block: remove blk_init_request_from_bio
[linux-block.git] / block / blk-mq-sched.c
CommitLineData
3dcf60bc 1// SPDX-License-Identifier: GPL-2.0
bd166ef1
JA
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
11#include <trace/events/block.h>
12
13#include "blk.h"
14#include "blk-mq.h"
d332ce09 15#include "blk-mq-debugfs.h"
bd166ef1
JA
16#include "blk-mq-sched.h"
17#include "blk-mq-tag.h"
18#include "blk-wbt.h"
19
20void blk_mq_sched_free_hctx_data(struct request_queue *q,
21 void (*exit)(struct blk_mq_hw_ctx *))
22{
23 struct blk_mq_hw_ctx *hctx;
24 int i;
25
26 queue_for_each_hw_ctx(q, hctx, i) {
27 if (exit && hctx->sched_data)
28 exit(hctx);
29 kfree(hctx->sched_data);
30 hctx->sched_data = NULL;
31 }
32}
33EXPORT_SYMBOL_GPL(blk_mq_sched_free_hctx_data);
34
e2b3fa5a 35void blk_mq_sched_assign_ioc(struct request *rq)
bd166ef1 36{
44e8c2bf 37 struct request_queue *q = rq->q;
0c62bff1 38 struct io_context *ioc;
bd166ef1
JA
39 struct io_cq *icq;
40
0c62bff1
JA
41 /*
42 * May not have an IO context if it's a passthrough request
43 */
44 ioc = current->io_context;
45 if (!ioc)
46 return;
47
0d945c1f 48 spin_lock_irq(&q->queue_lock);
bd166ef1 49 icq = ioc_lookup_icq(ioc, q);
0d945c1f 50 spin_unlock_irq(&q->queue_lock);
bd166ef1
JA
51
52 if (!icq) {
53 icq = ioc_create_icq(ioc, q, GFP_ATOMIC);
54 if (!icq)
55 return;
56 }
ea511e3c 57 get_io_context(icq->ioc);
44e8c2bf 58 rq->elv.icq = icq;
bd166ef1
JA
59}
60
8e8320c9
JA
61/*
62 * Mark a hardware queue as needing a restart. For shared queues, maintain
63 * a count of how many hardware queues are marked for restart.
64 */
7211aef8 65void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx)
8e8320c9
JA
66{
67 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
68 return;
69
97889f9a 70 set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
8e8320c9 71}
7211aef8 72EXPORT_SYMBOL_GPL(blk_mq_sched_mark_restart_hctx);
8e8320c9 73
97889f9a 74void blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx)
8e8320c9
JA
75{
76 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
97889f9a
ML
77 return;
78 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
8e8320c9 79
97889f9a 80 blk_mq_run_hw_queue(hctx, true);
8e8320c9
JA
81}
82
1f460b63
ML
83/*
84 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
85 * its queue by itself in its completion handler, so we don't need to
86 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
87 */
88static void blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
caf8eb0d
ML
89{
90 struct request_queue *q = hctx->queue;
91 struct elevator_queue *e = q->elevator;
92 LIST_HEAD(rq_list);
93
94 do {
de148297 95 struct request *rq;
caf8eb0d 96
f9cd4bfe 97 if (e->type->ops.has_work && !e->type->ops.has_work(hctx))
caf8eb0d 98 break;
de148297 99
88022d72 100 if (!blk_mq_get_dispatch_budget(hctx))
1f460b63 101 break;
de148297 102
f9cd4bfe 103 rq = e->type->ops.dispatch_request(hctx);
de148297
ML
104 if (!rq) {
105 blk_mq_put_dispatch_budget(hctx);
106 break;
de148297
ML
107 }
108
109 /*
110 * Now this rq owns the budget which has to be released
111 * if this rq won't be queued to driver via .queue_rq()
112 * in blk_mq_dispatch_rq_list().
113 */
caf8eb0d 114 list_add(&rq->queuelist, &rq_list);
de148297 115 } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
caf8eb0d
ML
116}
117
b347689f
ML
118static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
119 struct blk_mq_ctx *ctx)
120{
f31967f0 121 unsigned short idx = ctx->index_hw[hctx->type];
b347689f
ML
122
123 if (++idx == hctx->nr_ctx)
124 idx = 0;
125
126 return hctx->ctxs[idx];
127}
128
1f460b63
ML
129/*
130 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
131 * its queue by itself in its completion handler, so we don't need to
132 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
133 */
134static void blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
b347689f
ML
135{
136 struct request_queue *q = hctx->queue;
137 LIST_HEAD(rq_list);
138 struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from);
139
140 do {
141 struct request *rq;
b347689f
ML
142
143 if (!sbitmap_any_bit_set(&hctx->ctx_map))
144 break;
145
88022d72 146 if (!blk_mq_get_dispatch_budget(hctx))
1f460b63 147 break;
b347689f
ML
148
149 rq = blk_mq_dequeue_from_ctx(hctx, ctx);
150 if (!rq) {
151 blk_mq_put_dispatch_budget(hctx);
152 break;
b347689f
ML
153 }
154
155 /*
156 * Now this rq owns the budget which has to be released
157 * if this rq won't be queued to driver via .queue_rq()
158 * in blk_mq_dispatch_rq_list().
159 */
160 list_add(&rq->queuelist, &rq_list);
161
162 /* round robin for fair dispatch */
163 ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
164
165 } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
166
167 WRITE_ONCE(hctx->dispatch_from, ctx);
b347689f
ML
168}
169
1f460b63 170void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
bd166ef1 171{
81380ca1
OS
172 struct request_queue *q = hctx->queue;
173 struct elevator_queue *e = q->elevator;
f9cd4bfe 174 const bool has_sched_dispatch = e && e->type->ops.dispatch_request;
bd166ef1
JA
175 LIST_HEAD(rq_list);
176
f4560ffe
ML
177 /* RCU or SRCU read lock is needed before checking quiesced flag */
178 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
1f460b63 179 return;
bd166ef1
JA
180
181 hctx->run++;
182
183 /*
184 * If we have previous entries on our dispatch list, grab them first for
185 * more fair dispatch.
186 */
187 if (!list_empty_careful(&hctx->dispatch)) {
188 spin_lock(&hctx->lock);
189 if (!list_empty(&hctx->dispatch))
190 list_splice_init(&hctx->dispatch, &rq_list);
191 spin_unlock(&hctx->lock);
192 }
193
194 /*
195 * Only ask the scheduler for requests, if we didn't have residual
196 * requests from the dispatch list. This is to avoid the case where
197 * we only ever dispatch a fraction of the requests available because
198 * of low device queue depth. Once we pull requests out of the IO
199 * scheduler, we can no longer merge or sort them. So it's best to
200 * leave them there for as long as we can. Mark the hw queue as
201 * needing a restart in that case.
caf8eb0d
ML
202 *
203 * We want to dispatch from the scheduler if there was nothing
204 * on the dispatch list or we were able to dispatch from the
205 * dispatch list.
bd166ef1 206 */
c13660a0 207 if (!list_empty(&rq_list)) {
d38d3515 208 blk_mq_sched_mark_restart_hctx(hctx);
b347689f
ML
209 if (blk_mq_dispatch_rq_list(q, &rq_list, false)) {
210 if (has_sched_dispatch)
1f460b63 211 blk_mq_do_dispatch_sched(hctx);
b347689f 212 else
1f460b63 213 blk_mq_do_dispatch_ctx(hctx);
b347689f 214 }
caf8eb0d 215 } else if (has_sched_dispatch) {
1f460b63 216 blk_mq_do_dispatch_sched(hctx);
6e768717
ML
217 } else if (hctx->dispatch_busy) {
218 /* dequeue request one by one from sw queue if queue is busy */
1f460b63 219 blk_mq_do_dispatch_ctx(hctx);
caf8eb0d 220 } else {
c13660a0 221 blk_mq_flush_busy_ctxs(hctx, &rq_list);
de148297 222 blk_mq_dispatch_rq_list(q, &rq_list, false);
64765a75 223 }
bd166ef1
JA
224}
225
e4d750c9
JA
226bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
227 struct request **merged_request)
bd166ef1
JA
228{
229 struct request *rq;
bd166ef1 230
34fe7c05
CH
231 switch (elv_merge(q, &rq, bio)) {
232 case ELEVATOR_BACK_MERGE:
bd166ef1
JA
233 if (!blk_mq_sched_allow_merge(q, rq, bio))
234 return false;
34fe7c05
CH
235 if (!bio_attempt_back_merge(q, rq, bio))
236 return false;
237 *merged_request = attempt_back_merge(q, rq);
238 if (!*merged_request)
239 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
240 return true;
241 case ELEVATOR_FRONT_MERGE:
bd166ef1
JA
242 if (!blk_mq_sched_allow_merge(q, rq, bio))
243 return false;
34fe7c05
CH
244 if (!bio_attempt_front_merge(q, rq, bio))
245 return false;
246 *merged_request = attempt_front_merge(q, rq);
247 if (!*merged_request)
248 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
249 return true;
bea99a50
KB
250 case ELEVATOR_DISCARD_MERGE:
251 return bio_attempt_discard_merge(q, rq, bio);
34fe7c05
CH
252 default:
253 return false;
bd166ef1 254 }
bd166ef1
JA
255}
256EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
257
9bddeb2a 258/*
9c558734
JA
259 * Iterate list of requests and see if we can merge this bio with any
260 * of them.
9bddeb2a 261 */
9c558734
JA
262bool blk_mq_bio_list_merge(struct request_queue *q, struct list_head *list,
263 struct bio *bio)
9bddeb2a
ML
264{
265 struct request *rq;
266 int checked = 8;
267
9c558734 268 list_for_each_entry_reverse(rq, list, queuelist) {
9bddeb2a
ML
269 bool merged = false;
270
271 if (!checked--)
272 break;
273
274 if (!blk_rq_merge_ok(rq, bio))
275 continue;
276
277 switch (blk_try_merge(rq, bio)) {
278 case ELEVATOR_BACK_MERGE:
279 if (blk_mq_sched_allow_merge(q, rq, bio))
280 merged = bio_attempt_back_merge(q, rq, bio);
281 break;
282 case ELEVATOR_FRONT_MERGE:
283 if (blk_mq_sched_allow_merge(q, rq, bio))
284 merged = bio_attempt_front_merge(q, rq, bio);
285 break;
286 case ELEVATOR_DISCARD_MERGE:
287 merged = bio_attempt_discard_merge(q, rq, bio);
288 break;
289 default:
290 continue;
291 }
292
9bddeb2a
ML
293 return merged;
294 }
295
296 return false;
297}
9c558734
JA
298EXPORT_SYMBOL_GPL(blk_mq_bio_list_merge);
299
300/*
301 * Reverse check our software queue for entries that we could potentially
302 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
303 * too much time checking for merges.
304 */
305static bool blk_mq_attempt_merge(struct request_queue *q,
c16d6b5a 306 struct blk_mq_hw_ctx *hctx,
9c558734
JA
307 struct blk_mq_ctx *ctx, struct bio *bio)
308{
c16d6b5a
ML
309 enum hctx_type type = hctx->type;
310
9c558734
JA
311 lockdep_assert_held(&ctx->lock);
312
c16d6b5a 313 if (blk_mq_bio_list_merge(q, &ctx->rq_lists[type], bio)) {
9c558734
JA
314 ctx->rq_merged++;
315 return true;
316 }
317
318 return false;
319}
9bddeb2a 320
bd166ef1
JA
321bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio)
322{
323 struct elevator_queue *e = q->elevator;
9bddeb2a 324 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
8ccdf4a3 325 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, bio->bi_opf, ctx);
9bddeb2a 326 bool ret = false;
c16d6b5a 327 enum hctx_type type;
bd166ef1 328
f9cd4bfe 329 if (e && e->type->ops.bio_merge) {
bd166ef1 330 blk_mq_put_ctx(ctx);
f9cd4bfe 331 return e->type->ops.bio_merge(hctx, bio);
bd166ef1
JA
332 }
333
c16d6b5a 334 type = hctx->type;
b04f50ab 335 if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
c16d6b5a 336 !list_empty_careful(&ctx->rq_lists[type])) {
9bddeb2a
ML
337 /* default per sw-queue merge */
338 spin_lock(&ctx->lock);
c16d6b5a 339 ret = blk_mq_attempt_merge(q, hctx, ctx, bio);
9bddeb2a
ML
340 spin_unlock(&ctx->lock);
341 }
342
343 blk_mq_put_ctx(ctx);
344 return ret;
bd166ef1
JA
345}
346
347bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
348{
349 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
350}
351EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
352
353void blk_mq_sched_request_inserted(struct request *rq)
354{
355 trace_block_rq_insert(rq->q, rq);
356}
357EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
358
0cacba6c 359static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
a6a252e6 360 bool has_sched,
0cacba6c 361 struct request *rq)
bd166ef1 362{
a6a252e6
ML
363 /* dispatch flush rq directly */
364 if (rq->rq_flags & RQF_FLUSH_SEQ) {
365 spin_lock(&hctx->lock);
366 list_add(&rq->queuelist, &hctx->dispatch);
367 spin_unlock(&hctx->lock);
368 return true;
369 }
370
923218f6 371 if (has_sched)
bd166ef1 372 rq->rq_flags |= RQF_SORTED;
bd166ef1 373
a6a252e6 374 return false;
bd166ef1 375}
bd166ef1 376
bd6737f1 377void blk_mq_sched_insert_request(struct request *rq, bool at_head,
9e97d295 378 bool run_queue, bool async)
bd6737f1
JA
379{
380 struct request_queue *q = rq->q;
381 struct elevator_queue *e = q->elevator;
382 struct blk_mq_ctx *ctx = rq->mq_ctx;
ea4f995e 383 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
bd6737f1 384
a6a252e6
ML
385 /* flush rq in flush machinery need to be dispatched directly */
386 if (!(rq->rq_flags & RQF_FLUSH_SEQ) && op_is_flush(rq->cmd_flags)) {
923218f6
ML
387 blk_insert_flush(rq);
388 goto run;
bd6737f1
JA
389 }
390
923218f6
ML
391 WARN_ON(e && (rq->tag != -1));
392
a6a252e6 393 if (blk_mq_sched_bypass_insert(hctx, !!e, rq))
0cacba6c
OS
394 goto run;
395
f9cd4bfe 396 if (e && e->type->ops.insert_requests) {
bd6737f1
JA
397 LIST_HEAD(list);
398
399 list_add(&rq->queuelist, &list);
f9cd4bfe 400 e->type->ops.insert_requests(hctx, &list, at_head);
bd6737f1
JA
401 } else {
402 spin_lock(&ctx->lock);
403 __blk_mq_insert_request(hctx, rq, at_head);
404 spin_unlock(&ctx->lock);
405 }
406
0cacba6c 407run:
bd6737f1
JA
408 if (run_queue)
409 blk_mq_run_hw_queue(hctx, async);
410}
411
67cae4c9 412void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx *hctx,
bd6737f1
JA
413 struct blk_mq_ctx *ctx,
414 struct list_head *list, bool run_queue_async)
415{
f9afca4d 416 struct elevator_queue *e;
e87eb301
ML
417 struct request_queue *q = hctx->queue;
418
419 /*
420 * blk_mq_sched_insert_requests() is called from flush plug
421 * context only, and hold one usage counter to prevent queue
422 * from being released.
423 */
424 percpu_ref_get(&q->q_usage_counter);
bd6737f1 425
f9afca4d 426 e = hctx->queue->elevator;
f9cd4bfe
JA
427 if (e && e->type->ops.insert_requests)
428 e->type->ops.insert_requests(hctx, list, false);
6ce3dd6e
ML
429 else {
430 /*
431 * try to issue requests directly if the hw queue isn't
432 * busy in case of 'none' scheduler, and this way may save
433 * us one extra enqueue & dequeue to sw queue.
434 */
fd9c40f6 435 if (!hctx->dispatch_busy && !e && !run_queue_async) {
6ce3dd6e 436 blk_mq_try_issue_list_directly(hctx, list);
fd9c40f6 437 if (list_empty(list))
e87eb301 438 goto out;
fd9c40f6
BVA
439 }
440 blk_mq_insert_requests(hctx, ctx, list);
6ce3dd6e 441 }
bd6737f1
JA
442
443 blk_mq_run_hw_queue(hctx, run_queue_async);
e87eb301
ML
444 out:
445 percpu_ref_put(&q->q_usage_counter);
bd6737f1
JA
446}
447
bd166ef1
JA
448static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
449 struct blk_mq_hw_ctx *hctx,
450 unsigned int hctx_idx)
451{
452 if (hctx->sched_tags) {
453 blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx);
454 blk_mq_free_rq_map(hctx->sched_tags);
455 hctx->sched_tags = NULL;
456 }
457}
458
6917ff0b
OS
459static int blk_mq_sched_alloc_tags(struct request_queue *q,
460 struct blk_mq_hw_ctx *hctx,
461 unsigned int hctx_idx)
462{
463 struct blk_mq_tag_set *set = q->tag_set;
464 int ret;
465
466 hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
467 set->reserved_tags);
468 if (!hctx->sched_tags)
469 return -ENOMEM;
470
471 ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests);
472 if (ret)
473 blk_mq_sched_free_tags(set, hctx, hctx_idx);
474
475 return ret;
476}
477
c3e22192 478/* called in queue's release handler, tagset has gone away */
54d5329d 479static void blk_mq_sched_tags_teardown(struct request_queue *q)
bd166ef1 480{
bd166ef1 481 struct blk_mq_hw_ctx *hctx;
6917ff0b
OS
482 int i;
483
c3e22192
ML
484 queue_for_each_hw_ctx(q, hctx, i) {
485 if (hctx->sched_tags) {
486 blk_mq_free_rq_map(hctx->sched_tags);
487 hctx->sched_tags = NULL;
488 }
489 }
6917ff0b
OS
490}
491
492int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
493{
494 struct blk_mq_hw_ctx *hctx;
ee056f98 495 struct elevator_queue *eq;
6917ff0b
OS
496 unsigned int i;
497 int ret;
498
499 if (!e) {
500 q->elevator = NULL;
32a50fab 501 q->nr_requests = q->tag_set->queue_depth;
6917ff0b
OS
502 return 0;
503 }
bd166ef1
JA
504
505 /*
32825c45
ML
506 * Default to double of smaller one between hw queue_depth and 128,
507 * since we don't split into sync/async like the old code did.
508 * Additionally, this is a per-hw queue depth.
bd166ef1 509 */
32825c45
ML
510 q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
511 BLKDEV_MAX_RQ);
bd166ef1 512
bd166ef1 513 queue_for_each_hw_ctx(q, hctx, i) {
6917ff0b 514 ret = blk_mq_sched_alloc_tags(q, hctx, i);
bd166ef1 515 if (ret)
6917ff0b 516 goto err;
bd166ef1
JA
517 }
518
f9cd4bfe 519 ret = e->ops.init_sched(q, e);
6917ff0b
OS
520 if (ret)
521 goto err;
bd166ef1 522
d332ce09
OS
523 blk_mq_debugfs_register_sched(q);
524
525 queue_for_each_hw_ctx(q, hctx, i) {
f9cd4bfe
JA
526 if (e->ops.init_hctx) {
527 ret = e->ops.init_hctx(hctx, i);
ee056f98
OS
528 if (ret) {
529 eq = q->elevator;
c3e22192 530 blk_mq_sched_free_requests(q);
ee056f98
OS
531 blk_mq_exit_sched(q, eq);
532 kobject_put(&eq->kobj);
533 return ret;
534 }
535 }
d332ce09 536 blk_mq_debugfs_register_sched_hctx(q, hctx);
ee056f98
OS
537 }
538
bd166ef1 539 return 0;
bd166ef1 540
6917ff0b 541err:
c3e22192 542 blk_mq_sched_free_requests(q);
54d5329d
OS
543 blk_mq_sched_tags_teardown(q);
544 q->elevator = NULL;
6917ff0b 545 return ret;
bd166ef1 546}
d3484991 547
c3e22192
ML
548/*
549 * called in either blk_queue_cleanup or elevator_switch, tagset
550 * is required for freeing requests
551 */
552void blk_mq_sched_free_requests(struct request_queue *q)
553{
554 struct blk_mq_hw_ctx *hctx;
555 int i;
556
557 lockdep_assert_held(&q->sysfs_lock);
c3e22192
ML
558
559 queue_for_each_hw_ctx(q, hctx, i) {
560 if (hctx->sched_tags)
561 blk_mq_free_rqs(q->tag_set, hctx->sched_tags, i);
562 }
563}
564
54d5329d
OS
565void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
566{
ee056f98
OS
567 struct blk_mq_hw_ctx *hctx;
568 unsigned int i;
569
d332ce09
OS
570 queue_for_each_hw_ctx(q, hctx, i) {
571 blk_mq_debugfs_unregister_sched_hctx(hctx);
f9cd4bfe
JA
572 if (e->type->ops.exit_hctx && hctx->sched_data) {
573 e->type->ops.exit_hctx(hctx, i);
d332ce09 574 hctx->sched_data = NULL;
ee056f98
OS
575 }
576 }
d332ce09 577 blk_mq_debugfs_unregister_sched(q);
f9cd4bfe
JA
578 if (e->type->ops.exit_sched)
579 e->type->ops.exit_sched(e);
54d5329d
OS
580 blk_mq_sched_tags_teardown(q);
581 q->elevator = NULL;
582}