Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * blk-mq scheduling framework
4 *
5 * Copyright (C) 2016 Jens Axboe
6 */
7#include <linux/kernel.h>
8#include <linux/module.h>
9#include <linux/blk-mq.h>
10#include <linux/list_sort.h>
11
12#include <trace/events/block.h>
13
14#include "blk.h"
15#include "blk-mq.h"
16#include "blk-mq-debugfs.h"
17#include "blk-mq-sched.h"
18#include "blk-mq-tag.h"
19#include "blk-wbt.h"
20
21/*
22 * Mark a hardware queue as needing a restart. For shared queues, maintain
23 * a count of how many hardware queues are marked for restart.
24 */
25void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx)
26{
27 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
28 return;
29
30 set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
31}
32EXPORT_SYMBOL_GPL(blk_mq_sched_mark_restart_hctx);
33
34void __blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx)
35{
36 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
37
38 /*
39 * Order clearing SCHED_RESTART and list_empty_careful(&hctx->dispatch)
40 * in blk_mq_run_hw_queue(). Its pair is the barrier in
41 * blk_mq_dispatch_rq_list(). So dispatch code won't see SCHED_RESTART,
42 * meantime new request added to hctx->dispatch is missed to check in
43 * blk_mq_run_hw_queue().
44 */
45 smp_mb();
46
47 blk_mq_run_hw_queue(hctx, true);
48}
49
50static int sched_rq_cmp(void *priv, const struct list_head *a,
51 const struct list_head *b)
52{
53 struct request *rqa = container_of(a, struct request, queuelist);
54 struct request *rqb = container_of(b, struct request, queuelist);
55
56 return rqa->mq_hctx > rqb->mq_hctx;
57}
58
59static bool blk_mq_dispatch_hctx_list(struct list_head *rq_list)
60{
61 struct blk_mq_hw_ctx *hctx =
62 list_first_entry(rq_list, struct request, queuelist)->mq_hctx;
63 struct request *rq;
64 LIST_HEAD(hctx_list);
65 unsigned int count = 0;
66
67 list_for_each_entry(rq, rq_list, queuelist) {
68 if (rq->mq_hctx != hctx) {
69 list_cut_before(&hctx_list, rq_list, &rq->queuelist);
70 goto dispatch;
71 }
72 count++;
73 }
74 list_splice_tail_init(rq_list, &hctx_list);
75
76dispatch:
77 return blk_mq_dispatch_rq_list(hctx, &hctx_list, count);
78}
79
80#define BLK_MQ_BUDGET_DELAY 3 /* ms units */
81
82/*
83 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
84 * its queue by itself in its completion handler, so we don't need to
85 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
86 *
87 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
88 * be run again. This is necessary to avoid starving flushes.
89 */
90static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
91{
92 struct request_queue *q = hctx->queue;
93 struct elevator_queue *e = q->elevator;
94 bool multi_hctxs = false, run_queue = false;
95 bool dispatched = false, busy = false;
96 unsigned int max_dispatch;
97 LIST_HEAD(rq_list);
98 int count = 0;
99
100 if (hctx->dispatch_busy)
101 max_dispatch = 1;
102 else
103 max_dispatch = hctx->queue->nr_requests;
104
105 do {
106 struct request *rq;
107 int budget_token;
108
109 if (e->type->ops.has_work && !e->type->ops.has_work(hctx))
110 break;
111
112 if (!list_empty_careful(&hctx->dispatch)) {
113 busy = true;
114 break;
115 }
116
117 budget_token = blk_mq_get_dispatch_budget(q);
118 if (budget_token < 0)
119 break;
120
121 rq = e->type->ops.dispatch_request(hctx);
122 if (!rq) {
123 blk_mq_put_dispatch_budget(q, budget_token);
124 /*
125 * We're releasing without dispatching. Holding the
126 * budget could have blocked any "hctx"s with the
127 * same queue and if we didn't dispatch then there's
128 * no guarantee anyone will kick the queue. Kick it
129 * ourselves.
130 */
131 run_queue = true;
132 break;
133 }
134
135 blk_mq_set_rq_budget_token(rq, budget_token);
136
137 /*
138 * Now this rq owns the budget which has to be released
139 * if this rq won't be queued to driver via .queue_rq()
140 * in blk_mq_dispatch_rq_list().
141 */
142 list_add_tail(&rq->queuelist, &rq_list);
143 count++;
144 if (rq->mq_hctx != hctx)
145 multi_hctxs = true;
146
147 /*
148 * If we cannot get tag for the request, stop dequeueing
149 * requests from the IO scheduler. We are unlikely to be able
150 * to submit them anyway and it creates false impression for
151 * scheduling heuristics that the device can take more IO.
152 */
153 if (!blk_mq_get_driver_tag(rq))
154 break;
155 } while (count < max_dispatch);
156
157 if (!count) {
158 if (run_queue)
159 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
160 } else if (multi_hctxs) {
161 /*
162 * Requests from different hctx may be dequeued from some
163 * schedulers, such as bfq and deadline.
164 *
165 * Sort the requests in the list according to their hctx,
166 * dispatch batching requests from same hctx at a time.
167 */
168 list_sort(NULL, &rq_list, sched_rq_cmp);
169 do {
170 dispatched |= blk_mq_dispatch_hctx_list(&rq_list);
171 } while (!list_empty(&rq_list));
172 } else {
173 dispatched = blk_mq_dispatch_rq_list(hctx, &rq_list, count);
174 }
175
176 if (busy)
177 return -EAGAIN;
178 return !!dispatched;
179}
180
181static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
182{
183 unsigned long end = jiffies + HZ;
184 int ret;
185
186 do {
187 ret = __blk_mq_do_dispatch_sched(hctx);
188 if (ret != 1)
189 break;
190 if (need_resched() || time_is_before_jiffies(end)) {
191 blk_mq_delay_run_hw_queue(hctx, 0);
192 break;
193 }
194 } while (1);
195
196 return ret;
197}
198
199static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
200 struct blk_mq_ctx *ctx)
201{
202 unsigned short idx = ctx->index_hw[hctx->type];
203
204 if (++idx == hctx->nr_ctx)
205 idx = 0;
206
207 return hctx->ctxs[idx];
208}
209
210/*
211 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
212 * its queue by itself in its completion handler, so we don't need to
213 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
214 *
215 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
216 * be run again. This is necessary to avoid starving flushes.
217 */
218static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
219{
220 struct request_queue *q = hctx->queue;
221 LIST_HEAD(rq_list);
222 struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from);
223 int ret = 0;
224 struct request *rq;
225
226 do {
227 int budget_token;
228
229 if (!list_empty_careful(&hctx->dispatch)) {
230 ret = -EAGAIN;
231 break;
232 }
233
234 if (!sbitmap_any_bit_set(&hctx->ctx_map))
235 break;
236
237 budget_token = blk_mq_get_dispatch_budget(q);
238 if (budget_token < 0)
239 break;
240
241 rq = blk_mq_dequeue_from_ctx(hctx, ctx);
242 if (!rq) {
243 blk_mq_put_dispatch_budget(q, budget_token);
244 /*
245 * We're releasing without dispatching. Holding the
246 * budget could have blocked any "hctx"s with the
247 * same queue and if we didn't dispatch then there's
248 * no guarantee anyone will kick the queue. Kick it
249 * ourselves.
250 */
251 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
252 break;
253 }
254
255 blk_mq_set_rq_budget_token(rq, budget_token);
256
257 /*
258 * Now this rq owns the budget which has to be released
259 * if this rq won't be queued to driver via .queue_rq()
260 * in blk_mq_dispatch_rq_list().
261 */
262 list_add(&rq->queuelist, &rq_list);
263
264 /* round robin for fair dispatch */
265 ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
266
267 } while (blk_mq_dispatch_rq_list(rq->mq_hctx, &rq_list, 1));
268
269 WRITE_ONCE(hctx->dispatch_from, ctx);
270 return ret;
271}
272
273static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
274{
275 struct request_queue *q = hctx->queue;
276 const bool has_sched = q->elevator;
277 int ret = 0;
278 LIST_HEAD(rq_list);
279
280 /*
281 * If we have previous entries on our dispatch list, grab them first for
282 * more fair dispatch.
283 */
284 if (!list_empty_careful(&hctx->dispatch)) {
285 spin_lock(&hctx->lock);
286 if (!list_empty(&hctx->dispatch))
287 list_splice_init(&hctx->dispatch, &rq_list);
288 spin_unlock(&hctx->lock);
289 }
290
291 /*
292 * Only ask the scheduler for requests, if we didn't have residual
293 * requests from the dispatch list. This is to avoid the case where
294 * we only ever dispatch a fraction of the requests available because
295 * of low device queue depth. Once we pull requests out of the IO
296 * scheduler, we can no longer merge or sort them. So it's best to
297 * leave them there for as long as we can. Mark the hw queue as
298 * needing a restart in that case.
299 *
300 * We want to dispatch from the scheduler if there was nothing
301 * on the dispatch list or we were able to dispatch from the
302 * dispatch list.
303 */
304 if (!list_empty(&rq_list)) {
305 blk_mq_sched_mark_restart_hctx(hctx);
306 if (blk_mq_dispatch_rq_list(hctx, &rq_list, 0)) {
307 if (has_sched)
308 ret = blk_mq_do_dispatch_sched(hctx);
309 else
310 ret = blk_mq_do_dispatch_ctx(hctx);
311 }
312 } else if (has_sched) {
313 ret = blk_mq_do_dispatch_sched(hctx);
314 } else if (hctx->dispatch_busy) {
315 /* dequeue request one by one from sw queue if queue is busy */
316 ret = blk_mq_do_dispatch_ctx(hctx);
317 } else {
318 blk_mq_flush_busy_ctxs(hctx, &rq_list);
319 blk_mq_dispatch_rq_list(hctx, &rq_list, 0);
320 }
321
322 return ret;
323}
324
325void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
326{
327 struct request_queue *q = hctx->queue;
328
329 /* RCU or SRCU read lock is needed before checking quiesced flag */
330 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
331 return;
332
333 hctx->run++;
334
335 /*
336 * A return of -EAGAIN is an indication that hctx->dispatch is not
337 * empty and we must run again in order to avoid starving flushes.
338 */
339 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) {
340 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN)
341 blk_mq_run_hw_queue(hctx, true);
342 }
343}
344
345bool blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio,
346 unsigned int nr_segs)
347{
348 struct elevator_queue *e = q->elevator;
349 struct blk_mq_ctx *ctx;
350 struct blk_mq_hw_ctx *hctx;
351 bool ret = false;
352 enum hctx_type type;
353
354 if (e && e->type->ops.bio_merge) {
355 ret = e->type->ops.bio_merge(q, bio, nr_segs);
356 goto out_put;
357 }
358
359 ctx = blk_mq_get_ctx(q);
360 hctx = blk_mq_map_queue(q, bio->bi_opf, ctx);
361 type = hctx->type;
362 if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE) ||
363 list_empty_careful(&ctx->rq_lists[type]))
364 goto out_put;
365
366 /* default per sw-queue merge */
367 spin_lock(&ctx->lock);
368 /*
369 * Reverse check our software queue for entries that we could
370 * potentially merge with. Currently includes a hand-wavy stop
371 * count of 8, to not spend too much time checking for merges.
372 */
373 if (blk_bio_list_merge(q, &ctx->rq_lists[type], bio, nr_segs))
374 ret = true;
375
376 spin_unlock(&ctx->lock);
377out_put:
378 return ret;
379}
380
381bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq,
382 struct list_head *free)
383{
384 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq, free);
385}
386EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
387
388static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
389 struct request *rq)
390{
391 /*
392 * dispatch flush and passthrough rq directly
393 *
394 * passthrough request has to be added to hctx->dispatch directly.
395 * For some reason, device may be in one situation which can't
396 * handle FS request, so STS_RESOURCE is always returned and the
397 * FS request will be added to hctx->dispatch. However passthrough
398 * request may be required at that time for fixing the problem. If
399 * passthrough request is added to scheduler queue, there isn't any
400 * chance to dispatch it given we prioritize requests in hctx->dispatch.
401 */
402 if ((rq->rq_flags & RQF_FLUSH_SEQ) || blk_rq_is_passthrough(rq))
403 return true;
404
405 return false;
406}
407
408void blk_mq_sched_insert_request(struct request *rq, bool at_head,
409 bool run_queue, bool async)
410{
411 struct request_queue *q = rq->q;
412 struct elevator_queue *e = q->elevator;
413 struct blk_mq_ctx *ctx = rq->mq_ctx;
414 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
415
416 WARN_ON(e && (rq->tag != BLK_MQ_NO_TAG));
417
418 if (blk_mq_sched_bypass_insert(hctx, rq)) {
419 /*
420 * Firstly normal IO request is inserted to scheduler queue or
421 * sw queue, meantime we add flush request to dispatch queue(
422 * hctx->dispatch) directly and there is at most one in-flight
423 * flush request for each hw queue, so it doesn't matter to add
424 * flush request to tail or front of the dispatch queue.
425 *
426 * Secondly in case of NCQ, flush request belongs to non-NCQ
427 * command, and queueing it will fail when there is any
428 * in-flight normal IO request(NCQ command). When adding flush
429 * rq to the front of hctx->dispatch, it is easier to introduce
430 * extra time to flush rq's latency because of S_SCHED_RESTART
431 * compared with adding to the tail of dispatch queue, then
432 * chance of flush merge is increased, and less flush requests
433 * will be issued to controller. It is observed that ~10% time
434 * is saved in blktests block/004 on disk attached to AHCI/NCQ
435 * drive when adding flush rq to the front of hctx->dispatch.
436 *
437 * Simply queue flush rq to the front of hctx->dispatch so that
438 * intensive flush workloads can benefit in case of NCQ HW.
439 */
440 at_head = (rq->rq_flags & RQF_FLUSH_SEQ) ? true : at_head;
441 blk_mq_request_bypass_insert(rq, at_head, false);
442 goto run;
443 }
444
445 if (e) {
446 LIST_HEAD(list);
447
448 list_add(&rq->queuelist, &list);
449 e->type->ops.insert_requests(hctx, &list, at_head);
450 } else {
451 spin_lock(&ctx->lock);
452 __blk_mq_insert_request(hctx, rq, at_head);
453 spin_unlock(&ctx->lock);
454 }
455
456run:
457 if (run_queue)
458 blk_mq_run_hw_queue(hctx, async);
459}
460
461void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx *hctx,
462 struct blk_mq_ctx *ctx,
463 struct list_head *list, bool run_queue_async)
464{
465 struct elevator_queue *e;
466 struct request_queue *q = hctx->queue;
467
468 /*
469 * blk_mq_sched_insert_requests() is called from flush plug
470 * context only, and hold one usage counter to prevent queue
471 * from being released.
472 */
473 percpu_ref_get(&q->q_usage_counter);
474
475 e = hctx->queue->elevator;
476 if (e) {
477 e->type->ops.insert_requests(hctx, list, false);
478 } else {
479 /*
480 * try to issue requests directly if the hw queue isn't
481 * busy in case of 'none' scheduler, and this way may save
482 * us one extra enqueue & dequeue to sw queue.
483 */
484 if (!hctx->dispatch_busy && !run_queue_async) {
485 blk_mq_run_dispatch_ops(hctx->queue,
486 blk_mq_try_issue_list_directly(hctx, list));
487 if (list_empty(list))
488 goto out;
489 }
490 blk_mq_insert_requests(hctx, ctx, list);
491 }
492
493 blk_mq_run_hw_queue(hctx, run_queue_async);
494 out:
495 percpu_ref_put(&q->q_usage_counter);
496}
497
498static int blk_mq_sched_alloc_map_and_rqs(struct request_queue *q,
499 struct blk_mq_hw_ctx *hctx,
500 unsigned int hctx_idx)
501{
502 if (blk_mq_is_shared_tags(q->tag_set->flags)) {
503 hctx->sched_tags = q->sched_shared_tags;
504 return 0;
505 }
506
507 hctx->sched_tags = blk_mq_alloc_map_and_rqs(q->tag_set, hctx_idx,
508 q->nr_requests);
509
510 if (!hctx->sched_tags)
511 return -ENOMEM;
512 return 0;
513}
514
515static void blk_mq_exit_sched_shared_tags(struct request_queue *queue)
516{
517 blk_mq_free_rq_map(queue->sched_shared_tags);
518 queue->sched_shared_tags = NULL;
519}
520
521/* called in queue's release handler, tagset has gone away */
522static void blk_mq_sched_tags_teardown(struct request_queue *q, unsigned int flags)
523{
524 struct blk_mq_hw_ctx *hctx;
525 unsigned long i;
526
527 queue_for_each_hw_ctx(q, hctx, i) {
528 if (hctx->sched_tags) {
529 if (!blk_mq_is_shared_tags(flags))
530 blk_mq_free_rq_map(hctx->sched_tags);
531 hctx->sched_tags = NULL;
532 }
533 }
534
535 if (blk_mq_is_shared_tags(flags))
536 blk_mq_exit_sched_shared_tags(q);
537}
538
539static int blk_mq_init_sched_shared_tags(struct request_queue *queue)
540{
541 struct blk_mq_tag_set *set = queue->tag_set;
542
543 /*
544 * Set initial depth at max so that we don't need to reallocate for
545 * updating nr_requests.
546 */
547 queue->sched_shared_tags = blk_mq_alloc_map_and_rqs(set,
548 BLK_MQ_NO_HCTX_IDX,
549 MAX_SCHED_RQ);
550 if (!queue->sched_shared_tags)
551 return -ENOMEM;
552
553 blk_mq_tag_update_sched_shared_tags(queue);
554
555 return 0;
556}
557
558/* caller must have a reference to @e, will grab another one if successful */
559int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
560{
561 unsigned int flags = q->tag_set->flags;
562 struct blk_mq_hw_ctx *hctx;
563 struct elevator_queue *eq;
564 unsigned long i;
565 int ret;
566
567 /*
568 * Default to double of smaller one between hw queue_depth and 128,
569 * since we don't split into sync/async like the old code did.
570 * Additionally, this is a per-hw queue depth.
571 */
572 q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
573 BLKDEV_DEFAULT_RQ);
574
575 if (blk_mq_is_shared_tags(flags)) {
576 ret = blk_mq_init_sched_shared_tags(q);
577 if (ret)
578 return ret;
579 }
580
581 queue_for_each_hw_ctx(q, hctx, i) {
582 ret = blk_mq_sched_alloc_map_and_rqs(q, hctx, i);
583 if (ret)
584 goto err_free_map_and_rqs;
585 }
586
587 ret = e->ops.init_sched(q, e);
588 if (ret)
589 goto err_free_map_and_rqs;
590
591 mutex_lock(&q->debugfs_mutex);
592 blk_mq_debugfs_register_sched(q);
593 mutex_unlock(&q->debugfs_mutex);
594
595 queue_for_each_hw_ctx(q, hctx, i) {
596 if (e->ops.init_hctx) {
597 ret = e->ops.init_hctx(hctx, i);
598 if (ret) {
599 eq = q->elevator;
600 blk_mq_sched_free_rqs(q);
601 blk_mq_exit_sched(q, eq);
602 kobject_put(&eq->kobj);
603 return ret;
604 }
605 }
606 mutex_lock(&q->debugfs_mutex);
607 blk_mq_debugfs_register_sched_hctx(q, hctx);
608 mutex_unlock(&q->debugfs_mutex);
609 }
610
611 return 0;
612
613err_free_map_and_rqs:
614 blk_mq_sched_free_rqs(q);
615 blk_mq_sched_tags_teardown(q, flags);
616
617 q->elevator = NULL;
618 return ret;
619}
620
621/*
622 * called in either blk_queue_cleanup or elevator_switch, tagset
623 * is required for freeing requests
624 */
625void blk_mq_sched_free_rqs(struct request_queue *q)
626{
627 struct blk_mq_hw_ctx *hctx;
628 unsigned long i;
629
630 if (blk_mq_is_shared_tags(q->tag_set->flags)) {
631 blk_mq_free_rqs(q->tag_set, q->sched_shared_tags,
632 BLK_MQ_NO_HCTX_IDX);
633 } else {
634 queue_for_each_hw_ctx(q, hctx, i) {
635 if (hctx->sched_tags)
636 blk_mq_free_rqs(q->tag_set,
637 hctx->sched_tags, i);
638 }
639 }
640}
641
642void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
643{
644 struct blk_mq_hw_ctx *hctx;
645 unsigned long i;
646 unsigned int flags = 0;
647
648 queue_for_each_hw_ctx(q, hctx, i) {
649 mutex_lock(&q->debugfs_mutex);
650 blk_mq_debugfs_unregister_sched_hctx(hctx);
651 mutex_unlock(&q->debugfs_mutex);
652
653 if (e->type->ops.exit_hctx && hctx->sched_data) {
654 e->type->ops.exit_hctx(hctx, i);
655 hctx->sched_data = NULL;
656 }
657 flags = hctx->flags;
658 }
659
660 mutex_lock(&q->debugfs_mutex);
661 blk_mq_debugfs_unregister_sched(q);
662 mutex_unlock(&q->debugfs_mutex);
663
664 if (e->type->ops.exit_sched)
665 e->type->ops.exit_sched(e);
666 blk_mq_sched_tags_teardown(q, flags);
667 q->elevator = NULL;
668}
1/*
2 * blk-mq scheduling framework
3 *
4 * Copyright (C) 2016 Jens Axboe
5 */
6#include <linux/kernel.h>
7#include <linux/module.h>
8#include <linux/blk-mq.h>
9
10#include <trace/events/block.h>
11
12#include "blk.h"
13#include "blk-mq.h"
14#include "blk-mq-debugfs.h"
15#include "blk-mq-sched.h"
16#include "blk-mq-tag.h"
17#include "blk-wbt.h"
18
19void blk_mq_sched_free_hctx_data(struct request_queue *q,
20 void (*exit)(struct blk_mq_hw_ctx *))
21{
22 struct blk_mq_hw_ctx *hctx;
23 int i;
24
25 queue_for_each_hw_ctx(q, hctx, i) {
26 if (exit && hctx->sched_data)
27 exit(hctx);
28 kfree(hctx->sched_data);
29 hctx->sched_data = NULL;
30 }
31}
32EXPORT_SYMBOL_GPL(blk_mq_sched_free_hctx_data);
33
34void blk_mq_sched_assign_ioc(struct request *rq, struct bio *bio)
35{
36 struct request_queue *q = rq->q;
37 struct io_context *ioc = rq_ioc(bio);
38 struct io_cq *icq;
39
40 spin_lock_irq(q->queue_lock);
41 icq = ioc_lookup_icq(ioc, q);
42 spin_unlock_irq(q->queue_lock);
43
44 if (!icq) {
45 icq = ioc_create_icq(ioc, q, GFP_ATOMIC);
46 if (!icq)
47 return;
48 }
49 get_io_context(icq->ioc);
50 rq->elv.icq = icq;
51}
52
53/*
54 * Mark a hardware queue as needing a restart. For shared queues, maintain
55 * a count of how many hardware queues are marked for restart.
56 */
57static void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx)
58{
59 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
60 return;
61
62 if (hctx->flags & BLK_MQ_F_TAG_SHARED) {
63 struct request_queue *q = hctx->queue;
64
65 if (!test_and_set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
66 atomic_inc(&q->shared_hctx_restart);
67 } else
68 set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
69}
70
71static bool blk_mq_sched_restart_hctx(struct blk_mq_hw_ctx *hctx)
72{
73 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
74 return false;
75
76 if (hctx->flags & BLK_MQ_F_TAG_SHARED) {
77 struct request_queue *q = hctx->queue;
78
79 if (test_and_clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
80 atomic_dec(&q->shared_hctx_restart);
81 } else
82 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
83
84 return blk_mq_run_hw_queue(hctx, true);
85}
86
87/*
88 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
89 * its queue by itself in its completion handler, so we don't need to
90 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
91 */
92static void blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
93{
94 struct request_queue *q = hctx->queue;
95 struct elevator_queue *e = q->elevator;
96 LIST_HEAD(rq_list);
97
98 do {
99 struct request *rq;
100
101 if (e->type->ops.mq.has_work &&
102 !e->type->ops.mq.has_work(hctx))
103 break;
104
105 if (!blk_mq_get_dispatch_budget(hctx))
106 break;
107
108 rq = e->type->ops.mq.dispatch_request(hctx);
109 if (!rq) {
110 blk_mq_put_dispatch_budget(hctx);
111 break;
112 }
113
114 /*
115 * Now this rq owns the budget which has to be released
116 * if this rq won't be queued to driver via .queue_rq()
117 * in blk_mq_dispatch_rq_list().
118 */
119 list_add(&rq->queuelist, &rq_list);
120 } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
121}
122
123static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
124 struct blk_mq_ctx *ctx)
125{
126 unsigned idx = ctx->index_hw;
127
128 if (++idx == hctx->nr_ctx)
129 idx = 0;
130
131 return hctx->ctxs[idx];
132}
133
134/*
135 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
136 * its queue by itself in its completion handler, so we don't need to
137 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
138 */
139static void blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
140{
141 struct request_queue *q = hctx->queue;
142 LIST_HEAD(rq_list);
143 struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from);
144
145 do {
146 struct request *rq;
147
148 if (!sbitmap_any_bit_set(&hctx->ctx_map))
149 break;
150
151 if (!blk_mq_get_dispatch_budget(hctx))
152 break;
153
154 rq = blk_mq_dequeue_from_ctx(hctx, ctx);
155 if (!rq) {
156 blk_mq_put_dispatch_budget(hctx);
157 break;
158 }
159
160 /*
161 * Now this rq owns the budget which has to be released
162 * if this rq won't be queued to driver via .queue_rq()
163 * in blk_mq_dispatch_rq_list().
164 */
165 list_add(&rq->queuelist, &rq_list);
166
167 /* round robin for fair dispatch */
168 ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
169
170 } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
171
172 WRITE_ONCE(hctx->dispatch_from, ctx);
173}
174
175void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
176{
177 struct request_queue *q = hctx->queue;
178 struct elevator_queue *e = q->elevator;
179 const bool has_sched_dispatch = e && e->type->ops.mq.dispatch_request;
180 LIST_HEAD(rq_list);
181
182 /* RCU or SRCU read lock is needed before checking quiesced flag */
183 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
184 return;
185
186 hctx->run++;
187
188 /*
189 * If we have previous entries on our dispatch list, grab them first for
190 * more fair dispatch.
191 */
192 if (!list_empty_careful(&hctx->dispatch)) {
193 spin_lock(&hctx->lock);
194 if (!list_empty(&hctx->dispatch))
195 list_splice_init(&hctx->dispatch, &rq_list);
196 spin_unlock(&hctx->lock);
197 }
198
199 /*
200 * Only ask the scheduler for requests, if we didn't have residual
201 * requests from the dispatch list. This is to avoid the case where
202 * we only ever dispatch a fraction of the requests available because
203 * of low device queue depth. Once we pull requests out of the IO
204 * scheduler, we can no longer merge or sort them. So it's best to
205 * leave them there for as long as we can. Mark the hw queue as
206 * needing a restart in that case.
207 *
208 * We want to dispatch from the scheduler if there was nothing
209 * on the dispatch list or we were able to dispatch from the
210 * dispatch list.
211 */
212 if (!list_empty(&rq_list)) {
213 blk_mq_sched_mark_restart_hctx(hctx);
214 if (blk_mq_dispatch_rq_list(q, &rq_list, false)) {
215 if (has_sched_dispatch)
216 blk_mq_do_dispatch_sched(hctx);
217 else
218 blk_mq_do_dispatch_ctx(hctx);
219 }
220 } else if (has_sched_dispatch) {
221 blk_mq_do_dispatch_sched(hctx);
222 } else if (q->mq_ops->get_budget) {
223 /*
224 * If we need to get budget before queuing request, we
225 * dequeue request one by one from sw queue for avoiding
226 * to mess up I/O merge when dispatch runs out of resource.
227 *
228 * TODO: get more budgets, and dequeue more requests in
229 * one time.
230 */
231 blk_mq_do_dispatch_ctx(hctx);
232 } else {
233 blk_mq_flush_busy_ctxs(hctx, &rq_list);
234 blk_mq_dispatch_rq_list(q, &rq_list, false);
235 }
236}
237
238bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
239 struct request **merged_request)
240{
241 struct request *rq;
242
243 switch (elv_merge(q, &rq, bio)) {
244 case ELEVATOR_BACK_MERGE:
245 if (!blk_mq_sched_allow_merge(q, rq, bio))
246 return false;
247 if (!bio_attempt_back_merge(q, rq, bio))
248 return false;
249 *merged_request = attempt_back_merge(q, rq);
250 if (!*merged_request)
251 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
252 return true;
253 case ELEVATOR_FRONT_MERGE:
254 if (!blk_mq_sched_allow_merge(q, rq, bio))
255 return false;
256 if (!bio_attempt_front_merge(q, rq, bio))
257 return false;
258 *merged_request = attempt_front_merge(q, rq);
259 if (!*merged_request)
260 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
261 return true;
262 case ELEVATOR_DISCARD_MERGE:
263 return bio_attempt_discard_merge(q, rq, bio);
264 default:
265 return false;
266 }
267}
268EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
269
270/*
271 * Reverse check our software queue for entries that we could potentially
272 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
273 * too much time checking for merges.
274 */
275static bool blk_mq_attempt_merge(struct request_queue *q,
276 struct blk_mq_ctx *ctx, struct bio *bio)
277{
278 struct request *rq;
279 int checked = 8;
280
281 lockdep_assert_held(&ctx->lock);
282
283 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
284 bool merged = false;
285
286 if (!checked--)
287 break;
288
289 if (!blk_rq_merge_ok(rq, bio))
290 continue;
291
292 switch (blk_try_merge(rq, bio)) {
293 case ELEVATOR_BACK_MERGE:
294 if (blk_mq_sched_allow_merge(q, rq, bio))
295 merged = bio_attempt_back_merge(q, rq, bio);
296 break;
297 case ELEVATOR_FRONT_MERGE:
298 if (blk_mq_sched_allow_merge(q, rq, bio))
299 merged = bio_attempt_front_merge(q, rq, bio);
300 break;
301 case ELEVATOR_DISCARD_MERGE:
302 merged = bio_attempt_discard_merge(q, rq, bio);
303 break;
304 default:
305 continue;
306 }
307
308 if (merged)
309 ctx->rq_merged++;
310 return merged;
311 }
312
313 return false;
314}
315
316bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio)
317{
318 struct elevator_queue *e = q->elevator;
319 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
320 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
321 bool ret = false;
322
323 if (e && e->type->ops.mq.bio_merge) {
324 blk_mq_put_ctx(ctx);
325 return e->type->ops.mq.bio_merge(hctx, bio);
326 }
327
328 if (hctx->flags & BLK_MQ_F_SHOULD_MERGE) {
329 /* default per sw-queue merge */
330 spin_lock(&ctx->lock);
331 ret = blk_mq_attempt_merge(q, ctx, bio);
332 spin_unlock(&ctx->lock);
333 }
334
335 blk_mq_put_ctx(ctx);
336 return ret;
337}
338
339bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
340{
341 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
342}
343EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
344
345void blk_mq_sched_request_inserted(struct request *rq)
346{
347 trace_block_rq_insert(rq->q, rq);
348}
349EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
350
351static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
352 bool has_sched,
353 struct request *rq)
354{
355 /* dispatch flush rq directly */
356 if (rq->rq_flags & RQF_FLUSH_SEQ) {
357 spin_lock(&hctx->lock);
358 list_add(&rq->queuelist, &hctx->dispatch);
359 spin_unlock(&hctx->lock);
360 return true;
361 }
362
363 if (has_sched)
364 rq->rq_flags |= RQF_SORTED;
365
366 return false;
367}
368
369/**
370 * list_for_each_entry_rcu_rr - iterate in a round-robin fashion over rcu list
371 * @pos: loop cursor.
372 * @skip: the list element that will not be examined. Iteration starts at
373 * @skip->next.
374 * @head: head of the list to examine. This list must have at least one
375 * element, namely @skip.
376 * @member: name of the list_head structure within typeof(*pos).
377 */
378#define list_for_each_entry_rcu_rr(pos, skip, head, member) \
379 for ((pos) = (skip); \
380 (pos = (pos)->member.next != (head) ? list_entry_rcu( \
381 (pos)->member.next, typeof(*pos), member) : \
382 list_entry_rcu((pos)->member.next->next, typeof(*pos), member)), \
383 (pos) != (skip); )
384
385/*
386 * Called after a driver tag has been freed to check whether a hctx needs to
387 * be restarted. Restarts @hctx if its tag set is not shared. Restarts hardware
388 * queues in a round-robin fashion if the tag set of @hctx is shared with other
389 * hardware queues.
390 */
391void blk_mq_sched_restart(struct blk_mq_hw_ctx *const hctx)
392{
393 struct blk_mq_tags *const tags = hctx->tags;
394 struct blk_mq_tag_set *const set = hctx->queue->tag_set;
395 struct request_queue *const queue = hctx->queue, *q;
396 struct blk_mq_hw_ctx *hctx2;
397 unsigned int i, j;
398
399 if (set->flags & BLK_MQ_F_TAG_SHARED) {
400 /*
401 * If this is 0, then we know that no hardware queues
402 * have RESTART marked. We're done.
403 */
404 if (!atomic_read(&queue->shared_hctx_restart))
405 return;
406
407 rcu_read_lock();
408 list_for_each_entry_rcu_rr(q, queue, &set->tag_list,
409 tag_set_list) {
410 queue_for_each_hw_ctx(q, hctx2, i)
411 if (hctx2->tags == tags &&
412 blk_mq_sched_restart_hctx(hctx2))
413 goto done;
414 }
415 j = hctx->queue_num + 1;
416 for (i = 0; i < queue->nr_hw_queues; i++, j++) {
417 if (j == queue->nr_hw_queues)
418 j = 0;
419 hctx2 = queue->queue_hw_ctx[j];
420 if (hctx2->tags == tags &&
421 blk_mq_sched_restart_hctx(hctx2))
422 break;
423 }
424done:
425 rcu_read_unlock();
426 } else {
427 blk_mq_sched_restart_hctx(hctx);
428 }
429}
430
431void blk_mq_sched_insert_request(struct request *rq, bool at_head,
432 bool run_queue, bool async)
433{
434 struct request_queue *q = rq->q;
435 struct elevator_queue *e = q->elevator;
436 struct blk_mq_ctx *ctx = rq->mq_ctx;
437 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
438
439 /* flush rq in flush machinery need to be dispatched directly */
440 if (!(rq->rq_flags & RQF_FLUSH_SEQ) && op_is_flush(rq->cmd_flags)) {
441 blk_insert_flush(rq);
442 goto run;
443 }
444
445 WARN_ON(e && (rq->tag != -1));
446
447 if (blk_mq_sched_bypass_insert(hctx, !!e, rq))
448 goto run;
449
450 if (e && e->type->ops.mq.insert_requests) {
451 LIST_HEAD(list);
452
453 list_add(&rq->queuelist, &list);
454 e->type->ops.mq.insert_requests(hctx, &list, at_head);
455 } else {
456 spin_lock(&ctx->lock);
457 __blk_mq_insert_request(hctx, rq, at_head);
458 spin_unlock(&ctx->lock);
459 }
460
461run:
462 if (run_queue)
463 blk_mq_run_hw_queue(hctx, async);
464}
465
466void blk_mq_sched_insert_requests(struct request_queue *q,
467 struct blk_mq_ctx *ctx,
468 struct list_head *list, bool run_queue_async)
469{
470 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
471 struct elevator_queue *e = hctx->queue->elevator;
472
473 if (e && e->type->ops.mq.insert_requests)
474 e->type->ops.mq.insert_requests(hctx, list, false);
475 else
476 blk_mq_insert_requests(hctx, ctx, list);
477
478 blk_mq_run_hw_queue(hctx, run_queue_async);
479}
480
481static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
482 struct blk_mq_hw_ctx *hctx,
483 unsigned int hctx_idx)
484{
485 if (hctx->sched_tags) {
486 blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx);
487 blk_mq_free_rq_map(hctx->sched_tags);
488 hctx->sched_tags = NULL;
489 }
490}
491
492static int blk_mq_sched_alloc_tags(struct request_queue *q,
493 struct blk_mq_hw_ctx *hctx,
494 unsigned int hctx_idx)
495{
496 struct blk_mq_tag_set *set = q->tag_set;
497 int ret;
498
499 hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
500 set->reserved_tags);
501 if (!hctx->sched_tags)
502 return -ENOMEM;
503
504 ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests);
505 if (ret)
506 blk_mq_sched_free_tags(set, hctx, hctx_idx);
507
508 return ret;
509}
510
511static void blk_mq_sched_tags_teardown(struct request_queue *q)
512{
513 struct blk_mq_tag_set *set = q->tag_set;
514 struct blk_mq_hw_ctx *hctx;
515 int i;
516
517 queue_for_each_hw_ctx(q, hctx, i)
518 blk_mq_sched_free_tags(set, hctx, i);
519}
520
521int blk_mq_sched_init_hctx(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
522 unsigned int hctx_idx)
523{
524 struct elevator_queue *e = q->elevator;
525 int ret;
526
527 if (!e)
528 return 0;
529
530 ret = blk_mq_sched_alloc_tags(q, hctx, hctx_idx);
531 if (ret)
532 return ret;
533
534 if (e->type->ops.mq.init_hctx) {
535 ret = e->type->ops.mq.init_hctx(hctx, hctx_idx);
536 if (ret) {
537 blk_mq_sched_free_tags(q->tag_set, hctx, hctx_idx);
538 return ret;
539 }
540 }
541
542 blk_mq_debugfs_register_sched_hctx(q, hctx);
543
544 return 0;
545}
546
547void blk_mq_sched_exit_hctx(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
548 unsigned int hctx_idx)
549{
550 struct elevator_queue *e = q->elevator;
551
552 if (!e)
553 return;
554
555 blk_mq_debugfs_unregister_sched_hctx(hctx);
556
557 if (e->type->ops.mq.exit_hctx && hctx->sched_data) {
558 e->type->ops.mq.exit_hctx(hctx, hctx_idx);
559 hctx->sched_data = NULL;
560 }
561
562 blk_mq_sched_free_tags(q->tag_set, hctx, hctx_idx);
563}
564
565int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
566{
567 struct blk_mq_hw_ctx *hctx;
568 struct elevator_queue *eq;
569 unsigned int i;
570 int ret;
571
572 if (!e) {
573 q->elevator = NULL;
574 return 0;
575 }
576
577 /*
578 * Default to double of smaller one between hw queue_depth and 128,
579 * since we don't split into sync/async like the old code did.
580 * Additionally, this is a per-hw queue depth.
581 */
582 q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
583 BLKDEV_MAX_RQ);
584
585 queue_for_each_hw_ctx(q, hctx, i) {
586 ret = blk_mq_sched_alloc_tags(q, hctx, i);
587 if (ret)
588 goto err;
589 }
590
591 ret = e->ops.mq.init_sched(q, e);
592 if (ret)
593 goto err;
594
595 blk_mq_debugfs_register_sched(q);
596
597 queue_for_each_hw_ctx(q, hctx, i) {
598 if (e->ops.mq.init_hctx) {
599 ret = e->ops.mq.init_hctx(hctx, i);
600 if (ret) {
601 eq = q->elevator;
602 blk_mq_exit_sched(q, eq);
603 kobject_put(&eq->kobj);
604 return ret;
605 }
606 }
607 blk_mq_debugfs_register_sched_hctx(q, hctx);
608 }
609
610 return 0;
611
612err:
613 blk_mq_sched_tags_teardown(q);
614 q->elevator = NULL;
615 return ret;
616}
617
618void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
619{
620 struct blk_mq_hw_ctx *hctx;
621 unsigned int i;
622
623 queue_for_each_hw_ctx(q, hctx, i) {
624 blk_mq_debugfs_unregister_sched_hctx(hctx);
625 if (e->type->ops.mq.exit_hctx && hctx->sched_data) {
626 e->type->ops.mq.exit_hctx(hctx, i);
627 hctx->sched_data = NULL;
628 }
629 }
630 blk_mq_debugfs_unregister_sched(q);
631 if (e->type->ops.mq.exit_sched)
632 e->type->ops.mq.exit_sched(e);
633 blk_mq_sched_tags_teardown(q);
634 q->elevator = NULL;
635}
636
637int blk_mq_sched_init(struct request_queue *q)
638{
639 int ret;
640
641 mutex_lock(&q->sysfs_lock);
642 ret = elevator_init(q, NULL);
643 mutex_unlock(&q->sysfs_lock);
644
645 return ret;
646}