]> git.ipfire.org Git - thirdparty/kernel/stable.git/blob - block/blk-mq-sched.c
Merge tag 'lkdtm-next' of https://git.kernel.org/pub/scm/linux/kernel/git/kees/linux...
[thirdparty/kernel/stable.git] / block / blk-mq-sched.c
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
20 void 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 }
33 EXPORT_SYMBOL_GPL(blk_mq_sched_free_hctx_data);
34
35 void 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 */
65 void 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 }
72 EXPORT_SYMBOL_GPL(blk_mq_sched_mark_restart_hctx);
73
74 void 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 */
88 static 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
118 static 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 */
134 static 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
170 void 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
226 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
227 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(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:
242 if (!blk_mq_sched_allow_merge(q, rq, bio))
243 return false;
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;
250 case ELEVATOR_DISCARD_MERGE:
251 return bio_attempt_discard_merge(q, rq, bio);
252 default:
253 return false;
254 }
255 }
256 EXPORT_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 */
262 bool blk_mq_bio_list_merge(struct request_queue *q, struct list_head *list,
263 struct bio *bio)
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(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
293 return merged;
294 }
295
296 return false;
297 }
298 EXPORT_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 */
305 static bool blk_mq_attempt_merge(struct request_queue *q,
306 struct blk_mq_hw_ctx *hctx,
307 struct blk_mq_ctx *ctx, struct bio *bio)
308 {
309 enum hctx_type type = hctx->type;
310
311 lockdep_assert_held(&ctx->lock);
312
313 if (blk_mq_bio_list_merge(q, &ctx->rq_lists[type], bio)) {
314 ctx->rq_merged++;
315 return true;
316 }
317
318 return false;
319 }
320
321 bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio)
322 {
323 struct elevator_queue *e = q->elevator;
324 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
325 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, bio->bi_opf, ctx);
326 bool ret = false;
327 enum hctx_type type;
328
329 if (e && e->type->ops.bio_merge) {
330 blk_mq_put_ctx(ctx);
331 return e->type->ops.bio_merge(hctx, bio);
332 }
333
334 type = hctx->type;
335 if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
336 !list_empty_careful(&ctx->rq_lists[type])) {
337 /* default per sw-queue merge */
338 spin_lock(&ctx->lock);
339 ret = blk_mq_attempt_merge(q, hctx, ctx, bio);
340 spin_unlock(&ctx->lock);
341 }
342
343 blk_mq_put_ctx(ctx);
344 return ret;
345 }
346
347 bool 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 }
351 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
352
353 void blk_mq_sched_request_inserted(struct request *rq)
354 {
355 trace_block_rq_insert(rq->q, rq);
356 }
357 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
358
359 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
360 bool has_sched,
361 struct request *rq)
362 {
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
371 if (has_sched)
372 rq->rq_flags |= RQF_SORTED;
373
374 return false;
375 }
376
377 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
378 bool run_queue, bool async)
379 {
380 struct request_queue *q = rq->q;
381 struct elevator_queue *e = q->elevator;
382 struct blk_mq_ctx *ctx = rq->mq_ctx;
383 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
384
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)) {
387 blk_insert_flush(rq);
388 goto run;
389 }
390
391 WARN_ON(e && (rq->tag != -1));
392
393 if (blk_mq_sched_bypass_insert(hctx, !!e, rq))
394 goto run;
395
396 if (e && e->type->ops.insert_requests) {
397 LIST_HEAD(list);
398
399 list_add(&rq->queuelist, &list);
400 e->type->ops.insert_requests(hctx, &list, at_head);
401 } else {
402 spin_lock(&ctx->lock);
403 __blk_mq_insert_request(hctx, rq, at_head);
404 spin_unlock(&ctx->lock);
405 }
406
407 run:
408 if (run_queue)
409 blk_mq_run_hw_queue(hctx, async);
410 }
411
412 void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx *hctx,
413 struct blk_mq_ctx *ctx,
414 struct list_head *list, bool run_queue_async)
415 {
416 struct elevator_queue *e;
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);
425
426 e = hctx->queue->elevator;
427 if (e && e->type->ops.insert_requests)
428 e->type->ops.insert_requests(hctx, list, false);
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 */
435 if (!hctx->dispatch_busy && !e && !run_queue_async) {
436 blk_mq_try_issue_list_directly(hctx, list);
437 if (list_empty(list))
438 goto out;
439 }
440 blk_mq_insert_requests(hctx, ctx, list);
441 }
442
443 blk_mq_run_hw_queue(hctx, run_queue_async);
444 out:
445 percpu_ref_put(&q->q_usage_counter);
446 }
447
448 static 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
459 static 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
478 static void blk_mq_sched_tags_teardown(struct request_queue *q)
479 {
480 struct blk_mq_tag_set *set = q->tag_set;
481 struct blk_mq_hw_ctx *hctx;
482 int i;
483
484 queue_for_each_hw_ctx(q, hctx, i)
485 blk_mq_sched_free_tags(set, hctx, i);
486 }
487
488 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
489 {
490 struct blk_mq_hw_ctx *hctx;
491 struct elevator_queue *eq;
492 unsigned int i;
493 int ret;
494
495 if (!e) {
496 q->elevator = NULL;
497 q->nr_requests = q->tag_set->queue_depth;
498 return 0;
499 }
500
501 /*
502 * Default to double of smaller one between hw queue_depth and 128,
503 * since we don't split into sync/async like the old code did.
504 * Additionally, this is a per-hw queue depth.
505 */
506 q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
507 BLKDEV_MAX_RQ);
508
509 queue_for_each_hw_ctx(q, hctx, i) {
510 ret = blk_mq_sched_alloc_tags(q, hctx, i);
511 if (ret)
512 goto err;
513 }
514
515 ret = e->ops.init_sched(q, e);
516 if (ret)
517 goto err;
518
519 blk_mq_debugfs_register_sched(q);
520
521 queue_for_each_hw_ctx(q, hctx, i) {
522 if (e->ops.init_hctx) {
523 ret = e->ops.init_hctx(hctx, i);
524 if (ret) {
525 eq = q->elevator;
526 blk_mq_exit_sched(q, eq);
527 kobject_put(&eq->kobj);
528 return ret;
529 }
530 }
531 blk_mq_debugfs_register_sched_hctx(q, hctx);
532 }
533
534 return 0;
535
536 err:
537 blk_mq_sched_tags_teardown(q);
538 q->elevator = NULL;
539 return ret;
540 }
541
542 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
543 {
544 struct blk_mq_hw_ctx *hctx;
545 unsigned int i;
546
547 queue_for_each_hw_ctx(q, hctx, i) {
548 blk_mq_debugfs_unregister_sched_hctx(hctx);
549 if (e->type->ops.exit_hctx && hctx->sched_data) {
550 e->type->ops.exit_hctx(hctx, i);
551 hctx->sched_data = NULL;
552 }
553 }
554 blk_mq_debugfs_unregister_sched(q);
555 if (e->type->ops.exit_sched)
556 e->type->ops.exit_sched(e);
557 blk_mq_sched_tags_teardown(q);
558 q->elevator = NULL;
559 }