Loading...
Note: File does not exist in v3.5.6.
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
11#include <trace/events/block.h>
12
13#include "blk.h"
14#include "blk-mq.h"
15#include "blk-mq-debugfs.h"
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
35void blk_mq_sched_assign_ioc(struct request *rq)
36{
37 struct request_queue *q = rq->q;
38 struct io_context *ioc;
39 struct io_cq *icq;
40
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
48 spin_lock_irq(&q->queue_lock);
49 icq = ioc_lookup_icq(ioc, q);
50 spin_unlock_irq(&q->queue_lock);
51
52 if (!icq) {
53 icq = ioc_create_icq(ioc, q, GFP_ATOMIC);
54 if (!icq)
55 return;
56 }
57 get_io_context(icq->ioc);
58 rq->elv.icq = icq;
59}
60
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 */
65void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx)
66{
67 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
68 return;
69
70 set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
71}
72EXPORT_SYMBOL_GPL(blk_mq_sched_mark_restart_hctx);
73
74void blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx)
75{
76 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
77 return;
78 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
79
80 blk_mq_run_hw_queue(hctx, true);
81}
82
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)
89{
90 struct request_queue *q = hctx->queue;
91 struct elevator_queue *e = q->elevator;
92 LIST_HEAD(rq_list);
93
94 do {
95 struct request *rq;
96
97 if (e->type->ops.has_work && !e->type->ops.has_work(hctx))
98 break;
99
100 if (!blk_mq_get_dispatch_budget(hctx))
101 break;
102
103 rq = e->type->ops.dispatch_request(hctx);
104 if (!rq) {
105 blk_mq_put_dispatch_budget(hctx);
106 break;
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 */
114 list_add(&rq->queuelist, &rq_list);
115 } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
116}
117
118static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
119 struct blk_mq_ctx *ctx)
120{
121 unsigned short idx = ctx->index_hw[hctx->type];
122
123 if (++idx == hctx->nr_ctx)
124 idx = 0;
125
126 return hctx->ctxs[idx];
127}
128
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)
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;
142
143 if (!sbitmap_any_bit_set(&hctx->ctx_map))
144 break;
145
146 if (!blk_mq_get_dispatch_budget(hctx))
147 break;
148
149 rq = blk_mq_dequeue_from_ctx(hctx, ctx);
150 if (!rq) {
151 blk_mq_put_dispatch_budget(hctx);
152 break;
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);
168}
169
170void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
171{
172 struct request_queue *q = hctx->queue;
173 struct elevator_queue *e = q->elevator;
174 const bool has_sched_dispatch = e && e->type->ops.dispatch_request;
175 LIST_HEAD(rq_list);
176
177 /* RCU or SRCU read lock is needed before checking quiesced flag */
178 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
179 return;
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.
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.
206 */
207 if (!list_empty(&rq_list)) {
208 blk_mq_sched_mark_restart_hctx(hctx);
209 if (blk_mq_dispatch_rq_list(q, &rq_list, false)) {
210 if (has_sched_dispatch)
211 blk_mq_do_dispatch_sched(hctx);
212 else
213 blk_mq_do_dispatch_ctx(hctx);
214 }
215 } else if (has_sched_dispatch) {
216 blk_mq_do_dispatch_sched(hctx);
217 } else if (hctx->dispatch_busy) {
218 /* dequeue request one by one from sw queue if queue is busy */
219 blk_mq_do_dispatch_ctx(hctx);
220 } else {
221 blk_mq_flush_busy_ctxs(hctx, &rq_list);
222 blk_mq_dispatch_rq_list(q, &rq_list, false);
223 }
224}
225
226bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
227 unsigned int nr_segs, struct request **merged_request)
228{
229 struct request *rq;
230
231 switch (elv_merge(q, &rq, bio)) {
232 case ELEVATOR_BACK_MERGE:
233 if (!blk_mq_sched_allow_merge(q, rq, bio))
234 return false;
235 if (!bio_attempt_back_merge(rq, bio, nr_segs))
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:
242 if (!blk_mq_sched_allow_merge(q, rq, bio))
243 return false;
244 if (!bio_attempt_front_merge(rq, bio, nr_segs))
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;
250 case ELEVATOR_DISCARD_MERGE:
251 return bio_attempt_discard_merge(q, rq, bio);
252 default:
253 return false;
254 }
255}
256EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
257
258/*
259 * Iterate list of requests and see if we can merge this bio with any
260 * of them.
261 */
262bool blk_mq_bio_list_merge(struct request_queue *q, struct list_head *list,
263 struct bio *bio, unsigned int nr_segs)
264{
265 struct request *rq;
266 int checked = 8;
267
268 list_for_each_entry_reverse(rq, list, queuelist) {
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(rq, bio,
281 nr_segs);
282 break;
283 case ELEVATOR_FRONT_MERGE:
284 if (blk_mq_sched_allow_merge(q, rq, bio))
285 merged = bio_attempt_front_merge(rq, bio,
286 nr_segs);
287 break;
288 case ELEVATOR_DISCARD_MERGE:
289 merged = bio_attempt_discard_merge(q, rq, bio);
290 break;
291 default:
292 continue;
293 }
294
295 return merged;
296 }
297
298 return false;
299}
300EXPORT_SYMBOL_GPL(blk_mq_bio_list_merge);
301
302/*
303 * Reverse check our software queue for entries that we could potentially
304 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
305 * too much time checking for merges.
306 */
307static bool blk_mq_attempt_merge(struct request_queue *q,
308 struct blk_mq_hw_ctx *hctx,
309 struct blk_mq_ctx *ctx, struct bio *bio,
310 unsigned int nr_segs)
311{
312 enum hctx_type type = hctx->type;
313
314 lockdep_assert_held(&ctx->lock);
315
316 if (blk_mq_bio_list_merge(q, &ctx->rq_lists[type], bio, nr_segs)) {
317 ctx->rq_merged++;
318 return true;
319 }
320
321 return false;
322}
323
324bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio,
325 unsigned int nr_segs)
326{
327 struct elevator_queue *e = q->elevator;
328 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
329 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, bio->bi_opf, ctx);
330 bool ret = false;
331 enum hctx_type type;
332
333 if (e && e->type->ops.bio_merge)
334 return e->type->ops.bio_merge(hctx, bio, nr_segs);
335
336 type = hctx->type;
337 if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
338 !list_empty_careful(&ctx->rq_lists[type])) {
339 /* default per sw-queue merge */
340 spin_lock(&ctx->lock);
341 ret = blk_mq_attempt_merge(q, hctx, ctx, bio, nr_segs);
342 spin_unlock(&ctx->lock);
343 }
344
345 return ret;
346}
347
348bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
349{
350 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
351}
352EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
353
354void blk_mq_sched_request_inserted(struct request *rq)
355{
356 trace_block_rq_insert(rq->q, rq);
357}
358EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
359
360static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
361 bool has_sched,
362 struct request *rq)
363{
364 /* dispatch flush rq directly */
365 if (rq->rq_flags & RQF_FLUSH_SEQ) {
366 spin_lock(&hctx->lock);
367 list_add(&rq->queuelist, &hctx->dispatch);
368 spin_unlock(&hctx->lock);
369 return true;
370 }
371
372 if (has_sched)
373 rq->rq_flags |= RQF_SORTED;
374
375 return false;
376}
377
378void blk_mq_sched_insert_request(struct request *rq, bool at_head,
379 bool run_queue, bool async)
380{
381 struct request_queue *q = rq->q;
382 struct elevator_queue *e = q->elevator;
383 struct blk_mq_ctx *ctx = rq->mq_ctx;
384 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
385
386 /* flush rq in flush machinery need to be dispatched directly */
387 if (!(rq->rq_flags & RQF_FLUSH_SEQ) && op_is_flush(rq->cmd_flags)) {
388 blk_insert_flush(rq);
389 goto run;
390 }
391
392 WARN_ON(e && (rq->tag != -1));
393
394 if (blk_mq_sched_bypass_insert(hctx, !!e, rq))
395 goto run;
396
397 if (e && e->type->ops.insert_requests) {
398 LIST_HEAD(list);
399
400 list_add(&rq->queuelist, &list);
401 e->type->ops.insert_requests(hctx, &list, at_head);
402 } else {
403 spin_lock(&ctx->lock);
404 __blk_mq_insert_request(hctx, rq, at_head);
405 spin_unlock(&ctx->lock);
406 }
407
408run:
409 if (run_queue)
410 blk_mq_run_hw_queue(hctx, async);
411}
412
413void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx *hctx,
414 struct blk_mq_ctx *ctx,
415 struct list_head *list, bool run_queue_async)
416{
417 struct elevator_queue *e;
418 struct request_queue *q = hctx->queue;
419
420 /*
421 * blk_mq_sched_insert_requests() is called from flush plug
422 * context only, and hold one usage counter to prevent queue
423 * from being released.
424 */
425 percpu_ref_get(&q->q_usage_counter);
426
427 e = hctx->queue->elevator;
428 if (e && e->type->ops.insert_requests)
429 e->type->ops.insert_requests(hctx, list, false);
430 else {
431 /*
432 * try to issue requests directly if the hw queue isn't
433 * busy in case of 'none' scheduler, and this way may save
434 * us one extra enqueue & dequeue to sw queue.
435 */
436 if (!hctx->dispatch_busy && !e && !run_queue_async) {
437 blk_mq_try_issue_list_directly(hctx, list);
438 if (list_empty(list))
439 goto out;
440 }
441 blk_mq_insert_requests(hctx, ctx, list);
442 }
443
444 blk_mq_run_hw_queue(hctx, run_queue_async);
445 out:
446 percpu_ref_put(&q->q_usage_counter);
447}
448
449static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
450 struct blk_mq_hw_ctx *hctx,
451 unsigned int hctx_idx)
452{
453 if (hctx->sched_tags) {
454 blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx);
455 blk_mq_free_rq_map(hctx->sched_tags);
456 hctx->sched_tags = NULL;
457 }
458}
459
460static int blk_mq_sched_alloc_tags(struct request_queue *q,
461 struct blk_mq_hw_ctx *hctx,
462 unsigned int hctx_idx)
463{
464 struct blk_mq_tag_set *set = q->tag_set;
465 int ret;
466
467 hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
468 set->reserved_tags);
469 if (!hctx->sched_tags)
470 return -ENOMEM;
471
472 ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests);
473 if (ret)
474 blk_mq_sched_free_tags(set, hctx, hctx_idx);
475
476 return ret;
477}
478
479/* called in queue's release handler, tagset has gone away */
480static void blk_mq_sched_tags_teardown(struct request_queue *q)
481{
482 struct blk_mq_hw_ctx *hctx;
483 int i;
484
485 queue_for_each_hw_ctx(q, hctx, i) {
486 if (hctx->sched_tags) {
487 blk_mq_free_rq_map(hctx->sched_tags);
488 hctx->sched_tags = NULL;
489 }
490 }
491}
492
493int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
494{
495 struct blk_mq_hw_ctx *hctx;
496 struct elevator_queue *eq;
497 unsigned int i;
498 int ret;
499
500 if (!e) {
501 q->elevator = NULL;
502 q->nr_requests = q->tag_set->queue_depth;
503 return 0;
504 }
505
506 /*
507 * Default to double of smaller one between hw queue_depth and 128,
508 * since we don't split into sync/async like the old code did.
509 * Additionally, this is a per-hw queue depth.
510 */
511 q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
512 BLKDEV_MAX_RQ);
513
514 queue_for_each_hw_ctx(q, hctx, i) {
515 ret = blk_mq_sched_alloc_tags(q, hctx, i);
516 if (ret)
517 goto err;
518 }
519
520 ret = e->ops.init_sched(q, e);
521 if (ret)
522 goto err;
523
524 blk_mq_debugfs_register_sched(q);
525
526 queue_for_each_hw_ctx(q, hctx, i) {
527 if (e->ops.init_hctx) {
528 ret = e->ops.init_hctx(hctx, i);
529 if (ret) {
530 eq = q->elevator;
531 blk_mq_sched_free_requests(q);
532 blk_mq_exit_sched(q, eq);
533 kobject_put(&eq->kobj);
534 return ret;
535 }
536 }
537 blk_mq_debugfs_register_sched_hctx(q, hctx);
538 }
539
540 return 0;
541
542err:
543 blk_mq_sched_free_requests(q);
544 blk_mq_sched_tags_teardown(q);
545 q->elevator = NULL;
546 return ret;
547}
548
549/*
550 * called in either blk_queue_cleanup or elevator_switch, tagset
551 * is required for freeing requests
552 */
553void blk_mq_sched_free_requests(struct request_queue *q)
554{
555 struct blk_mq_hw_ctx *hctx;
556 int i;
557
558 queue_for_each_hw_ctx(q, hctx, i) {
559 if (hctx->sched_tags)
560 blk_mq_free_rqs(q->tag_set, hctx->sched_tags, i);
561 }
562}
563
564void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
565{
566 struct blk_mq_hw_ctx *hctx;
567 unsigned int i;
568
569 queue_for_each_hw_ctx(q, hctx, i) {
570 blk_mq_debugfs_unregister_sched_hctx(hctx);
571 if (e->type->ops.exit_hctx && hctx->sched_data) {
572 e->type->ops.exit_hctx(hctx, i);
573 hctx->sched_data = NULL;
574 }
575 }
576 blk_mq_debugfs_unregister_sched(q);
577 if (e->type->ops.exit_sched)
578 e->type->ops.exit_sched(e);
579 blk_mq_sched_tags_teardown(q);
580 q->elevator = NULL;
581}