xref: /openbmc/linux/block/blk-mq-sched.c (revision 2359ccdd)
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 void 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 
238 bool 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 }
268 EXPORT_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  */
275 static 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 
316 bool __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 
339 bool 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 }
343 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
344 
345 void blk_mq_sched_request_inserted(struct request *rq)
346 {
347 	trace_block_rq_insert(rq->q, rq);
348 }
349 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
350 
351 static 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  */
391 void 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 		}
424 done:
425 		rcu_read_unlock();
426 	} else {
427 		blk_mq_sched_restart_hctx(hctx);
428 	}
429 }
430 
431 void 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 
461 run:
462 	if (run_queue)
463 		blk_mq_run_hw_queue(hctx, async);
464 }
465 
466 void 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 
481 static 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 
492 static 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 
511 static 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 
521 int 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 
547 void 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 
565 int 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 
612 err:
613 	blk_mq_sched_tags_teardown(q);
614 	q->elevator = NULL;
615 	return ret;
616 }
617 
618 void 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 
637 int 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 }
647