]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - block/blk-mq-sched.c
blk-mq: dequeue request one by one from sw queue if hctx is busy
[mirror_ubuntu-bionic-kernel.git] / block / blk-mq-sched.c
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
19 void 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 }
32 EXPORT_SYMBOL_GPL(blk_mq_sched_free_hctx_data);
33
34 void 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 */
57 static 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
71 static 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 */
92 static 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
123 static 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 */
139 static 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
175 /* return true if hw queue need to be run again */
176 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
177 {
178 struct request_queue *q = hctx->queue;
179 struct elevator_queue *e = q->elevator;
180 const bool has_sched_dispatch = e && e->type->ops.mq.dispatch_request;
181 LIST_HEAD(rq_list);
182
183 /* RCU or SRCU read lock is needed before checking quiesced flag */
184 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
185 return;
186
187 hctx->run++;
188
189 /*
190 * If we have previous entries on our dispatch list, grab them first for
191 * more fair dispatch.
192 */
193 if (!list_empty_careful(&hctx->dispatch)) {
194 spin_lock(&hctx->lock);
195 if (!list_empty(&hctx->dispatch))
196 list_splice_init(&hctx->dispatch, &rq_list);
197 spin_unlock(&hctx->lock);
198 }
199
200 /*
201 * Only ask the scheduler for requests, if we didn't have residual
202 * requests from the dispatch list. This is to avoid the case where
203 * we only ever dispatch a fraction of the requests available because
204 * of low device queue depth. Once we pull requests out of the IO
205 * scheduler, we can no longer merge or sort them. So it's best to
206 * leave them there for as long as we can. Mark the hw queue as
207 * needing a restart in that case.
208 *
209 * We want to dispatch from the scheduler if there was nothing
210 * on the dispatch list or we were able to dispatch from the
211 * dispatch list.
212 */
213 if (!list_empty(&rq_list)) {
214 blk_mq_sched_mark_restart_hctx(hctx);
215 if (blk_mq_dispatch_rq_list(q, &rq_list, false)) {
216 if (has_sched_dispatch)
217 blk_mq_do_dispatch_sched(hctx);
218 else
219 blk_mq_do_dispatch_ctx(hctx);
220 }
221 } else if (has_sched_dispatch) {
222 blk_mq_do_dispatch_sched(hctx);
223 } else if (hctx->dispatch_busy) {
224 /* dequeue request one by one from sw queue if queue is busy */
225 blk_mq_do_dispatch_ctx(hctx);
226 } else {
227 blk_mq_flush_busy_ctxs(hctx, &rq_list);
228 blk_mq_dispatch_rq_list(q, &rq_list, false);
229 }
230 }
231
232 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
233 struct request **merged_request)
234 {
235 struct request *rq;
236
237 switch (elv_merge(q, &rq, bio)) {
238 case ELEVATOR_BACK_MERGE:
239 if (!blk_mq_sched_allow_merge(q, rq, bio))
240 return false;
241 if (!bio_attempt_back_merge(q, rq, bio))
242 return false;
243 *merged_request = attempt_back_merge(q, rq);
244 if (!*merged_request)
245 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
246 return true;
247 case ELEVATOR_FRONT_MERGE:
248 if (!blk_mq_sched_allow_merge(q, rq, bio))
249 return false;
250 if (!bio_attempt_front_merge(q, rq, bio))
251 return false;
252 *merged_request = attempt_front_merge(q, rq);
253 if (!*merged_request)
254 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
255 return true;
256 default:
257 return false;
258 }
259 }
260 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
261
262 /*
263 * Reverse check our software queue for entries that we could potentially
264 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
265 * too much time checking for merges.
266 */
267 static bool blk_mq_attempt_merge(struct request_queue *q,
268 struct blk_mq_ctx *ctx, struct bio *bio)
269 {
270 struct request *rq;
271 int checked = 8;
272
273 lockdep_assert_held(&ctx->lock);
274
275 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
276 bool merged = false;
277
278 if (!checked--)
279 break;
280
281 if (!blk_rq_merge_ok(rq, bio))
282 continue;
283
284 switch (blk_try_merge(rq, bio)) {
285 case ELEVATOR_BACK_MERGE:
286 if (blk_mq_sched_allow_merge(q, rq, bio))
287 merged = bio_attempt_back_merge(q, rq, bio);
288 break;
289 case ELEVATOR_FRONT_MERGE:
290 if (blk_mq_sched_allow_merge(q, rq, bio))
291 merged = bio_attempt_front_merge(q, rq, bio);
292 break;
293 case ELEVATOR_DISCARD_MERGE:
294 merged = bio_attempt_discard_merge(q, rq, bio);
295 break;
296 default:
297 continue;
298 }
299
300 if (merged)
301 ctx->rq_merged++;
302 return merged;
303 }
304
305 return false;
306 }
307
308 bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio)
309 {
310 struct elevator_queue *e = q->elevator;
311 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
312 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
313 bool ret = false;
314
315 if (e && e->type->ops.mq.bio_merge) {
316 blk_mq_put_ctx(ctx);
317 return e->type->ops.mq.bio_merge(hctx, bio);
318 }
319
320 if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
321 !list_empty_careful(&ctx->rq_list)) {
322 /* default per sw-queue merge */
323 spin_lock(&ctx->lock);
324 ret = blk_mq_attempt_merge(q, ctx, bio);
325 spin_unlock(&ctx->lock);
326 }
327
328 blk_mq_put_ctx(ctx);
329 return ret;
330 }
331
332 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
333 {
334 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
335 }
336 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
337
338 void blk_mq_sched_request_inserted(struct request *rq)
339 {
340 trace_block_rq_insert(rq->q, rq);
341 }
342 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
343
344 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
345 bool has_sched,
346 struct request *rq)
347 {
348 /* dispatch flush rq directly */
349 if (rq->rq_flags & RQF_FLUSH_SEQ) {
350 spin_lock(&hctx->lock);
351 list_add(&rq->queuelist, &hctx->dispatch);
352 spin_unlock(&hctx->lock);
353 return true;
354 }
355
356 if (has_sched)
357 rq->rq_flags |= RQF_SORTED;
358
359 return false;
360 }
361
362 /**
363 * list_for_each_entry_rcu_rr - iterate in a round-robin fashion over rcu list
364 * @pos: loop cursor.
365 * @skip: the list element that will not be examined. Iteration starts at
366 * @skip->next.
367 * @head: head of the list to examine. This list must have at least one
368 * element, namely @skip.
369 * @member: name of the list_head structure within typeof(*pos).
370 */
371 #define list_for_each_entry_rcu_rr(pos, skip, head, member) \
372 for ((pos) = (skip); \
373 (pos = (pos)->member.next != (head) ? list_entry_rcu( \
374 (pos)->member.next, typeof(*pos), member) : \
375 list_entry_rcu((pos)->member.next->next, typeof(*pos), member)), \
376 (pos) != (skip); )
377
378 /*
379 * Called after a driver tag has been freed to check whether a hctx needs to
380 * be restarted. Restarts @hctx if its tag set is not shared. Restarts hardware
381 * queues in a round-robin fashion if the tag set of @hctx is shared with other
382 * hardware queues.
383 */
384 void blk_mq_sched_restart(struct blk_mq_hw_ctx *const hctx)
385 {
386 struct blk_mq_tags *const tags = hctx->tags;
387 struct blk_mq_tag_set *const set = hctx->queue->tag_set;
388 struct request_queue *const queue = hctx->queue, *q;
389 struct blk_mq_hw_ctx *hctx2;
390 unsigned int i, j;
391
392 if (set->flags & BLK_MQ_F_TAG_SHARED) {
393 /*
394 * If this is 0, then we know that no hardware queues
395 * have RESTART marked. We're done.
396 */
397 if (!atomic_read(&queue->shared_hctx_restart))
398 return;
399
400 rcu_read_lock();
401 list_for_each_entry_rcu_rr(q, queue, &set->tag_list,
402 tag_set_list) {
403 queue_for_each_hw_ctx(q, hctx2, i)
404 if (hctx2->tags == tags &&
405 blk_mq_sched_restart_hctx(hctx2))
406 goto done;
407 }
408 j = hctx->queue_num + 1;
409 for (i = 0; i < queue->nr_hw_queues; i++, j++) {
410 if (j == queue->nr_hw_queues)
411 j = 0;
412 hctx2 = queue->queue_hw_ctx[j];
413 if (hctx2->tags == tags &&
414 blk_mq_sched_restart_hctx(hctx2))
415 break;
416 }
417 done:
418 rcu_read_unlock();
419 } else {
420 blk_mq_sched_restart_hctx(hctx);
421 }
422 }
423
424 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
425 bool run_queue, bool async)
426 {
427 struct request_queue *q = rq->q;
428 struct elevator_queue *e = q->elevator;
429 struct blk_mq_ctx *ctx = rq->mq_ctx;
430 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
431
432 /* flush rq in flush machinery need to be dispatched directly */
433 if (!(rq->rq_flags & RQF_FLUSH_SEQ) && op_is_flush(rq->cmd_flags)) {
434 blk_insert_flush(rq);
435 goto run;
436 }
437
438 WARN_ON(e && (rq->tag != -1));
439
440 if (blk_mq_sched_bypass_insert(hctx, !!e, rq))
441 goto run;
442
443 if (e && e->type->ops.mq.insert_requests) {
444 LIST_HEAD(list);
445
446 list_add(&rq->queuelist, &list);
447 e->type->ops.mq.insert_requests(hctx, &list, at_head);
448 } else {
449 spin_lock(&ctx->lock);
450 __blk_mq_insert_request(hctx, rq, at_head);
451 spin_unlock(&ctx->lock);
452 }
453
454 run:
455 if (run_queue)
456 blk_mq_run_hw_queue(hctx, async);
457 }
458
459 void blk_mq_sched_insert_requests(struct request_queue *q,
460 struct blk_mq_ctx *ctx,
461 struct list_head *list, bool run_queue_async)
462 {
463 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
464 struct elevator_queue *e = hctx->queue->elevator;
465
466 if (e && e->type->ops.mq.insert_requests)
467 e->type->ops.mq.insert_requests(hctx, list, false);
468 else
469 blk_mq_insert_requests(hctx, ctx, list);
470
471 blk_mq_run_hw_queue(hctx, run_queue_async);
472 }
473
474 static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
475 struct blk_mq_hw_ctx *hctx,
476 unsigned int hctx_idx)
477 {
478 if (hctx->sched_tags) {
479 blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx);
480 blk_mq_free_rq_map(hctx->sched_tags);
481 hctx->sched_tags = NULL;
482 }
483 }
484
485 static int blk_mq_sched_alloc_tags(struct request_queue *q,
486 struct blk_mq_hw_ctx *hctx,
487 unsigned int hctx_idx)
488 {
489 struct blk_mq_tag_set *set = q->tag_set;
490 int ret;
491
492 hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
493 set->reserved_tags);
494 if (!hctx->sched_tags)
495 return -ENOMEM;
496
497 ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests);
498 if (ret)
499 blk_mq_sched_free_tags(set, hctx, hctx_idx);
500
501 return ret;
502 }
503
504 static void blk_mq_sched_tags_teardown(struct request_queue *q)
505 {
506 struct blk_mq_tag_set *set = q->tag_set;
507 struct blk_mq_hw_ctx *hctx;
508 int i;
509
510 queue_for_each_hw_ctx(q, hctx, i)
511 blk_mq_sched_free_tags(set, hctx, i);
512 }
513
514 int blk_mq_sched_init_hctx(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
515 unsigned int hctx_idx)
516 {
517 struct elevator_queue *e = q->elevator;
518 int ret;
519
520 if (!e)
521 return 0;
522
523 ret = blk_mq_sched_alloc_tags(q, hctx, hctx_idx);
524 if (ret)
525 return ret;
526
527 if (e->type->ops.mq.init_hctx) {
528 ret = e->type->ops.mq.init_hctx(hctx, hctx_idx);
529 if (ret) {
530 blk_mq_sched_free_tags(q->tag_set, hctx, hctx_idx);
531 return ret;
532 }
533 }
534
535 blk_mq_debugfs_register_sched_hctx(q, hctx);
536
537 return 0;
538 }
539
540 void blk_mq_sched_exit_hctx(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
541 unsigned int hctx_idx)
542 {
543 struct elevator_queue *e = q->elevator;
544
545 if (!e)
546 return;
547
548 blk_mq_debugfs_unregister_sched_hctx(hctx);
549
550 if (e->type->ops.mq.exit_hctx && hctx->sched_data) {
551 e->type->ops.mq.exit_hctx(hctx, hctx_idx);
552 hctx->sched_data = NULL;
553 }
554
555 blk_mq_sched_free_tags(q->tag_set, hctx, hctx_idx);
556 }
557
558 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
559 {
560 struct blk_mq_hw_ctx *hctx;
561 struct elevator_queue *eq;
562 unsigned int i;
563 int ret;
564
565 if (!e) {
566 q->elevator = NULL;
567 return 0;
568 }
569
570 /*
571 * Default to double of smaller one between hw queue_depth and 128,
572 * since we don't split into sync/async like the old code did.
573 * Additionally, this is a per-hw queue depth.
574 */
575 q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
576 BLKDEV_MAX_RQ);
577
578 queue_for_each_hw_ctx(q, hctx, i) {
579 ret = blk_mq_sched_alloc_tags(q, hctx, i);
580 if (ret)
581 goto err;
582 }
583
584 ret = e->ops.mq.init_sched(q, e);
585 if (ret)
586 goto err;
587
588 blk_mq_debugfs_register_sched(q);
589
590 queue_for_each_hw_ctx(q, hctx, i) {
591 if (e->ops.mq.init_hctx) {
592 ret = e->ops.mq.init_hctx(hctx, i);
593 if (ret) {
594 eq = q->elevator;
595 blk_mq_exit_sched(q, eq);
596 kobject_put(&eq->kobj);
597 return ret;
598 }
599 }
600 blk_mq_debugfs_register_sched_hctx(q, hctx);
601 }
602
603 return 0;
604
605 err:
606 blk_mq_sched_tags_teardown(q);
607 q->elevator = NULL;
608 return ret;
609 }
610
611 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
612 {
613 struct blk_mq_hw_ctx *hctx;
614 unsigned int i;
615
616 queue_for_each_hw_ctx(q, hctx, i) {
617 blk_mq_debugfs_unregister_sched_hctx(hctx);
618 if (e->type->ops.mq.exit_hctx && hctx->sched_data) {
619 e->type->ops.mq.exit_hctx(hctx, i);
620 hctx->sched_data = NULL;
621 }
622 }
623 blk_mq_debugfs_unregister_sched(q);
624 if (e->type->ops.mq.exit_sched)
625 e->type->ops.mq.exit_sched(e);
626 blk_mq_sched_tags_teardown(q);
627 q->elevator = NULL;
628 }
629
630 int blk_mq_sched_init(struct request_queue *q)
631 {
632 int ret;
633
634 mutex_lock(&q->sysfs_lock);
635 ret = elevator_init(q, NULL);
636 mutex_unlock(&q->sysfs_lock);
637
638 return ret;
639 }