q_usage_counter, otherwise this increases the freeze depth
171 * and waits for it to return to zero. For this reason there is
172 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
173 * exported to drivers as the only user for unfreeze is blk_mq.
174 */
175 blk_freeze_queue_start(q);
176 blk_mq_freeze_queue_wait(q);
177}
178
179void blk_mq_freeze_queue(struct request_queue *q)
180{
181 /*
182 * ...just an alias to keep freeze and unfreeze actions balanced
183 * in the blk_mq_* namespace
184 */
185 blk_freeze_queue(q);
186}
187EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
188
189void blk_mq_unfreeze_queue(struct request_queue *q)
190{
191 mutex_lock(&q->mq_freeze_lock);
192 q->mq_freeze_depth--;
193 WARN_ON_ONCE(q->mq_freeze_depth < 0);
194 if (!q->mq_freeze_depth) {
195 percpu_ref_resurrect(&q->q_usage_counter);
196 wake_up_all(&q->mq_freeze_wq);
197 }
198 mutex_unlock(&q->mq_freeze_lock);
199}
200EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
201
202/*
203 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
204 * mpt3sas driver such that this function can be removed.
205 */
206void blk_mq_quiesce_queue_nowait(struct request_queue *q)
207{
208 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
209}
210EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
211
212/**
213 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
214 * @q: request queue.
215 *
216 * Note: this function does not prevent that the struct request end_io()
217 * callback function is invoked. Once this function is returned, we make
218 * sure no dispatch can happen until the queue is unquiesced via
219 * blk_mq_unquiesce_queue().
220 */
221void blk_mq_quiesce_queue(struct request_queue *q)
222{
223 struct blk_mq_hw_ctx *hctx;
224 unsigned int i;
225 bool rcu = false;
226
227 blk_mq_quiesce_queue_nowait(q);
228
229 queue_for_each_hw_ctx(q, hctx, i) {
230 if (hctx->flags & BLK_MQ_F_BLOCKING)
231 synchronize_srcu(hctx->srcu);
232 else
233 rcu = true;
234 }
235 if (rcu)
236 synchronize_rcu();
237}
238EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
239
240/*
241 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
242 * @q: request queue.
243 *
244 * This function recovers queue into the state before quiescing
245 * which is done by blk_mq_quiesce_queue.
246 */
247void blk_mq_unquiesce_queue(struct request_queue *q)
248{
249 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
250
251 /* dispatch requests which are inserted during quiescing */
252 blk_mq_run_hw_queues(q, true);
253}
254EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
255
256void blk_mq_wake_waiters(struct request_queue *q)
257{
258 struct blk_mq_hw_ctx *hctx;
259 unsigned int i;
260
261 queue_for_each_hw_ctx(q, hctx, i)
262 if (blk_mq_hw_queue_mapped(hctx))
263 blk_mq_tag_wakeup_all(hctx->tags, true);
264}
265
266/*
267 * Only need start/end time stamping if we have iostat or
268 * blk stats enabled, or using an IO scheduler.
269 */
270static inline bool blk_mq_need_time_stamp(struct request *rq)
271{
272 return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator;
273}
274
275static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
276 unsigned int tag, u64 alloc_time_ns)
277{
278 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
279 struct request *rq = tags->static_rqs[tag];
280
281 if (data->q->elevator) {
282 rq->tag = BLK_MQ_NO_TAG;
283 rq->internal_tag = tag;
284 } else {
285 rq->tag = tag;
286 rq->internal_tag = BLK_MQ_NO_TAG;
287 }
288
289 /* csd/requeue_work/fifo_time is initialized before use */
290 rq->q = data->q;
291 rq->mq_ctx = data->ctx;
292 rq->mq_hctx = data->hctx;
293 rq->rq_flags = 0;
294 rq->cmd_flags = data->cmd_flags;
295 if (data->flags & BLK_MQ_REQ_PREEMPT)
296 rq->rq_flags |= RQF_PREEMPT;
297 if (blk_queue_io_stat(data->q))
298 rq->rq_flags |= RQF_IO_STAT;
299 INIT_LIST_HEAD(&rq->queuelist);
300 INIT_HLIST_NODE(&rq->hash);
301 RB_CLEAR_NODE(&rq->rb_node);
302 rq->rq_disk = NULL;
303 rq->part = NULL;
304#ifdef CONFIG_BLK_RQ_ALLOC_TIME
305 rq->alloc_time_ns = alloc_time_ns;
306#endif
307 if (blk_mq_need_time_stamp(rq))
308 rq->start_time_ns = ktime_get_ns();
309 else
310 rq->start_time_ns = 0;
311 rq->io_start_time_ns = 0;
312 rq->stats_sectors = 0;
313 rq->nr_phys_segments = 0;
314#if defined(CONFIG_BLK_DEV_INTEGRITY)
315 rq->nr_integrity_segments = 0;
316#endif
317 blk_crypto_rq_set_defaults(rq);
318 /* tag was already set */
319 WRITE_ONCE(rq->deadline, 0);
320
321 rq->timeout = 0;
322
323 rq->end_io = NULL;
324 rq->end_io_data = NULL;
325
326 data->ctx->rq_dispatched[op_is_sync(data->cmd_flags)]++;
327 refcount_set(&rq->ref, 1);
328
329 if (!op_is_flush(data->cmd_flags)) {
330 struct elevator_queue *e = data->q->elevator;
331
332 rq->elv.icq = NULL;
333 if (e && e->type->ops.prepare_request) {
334 if (e->type->icq_cache)
335 blk_mq_sched_assign_ioc(rq);
336
337 e->type->ops.prepare_request(rq);
338 rq->rq_flags |= RQF_ELVPRIV;
339 }
340 }
341
342 data->hctx->queued++;
343 return rq;
344}
345
346static struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data)
347{
348 struct request_queue *q = data->q;
349 struct elevator_queue *e = q->elevator;
350 u64 alloc_time_ns = 0;
351 unsigned int tag;
352
353 /* alloc_time includes depth and tag waits */
354 if (blk_queue_rq_alloc_time(q))
355 alloc_time_ns = ktime_get_ns();
356
357 if (data->cmd_flags & REQ_NOWAIT)
358 data->flags |= BLK_MQ_REQ_NOWAIT;
359
360 if (e) {
361 /*
362 * Flush requests are special and go directly to the
363 * dispatch list. Don't include reserved tags in the
364 * limiting, as it isn't useful.
365 */
366 if (!op_is_flush(data->cmd_flags) &&
367 e->type->ops.limit_depth &&
368 !(data->flags & BLK_MQ_REQ_RESERVED))
369 e->type->ops.limit_depth(data->cmd_flags, data);
370 }
371
372retry:
373 data->ctx = blk_mq_get_ctx(q);
374 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
375 if (!e)
376 blk_mq_tag_busy(data->hctx);
377
378 /*
379 * Waiting allocations only fail because of an inactive hctx. In that
380 * case just retry the hctx assignment and tag allocation as CPU hotplug
381 * should have migrated us to an online CPU by now.
382 */
383 tag = blk_mq_get_tag(data);
384 if (tag == BLK_MQ_NO_TAG) {
385 if (data->flags & BLK_MQ_REQ_NOWAIT)
386 return NULL;
387
388 /*
389 * Give up the CPU and sleep for a random short time to ensure
390 * that thread using a realtime scheduling class are migrated
391 * off the CPU, and thus off the hctx that is going away.
392 */
393 msleep(3);
394 goto retry;
395 }
396 return blk_mq_rq_ctx_init(data, tag, alloc_time_ns);
397}
398
399struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
400 blk_mq_req_flags_t flags)
401{
402 struct blk_mq_alloc_data data = {
403 .q = q,
404 .flags = flags,
405 .cmd_flags = op,
406 };
407 struct request *rq;
408 int ret;
409
410 ret = blk_queue_enter(q, flags);
411 if (ret)
412 return ERR_PTR(ret);
413
414 rq = __blk_mq_alloc_request(&data);
415 if (!rq)
416 goto out_queue_exit;
417 rq->__data_len = 0;
418 rq->__sector = (sector_t) -1;
419 rq->bio = rq->biotail = NULL;
420 return rq;
421out_queue_exit:
422 blk_queue_exit(q);
423 return ERR_PTR(-EWOULDBLOCK);
424}
425EXPORT_SYMBOL(blk_mq_alloc_request);
426
427struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
428 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
429{
430 struct blk_mq_alloc_data data = {
431 .q = q,
432 .flags = flags,
433 .cmd_flags = op,
434 };
435 u64 alloc_time_ns = 0;
436 unsigned int cpu;
437 unsigned int tag;
438 int ret;
439
440 /* alloc_time includes depth and tag waits */
441 if (blk_queue_rq_alloc_time(q))
442 alloc_time_ns = ktime_get_ns();
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 | BLK_MQ_REQ_RESERVED))))
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 ret = -EXDEV;
465 data.hctx = q->queue_hw_ctx[hctx_idx];
466 if (!blk_mq_hw_queue_mapped(data.hctx))
467 goto out_queue_exit;
468 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
469 data.ctx = __blk_mq_get_ctx(q, cpu);
470
471 if (!q->elevator)
472 blk_mq_tag_busy(data.hctx);
473
474 ret = -EWOULDBLOCK;
475 tag = blk_mq_get_tag(&data);
476 if (tag == BLK_MQ_NO_TAG)
477 goto out_queue_exit;
478 return blk_mq_rq_ctx_init(&data, tag, alloc_time_ns);
479
480out_queue_exit:
481 blk_queue_exit(q);
482 return ERR_PTR(ret);
483}
484EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
485
486static void __blk_mq_free_request(struct request *rq)
487{
488 struct request_queue *q = rq->q;
489 struct blk_mq_ctx *ctx = rq->mq_ctx;
490 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
491 const int sched_tag = rq->internal_tag;
492
493 blk_crypto_free_request(rq);
494 blk_pm_mark_last_busy(rq);
495 rq->mq_hctx = NULL;
496 if (rq->tag != BLK_MQ_NO_TAG)
497 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
498 if (sched_tag != BLK_MQ_NO_TAG)
499 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
500 blk_mq_sched_restart(hctx);
501 blk_queue_exit(q);
502}
503
504void blk_mq_free_request(struct request *rq)
505{
506 struct request_queue *q = rq->q;
507 struct elevator_queue *e = q->elevator;
508 struct blk_mq_ctx *ctx = rq->mq_ctx;
509 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
510
511 if (rq->rq_flags & RQF_ELVPRIV) {
512 if (e && e->type->ops.finish_request)
513 e->type->ops.finish_request(rq);
514 if (rq->elv.icq) {
515 put_io_context(rq->elv.icq->ioc);
516 rq->elv.icq = NULL;
517 }
518 }
519
520 ctx->rq_completed[rq_is_sync(rq)]++;
521 if (rq->rq_flags & RQF_MQ_INFLIGHT)
522 atomic_dec(&hctx->nr_active);
523
524 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
525 laptop_io_completion(q->backing_dev_info);
526
527 rq_qos_done(q, rq);
528
529 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
530 if (refcount_dec_and_test(&rq->ref))
531 __blk_mq_free_request(rq);
532}
533EXPORT_SYMBOL_GPL(blk_mq_free_request);
534
535inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
536{
537 u64 now = 0;
538
539 if (blk_mq_need_time_stamp(rq))
540 now = ktime_get_ns();
541
542 if (rq->rq_flags & RQF_STATS) {
543 blk_mq_poll_stats_start(rq->q);
544 blk_stat_add(rq, now);
545 }
546
547 blk_mq_sched_completed_request(rq, now);
548
549 blk_account_io_done(rq, now);
550
551 if (rq->end_io) {
552 rq_qos_done(rq->q, rq);
553 rq->end_io(rq, error);
554 } else {
555 blk_mq_free_request(rq);
556 }
557}
558EXPORT_SYMBOL(__blk_mq_end_request);
559
560void blk_mq_end_request(struct request *rq, blk_status_t error)
561{
562 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
563 BUG();
564 __blk_mq_end_request(rq, error);
565}
566EXPORT_SYMBOL(blk_mq_end_request);
567
568/*
569 * Softirq action handler - move entries to local list and loop over them
570 * while passing them to the queue registered handler.
571 */
572static __latent_entropy void blk_done_softirq(struct softirq_action *h)
573{
574 struct list_head *cpu_list, local_list;
575
576 local_irq_disable();
577 cpu_list = this_cpu_ptr(&blk_cpu_done);
578 list_replace_init(cpu_list, &local_list);
579 local_irq_enable();
580
581 while (!list_empty(&local_list)) {
582 struct request *rq;
583
584 rq = list_entry(local_list.next, struct request, ipi_list);
585 list_del_init(&rq->ipi_list);
586 rq->q->mq_ops->complete(rq);
587 }
588}
589
590static void blk_mq_trigger_softirq(struct request *rq)
591{
592 struct list_head *list;
593 unsigned long flags;
594
595 local_irq_save(flags);
596 list = this_cpu_ptr(&blk_cpu_done);
597 list_add_tail(&rq->ipi_list, list);
598
599 /*
600 * If the list only contains our just added request, signal a raise of
601 * the softirq. If there are already entries there, someone already
602 * raised the irq but it hasn't run yet.
603 */
604 if (list->next == &rq->ipi_list)
605 raise_softirq_irqoff(BLOCK_SOFTIRQ);
606 local_irq_restore(flags);
607}
608
609static int blk_softirq_cpu_dead(unsigned int cpu)
610{
611 /*
612 * If a CPU goes away, splice its entries to the current CPU
613 * and trigger a run of the softirq
614 */
615 local_irq_disable();
616 list_splice_init(&per_cpu(blk_cpu_done, cpu),
617 this_cpu_ptr(&blk_cpu_done));
618 raise_softirq_irqoff(BLOCK_SOFTIRQ);
619 local_irq_enable();
620
621 return 0;
622}
623
624
625static void __blk_mq_complete_request_remote(void *data)
626{
627 struct request *rq = data;
628
629 /*
630 * For most of single queue controllers, there is only one irq vector
631 * for handling I/O completion, and the only irq's affinity is set
632 * to all possible CPUs. On most of ARCHs, this affinity means the irq
633 * is handled on one specific CPU.
634 *
635 * So complete I/O requests in softirq context in case of single queue
636 * devices to avoid degrading I/O performance due to irqsoff latency.
637 */
638 if (rq->q->nr_hw_queues == 1)
639 blk_mq_trigger_softirq(rq);
640 else
641 rq->q->mq_ops->complete(rq);
642}
643
644static inline bool blk_mq_complete_need_ipi(struct request *rq)
645{
646 int cpu = raw_smp_processor_id();
647
648 if (!IS_ENABLED(CONFIG_SMP) ||
649 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
650 return false;
651
652 /* same CPU or cache domain? Complete locally */
653 if (cpu == rq->mq_ctx->cpu ||
654 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
655 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
656 return false;
657
658 /* don't try to IPI to an offline CPU */
659 return cpu_online(rq->mq_ctx->cpu);
660}
661
662bool blk_mq_complete_request_remote(struct request *rq)
663{
664 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
665
666 /*
667 * For a polled request, always complete locallly, it's pointless
668 * to redirect the completion.
669 */
670 if (rq->cmd_flags & REQ_HIPRI)
671 return false;
672
673 if (blk_mq_complete_need_ipi(rq)) {
674 rq->csd.func = __blk_mq_complete_request_remote;
675 rq->csd.info = rq;
676 rq->csd.flags = 0;
677 smp_call_function_single_async(rq->mq_ctx->cpu, &rq->csd);
678 } else {
679 if (rq->q->nr_hw_queues > 1)
680 return false;
681 blk_mq_trigger_softirq(rq);
682 }
683
684 return true;
685}
686EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
687
688/**
689 * blk_mq_complete_request - end I/O on a request
690 * @rq: the request being processed
691 *
692 * Description:
693 * Complete a request by scheduling the ->complete_rq operation.
694 **/
695void blk_mq_complete_request(struct request *rq)
696{
697 if (!blk_mq_complete_request_remote(rq))
698 rq->q->mq_ops->complete(rq);
699}
700EXPORT_SYMBOL(blk_mq_complete_request);
701
702static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
703 __releases(hctx->srcu)
704{
705 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
706 rcu_read_unlock();
707 else
708 srcu_read_unlock(hctx->srcu, srcu_idx);
709}
710
711static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
712 __acquires(hctx->srcu)
713{
714 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
715 /* shut up gcc false positive */
716 *srcu_idx = 0;
717 rcu_read_lock();
718 } else
719 *srcu_idx = srcu_read_lock(hctx->srcu);
720}
721
722/**
723 * blk_mq_start_request - Start processing a request
724 * @rq: Pointer to request to be started
725 *
726 * Function used by device drivers to notify the block layer that a request
727 * is going to be processed now, so blk layer can do proper initializations
728 * such as starting the timeout timer.
729 */
730void blk_mq_start_request(struct request *rq)
731{
732 struct request_queue *q = rq->q;
733
734 trace_block_rq_issue(q, rq);
735
736 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
737 rq->io_start_time_ns = ktime_get_ns();
738 rq->stats_sectors = blk_rq_sectors(rq);
739 rq->rq_flags |= RQF_STATS;
740 rq_qos_issue(q, rq);
741 }
742
743 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
744
745 blk_add_timer(rq);
746 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
747
748#ifdef CONFIG_BLK_DEV_INTEGRITY
749 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
750 q->integrity.profile->prepare_fn(rq);
751#endif
752}
753EXPORT_SYMBOL(blk_mq_start_request);
754
755static void __blk_mq_requeue_request(struct request *rq)
756{
757 struct request_queue *q = rq->q;
758
759 blk_mq_put_driver_tag(rq);
760
761 trace_block_rq_requeue(q, rq);
762 rq_qos_requeue(q, rq);
763
764 if (blk_mq_request_started(rq)) {
765 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
766 rq->rq_flags &= ~RQF_TIMED_OUT;
767 }
768}
769
770void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
771{
772 __blk_mq_requeue_request(rq);
773
774 /* this request will be re-inserted to io scheduler queue */
775 blk_mq_sched_requeue_request(rq);
776
777 BUG_ON(!list_empty(&rq->queuelist));
778 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
779}
780EXPORT_SYMBOL(blk_mq_requeue_request);
781
782static void blk_mq_requeue_work(struct work_struct *work)
783{
784 struct request_queue *q =
785 container_of(work, struct request_queue, requeue_work.work);
786 LIST_HEAD(rq_list);
787 struct request *rq, *next;
788
789 spin_lock_irq(&q->requeue_lock);
790 list_splice_init(&q->requeue_list, &rq_list);
791 spin_unlock_irq(&q->requeue_lock);
792
793 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
794 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
795 continue;
796
797 rq->rq_flags &= ~RQF_SOFTBARRIER;
798 list_del_init(&rq->queuelist);
799 /*
800 * If RQF_DONTPREP, rq has contained some driver specific
801 * data, so insert it to hctx dispatch list to avoid any
802 * merge.
803 */
804 if (rq->rq_flags & RQF_DONTPREP)
805 blk_mq_request_bypass_insert(rq, false, false);
806 else
807 blk_mq_sched_insert_request(rq, true, false, false);
808 }
809
810 while (!list_empty(&rq_list)) {
811 rq = list_entry(rq_list.next, struct request, queuelist);
812 list_del_init(&rq->queuelist);
813 blk_mq_sched_insert_request(rq, false, false, false);
814 }
815
816 blk_mq_run_hw_queues(q, false);
817}
818
819void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
820 bool kick_requeue_list)
821{
822 struct request_queue *q = rq->q;
823 unsigned long flags;
824
825 /*
826 * We abuse this flag that is otherwise used by the I/O scheduler to
827 * request head insertion from the workqueue.
828 */
829 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
830
831 spin_lock_irqsave(&q->requeue_lock, flags);
832 if (at_head) {
833 rq->rq_flags |= RQF_SOFTBARRIER;
834 list_add(&rq->queuelist, &q->requeue_list);
835 } else {
836 list_add_tail(&rq->queuelist, &q->requeue_list);
837 }
838 spin_unlock_irqrestore(&q->requeue_lock, flags);
839
840 if (kick_requeue_list)
841 blk_mq_kick_requeue_list(q);
842}
843
844void blk_mq_kick_requeue_list(struct request_queue *q)
845{
846 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
847}
848EXPORT_SYMBOL(blk_mq_kick_requeue_list);
849
850void blk_mq_delay_kick_requeue_list(struct request_queue *q,
851 unsigned long msecs)
852{
853 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
854 msecs_to_jiffies(msecs));
855}
856EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
857
858struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
859{
860 if (tag < tags->nr_tags) {
861 prefetch(tags->rqs[tag]);
862 return tags->rqs[tag];
863 }
864
865 return NULL;
866}
867EXPORT_SYMBOL(blk_mq_tag_to_rq);
868
869static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
870 void *priv, bool reserved)
871{
872 /*
873 * If we find a request that isn't idle and the queue matches,
874 * we know the queue is busy. Return false to stop the iteration.
875 */
876 if (blk_mq_request_started(rq) && rq->q == hctx->queue) {
877 bool *busy = priv;
878
879 *busy = true;
880 return false;
881 }
882
883 return true;
884}
885
886bool blk_mq_queue_inflight(struct request_queue *q)
887{
888 bool busy = false;
889
890 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
891 return busy;
892}
893EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
894
895static void blk_mq_rq_timed_out(struct request *req, bool reserved)
896{
897 req->rq_flags |= RQF_TIMED_OUT;
898 if (req->q->mq_ops->timeout) {
899 enum blk_eh_timer_return ret;
900
901 ret = req->q->mq_ops->timeout(req, reserved);
902 if (ret == BLK_EH_DONE)
903 return;
904 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
905 }
906
907 blk_add_timer(req);
908}
909
910static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
911{
912 unsigned long deadline;
913
914 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
915 return false;
916 if (rq->rq_flags & RQF_TIMED_OUT)
917 return false;
918
919 deadline = READ_ONCE(rq->deadline);
920 if (time_after_eq(jiffies, deadline))
921 return true;
922
923 if (*next == 0)
924 *next = deadline;
925 else if (time_after(*next, deadline))
926 *next = deadline;
927 return false;
928}
929
930static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
931 struct request *rq, void *priv, bool reserved)
932{
933 unsigned long *next = priv;
934
935 /*
936 * Just do a quick check if it is expired before locking the request in
937 * so we're not unnecessarilly synchronizing across CPUs.
938 */
939 if (!blk_mq_req_expired(rq, next))
940 return true;
941
942 /*
943 * We have reason to believe the request may be expired. Take a
944 * reference on the request to lock this request lifetime into its
945 * currently allocated context to prevent it from being reallocated in
946 * the event the completion by-passes this timeout handler.
947 *
948 * If the reference was already released, then the driver beat the
949 * timeout handler to posting a natural completion.
950 */
951 if (!refcount_inc_not_zero(&rq->ref))
952 return true;
953
954 /*
955 * The request is now locked and cannot be reallocated underneath the
956 * timeout handler's processing. Re-verify this exact request is truly
957 * expired; if it is not expired, then the request was completed and
958 * reallocated as a new request.
959 */
960 if (blk_mq_req_expired(rq, next))
961 blk_mq_rq_timed_out(rq, reserved);
962
963 if (is_flush_rq(rq, hctx))
964 rq->end_io(rq, 0);
965 else if (refcount_dec_and_test(&rq->ref))
966 __blk_mq_free_request(rq);
967
968 return true;
969}
970
971static void blk_mq_timeout_work(struct work_struct *work)
972{
973 struct request_queue *q =
974 container_of(work, struct request_queue, timeout_work);
975 unsigned long next = 0;
976 struct blk_mq_hw_ctx *hctx;
977 int i;
978
979 /* A deadlock might occur if a request is stuck requiring a
980 * timeout at the same time a queue freeze is waiting
981 * completion, since the timeout code would not be able to
982 * acquire the queue reference here.
983 *
984 * That's why we don't use blk_queue_enter here; instead, we use
985 * percpu_ref_tryget directly, because we need to be able to
986 * obtain a reference even in the short window between the queue
987 * starting to freeze, by dropping the first reference in
988 * blk_freeze_queue_start, and the moment the last request is
989 * consumed, marked by the instant q_usage_counter reaches
990 * zero.
991 */
992 if (!percpu_ref_tryget(&q->q_usage_counter))
993 return;
994
995 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
996
997 if (next != 0) {
998 mod_timer(&q->timeout, next);
999 } else {
1000 /*
1001 * Request timeouts are handled as a forward rolling timer. If
1002 * we end up here it means that no requests are pending and
1003 * also that no request has been pending for a while. Mark
1004 * each hctx as idle.
1005 */
1006 queue_for_each_hw_ctx(q, hctx, i) {
1007 /* the hctx may be unmapped, so check it here */
1008 if (blk_mq_hw_queue_mapped(hctx))
1009 blk_mq_tag_idle(hctx);
1010 }
1011 }
1012 blk_queue_exit(q);
1013}
1014
1015struct flush_busy_ctx_data {
1016 struct blk_mq_hw_ctx *hctx;
1017 struct list_head *list;
1018};
1019
1020static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1021{
1022 struct flush_busy_ctx_data *flush_data = data;
1023 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1024 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1025 enum hctx_type type = hctx->type;
1026
1027 spin_lock(&ctx->lock);
1028 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1029 sbitmap_clear_bit(sb, bitnr);
1030 spin_unlock(&ctx->lock);
1031 return true;
1032}
1033
1034/*
1035 * Process software queues that have been marked busy, splicing them
1036 * to the for-dispatch
1037 */
1038void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1039{
1040 struct flush_busy_ctx_data data = {
1041 .hctx = hctx,
1042 .list = list,
1043 };
1044
1045 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1046}
1047EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1048
1049struct dispatch_rq_data {
1050 struct blk_mq_hw_ctx *hctx;
1051 struct request *rq;
1052};
1053
1054static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1055 void *data)
1056{
1057 struct dispatch_rq_data *dispatch_data = data;
1058 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1059 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1060 enum hctx_type type = hctx->type;
1061
1062 spin_lock(&ctx->lock);
1063 if (!list_empty(&ctx->rq_lists[type])) {
1064 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1065 list_del_init(&dispatch_data->rq->queuelist);
1066 if (list_empty(&ctx->rq_lists[type]))
1067 sbitmap_clear_bit(sb, bitnr);
1068 }
1069 spin_unlock(&ctx->lock);
1070
1071 return !dispatch_data->rq;
1072}
1073
1074struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1075 struct blk_mq_ctx *start)
1076{
1077 unsigned off = start ? start->index_hw[hctx->type] : 0;
1078 struct dispatch_rq_data data = {
1079 .hctx = hctx,
1080 .rq = NULL,
1081 };
1082
1083 __sbitmap_for_each_set(&hctx->ctx_map, off,
1084 dispatch_rq_from_ctx, &data);
1085
1086 return data.rq;
1087}
1088
1089static inline unsigned int queued_to_index(unsigned int queued)
1090{
1091 if (!queued)
1092 return 0;
1093
1094 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1095}
1096
1097static bool __blk_mq_get_driver_tag(struct request *rq)
1098{
1099 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1100 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1101 int tag;
1102
1103 blk_mq_tag_busy(rq->mq_hctx);
1104
1105 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1106 bt = &rq->mq_hctx->tags->breserved_tags;
1107 tag_offset = 0;
1108 }
1109
1110 if (!hctx_may_queue(rq->mq_hctx, bt))
1111 return false;
1112 tag = __sbitmap_queue_get(bt);
1113 if (tag == BLK_MQ_NO_TAG)
1114 return false;
1115
1116 rq->tag = tag + tag_offset;
1117 return true;
1118}
1119
1120static bool blk_mq_get_driver_tag(struct request *rq)
1121{
1122 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1123
1124 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_get_driver_tag(rq))
1125 return false;
1126
1127 if ((hctx->flags & BLK_MQ_F_TAG_SHARED) &&
1128 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1129 rq->rq_flags |= RQF_MQ_INFLIGHT;
1130 atomic_inc(&hctx->nr_active);
1131 }
1132 hctx->tags->rqs[rq->tag] = rq;
1133 return true;
1134}
1135
1136static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1137 int flags, void *key)
1138{
1139 struct blk_mq_hw_ctx *hctx;
1140
1141 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1142
1143 spin_lock(&hctx->dispatch_wait_lock);
1144 if (!list_empty(&wait->entry)) {
1145 struct sbitmap_queue *sbq;
1146
1147 list_del_init(&wait->entry);
1148 sbq = &hctx->tags->bitmap_tags;
1149 atomic_dec(&sbq->ws_active);
1150 }
1151 spin_unlock(&hctx->dispatch_wait_lock);
1152
1153 blk_mq_run_hw_queue(hctx, true);
1154 return 1;
1155}
1156
1157/*
1158 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1159 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1160 * restart. For both cases, take care to check the condition again after
1161 * marking us as waiting.
1162 */
1163static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1164 struct request *rq)
1165{
1166 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1167 struct wait_queue_head *wq;
1168 wait_queue_entry_t *wait;
1169 bool ret;
1170
1171 if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1172 blk_mq_sched_mark_restart_hctx(hctx);
1173
1174 /*
1175 * It's possible that a tag was freed in the window between the
1176 * allocation failure and adding the hardware queue to the wait
1177 * queue.
1178 *
1179 * Don't clear RESTART here, someone else could have set it.
1180 * At most this will cost an extra queue run.
1181 */
1182 return blk_mq_get_driver_tag(rq);
1183 }
1184
1185 wait = &hctx->dispatch_wait;
1186 if (!list_empty_careful(&wait->entry))
1187 return false;
1188
1189 wq = &bt_wait_ptr(sbq, hctx)->wait;
1190
1191 spin_lock_irq(&wq->lock);
1192 spin_lock(&hctx->dispatch_wait_lock);
1193 if (!list_empty(&wait->entry)) {
1194 spin_unlock(&hctx->dispatch_wait_lock);
1195 spin_unlock_irq(&wq->lock);
1196 return false;
1197 }
1198
1199 atomic_inc(&sbq->ws_active);
1200 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1201 __add_wait_queue(wq, wait);
1202
1203 /*
1204 * It's possible that a tag was freed in the window between the
1205 * allocation failure and adding the hardware queue to the wait
1206 * queue.
1207 */
1208 ret = blk_mq_get_driver_tag(rq);
1209 if (!ret) {
1210 spin_unlock(&hctx->dispatch_wait_lock);
1211 spin_unlock_irq(&wq->lock);
1212 return false;
1213 }
1214
1215 /*
1216 * We got a tag, remove ourselves from the wait queue to ensure
1217 * someone else gets the wakeup.
1218 */
1219 list_del_init(&wait->entry);
1220 atomic_dec(&sbq->ws_active);
1221 spin_unlock(&hctx->dispatch_wait_lock);
1222 spin_unlock_irq(&wq->lock);
1223
1224 return true;
1225}
1226
1227#define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1228#define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1229/*
1230 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1231 * - EWMA is one simple way to compute running average value
1232 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1233 * - take 4 as factor for avoiding to get too small(0) result, and this
1234 * factor doesn't matter because EWMA decreases exponentially
1235 */
1236static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1237{
1238 unsigned int ewma;
1239
1240 if (hctx->queue->elevator)
1241 return;
1242
1243 ewma = hctx->dispatch_busy;
1244
1245 if (!ewma && !busy)
1246 return;
1247
1248 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1249 if (busy)
1250 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1251 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1252
1253 hctx->dispatch_busy = ewma;
1254}
1255
1256#define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1257
1258static void blk_mq_handle_dev_resource(struct request *rq,
1259 struct list_head *list)
1260{
1261 struct request *next =
1262 list_first_entry_or_null(list, struct request, queuelist);
1263
1264 /*
1265 * If an I/O scheduler has been configured and we got a driver tag for
1266 * the next request already, free it.
1267 */
1268 if (next)
1269 blk_mq_put_driver_tag(next);
1270
1271 list_add(&rq->queuelist, list);
1272 __blk_mq_requeue_request(rq);
1273}
1274
1275static void blk_mq_handle_zone_resource(struct request *rq,
1276 struct list_head *zone_list)
1277{
1278 /*
1279 * If we end up here it is because we cannot dispatch a request to a
1280 * specific zone due to LLD level zone-write locking or other zone
1281 * related resource not being available. In this case, set the request
1282 * aside in zone_list for retrying it later.
1283 */
1284 list_add(&rq->queuelist, zone_list);
1285 __blk_mq_requeue_request(rq);
1286}
1287
1288enum prep_dispatch {
1289 PREP_DISPATCH_OK,
1290 PREP_DISPATCH_NO_TAG,
1291 PREP_DISPATCH_NO_BUDGET,
1292};
1293
1294static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1295 bool need_budget)
1296{
1297 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1298
1299 if (need_budget && !blk_mq_get_dispatch_budget(rq->q)) {
1300 blk_mq_put_driver_tag(rq);
1301 return PREP_DISPATCH_NO_BUDGET;
1302 }
1303
1304 if (!blk_mq_get_driver_tag(rq)) {
1305 /*
1306 * The initial allocation attempt failed, so we need to
1307 * rerun the hardware queue when a tag is freed. The
1308 * waitqueue takes care of that. If the queue is run
1309 * before we add this entry back on the dispatch list,
1310 * we'll re-run it below.
1311 */
1312 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1313 /*
1314 * All budgets not got from this function will be put
1315 * together during handling partial dispatch
1316 */
1317 if (need_budget)
1318 blk_mq_put_dispatch_budget(rq->q);
1319 return PREP_DISPATCH_NO_TAG;
1320 }
1321 }
1322
1323 return PREP_DISPATCH_OK;
1324}
1325
1326/* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1327static void blk_mq_release_budgets(struct request_queue *q,
1328 unsigned int nr_budgets)
1329{
1330 int i;
1331
1332 for (i = 0; i < nr_budgets; i++)
1333 blk_mq_put_dispatch_budget(q);
1334}
1335
1336/*
1337 * Returns true if we did some work AND can potentially do more.
1338 */
1339bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1340 unsigned int nr_budgets)
1341{
1342 enum prep_dispatch prep;
1343 struct request_queue *q = hctx->queue;
1344 struct request *rq, *nxt;
1345 int errors, queued;
1346 blk_status_t ret = BLK_STS_OK;
1347 LIST_HEAD(zone_list);
1348
1349 if (list_empty(list))
1350 return false;
1351
1352 /*
1353 * Now process all the entries, sending them to the driver.
1354 */
1355 errors = queued = 0;
1356 do {
1357 struct blk_mq_queue_data bd;
1358
1359 rq = list_first_entry(list, struct request, queuelist);
1360
1361 WARN_ON_ONCE(hctx != rq->mq_hctx);
1362 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1363 if (prep != PREP_DISPATCH_OK)
1364 break;
1365
1366 list_del_init(&rq->queuelist);
1367
1368 bd.rq = rq;
1369
1370 /*
1371 * Flag last if we have no more requests, or if we have more
1372 * but can't assign a driver tag to it.
1373 */
1374 if (list_empty(list))
1375 bd.last = true;
1376 else {
1377 nxt = list_first_entry(list, struct request, queuelist);
1378 bd.last = !blk_mq_get_driver_tag(nxt);
1379 }
1380
1381 /*
1382 * once the request is queued to lld, no need to cover the
1383 * budget any more
1384 */
1385 if (nr_budgets)
1386 nr_budgets--;
1387 ret = q->mq_ops->queue_rq(hctx, &bd);
1388 switch (ret) {
1389 case BLK_STS_OK:
1390 queued++;
1391 break;
1392 case BLK_STS_RESOURCE:
1393 case BLK_STS_DEV_RESOURCE:
1394 blk_mq_handle_dev_resource(rq, list);
1395 goto out;
1396 case BLK_STS_ZONE_RESOURCE:
1397 /*
1398 * Move the request to zone_list and keep going through
1399 * the dispatch list to find more requests the drive can
1400 * accept.
1401 */
1402 blk_mq_handle_zone_resource(rq, &zone_list);
1403 break;
1404 default:
1405 errors++;
1406 blk_mq_end_request(rq, BLK_STS_IOERR);
1407 }
1408 } while (!list_empty(list));
1409out:
1410 if (!list_empty(&zone_list))
1411 list_splice_tail_init(&zone_list, list);
1412
1413 hctx->dispatched[queued_to_index(queued)]++;
1414
1415 /* If we didn't flush the entire list, we could have told the driver
1416 * there was more coming, but that turned out to be a lie.
1417 */
1418 if ((!list_empty(list) || errors) && q->mq_ops->commit_rqs && queued)
1419 q->mq_ops->commit_rqs(hctx);
1420 /*
1421 * Any items that need requeuing? Stuff them into hctx->dispatch,
1422 * that is where we will continue on next queue run.
1423 */
1424 if (!list_empty(list)) {
1425 bool needs_restart;
1426 /* For non-shared tags, the RESTART check will suffice */
1427 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
1428 (hctx->flags & BLK_MQ_F_TAG_SHARED);
1429 bool no_budget_avail = prep == PREP_DISPATCH_NO_BUDGET;
1430
1431 blk_mq_release_budgets(q, nr_budgets);
1432
1433 spin_lock(&hctx->lock);
1434 list_splice_tail_init(list, &hctx->dispatch);
1435 spin_unlock(&hctx->lock);
1436
1437 /*
1438 * Order adding requests to hctx->dispatch and checking
1439 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1440 * in blk_mq_sched_restart(). Avoid restart code path to
1441 * miss the new added requests to hctx->dispatch, meantime
1442 * SCHED_RESTART is observed here.
1443 */
1444 smp_mb();
1445
1446 /*
1447 * If SCHED_RESTART was set by the caller of this function and
1448 * it is no longer set that means that it was cleared by another
1449 * thread and hence that a queue rerun is needed.
1450 *
1451 * If 'no_tag' is set, that means that we failed getting
1452 * a driver tag with an I/O scheduler attached. If our dispatch
1453 * waitqueue is no longer active, ensure that we run the queue
1454 * AFTER adding our entries back to the list.
1455 *
1456 * If no I/O scheduler has been configured it is possible that
1457 * the hardware queue got stopped and restarted before requests
1458 * were pushed back onto the dispatch list. Rerun the queue to
1459 * avoid starvation. Notes:
1460 * - blk_mq_run_hw_queue() checks whether or not a queue has
1461 * been stopped before rerunning a queue.
1462 * - Some but not all block drivers stop a queue before
1463 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1464 * and dm-rq.
1465 *
1466 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1467 * bit is set, run queue after a delay to avoid IO stalls
1468 * that could otherwise occur if the queue is idle. We'll do
1469 * similar if we couldn't get budget and SCHED_RESTART is set.
1470 */
1471 needs_restart = blk_mq_sched_needs_restart(hctx);
1472 if (!needs_restart ||
1473 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1474 blk_mq_run_hw_queue(hctx, true);
1475 else if (needs_restart && (ret == BLK_STS_RESOURCE ||
1476 no_budget_avail))
1477 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1478
1479 blk_mq_update_dispatch_busy(hctx, true);
1480 return false;
1481 } else
1482 blk_mq_update_dispatch_busy(hctx, false);
1483
1484 return (queued + errors) != 0;
1485}
1486
1487/**
1488 * __blk_mq_run_hw_queue - Run a hardware queue.
1489 * @hctx: Pointer to the hardware queue to run.
1490 *
1491 * Send pending requests to the hardware.
1492 */
1493static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1494{
1495 int srcu_idx;
1496
1497 /*
1498 * We should be running this queue from one of the CPUs that
1499 * are mapped to it.
1500 *
1501 * There are at least two related races now between setting
1502 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1503 * __blk_mq_run_hw_queue():
1504 *
1505 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1506 * but later it becomes online, then this warning is harmless
1507 * at all
1508 *
1509 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1510 * but later it becomes offline, then the warning can't be
1511 * triggered, and we depend on blk-mq timeout handler to
1512 * handle dispatched requests to this hctx
1513 */
1514 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1515 cpu_online(hctx->next_cpu)) {
1516 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1517 raw_smp_processor_id(),
1518 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1519 dump_stack();
1520 }
1521
1522 /*
1523 * We can't run the queue inline with ints disabled. Ensure that
1524 * we catch bad users of this early.
1525 */
1526 WARN_ON_ONCE(in_interrupt());
1527
1528 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1529
1530 hctx_lock(hctx, &srcu_idx);
1531 blk_mq_sched_dispatch_requests(hctx);
1532 hctx_unlock(hctx, srcu_idx);
1533}
1534
1535static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1536{
1537 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1538
1539 if (cpu >= nr_cpu_ids)
1540 cpu = cpumask_first(hctx->cpumask);
1541 return cpu;
1542}
1543
1544/*
1545 * It'd be great if the workqueue API had a way to pass
1546 * in a mask and had some smarts for more clever placement.
1547 * For now we just round-robin here, switching for every
1548 * BLK_MQ_CPU_WORK_BATCH queued items.
1549 */
1550static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1551{
1552 bool tried = false;
1553 int next_cpu = hctx->next_cpu;
1554
1555 if (hctx->queue->nr_hw_queues == 1)
1556 return WORK_CPU_UNBOUND;
1557
1558 if (--hctx->next_cpu_batch <= 0) {
1559select_cpu:
1560 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1561 cpu_online_mask);
1562 if (next_cpu >= nr_cpu_ids)
1563 next_cpu = blk_mq_first_mapped_cpu(hctx);
1564 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1565 }
1566
1567 /*
1568 * Do unbound schedule if we can't find a online CPU for this hctx,
1569 * and it should only happen in the path of handling CPU DEAD.
1570 */
1571 if (!cpu_online(next_cpu)) {
1572 if (!tried) {
1573 tried = true;
1574 goto select_cpu;
1575 }
1576
1577 /*
1578 * Make sure to re-select CPU next time once after CPUs
1579 * in hctx->cpumask become online again.
1580 */
1581 hctx->next_cpu = next_cpu;
1582 hctx->next_cpu_batch = 1;
1583 return WORK_CPU_UNBOUND;
1584 }
1585
1586 hctx->next_cpu = next_cpu;
1587 return next_cpu;
1588}
1589
1590/**
1591 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1592 * @hctx: Pointer to the hardware queue to run.
1593 * @async: If we want to run the queue asynchronously.
1594 * @msecs: Microseconds of delay to wait before running the queue.
1595 *
1596 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1597 * with a delay of @msecs.
1598 */
1599static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1600 unsigned long msecs)
1601{
1602 if (unlikely(blk_mq_hctx_stopped(hctx)))
1603 return;
1604
1605 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1606 int cpu = get_cpu();
1607 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1608 __blk_mq_run_hw_queue(hctx);
1609 put_cpu();
1610 return;
1611 }
1612
1613 put_cpu();
1614 }
1615
1616 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1617 msecs_to_jiffies(msecs));
1618}
1619
1620/**
1621 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1622 * @hctx: Pointer to the hardware queue to run.
1623 * @msecs: Microseconds of delay to wait before running the queue.
1624 *
1625 * Run a hardware queue asynchronously with a delay of @msecs.
1626 */
1627void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1628{
1629 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1630}
1631EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1632
1633/**
1634 * blk_mq_run_hw_queue - Start to run a hardware queue.
1635 * @hctx: Pointer to the hardware queue to run.
1636 * @async: If we want to run the queue asynchronously.
1637 *
1638 * Check if the request queue is not in a quiesced state and if there are
1639 * pending requests to be sent. If this is true, run the queue to send requests
1640 * to hardware.
1641 */
1642void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1643{
1644 int srcu_idx;
1645 bool need_run;
1646
1647 /*
1648 * When queue is quiesced, we may be switching io scheduler, or
1649 * updating nr_hw_queues, or other things, and we can't run queue
1650 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1651 *
1652 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1653 * quiesced.
1654 */
1655 hctx_lock(hctx, &srcu_idx);
1656 need_run = !blk_queue_quiesced(hctx->queue) &&
1657 blk_mq_hctx_has_pending(hctx);
1658 hctx_unlock(hctx, srcu_idx);
1659
1660 if (need_run)
1661 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1662}
1663EXPORT_SYMBOL(blk_mq_run_hw_queue);
1664
1665/**
1666 * blk_mq_run_hw_queue - Run all hardware queues in a request queue.
1667 * @q: Pointer to the request queue to run.
1668 * @async: If we want to run the queue asynchronously.
1669 */
1670void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1671{
1672 struct blk_mq_hw_ctx *hctx;
1673 int i;
1674
1675 queue_for_each_hw_ctx(q, hctx, i) {
1676 if (blk_mq_hctx_stopped(hctx))
1677 continue;
1678
1679 blk_mq_run_hw_queue(hctx, async);
1680 }
1681}
1682EXPORT_SYMBOL(blk_mq_run_hw_queues);
1683
1684/**
1685 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
1686 * @q: Pointer to the request queue to run.
1687 * @msecs: Microseconds of delay to wait before running the queues.
1688 */
1689void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
1690{
1691 struct blk_mq_hw_ctx *hctx;
1692 int i;
1693
1694 queue_for_each_hw_ctx(q, hctx, i) {
1695 if (blk_mq_hctx_stopped(hctx))
1696 continue;
1697
1698 blk_mq_delay_run_hw_queue(hctx, msecs);
1699 }
1700}
1701EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
1702
1703/**
1704 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1705 * @q: request queue.
1706 *
1707 * The caller is responsible for serializing this function against
1708 * blk_mq_{start,stop}_hw_queue().
1709 */
1710bool blk_mq_queue_stopped(struct request_queue *q)
1711{
1712 struct blk_mq_hw_ctx *hctx;
1713 int i;
1714
1715 queue_for_each_hw_ctx(q, hctx, i)
1716 if (blk_mq_hctx_stopped(hctx))
1717 return true;
1718
1719 return false;
1720}
1721EXPORT_SYMBOL(blk_mq_queue_stopped);
1722
1723/*
1724 * This function is often used for pausing .queue_rq() by driver when
1725 * there isn't enough resource or some conditions aren't satisfied, and
1726 * BLK_STS_RESOURCE is usually returned.
1727 *
1728 * We do not guarantee that dispatch can be drained or blocked
1729 * after blk_mq_stop_hw_queue() returns. Please use
1730 * blk_mq_quiesce_queue() for that requirement.
1731 */
1732void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1733{
1734 cancel_delayed_work(&hctx->run_work);
1735
1736 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1737}
1738EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1739
1740/*
1741 * This function is often used for pausing .queue_rq() by driver when
1742 * there isn't enough resource or some conditions aren't satisfied, and
1743 * BLK_STS_RESOURCE is usually returned.
1744 *
1745 * We do not guarantee that dispatch can be drained or blocked
1746 * after blk_mq_stop_hw_queues() returns. Please use
1747 * blk_mq_quiesce_queue() for that requirement.
1748 */
1749void blk_mq_stop_hw_queues(struct request_queue *q)
1750{
1751 struct blk_mq_hw_ctx *hctx;
1752 int i;
1753
1754 queue_for_each_hw_ctx(q, hctx, i)
1755 blk_mq_stop_hw_queue(hctx);
1756}
1757EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1758
1759void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1760{
1761 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1762
1763 blk_mq_run_hw_queue(hctx, false);
1764}
1765EXPORT_SYMBOL(blk_mq_start_hw_queue);
1766
1767void blk_mq_start_hw_queues(struct request_queue *q)
1768{
1769 struct blk_mq_hw_ctx *hctx;
1770 int i;
1771
1772 queue_for_each_hw_ctx(q, hctx, i)
1773 blk_mq_start_hw_queue(hctx);
1774}
1775EXPORT_SYMBOL(blk_mq_start_hw_queues);
1776
1777void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1778{
1779 if (!blk_mq_hctx_stopped(hctx))
1780 return;
1781
1782 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1783 blk_mq_run_hw_queue(hctx, async);
1784}
1785EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1786
1787void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1788{
1789 struct blk_mq_hw_ctx *hctx;
1790 int i;
1791
1792 queue_for_each_hw_ctx(q, hctx, i)
1793 blk_mq_start_stopped_hw_queue(hctx, async);
1794}
1795EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1796
1797static void blk_mq_run_work_fn(struct work_struct *work)
1798{
1799 struct blk_mq_hw_ctx *hctx;
1800
1801 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1802
1803 /*
1804 * If we are stopped, don't run the queue.
1805 */
1806 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1807 return;
1808
1809 __blk_mq_run_hw_queue(hctx);
1810}
1811
1812static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1813 struct request *rq,
1814 bool at_head)
1815{
1816 struct blk_mq_ctx *ctx = rq->mq_ctx;
1817 enum hctx_type type = hctx->type;
1818
1819 lockdep_assert_held(&ctx->lock);
1820
1821 trace_block_rq_insert(hctx->queue, rq);
1822
1823 if (at_head)
1824 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1825 else
1826 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1827}
1828
1829void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1830 bool at_head)
1831{
1832 struct blk_mq_ctx *ctx = rq->mq_ctx;
1833
1834 lockdep_assert_held(&ctx->lock);
1835
1836 __blk_mq_insert_req_list(hctx, rq, at_head);
1837 blk_mq_hctx_mark_pending(hctx, ctx);
1838}
1839
1840/**
1841 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1842 * @rq: Pointer to request to be inserted.
1843 * @at_head: true if the request should be inserted at the head of the list.
1844 * @run_queue: If we should run the hardware queue after inserting the request.
1845 *
1846 * Should only be used carefully, when the caller knows we want to
1847 * bypass a potential IO scheduler on the target device.
1848 */
1849void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
1850 bool run_queue)
1851{
1852 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1853
1854 spin_lock(&hctx->lock);
1855 if (at_head)
1856 list_add(&rq->queuelist, &hctx->dispatch);
1857 else
1858 list_add_tail(&rq->queuelist, &hctx->dispatch);
1859 spin_unlock(&hctx->lock);
1860
1861 if (run_queue)
1862 blk_mq_run_hw_queue(hctx, false);
1863}
1864
1865void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1866 struct list_head *list)
1867
1868{
1869 struct request *rq;
1870 enum hctx_type type = hctx->type;
1871
1872 /*
1873 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1874 * offline now
1875 */
1876 list_for_each_entry(rq, list, queuelist) {
1877 BUG_ON(rq->mq_ctx != ctx);
1878 trace_block_rq_insert(hctx->queue, rq);
1879 }
1880
1881 spin_lock(&ctx->lock);
1882 list_splice_tail_init(list, &ctx->rq_lists[type]);
1883 blk_mq_hctx_mark_pending(hctx, ctx);
1884 spin_unlock(&ctx->lock);
1885}
1886
1887static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1888{
1889 struct request *rqa = container_of(a, struct request, queuelist);
1890 struct request *rqb = container_of(b, struct request, queuelist);
1891
1892 if (rqa->mq_ctx != rqb->mq_ctx)
1893 return rqa->mq_ctx > rqb->mq_ctx;
1894 if (rqa->mq_hctx != rqb->mq_hctx)
1895 return rqa->mq_hctx > rqb->mq_hctx;
1896
1897 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1898}
1899
1900void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1901{
1902 LIST_HEAD(list);
1903
1904 if (list_empty(&plug->mq_list))
1905 return;
1906 list_splice_init(&plug->mq_list, &list);
1907
1908 if (plug->rq_count > 2 && plug->multiple_queues)
1909 list_sort(NULL, &list, plug_rq_cmp);
1910
1911 plug->rq_count = 0;
1912
1913 do {
1914 struct list_head rq_list;
1915 struct request *rq, *head_rq = list_entry_rq(list.next);
1916 struct list_head *pos = &head_rq->queuelist; /* skip first */
1917 struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx;
1918 struct blk_mq_ctx *this_ctx = head_rq->mq_ctx;
1919 unsigned int depth = 1;
1920
1921 list_for_each_continue(pos, &list) {
1922 rq = list_entry_rq(pos);
1923 BUG_ON(!rq->q);
1924 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx)
1925 break;
1926 depth++;
1927 }
1928
1929 list_cut_before(&rq_list, &list, pos);
1930 trace_block_unplug(head_rq->q, depth, !from_schedule);
1931 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1932 from_schedule);
1933 } while(!list_empty(&list));
1934}
1935
1936static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1937 unsigned int nr_segs)
1938{
1939 if (bio->bi_opf & REQ_RAHEAD)
1940 rq->cmd_flags |= REQ_FAILFAST_MASK;
1941
1942 rq->__sector = bio->bi_iter.bi_sector;
1943 rq->write_hint = bio->bi_write_hint;
1944 blk_rq_bio_prep(rq, bio, nr_segs);
1945 blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
1946
1947 blk_account_io_start(rq);
1948}
1949
1950static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1951 struct request *rq,
1952 blk_qc_t *cookie, bool last)
1953{
1954 struct request_queue *q = rq->q;
1955 struct blk_mq_queue_data bd = {
1956 .rq = rq,
1957 .last = last,
1958 };
1959 blk_qc_t new_cookie;
1960 blk_status_t ret;
1961
1962 new_cookie = request_to_qc_t(hctx, rq);
1963
1964 /*
1965 * For OK queue, we are done. For error, caller may kill it.
1966 * Any other error (busy), just add it to our list as we
1967 * previously would have done.
1968 */
1969 ret = q->mq_ops->queue_rq(hctx, &bd);
1970 switch (ret) {
1971 case BLK_STS_OK:
1972 blk_mq_update_dispatch_busy(hctx, false);
1973 *cookie = new_cookie;
1974 break;
1975 case BLK_STS_RESOURCE:
1976 case BLK_STS_DEV_RESOURCE:
1977 blk_mq_update_dispatch_busy(hctx, true);
1978 __blk_mq_requeue_request(rq);
1979 break;
1980 default:
1981 blk_mq_update_dispatch_busy(hctx, false);
1982 *cookie = BLK_QC_T_NONE;
1983 break;
1984 }
1985
1986 return ret;
1987}
1988
1989static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1990 struct request *rq,
1991 blk_qc_t *cookie,
1992 bool bypass_insert, bool last)
1993{
1994 struct request_queue *q = rq->q;
1995 bool run_queue = true;
1996
1997 /*
1998 * RCU or SRCU read lock is needed before checking quiesced flag.
1999 *
2000 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2001 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2002 * and avoid driver to try to dispatch again.
2003 */
2004 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2005 run_queue = false;
2006 bypass_insert = false;
2007 goto insert;
2008 }
2009
2010 if (q->elevator && !bypass_insert)
2011 goto insert;
2012
2013 if (!blk_mq_get_dispatch_budget(q))
2014 goto insert;
2015
2016 if (!blk_mq_get_driver_tag(rq)) {
2017 blk_mq_put_dispatch_budget(q);
2018 goto insert;
2019 }
2020
2021 return __blk_mq_issue_directly(hctx, rq, cookie, last);
2022insert:
2023 if (bypass_insert)
2024 return BLK_STS_RESOURCE;
2025
2026 blk_mq_sched_insert_request(rq, false, run_queue, false);
2027
2028 return BLK_STS_OK;
2029}
2030
2031/**
2032 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2033 * @hctx: Pointer of the associated hardware queue.
2034 * @rq: Pointer to request to be sent.
2035 * @cookie: Request queue cookie.
2036 *
2037 * If the device has enough resources to accept a new request now, send the
2038 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2039 * we can try send it another time in the future. Requests inserted at this
2040 * queue have higher priority.
2041 */
2042static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2043 struct request *rq, blk_qc_t *cookie)
2044{
2045 blk_status_t ret;
2046 int srcu_idx;
2047
2048 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
2049
2050 hctx_lock(hctx, &srcu_idx);
2051
2052 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
2053 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2054 blk_mq_request_bypass_insert(rq, false, true);
2055 else if (ret != BLK_STS_OK)
2056 blk_mq_end_request(rq, ret);
2057
2058 hctx_unlock(hctx, srcu_idx);
2059}
2060
2061blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2062{
2063 blk_status_t ret;
2064 int srcu_idx;
2065 blk_qc_t unused_cookie;
2066 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2067
2068 hctx_lock(hctx, &srcu_idx);
2069 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
2070 hctx_unlock(hctx, srcu_idx);
2071
2072 return ret;
2073}
2074
2075void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2076 struct list_head *list)
2077{
2078 int queued = 0;
2079 int errors = 0;
2080
2081 while (!list_empty(list)) {
2082 blk_status_t ret;
2083 struct request *rq = list_first_entry(list, struct request,
2084 queuelist);
2085
2086 list_del_init(&rq->queuelist);
2087 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2088 if (ret != BLK_STS_OK) {
2089 if (ret == BLK_STS_RESOURCE ||
2090 ret == BLK_STS_DEV_RESOURCE) {
2091 blk_mq_request_bypass_insert(rq, false,
2092 list_empty(list));
2093 break;
2094 }
2095 blk_mq_end_request(rq, ret);
2096 errors++;
2097 } else
2098 queued++;
2099 }
2100
2101 /*
2102 * If we didn't flush the entire list, we could have told
2103 * the driver there was more coming, but that turned out to
2104 * be a lie.
2105 */
2106 if ((!list_empty(list) || errors) &&
2107 hctx->queue->mq_ops->commit_rqs && queued)
2108 hctx->queue->mq_ops->commit_rqs(hctx);
2109}
2110
2111static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
2112{
2113 list_add_tail(&rq->queuelist, &plug->mq_list);
2114 plug->rq_count++;
2115 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
2116 struct request *tmp;
2117
2118 tmp = list_first_entry(&plug->mq_list, struct request,
2119 queuelist);
2120 if (tmp->q != rq->q)
2121 plug->multiple_queues = true;
2122 }
2123}
2124
2125/**
2126 * blk_mq_submit_bio - Create and send a request to block device.
2127 * @bio: Bio pointer.
2128 *
2129 * Builds up a request structure from @q and @bio and send to the device. The
2130 * request may not be queued directly to hardware if:
2131 * * This request can be merged with another one
2132 * * We want to place request at plug queue for possible future merging
2133 * * There is an IO scheduler active at this queue
2134 *
2135 * It will not queue the request if there is an error with the bio, or at the
2136 * request creation.
2137 *
2138 * Returns: Request queue cookie.
2139 */
2140blk_qc_t blk_mq_submit_bio(struct bio *bio)
2141{
2142 struct request_queue *q = bio->bi_disk->queue;
2143 const int is_sync = op_is_sync(bio->bi_opf);
2144 const int is_flush_fua = op_is_flush(bio->bi_opf);
2145 struct blk_mq_alloc_data data = {
2146 .q = q,
2147 };
2148 struct request *rq;
2149 struct blk_plug *plug;
2150 struct request *same_queue_rq = NULL;
2151 unsigned int nr_segs;
2152 blk_qc_t cookie;
2153 blk_status_t ret;
2154
2155 blk_queue_bounce(q, &bio);
2156 __blk_queue_split(&bio, &nr_segs);
2157
2158 if (!bio_integrity_prep(bio))
2159 goto queue_exit;
2160
2161 if (!is_flush_fua && !blk_queue_nomerges(q) &&
2162 blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
2163 goto queue_exit;
2164
2165 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2166 goto queue_exit;
2167
2168 rq_qos_throttle(q, bio);
2169
2170 data.cmd_flags = bio->bi_opf;
2171 rq = __blk_mq_alloc_request(&data);
2172 if (unlikely(!rq)) {
2173 rq_qos_cleanup(q, bio);
2174 if (bio->bi_opf & REQ_NOWAIT)
2175 bio_wouldblock_error(bio);
2176 goto queue_exit;
2177 }
2178
2179 trace_block_getrq(q, bio, bio->bi_opf);
2180
2181 rq_qos_track(q, rq, bio);
2182
2183 cookie = request_to_qc_t(data.hctx, rq);
2184
2185 blk_mq_bio_to_request(rq, bio, nr_segs);
2186
2187 ret = blk_crypto_init_request(rq);
2188 if (ret != BLK_STS_OK) {
2189 bio->bi_status = ret;
2190 bio_endio(bio);
2191 blk_mq_free_request(rq);
2192 return BLK_QC_T_NONE;
2193 }
2194
2195 plug = blk_mq_plug(q, bio);
2196 if (unlikely(is_flush_fua)) {
2197 /* Bypass scheduler for flush requests */
2198 blk_insert_flush(rq);
2199 blk_mq_run_hw_queue(data.hctx, true);
2200 } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs ||
2201 !blk_queue_nonrot(q))) {
2202 /*
2203 * Use plugging if we have a ->commit_rqs() hook as well, as
2204 * we know the driver uses bd->last in a smart fashion.
2205 *
2206 * Use normal plugging if this disk is slow HDD, as sequential
2207 * IO may benefit a lot from plug merging.
2208 */
2209 unsigned int request_count = plug->rq_count;
2210 struct request *last = NULL;
2211
2212 if (!request_count)
2213 trace_block_plug(q);
2214 else
2215 last = list_entry_rq(plug->mq_list.prev);
2216
2217 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
2218 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2219 blk_flush_plug_list(plug, false);
2220 trace_block_plug(q);
2221 }
2222
2223 blk_add_rq_to_plug(plug, rq);
2224 } else if (q->elevator) {
2225 /* Insert the request at the IO scheduler queue */
2226 blk_mq_sched_insert_request(rq, false, true, true);
2227 } else if (plug && !blk_queue_nomerges(q)) {
2228 /*
2229 * We do limited plugging. If the bio can be merged, do that.
2230 * Otherwise the existing request in the plug list will be
2231 * issued. So the plug list will have one request at most
2232 * The plug list might get flushed before this. If that happens,
2233 * the plug list is empty, and same_queue_rq is invalid.
2234 */
2235 if (list_empty(&plug->mq_list))
2236 same_queue_rq = NULL;
2237 if (same_queue_rq) {
2238 list_del_init(&same_queue_rq->queuelist);
2239 plug->rq_count--;
2240 }
2241 blk_add_rq_to_plug(plug, rq);
2242 trace_block_plug(q);
2243
2244 if (same_queue_rq) {
2245 data.hctx = same_queue_rq->mq_hctx;
2246 trace_block_unplug(q, 1, true);
2247 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2248 &cookie);
2249 }
2250 } else if ((q->nr_hw_queues > 1 && is_sync) ||
2251 !data.hctx->dispatch_busy) {
2252 /*
2253 * There is no scheduler and we can try to send directly
2254 * to the hardware.
2255 */
2256 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2257 } else {
2258 /* Default case. */
2259 blk_mq_sched_insert_request(rq, false, true, true);
2260 }
2261
2262 return cookie;
2263queue_exit:
2264 blk_queue_exit(q);
2265 return BLK_QC_T_NONE;
2266}
2267EXPORT_SYMBOL_GPL(blk_mq_submit_bio); /* only for request based dm */
2268
2269void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2270 unsigned int hctx_idx)
2271{
2272 struct page *page;
2273
2274 if (tags->rqs && set->ops->exit_request) {
2275 int i;
2276
2277 for (i = 0; i < tags->nr_tags; i++) {
2278 struct request *rq = tags->static_rqs[i];
2279
2280 if (!rq)
2281 continue;
2282 set->ops->exit_request(set, rq, hctx_idx);
2283 tags->static_rqs[i] = NULL;
2284 }
2285 }
2286
2287 while (!list_empty(&tags->page_list)) {
2288 page = list_first_entry(&tags->page_list, struct page, lru);
2289 list_del_init(&page->lru);
2290 /*
2291 * Remove kmemleak object previously allocated in
2292 * blk_mq_alloc_rqs().
2293 */
2294 kmemleak_free(page_address(page));
2295 __free_pages(page, page->private);
2296 }
2297}
2298
2299void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2300{
2301 kfree(tags->rqs);
2302 tags->rqs = NULL;
2303 kfree(tags->static_rqs);
2304 tags->static_rqs = NULL;
2305
2306 blk_mq_free_tags(tags);
2307}
2308
2309struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2310 unsigned int hctx_idx,
2311 unsigned int nr_tags,
2312 unsigned int reserved_tags)
2313{
2314 struct blk_mq_tags *tags;
2315 int node;
2316
2317 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2318 if (node == NUMA_NO_NODE)
2319 node = set->numa_node;
2320
2321 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2322 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2323 if (!tags)
2324 return NULL;
2325
2326 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2327 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2328 node);
2329 if (!tags->rqs) {
2330 blk_mq_free_tags(tags);
2331 return NULL;
2332 }
2333
2334 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2335 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2336 node);
2337 if (!tags->static_rqs) {
2338 kfree(tags->rqs);
2339 blk_mq_free_tags(tags);
2340 return NULL;
2341 }
2342
2343 return tags;
2344}
2345
2346static size_t order_to_size(unsigned int order)
2347{
2348 return (size_t)PAGE_SIZE << order;
2349}
2350
2351static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2352 unsigned int hctx_idx, int node)
2353{
2354 int ret;
2355
2356 if (set->ops->init_request) {
2357 ret = set->ops->init_request(set, rq, hctx_idx, node);
2358 if (ret)
2359 return ret;
2360 }
2361
2362 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2363 return 0;
2364}
2365
2366int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2367 unsigned int hctx_idx, unsigned int depth)
2368{
2369 unsigned int i, j, entries_per_page, max_order = 4;
2370 size_t rq_size, left;
2371 int node;
2372
2373 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2374 if (node == NUMA_NO_NODE)
2375 node = set->numa_node;
2376
2377 INIT_LIST_HEAD(&tags->page_list);
2378
2379 /*
2380 * rq_size is the size of the request plus driver payload, rounded
2381 * to the cacheline size
2382 */
2383 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2384 cache_line_size());
2385 left = rq_size * depth;
2386
2387 for (i = 0; i < depth; ) {
2388 int this_order = max_order;
2389 struct page *page;
2390 int to_do;
2391 void *p;
2392
2393 while (this_order && left < order_to_size(this_order - 1))
2394 this_order--;
2395
2396 do {
2397 page = alloc_pages_node(node,
2398 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2399 this_order);
2400 if (page)
2401 break;
2402 if (!this_order--)
2403 break;
2404 if (order_to_size(this_order) < rq_size)
2405 break;
2406 } while (1);
2407
2408 if (!page)
2409 goto fail;
2410
2411 page->private = this_order;
2412 list_add_tail(&page->lru, &tags->page_list);
2413
2414 p = page_address(page);
2415 /*
2416 * Allow kmemleak to scan these pages as they contain pointers
2417 * to additional allocations like via ops->init_request().
2418 */
2419 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2420 entries_per_page = order_to_size(this_order) / rq_size;
2421 to_do = min(entries_per_page, depth - i);
2422 left -= to_do * rq_size;
2423 for (j = 0; j < to_do; j++) {
2424 struct request *rq = p;
2425
2426 tags->static_rqs[i] = rq;
2427 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2428 tags->static_rqs[i] = NULL;
2429 goto fail;
2430 }
2431
2432 p += rq_size;
2433 i++;
2434 }
2435 }
2436 return 0;
2437
2438fail:
2439 blk_mq_free_rqs(set, tags, hctx_idx);
2440 return -ENOMEM;
2441}
2442
2443struct rq_iter_data {
2444 struct blk_mq_hw_ctx *hctx;
2445 bool has_rq;
2446};
2447
2448static bool blk_mq_has_request(struct request *rq, void *data, bool reserved)
2449{
2450 struct rq_iter_data *iter_data = data;
2451
2452 if (rq->mq_hctx != iter_data->hctx)
2453 return true;
2454 iter_data->has_rq = true;
2455 return false;
2456}
2457
2458static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
2459{
2460 struct blk_mq_tags *tags = hctx->sched_tags ?
2461 hctx->sched_tags : hctx->tags;
2462 struct rq_iter_data data = {
2463 .hctx = hctx,
2464 };
2465
2466 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
2467 return data.has_rq;
2468}
2469
2470static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
2471 struct blk_mq_hw_ctx *hctx)
2472{
2473 if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu)
2474 return false;
2475 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
2476 return false;
2477 return true;
2478}
2479
2480static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
2481{
2482 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2483 struct blk_mq_hw_ctx, cpuhp_online);
2484
2485 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
2486 !blk_mq_last_cpu_in_hctx(cpu, hctx))
2487 return 0;
2488
2489 /*
2490 * Prevent new request from being allocated on the current hctx.
2491 *
2492 * The smp_mb__after_atomic() Pairs with the implied barrier in
2493 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
2494 * seen once we return from the tag allocator.
2495 */
2496 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2497 smp_mb__after_atomic();
2498
2499 /*
2500 * Try to grab a reference to the queue and wait for any outstanding
2501 * requests. If we could not grab a reference the queue has been
2502 * frozen and there are no requests.
2503 */
2504 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
2505 while (blk_mq_hctx_has_requests(hctx))
2506 msleep(5);
2507 percpu_ref_put(&hctx->queue->q_usage_counter);
2508 }
2509
2510 return 0;
2511}
2512
2513static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
2514{
2515 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2516 struct blk_mq_hw_ctx, cpuhp_online);
2517
2518 if (cpumask_test_cpu(cpu, hctx->cpumask))
2519 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2520 return 0;
2521}
2522
2523/*
2524 * 'cpu' is going away. splice any existing rq_list entries from this
2525 * software queue to the hw queue dispatch list, and ensure that it
2526 * gets run.
2527 */
2528static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2529{
2530 struct blk_mq_hw_ctx *hctx;
2531 struct blk_mq_ctx *ctx;
2532 LIST_HEAD(tmp);
2533 enum hctx_type type;
2534
2535 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2536 if (!cpumask_test_cpu(cpu, hctx->cpumask))
2537 return 0;
2538
2539 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2540 type = hctx->type;
2541
2542 spin_lock(&ctx->lock);
2543 if (!list_empty(&ctx->rq_lists[type])) {
2544 list_splice_init(&ctx->rq_lists[type], &tmp);
2545 blk_mq_hctx_clear_pending(hctx, ctx);
2546 }
2547 spin_unlock(&ctx->lock);
2548
2549 if (list_empty(&tmp))
2550 return 0;
2551
2552 spin_lock(&hctx->lock);
2553 list_splice_tail_init(&tmp, &hctx->dispatch);
2554 spin_unlock(&hctx->lock);
2555
2556 blk_mq_run_hw_queue(hctx, true);
2557 return 0;
2558}
2559
2560static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2561{
2562 if (!(hctx->flags & BLK_MQ_F_STACKING))
2563 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2564 &hctx->cpuhp_online);
2565 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2566 &hctx->cpuhp_dead);
2567}
2568
2569/* hctx->ctxs will be freed in queue's release handler */
2570static void blk_mq_exit_hctx(struct request_queue *q,
2571 struct blk_mq_tag_set *set,
2572 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2573{
2574 if (blk_mq_hw_queue_mapped(hctx))
2575 blk_mq_tag_idle(hctx);
2576
2577 if (set->ops->exit_request)
2578 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2579
2580 if (set->ops->exit_hctx)
2581 set->ops->exit_hctx(hctx, hctx_idx);
2582
2583 blk_mq_remove_cpuhp(hctx);
2584
2585 spin_lock(&q->unused_hctx_lock);
2586 list_add(&hctx->hctx_list, &q->unused_hctx_list);
2587 spin_unlock(&q->unused_hctx_lock);
2588}
2589
2590static void blk_mq_exit_hw_queues(struct request_queue *q,
2591 struct blk_mq_tag_set *set, int nr_queue)
2592{
2593 struct blk_mq_hw_ctx *hctx;
2594 unsigned int i;
2595
2596 queue_for_each_hw_ctx(q, hctx, i) {
2597 if (i == nr_queue)
2598 break;
2599 blk_mq_debugfs_unregister_hctx(hctx);
2600 blk_mq_exit_hctx(q, set, hctx, i);
2601 }
2602}
2603
2604static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2605{
2606 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2607
2608 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2609 __alignof__(struct blk_mq_hw_ctx)) !=
2610 sizeof(struct blk_mq_hw_ctx));
2611
2612 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2613 hw_ctx_size += sizeof(struct srcu_struct);
2614
2615 return hw_ctx_size;
2616}
2617
2618static int blk_mq_init_hctx(struct request_queue *q,
2619 struct blk_mq_tag_set *set,
2620 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2621{
2622 hctx->queue_num = hctx_idx;
2623
2624 if (!(hctx->flags & BLK_MQ_F_STACKING))
2625 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2626 &hctx->cpuhp_online);
2627 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2628
2629 hctx->tags = set->tags[hctx_idx];
2630
2631 if (set->ops->init_hctx &&
2632 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2633 goto unregister_cpu_notifier;
2634
2635 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2636 hctx->numa_node))
2637 goto exit_hctx;
2638 return 0;
2639
2640 exit_hctx:
2641 if (set->ops->exit_hctx)
2642 set->ops->exit_hctx(hctx, hctx_idx);
2643 unregister_cpu_notifier:
2644 blk_mq_remove_cpuhp(hctx);
2645 return -1;
2646}
2647
2648static struct blk_mq_hw_ctx *
2649blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2650 int node)
2651{
2652 struct blk_mq_hw_ctx *hctx;
2653 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2654
2655 hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2656 if (!hctx)
2657 goto fail_alloc_hctx;
2658
2659 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2660 goto free_hctx;
2661
2662 atomic_set(&hctx->nr_active, 0);
2663 if (node == NUMA_NO_NODE)
2664 node = set->numa_node;
2665 hctx->numa_node = node;
2666
2667 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2668 spin_lock_init(&hctx->lock);
2669 INIT_LIST_HEAD(&hctx->dispatch);
2670 hctx->queue = q;
2671 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2672
2673 INIT_LIST_HEAD(&hctx->hctx_list);
2674
2675 /*
2676 * Allocate space for all possible cpus to avoid allocation at
2677 * runtime
2678 */
2679 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2680 gfp, node);
2681 if (!hctx->ctxs)
2682 goto free_cpumask;
2683
2684 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2685 gfp, node))
2686 goto free_ctxs;
2687 hctx->nr_ctx = 0;
2688
2689 spin_lock_init(&hctx->dispatch_wait_lock);
2690 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2691 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2692
2693 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
2694 if (!hctx->fq)
2695 goto free_bitmap;
2696
2697 if (hctx->flags & BLK_MQ_F_BLOCKING)
2698 init_srcu_struct(hctx->srcu);
2699 blk_mq_hctx_kobj_init(hctx);
2700
2701 return hctx;
2702
2703 free_bitmap:
2704 sbitmap_free(&hctx->ctx_map);
2705 free_ctxs:
2706 kfree(hctx->ctxs);
2707 free_cpumask:
2708 free_cpumask_var(hctx->cpumask);
2709 free_hctx:
2710 kfree(hctx);
2711 fail_alloc_hctx:
2712 return NULL;
2713}
2714
2715static void blk_mq_init_cpu_queues(struct request_queue *q,
2716 unsigned int nr_hw_queues)
2717{
2718 struct blk_mq_tag_set *set = q->tag_set;
2719 unsigned int i, j;
2720
2721 for_each_possible_cpu(i) {
2722 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2723 struct blk_mq_hw_ctx *hctx;
2724 int k;
2725
2726 __ctx->cpu = i;
2727 spin_lock_init(&__ctx->lock);
2728 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2729 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2730
2731 __ctx->queue = q;
2732
2733 /*
2734 * Set local node, IFF we have more than one hw queue. If
2735 * not, we remain on the home node of the device
2736 */
2737 for (j = 0; j < set->nr_maps; j++) {
2738 hctx = blk_mq_map_queue_type(q, j, i);
2739 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2740 hctx->numa_node = local_memory_node(cpu_to_node(i));
2741 }
2742 }
2743}
2744
2745static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set *set,
2746 int hctx_idx)
2747{
2748 int ret = 0;
2749
2750 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2751 set->queue_depth, set->reserved_tags);
2752 if (!set->tags[hctx_idx])
2753 return false;
2754
2755 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2756 set->queue_depth);
2757 if (!ret)
2758 return true;
2759
2760 blk_mq_free_rq_map(set->tags[hctx_idx]);
2761 set->tags[hctx_idx] = NULL;
2762 return false;
2763}
2764
2765static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2766 unsigned int hctx_idx)
2767{
2768 if (set->tags && set->tags[hctx_idx]) {
2769 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2770 blk_mq_free_rq_map(set->tags[hctx_idx]);
2771 set->tags[hctx_idx] = NULL;
2772 }
2773}
2774
2775static void blk_mq_map_swqueue(struct request_queue *q)
2776{
2777 unsigned int i, j, hctx_idx;
2778 struct blk_mq_hw_ctx *hctx;
2779 struct blk_mq_ctx *ctx;
2780 struct blk_mq_tag_set *set = q->tag_set;
2781
2782 queue_for_each_hw_ctx(q, hctx, i) {
2783 cpumask_clear(hctx->cpumask);
2784 hctx->nr_ctx = 0;
2785 hctx->dispatch_from = NULL;
2786 }
2787
2788 /*
2789 * Map software to hardware queues.
2790 *
2791 * If the cpu isn't present, the cpu is mapped to first hctx.
2792 */
2793 for_each_possible_cpu(i) {
2794
2795 ctx = per_cpu_ptr(q->queue_ctx, i);
2796 for (j = 0; j < set->nr_maps; j++) {
2797 if (!set->map[j].nr_queues) {
2798 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2799 HCTX_TYPE_DEFAULT, i);
2800 continue;
2801 }
2802 hctx_idx = set->map[j].mq_map[i];
2803 /* unmapped hw queue can be remapped after CPU topo changed */
2804 if (!set->tags[hctx_idx] &&
2805 !__blk_mq_alloc_map_and_request(set, hctx_idx)) {
2806 /*
2807 * If tags initialization fail for some hctx,
2808 * that hctx won't be brought online. In this
2809 * case, remap the current ctx to hctx[0] which
2810 * is guaranteed to always have tags allocated
2811 */
2812 set->map[j].mq_map[i] = 0;
2813 }
2814
2815 hctx = blk_mq_map_queue_type(q, j, i);
2816 ctx->hctxs[j] = hctx;
2817 /*
2818 * If the CPU is already set in the mask, then we've
2819 * mapped this one already. This can happen if
2820 * devices share queues across queue maps.
2821 */
2822 if (cpumask_test_cpu(i, hctx->cpumask))
2823 continue;
2824
2825 cpumask_set_cpu(i, hctx->cpumask);
2826 hctx->type = j;
2827 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2828 hctx->ctxs[hctx->nr_ctx++] = ctx;
2829
2830 /*
2831 * If the nr_ctx type overflows, we have exceeded the
2832 * amount of sw queues we can support.
2833 */
2834 BUG_ON(!hctx->nr_ctx);
2835 }
2836
2837 for (; j < HCTX_MAX_TYPES; j++)
2838 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2839 HCTX_TYPE_DEFAULT, i);
2840 }
2841
2842 queue_for_each_hw_ctx(q, hctx, i) {
2843 /*
2844 * If no software queues are mapped to this hardware queue,
2845 * disable it and free the request entries.
2846 */
2847 if (!hctx->nr_ctx) {
2848 /* Never unmap queue 0. We need it as a
2849 * fallback in case of a new remap fails
2850 * allocation
2851 */
2852 if (i && set->tags[i])
2853 blk_mq_free_map_and_requests(set, i);
2854
2855 hctx->tags = NULL;
2856 continue;
2857 }
2858
2859 hctx->tags = set->tags[i];
2860 WARN_ON(!hctx->tags);
2861
2862 /*
2863 * Set the map size to the number of mapped software queues.
2864 * This is more accurate and more efficient than looping
2865 * over all possibly mapped software queues.
2866 */
2867 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2868
2869 /*
2870 * Initialize batch roundrobin counts
2871 */
2872 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2873 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2874 }
2875}
2876
2877/*
2878 * Caller needs to ensure that we're either frozen/quiesced, or that
2879 * the queue isn't live yet.
2880 */
2881static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2882{
2883 struct blk_mq_hw_ctx *hctx;
2884 int i;
2885
2886 queue_for_each_hw_ctx(q, hctx, i) {
2887 if (shared)
2888 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2889 else
2890 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2891 }
2892}
2893
2894static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2895 bool shared)
2896{
2897 struct request_queue *q;
2898
2899 lockdep_assert_held(&set->tag_list_lock);
2900
2901 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2902 blk_mq_freeze_queue(q);
2903 queue_set_hctx_shared(q, shared);
2904 blk_mq_unfreeze_queue(q);
2905 }
2906}
2907
2908static void blk_mq_del_queue_tag_set(struct request_queue *q)
2909{
2910 struct blk_mq_tag_set *set = q->tag_set;
2911
2912 mutex_lock(&set->tag_list_lock);
2913 list_del(&q->tag_set_list);
2914 if (list_is_singular(&set->tag_list)) {
2915 /* just transitioned to unshared */
2916 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2917 /* update existing queue */
2918 blk_mq_update_tag_set_depth(set, false);
2919 }
2920 mutex_unlock(&set->tag_list_lock);
2921 INIT_LIST_HEAD(&q->tag_set_list);
2922}
2923
2924static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2925 struct request_queue *q)
2926{
2927 mutex_lock(&set->tag_list_lock);
2928
2929 /*
2930 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2931 */
2932 if (!list_empty(&set->tag_list) &&
2933 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2934 set->flags |= BLK_MQ_F_TAG_SHARED;
2935 /* update existing queue */
2936 blk_mq_update_tag_set_depth(set, true);
2937 }
2938 if (set->flags & BLK_MQ_F_TAG_SHARED)
2939 queue_set_hctx_shared(q, true);
2940 list_add_tail(&q->tag_set_list, &set->tag_list);
2941
2942 mutex_unlock(&set->tag_list_lock);
2943}
2944
2945/* All allocations will be freed in release handler of q->mq_kobj */
2946static int blk_mq_alloc_ctxs(struct request_queue *q)
2947{
2948 struct blk_mq_ctxs *ctxs;
2949 int cpu;
2950
2951 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2952 if (!ctxs)
2953 return -ENOMEM;
2954
2955 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2956 if (!ctxs->queue_ctx)
2957 goto fail;
2958
2959 for_each_possible_cpu(cpu) {
2960 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2961 ctx->ctxs = ctxs;
2962 }
2963
2964 q->mq_kobj = &ctxs->kobj;
2965 q->queue_ctx = ctxs->queue_ctx;
2966
2967 return 0;
2968 fail:
2969 kfree(ctxs);
2970 return -ENOMEM;
2971}
2972
2973/*
2974 * It is the actual release handler for mq, but we do it from
2975 * request queue's release handler for avoiding use-after-free
2976 * and headache because q->mq_kobj shouldn't have been introduced,
2977 * but we can't group ctx/kctx kobj without it.
2978 */
2979void blk_mq_release(struct request_queue *q)
2980{
2981 struct blk_mq_hw_ctx *hctx, *next;
2982 int i;
2983
2984 queue_for_each_hw_ctx(q, hctx, i)
2985 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
2986
2987 /* all hctx are in .unused_hctx_list now */
2988 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
2989 list_del_init(&hctx->hctx_list);
2990 kobject_put(&hctx->kobj);
2991 }
2992
2993 kfree(q->queue_hw_ctx);
2994
2995 /*
2996 * release .mq_kobj and sw queue's kobject now because
2997 * both share lifetime with request queue.
2998 */
2999 blk_mq_sysfs_deinit(q);
3000}
3001
3002struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3003 void *queuedata)
3004{
3005 struct request_queue *uninit_q, *q;
3006
3007 uninit_q = blk_alloc_queue(set->numa_node);
3008 if (!uninit_q)
3009 return ERR_PTR(-ENOMEM);
3010 uninit_q->queuedata = queuedata;
3011
3012 /*
3013 * Initialize the queue without an elevator. device_add_disk() will do
3014 * the initialization.
3015 */
3016 q = blk_mq_init_allocated_queue(set, uninit_q, false);
3017 if (IS_ERR(q))
3018 blk_cleanup_queue(uninit_q);
3019
3020 return q;
3021}
3022EXPORT_SYMBOL_GPL(blk_mq_init_queue_data);
3023
3024struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3025{
3026 return blk_mq_init_queue_data(set, NULL);
3027}
3028EXPORT_SYMBOL(blk_mq_init_queue);
3029
3030/*
3031 * Helper for setting up a queue with mq ops, given queue depth, and
3032 * the passed in mq ops flags.
3033 */
3034struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
3035 const struct blk_mq_ops *ops,
3036 unsigned int queue_depth,
3037 unsigned int set_flags)
3038{
3039 struct request_queue *q;
3040 int ret;
3041
3042 memset(set, 0, sizeof(*set));
3043 set->ops = ops;
3044 set->nr_hw_queues = 1;
3045 set->nr_maps = 1;
3046 set->queue_depth = queue_depth;
3047 set->numa_node = NUMA_NO_NODE;
3048 set->flags = set_flags;
3049
3050 ret = blk_mq_alloc_tag_set(set);
3051 if (ret)
3052 return ERR_PTR(ret);
3053
3054 q = blk_mq_init_queue(set);
3055 if (IS_ERR(q)) {
3056 blk_mq_free_tag_set(set);
3057 return q;
3058 }
3059
3060 return q;
3061}
3062EXPORT_SYMBOL(blk_mq_init_sq_queue);
3063
3064static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
3065 struct blk_mq_tag_set *set, struct request_queue *q,
3066 int hctx_idx, int node)
3067{
3068 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
3069
3070 /* reuse dead hctx first */
3071 spin_lock(&q->unused_hctx_lock);
3072 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
3073 if (tmp->numa_node == node) {
3074 hctx = tmp;
3075 break;
3076 }
3077 }
3078 if (hctx)
3079 list_del_init(&hctx->hctx_list);
3080 spin_unlock(&q->unused_hctx_lock);
3081
3082 if (!hctx)
3083 hctx = blk_mq_alloc_hctx(q, set, node);
3084 if (!hctx)
3085 goto fail;
3086
3087 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
3088 goto free_hctx;
3089
3090 return hctx;
3091
3092 free_hctx:
3093 kobject_put(&hctx->kobj);
3094 fail:
3095 return NULL;
3096}
3097
3098static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
3099 struct request_queue *q)
3100{
3101 int i, j, end;
3102 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
3103
3104 if (q->nr_hw_queues < set->nr_hw_queues) {
3105 struct blk_mq_hw_ctx **new_hctxs;
3106
3107 new_hctxs = kcalloc_node(set->nr_hw_queues,
3108 sizeof(*new_hctxs), GFP_KERNEL,
3109 set->numa_node);
3110 if (!new_hctxs)
3111 return;
3112 if (hctxs)
3113 memcpy(new_hctxs, hctxs, q->nr_hw_queues *
3114 sizeof(*hctxs));
3115 q->queue_hw_ctx = new_hctxs;
3116 kfree(hctxs);
3117 hctxs = new_hctxs;
3118 }
3119
3120 /* protect against switching io scheduler */
3121 mutex_lock(&q->sysfs_lock);
3122 for (i = 0; i < set->nr_hw_queues; i++) {
3123 int node;
3124 struct blk_mq_hw_ctx *hctx;
3125
3126 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
3127 /*
3128 * If the hw queue has been mapped to another numa node,
3129 * we need to realloc the hctx. If allocation fails, fallback
3130 * to use the previous one.
3131 */
3132 if (hctxs[i] && (hctxs[i]->numa_node == node))
3133 continue;
3134
3135 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
3136 if (hctx) {
3137 if (hctxs[i])
3138 blk_mq_exit_hctx(q, set, hctxs[i], i);
3139 hctxs[i] = hctx;
3140 } else {
3141 if (hctxs[i])
3142 pr_warn("Allocate new hctx on node %d fails,\
3143 fallback to previous one on node %d\n",
3144 node, hctxs[i]->numa_node);
3145 else
3146 break;
3147 }
3148 }
3149 /*
3150 * Increasing nr_hw_queues fails. Free the newly allocated
3151 * hctxs and keep the previous q->nr_hw_queues.
3152 */
3153 if (i != set->nr_hw_queues) {
3154 j = q->nr_hw_queues;
3155 end = i;
3156 } else {
3157 j = i;
3158 end = q->nr_hw_queues;
3159 q->nr_hw_queues = set->nr_hw_queues;
3160 }
3161
3162 for (; j < end; j++) {
3163 struct blk_mq_hw_ctx *hctx = hctxs[j];
3164
3165 if (hctx) {
3166 if (hctx->tags)
3167 blk_mq_free_map_and_requests(set, j);
3168 blk_mq_exit_hctx(q, set, hctx, j);
3169 hctxs[j] = NULL;
3170 }
3171 }
3172 mutex_unlock(&q->sysfs_lock);
3173}
3174
3175struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
3176 struct request_queue *q,
3177 bool elevator_init)
3178{
3179 /* mark the queue as mq asap */
3180 q->mq_ops = set->ops;
3181
3182 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
3183 blk_mq_poll_stats_bkt,
3184 BLK_MQ_POLL_STATS_BKTS, q);
3185 if (!q->poll_cb)
3186 goto err_exit;
3187
3188 if (blk_mq_alloc_ctxs(q))
3189 goto err_poll;
3190
3191 /* init q->mq_kobj and sw queues' kobjects */
3192 blk_mq_sysfs_init(q);
3193
3194 INIT_LIST_HEAD(&q->unused_hctx_list);
3195 spin_lock_init(&q->unused_hctx_lock);
3196
3197 blk_mq_realloc_hw_ctxs(set, q);
3198 if (!q->nr_hw_queues)
3199 goto err_hctxs;
3200
3201 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
3202 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
3203
3204 q->tag_set = set;
3205
3206 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
3207 if (set->nr_maps > HCTX_TYPE_POLL &&
3208 set->map[HCTX_TYPE_POLL].nr_queues)
3209 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
3210
3211 q->sg_reserved_size = INT_MAX;
3212
3213 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
3214 INIT_LIST_HEAD(&q->requeue_list);
3215 spin_lock_init(&q->requeue_lock);
3216
3217 q->nr_requests = set->queue_depth;
3218
3219 /*
3220 * Default to classic polling
3221 */
3222 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
3223
3224 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
3225 blk_mq_add_queue_tag_set(set, q);
3226 blk_mq_map_swqueue(q);
3227
3228 if (elevator_init)
3229 elevator_init_mq(q);
3230
3231 return q;
3232
3233err_hctxs:
3234 kfree(q->queue_hw_ctx);
3235 q->nr_hw_queues = 0;
3236 blk_mq_sysfs_deinit(q);
3237err_poll:
3238 blk_stat_free_callback(q->poll_cb);
3239 q->poll_cb = NULL;
3240err_exit:
3241 q->mq_ops = NULL;
3242 return ERR_PTR(-ENOMEM);
3243}
3244EXPORT_SYMBOL(blk_mq_init_allocated_queue);
3245
3246/* tags can _not_ be used after returning from blk_mq_exit_queue */
3247void blk_mq_exit_queue(struct request_queue *q)
3248{
3249 struct blk_mq_tag_set *set = q->tag_set;
3250
3251 blk_mq_del_queue_tag_set(q);
3252 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
3253}
3254
3255static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
3256{
3257 int i;
3258
3259 for (i = 0; i < set->nr_hw_queues; i++)
3260 if (!__blk_mq_alloc_map_and_request(set, i))
3261 goto out_unwind;
3262
3263 return 0;
3264
3265out_unwind:
3266 while (--i >= 0)
3267 blk_mq_free_map_and_requests(set, i);
3268
3269 return -ENOMEM;
3270}
3271
3272/*
3273 * Allocate the request maps associated with this tag_set. Note that this
3274 * may reduce the depth asked for, if memory is tight. set->queue_depth
3275 * will be updated to reflect the allocated depth.
3276 */
3277static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set *set)
3278{
3279 unsigned int depth;
3280 int err;
3281
3282 depth = set->queue_depth;
3283 do {
3284 err = __blk_mq_alloc_rq_maps(set);
3285 if (!err)
3286 break;
3287
3288 set->queue_depth >>= 1;
3289 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
3290 err = -ENOMEM;
3291 break;
3292 }
3293 } while (set->queue_depth);
3294
3295 if (!set->queue_depth || err) {
3296 pr_err("blk-mq: failed to allocate request map\n");
3297 return -ENOMEM;
3298 }
3299
3300 if (depth != set->queue_depth)
3301 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3302 depth, set->queue_depth);
3303
3304 return 0;
3305}
3306
3307static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3308{
3309 /*
3310 * blk_mq_map_queues() and multiple .map_queues() implementations
3311 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3312 * number of hardware queues.
3313 */
3314 if (set->nr_maps == 1)
3315 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
3316
3317 if (set->ops->map_queues && !is_kdump_kernel()) {
3318 int i;
3319
3320 /*
3321 * transport .map_queues is usually done in the following
3322 * way:
3323 *
3324 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3325 * mask = get_cpu_mask(queue)
3326 * for_each_cpu(cpu, mask)
3327 * set->map[x].mq_map[cpu] = queue;
3328 * }
3329 *
3330 * When we need to remap, the table has to be cleared for
3331 * killing stale mapping since one CPU may not be mapped
3332 * to any hw queue.
3333 */
3334 for (i = 0; i < set->nr_maps; i++)
3335 blk_mq_clear_mq_map(&set->map[i]);
3336
3337 return set->ops->map_queues(set);
3338 } else {
3339 BUG_ON(set->nr_maps > 1);
3340 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3341 }
3342}
3343
3344static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
3345 int cur_nr_hw_queues, int new_nr_hw_queues)
3346{
3347 struct blk_mq_tags **new_tags;
3348
3349 if (cur_nr_hw_queues >= new_nr_hw_queues)
3350 return 0;
3351
3352 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
3353 GFP_KERNEL, set->numa_node);
3354 if (!new_tags)
3355 return -ENOMEM;
3356
3357 if (set->tags)
3358 memcpy(new_tags, set->tags, cur_nr_hw_queues *
3359 sizeof(*set->tags));
3360 kfree(set->tags);
3361 set->tags = new_tags;
3362 set->nr_hw_queues = new_nr_hw_queues;
3363
3364 return 0;
3365}
3366
3367/*
3368 * Alloc a tag set to be associated with one or more request queues.
3369 * May fail with EINVAL for various error conditions. May adjust the
3370 * requested depth down, if it's too large. In that case, the set
3371 * value will be stored in set->queue_depth.
3372 */
3373int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3374{
3375 int i, ret;
3376
3377 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3378
3379 if (!set->nr_hw_queues)
3380 return -EINVAL;
3381 if (!set->queue_depth)
3382 return -EINVAL;
3383 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3384 return -EINVAL;
3385
3386 if (!set->ops->queue_rq)
3387 return -EINVAL;
3388
3389 if (!set->ops->get_budget ^ !set->ops->put_budget)
3390 return -EINVAL;
3391
3392 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3393 pr_info("blk-mq: reduced tag depth to %u\n",
3394 BLK_MQ_MAX_DEPTH);
3395 set->queue_depth = BLK_MQ_MAX_DEPTH;
3396 }
3397
3398 if (!set->nr_maps)
3399 set->nr_maps = 1;
3400 else if (set->nr_maps > HCTX_MAX_TYPES)
3401 return -EINVAL;
3402
3403 /*
3404 * If a crashdump is active, then we are potentially in a very
3405 * memory constrained environment. Limit us to 1 queue and
3406 * 64 tags to prevent using too much memory.
3407 */
3408 if (is_kdump_kernel()) {
3409 set->nr_hw_queues = 1;
3410 set->nr_maps = 1;
3411 set->queue_depth = min(64U, set->queue_depth);
3412 }
3413 /*
3414 * There is no use for more h/w queues than cpus if we just have
3415 * a single map
3416 */
3417 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3418 set->nr_hw_queues = nr_cpu_ids;
3419
3420 if (blk_mq_realloc_tag_set_tags(set, 0, set->nr_hw_queues) < 0)
3421 return -ENOMEM;
3422
3423 ret = -ENOMEM;
3424 for (i = 0; i < set->nr_maps; i++) {
3425 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3426 sizeof(set->map[i].mq_map[0]),
3427 GFP_KERNEL, set->numa_node);
3428 if (!set->map[i].mq_map)
3429 goto out_free_mq_map;
3430 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3431 }
3432
3433 ret = blk_mq_update_queue_map(set);
3434 if (ret)
3435 goto out_free_mq_map;
3436
3437 ret = blk_mq_alloc_map_and_requests(set);
3438 if (ret)
3439 goto out_free_mq_map;
3440
3441 mutex_init(&set->tag_list_lock);
3442 INIT_LIST_HEAD(&set->tag_list);
3443
3444 return 0;
3445
3446out_free_mq_map:
3447 for (i = 0; i < set->nr_maps; i++) {
3448 kfree(set->map[i].mq_map);
3449 set->map[i].mq_map = NULL;
3450 }
3451 kfree(set->tags);
3452 set->tags = NULL;
3453 return ret;
3454}
3455EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3456
3457void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3458{
3459 int i, j;
3460
3461 for (i = 0; i < set->nr_hw_queues; i++)
3462 blk_mq_free_map_and_requests(set, i);
3463
3464 for (j = 0; j < set->nr_maps; j++) {
3465 kfree(set->map[j].mq_map);
3466 set->map[j].mq_map = NULL;
3467 }
3468
3469 kfree(set->tags);
3470 set->tags = NULL;
3471}
3472EXPORT_SYMBOL(blk_mq_free_tag_set);
3473
3474int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3475{
3476 struct blk_mq_tag_set *set = q->tag_set;
3477 struct blk_mq_hw_ctx *hctx;
3478 int i, ret;
3479
3480 if (!set)
3481 return -EINVAL;
3482
3483 if (q->nr_requests == nr)
3484 return 0;
3485
3486 blk_mq_freeze_queue(q);
3487 blk_mq_quiesce_queue(q);
3488
3489 ret = 0;
3490 queue_for_each_hw_ctx(q, hctx, i) {
3491 if (!hctx->tags)
3492 continue;
3493 /*
3494 * If we're using an MQ scheduler, just update the scheduler
3495 * queue depth. This is similar to what the old code would do.
3496 */
3497 if (!hctx->sched_tags) {
3498 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3499 false);
3500 } else {
3501 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3502 nr, true);
3503 }
3504 if (ret)
3505 break;
3506 if (q->elevator && q->elevator->type->ops.depth_updated)
3507 q->elevator->type->ops.depth_updated(hctx);
3508 }
3509
3510 if (!ret)
3511 q->nr_requests = nr;
3512
3513 blk_mq_unquiesce_queue(q);
3514 blk_mq_unfreeze_queue(q);
3515
3516 return ret;
3517}
3518
3519/*
3520 * request_queue and elevator_type pair.
3521 * It is just used by __blk_mq_update_nr_hw_queues to cache
3522 * the elevator_type associated with a request_queue.
3523 */
3524struct blk_mq_qe_pair {
3525 struct list_head node;
3526 struct request_queue *q;
3527 struct elevator_type *type;
3528};
3529
3530/*
3531 * Cache the elevator_type in qe pair list and switch the
3532 * io scheduler to 'none'
3533 */
3534static bool blk_mq_elv_switch_none(struct list_head *head,
3535 struct request_queue *q)
3536{
3537 struct blk_mq_qe_pair *qe;
3538
3539 if (!q->elevator)
3540 return true;
3541
3542 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3543 if (!qe)
3544 return false;
3545
3546 INIT_LIST_HEAD(&qe->node);
3547 qe->q = q;
3548 qe->type = q->elevator->type;
3549 list_add(&qe->node, head);
3550
3551 mutex_lock(&q->sysfs_lock);
3552 /*
3553 * After elevator_switch_mq, the previous elevator_queue will be
3554 * released by elevator_release. The reference of the io scheduler
3555 * module get by elevator_get will also be put. So we need to get
3556 * a reference of the io scheduler module here to prevent it to be
3557 * removed.
3558 */
3559 __module_get(qe->type->elevator_owner);
3560 elevator_switch_mq(q, NULL);
3561 mutex_unlock(&q->sysfs_lock);
3562
3563 return true;
3564}
3565
3566static void blk_mq_elv_switch_back(struct list_head *head,
3567 struct request_queue *q)
3568{
3569 struct blk_mq_qe_pair *qe;
3570 struct elevator_type *t = NULL;
3571
3572 list_for_each_entry(qe, head, node)
3573 if (qe->q == q) {
3574 t = qe->type;
3575 break;
3576 }
3577
3578 if (!t)
3579 return;
3580
3581 list_del(&qe->node);
3582 kfree(qe);
3583
3584 mutex_lock(&q->sysfs_lock);
3585 elevator_switch_mq(q, t);
3586 mutex_unlock(&q->sysfs_lock);
3587}
3588
3589static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3590 int nr_hw_queues)
3591{
3592 struct request_queue *q;
3593 LIST_HEAD(head);
3594 int prev_nr_hw_queues;
3595
3596 lockdep_assert_held(&set->tag_list_lock);
3597
3598 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3599 nr_hw_queues = nr_cpu_ids;
3600 if (nr_hw_queues < 1)
3601 return;
3602 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
3603 return;
3604
3605 list_for_each_entry(q, &set->tag_list, tag_set_list)
3606 blk_mq_freeze_queue(q);
3607 /*
3608 * Switch IO scheduler to 'none', cleaning up the data associated
3609 * with the previous scheduler. We will switch back once we are done
3610 * updating the new sw to hw queue mappings.
3611 */
3612 list_for_each_entry(q, &set->tag_list, tag_set_list)
3613 if (!blk_mq_elv_switch_none(&head, q))
3614 goto switch_back;
3615
3616 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3617 blk_mq_debugfs_unregister_hctxs(q);
3618 blk_mq_sysfs_unregister(q);
3619 }
3620
3621 prev_nr_hw_queues = set->nr_hw_queues;
3622 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
3623 0)
3624 goto reregister;
3625
3626 set->nr_hw_queues = nr_hw_queues;
3627fallback:
3628 blk_mq_update_queue_map(set);
3629 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3630 blk_mq_realloc_hw_ctxs(set, q);
3631 if (q->nr_hw_queues != set->nr_hw_queues) {
3632 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3633 nr_hw_queues, prev_nr_hw_queues);
3634 set->nr_hw_queues = prev_nr_hw_queues;
3635 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3636 goto fallback;
3637 }
3638 blk_mq_map_swqueue(q);
3639 }
3640
3641reregister:
3642 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3643 blk_mq_sysfs_register(q);
3644 blk_mq_debugfs_register_hctxs(q);
3645 }
3646
3647switch_back:
3648 list_for_each_entry(q, &set->tag_list, tag_set_list)
3649 blk_mq_elv_switch_back(&head, q);
3650
3651 list_for_each_entry(q, &set->tag_list, tag_set_list)
3652 blk_mq_unfreeze_queue(q);
3653}
3654
3655void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3656{
3657 mutex_lock(&set->tag_list_lock);
3658 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3659 mutex_unlock(&set->tag_list_lock);
3660}
3661EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3662
3663/* Enable polling stats and return whether they were already enabled. */
3664static bool blk_poll_stats_enable(struct request_queue *q)
3665{
3666 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3667 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3668 return true;
3669 blk_stat_add_callback(q, q->poll_cb);
3670 return false;
3671}
3672
3673static void blk_mq_poll_stats_start(struct request_queue *q)
3674{
3675 /*
3676 * We don't arm the callback if polling stats are not enabled or the
3677 * callback is already active.
3678 */
3679 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3680 blk_stat_is_active(q->poll_cb))
3681 return;
3682
3683 blk_stat_activate_msecs(q->poll_cb, 100);
3684}
3685
3686static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3687{
3688 struct request_queue *q = cb->data;
3689 int bucket;
3690
3691 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3692 if (cb->stat[bucket].nr_samples)
3693 q->poll_stat[bucket] = cb->stat[bucket];
3694 }
3695}
3696
3697static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3698 struct request *rq)
3699{
3700 unsigned long ret = 0;
3701 int bucket;
3702
3703 /*
3704 * If stats collection isn't on, don't sleep but turn it on for
3705 * future users
3706 */
3707 if (!blk_poll_stats_enable(q))
3708 return 0;
3709
3710 /*
3711 * As an optimistic guess, use half of the mean service time
3712 * for this type of request. We can (and should) make this smarter.
3713 * For instance, if the completion latencies are tight, we can
3714 * get closer than just half the mean. This is especially
3715 * important on devices where the completion latencies are longer
3716 * than ~10 usec. We do use the stats for the relevant IO size
3717 * if available which does lead to better estimates.
3718 */
3719 bucket = blk_mq_poll_stats_bkt(rq);
3720 if (bucket < 0)
3721 return ret;
3722
3723 if (q->poll_stat[bucket].nr_samples)
3724 ret = (q->poll_stat[bucket].mean + 1) / 2;
3725
3726 return ret;
3727}
3728
3729static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3730 struct request *rq)
3731{
3732 struct hrtimer_sleeper hs;
3733 enum hrtimer_mode mode;
3734 unsigned int nsecs;
3735 ktime_t kt;
3736
3737 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3738 return false;
3739
3740 /*
3741 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3742 *
3743 * 0: use half of prev avg
3744 * >0: use this specific value
3745 */
3746 if (q->poll_nsec > 0)
3747 nsecs = q->poll_nsec;
3748 else
3749 nsecs = blk_mq_poll_nsecs(q, rq);
3750
3751 if (!nsecs)
3752 return false;
3753
3754 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3755
3756 /*
3757 * This will be replaced with the stats tracking code, using
3758 * 'avg_completion_time / 2' as the pre-sleep target.
3759 */
3760 kt = nsecs;
3761
3762 mode = HRTIMER_MODE_REL;
3763 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3764 hrtimer_set_expires(&hs.timer, kt);
3765
3766 do {
3767 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3768 break;
3769 set_current_state(TASK_UNINTERRUPTIBLE);
3770 hrtimer_sleeper_start_expires(&hs, mode);
3771 if (hs.task)
3772 io_schedule();
3773 hrtimer_cancel(&hs.timer);
3774 mode = HRTIMER_MODE_ABS;
3775 } while (hs.task && !signal_pending(current));
3776
3777 __set_current_state(TASK_RUNNING);
3778 destroy_hrtimer_on_stack(&hs.timer);
3779 return true;
3780}
3781
3782static bool blk_mq_poll_hybrid(struct request_queue *q,
3783 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3784{
3785 struct request *rq;
3786
3787 if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3788 return false;
3789
3790 if (!blk_qc_t_is_internal(cookie))
3791 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3792 else {
3793 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3794 /*
3795 * With scheduling, if the request has completed, we'll
3796 * get a NULL return here, as we clear the sched tag when
3797 * that happens. The request still remains valid, like always,
3798 * so we should be safe with just the NULL check.
3799 */
3800 if (!rq)
3801 return false;
3802 }
3803
3804 return blk_mq_poll_hybrid_sleep(q, rq);
3805}
3806
3807/**
3808 * blk_poll - poll for IO completions
3809 * @q: the queue
3810 * @cookie: cookie passed back at IO submission time
3811 * @spin: whether to spin for completions
3812 *
3813 * Description:
3814 * Poll for completions on the passed in queue. Returns number of
3815 * completed entries found. If @spin is true, then blk_poll will continue
3816 * looping until at least one completion is found, unless the task is
3817 * otherwise marked running (or we need to reschedule).
3818 */
3819int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3820{
3821 struct blk_mq_hw_ctx *hctx;
3822 long state;
3823
3824 if (!blk_qc_t_valid(cookie) ||
3825 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3826 return 0;
3827
3828 if (current->plug)
3829 blk_flush_plug_list(current->plug, false);
3830
3831 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3832
3833 /*
3834 * If we sleep, have the caller restart the poll loop to reset
3835 * the state. Like for the other success return cases, the
3836 * caller is responsible for checking if the IO completed. If
3837 * the IO isn't complete, we'll get called again and will go
3838 * straight to the busy poll loop.
3839 */
3840 if (blk_mq_poll_hybrid(q, hctx, cookie))
3841 return 1;
3842
3843 hctx->poll_considered++;
3844
3845 state = current->state;
3846 do {
3847 int ret;
3848
3849 hctx->poll_invoked++;
3850
3851 ret = q->mq_ops->poll(hctx);
3852 if (ret > 0) {
3853 hctx->poll_success++;
3854 __set_current_state(TASK_RUNNING);
3855 return ret;
3856 }
3857
3858 if (signal_pending_state(state, current))
3859 __set_current_state(TASK_RUNNING);
3860
3861 if (current->state == TASK_RUNNING)
3862 return 1;
3863 if (ret < 0 || !spin)
3864 break;
3865 cpu_relax();
3866 } while (!need_resched());
3867
3868 __set_current_state(TASK_RUNNING);
3869 return 0;
3870}
3871EXPORT_SYMBOL_GPL(blk_poll);
3872
3873unsigned int blk_mq_rq_cpu(struct request *rq)
3874{
3875 return rq->mq_ctx->cpu;
3876}
3877EXPORT_SYMBOL(blk_mq_rq_cpu);
3878
3879static int __init blk_mq_init(void)
3880{
3881 int i;
3882
3883 for_each_possible_cpu(i)
3884 INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
3885 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
3886
3887 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
3888 "block/softirq:dead", NULL,
3889 blk_softirq_cpu_dead);
3890 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3891 blk_mq_hctx_notify_dead);
3892 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
3893 blk_mq_hctx_notify_online,
3894 blk_mq_hctx_notify_offline);
3895 return 0;
3896}
3897subsys_initcall(blk_mq_init);