xref: /openbmc/linux/block/blk-mq-sched.c (revision dea54fba)
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 	if (blk_mq_hctx_has_pending(hctx)) {
85 		blk_mq_run_hw_queue(hctx, true);
86 		return true;
87 	}
88 
89 	return false;
90 }
91 
92 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
93 {
94 	struct request_queue *q = hctx->queue;
95 	struct elevator_queue *e = q->elevator;
96 	const bool has_sched_dispatch = e && e->type->ops.mq.dispatch_request;
97 	bool did_work = false;
98 	LIST_HEAD(rq_list);
99 
100 	/* RCU or SRCU read lock is needed before checking quiesced flag */
101 	if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
102 		return;
103 
104 	hctx->run++;
105 
106 	/*
107 	 * If we have previous entries on our dispatch list, grab them first for
108 	 * more fair dispatch.
109 	 */
110 	if (!list_empty_careful(&hctx->dispatch)) {
111 		spin_lock(&hctx->lock);
112 		if (!list_empty(&hctx->dispatch))
113 			list_splice_init(&hctx->dispatch, &rq_list);
114 		spin_unlock(&hctx->lock);
115 	}
116 
117 	/*
118 	 * Only ask the scheduler for requests, if we didn't have residual
119 	 * requests from the dispatch list. This is to avoid the case where
120 	 * we only ever dispatch a fraction of the requests available because
121 	 * of low device queue depth. Once we pull requests out of the IO
122 	 * scheduler, we can no longer merge or sort them. So it's best to
123 	 * leave them there for as long as we can. Mark the hw queue as
124 	 * needing a restart in that case.
125 	 */
126 	if (!list_empty(&rq_list)) {
127 		blk_mq_sched_mark_restart_hctx(hctx);
128 		did_work = blk_mq_dispatch_rq_list(q, &rq_list);
129 	} else if (!has_sched_dispatch) {
130 		blk_mq_flush_busy_ctxs(hctx, &rq_list);
131 		blk_mq_dispatch_rq_list(q, &rq_list);
132 	}
133 
134 	/*
135 	 * We want to dispatch from the scheduler if we had no work left
136 	 * on the dispatch list, OR if we did have work but weren't able
137 	 * to make progress.
138 	 */
139 	if (!did_work && has_sched_dispatch) {
140 		do {
141 			struct request *rq;
142 
143 			rq = e->type->ops.mq.dispatch_request(hctx);
144 			if (!rq)
145 				break;
146 			list_add(&rq->queuelist, &rq_list);
147 		} while (blk_mq_dispatch_rq_list(q, &rq_list));
148 	}
149 }
150 
151 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
152 			    struct request **merged_request)
153 {
154 	struct request *rq;
155 
156 	switch (elv_merge(q, &rq, bio)) {
157 	case ELEVATOR_BACK_MERGE:
158 		if (!blk_mq_sched_allow_merge(q, rq, bio))
159 			return false;
160 		if (!bio_attempt_back_merge(q, rq, bio))
161 			return false;
162 		*merged_request = attempt_back_merge(q, rq);
163 		if (!*merged_request)
164 			elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
165 		return true;
166 	case ELEVATOR_FRONT_MERGE:
167 		if (!blk_mq_sched_allow_merge(q, rq, bio))
168 			return false;
169 		if (!bio_attempt_front_merge(q, rq, bio))
170 			return false;
171 		*merged_request = attempt_front_merge(q, rq);
172 		if (!*merged_request)
173 			elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
174 		return true;
175 	default:
176 		return false;
177 	}
178 }
179 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
180 
181 /*
182  * Reverse check our software queue for entries that we could potentially
183  * merge with. Currently includes a hand-wavy stop count of 8, to not spend
184  * too much time checking for merges.
185  */
186 static bool blk_mq_attempt_merge(struct request_queue *q,
187 				 struct blk_mq_ctx *ctx, struct bio *bio)
188 {
189 	struct request *rq;
190 	int checked = 8;
191 
192 	lockdep_assert_held(&ctx->lock);
193 
194 	list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
195 		bool merged = false;
196 
197 		if (!checked--)
198 			break;
199 
200 		if (!blk_rq_merge_ok(rq, bio))
201 			continue;
202 
203 		switch (blk_try_merge(rq, bio)) {
204 		case ELEVATOR_BACK_MERGE:
205 			if (blk_mq_sched_allow_merge(q, rq, bio))
206 				merged = bio_attempt_back_merge(q, rq, bio);
207 			break;
208 		case ELEVATOR_FRONT_MERGE:
209 			if (blk_mq_sched_allow_merge(q, rq, bio))
210 				merged = bio_attempt_front_merge(q, rq, bio);
211 			break;
212 		case ELEVATOR_DISCARD_MERGE:
213 			merged = bio_attempt_discard_merge(q, rq, bio);
214 			break;
215 		default:
216 			continue;
217 		}
218 
219 		if (merged)
220 			ctx->rq_merged++;
221 		return merged;
222 	}
223 
224 	return false;
225 }
226 
227 bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio)
228 {
229 	struct elevator_queue *e = q->elevator;
230 	struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
231 	struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
232 	bool ret = false;
233 
234 	if (e && e->type->ops.mq.bio_merge) {
235 		blk_mq_put_ctx(ctx);
236 		return e->type->ops.mq.bio_merge(hctx, bio);
237 	}
238 
239 	if (hctx->flags & BLK_MQ_F_SHOULD_MERGE) {
240 		/* default per sw-queue merge */
241 		spin_lock(&ctx->lock);
242 		ret = blk_mq_attempt_merge(q, ctx, bio);
243 		spin_unlock(&ctx->lock);
244 	}
245 
246 	blk_mq_put_ctx(ctx);
247 	return ret;
248 }
249 
250 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
251 {
252 	return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
253 }
254 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
255 
256 void blk_mq_sched_request_inserted(struct request *rq)
257 {
258 	trace_block_rq_insert(rq->q, rq);
259 }
260 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
261 
262 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
263 				       struct request *rq)
264 {
265 	if (rq->tag == -1) {
266 		rq->rq_flags |= RQF_SORTED;
267 		return false;
268 	}
269 
270 	/*
271 	 * If we already have a real request tag, send directly to
272 	 * the dispatch list.
273 	 */
274 	spin_lock(&hctx->lock);
275 	list_add(&rq->queuelist, &hctx->dispatch);
276 	spin_unlock(&hctx->lock);
277 	return true;
278 }
279 
280 /**
281  * list_for_each_entry_rcu_rr - iterate in a round-robin fashion over rcu list
282  * @pos:    loop cursor.
283  * @skip:   the list element that will not be examined. Iteration starts at
284  *          @skip->next.
285  * @head:   head of the list to examine. This list must have at least one
286  *          element, namely @skip.
287  * @member: name of the list_head structure within typeof(*pos).
288  */
289 #define list_for_each_entry_rcu_rr(pos, skip, head, member)		\
290 	for ((pos) = (skip);						\
291 	     (pos = (pos)->member.next != (head) ? list_entry_rcu(	\
292 			(pos)->member.next, typeof(*pos), member) :	\
293 	      list_entry_rcu((pos)->member.next->next, typeof(*pos), member)), \
294 	     (pos) != (skip); )
295 
296 /*
297  * Called after a driver tag has been freed to check whether a hctx needs to
298  * be restarted. Restarts @hctx if its tag set is not shared. Restarts hardware
299  * queues in a round-robin fashion if the tag set of @hctx is shared with other
300  * hardware queues.
301  */
302 void blk_mq_sched_restart(struct blk_mq_hw_ctx *const hctx)
303 {
304 	struct blk_mq_tags *const tags = hctx->tags;
305 	struct blk_mq_tag_set *const set = hctx->queue->tag_set;
306 	struct request_queue *const queue = hctx->queue, *q;
307 	struct blk_mq_hw_ctx *hctx2;
308 	unsigned int i, j;
309 
310 	if (set->flags & BLK_MQ_F_TAG_SHARED) {
311 		/*
312 		 * If this is 0, then we know that no hardware queues
313 		 * have RESTART marked. We're done.
314 		 */
315 		if (!atomic_read(&queue->shared_hctx_restart))
316 			return;
317 
318 		rcu_read_lock();
319 		list_for_each_entry_rcu_rr(q, queue, &set->tag_list,
320 					   tag_set_list) {
321 			queue_for_each_hw_ctx(q, hctx2, i)
322 				if (hctx2->tags == tags &&
323 				    blk_mq_sched_restart_hctx(hctx2))
324 					goto done;
325 		}
326 		j = hctx->queue_num + 1;
327 		for (i = 0; i < queue->nr_hw_queues; i++, j++) {
328 			if (j == queue->nr_hw_queues)
329 				j = 0;
330 			hctx2 = queue->queue_hw_ctx[j];
331 			if (hctx2->tags == tags &&
332 			    blk_mq_sched_restart_hctx(hctx2))
333 				break;
334 		}
335 done:
336 		rcu_read_unlock();
337 	} else {
338 		blk_mq_sched_restart_hctx(hctx);
339 	}
340 }
341 
342 /*
343  * Add flush/fua to the queue. If we fail getting a driver tag, then
344  * punt to the requeue list. Requeue will re-invoke us from a context
345  * that's safe to block from.
346  */
347 static void blk_mq_sched_insert_flush(struct blk_mq_hw_ctx *hctx,
348 				      struct request *rq, bool can_block)
349 {
350 	if (blk_mq_get_driver_tag(rq, &hctx, can_block)) {
351 		blk_insert_flush(rq);
352 		blk_mq_run_hw_queue(hctx, true);
353 	} else
354 		blk_mq_add_to_requeue_list(rq, false, true);
355 }
356 
357 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
358 				 bool run_queue, bool async, bool can_block)
359 {
360 	struct request_queue *q = rq->q;
361 	struct elevator_queue *e = q->elevator;
362 	struct blk_mq_ctx *ctx = rq->mq_ctx;
363 	struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
364 
365 	if (rq->tag == -1 && op_is_flush(rq->cmd_flags)) {
366 		blk_mq_sched_insert_flush(hctx, rq, can_block);
367 		return;
368 	}
369 
370 	if (e && blk_mq_sched_bypass_insert(hctx, rq))
371 		goto run;
372 
373 	if (e && e->type->ops.mq.insert_requests) {
374 		LIST_HEAD(list);
375 
376 		list_add(&rq->queuelist, &list);
377 		e->type->ops.mq.insert_requests(hctx, &list, at_head);
378 	} else {
379 		spin_lock(&ctx->lock);
380 		__blk_mq_insert_request(hctx, rq, at_head);
381 		spin_unlock(&ctx->lock);
382 	}
383 
384 run:
385 	if (run_queue)
386 		blk_mq_run_hw_queue(hctx, async);
387 }
388 
389 void blk_mq_sched_insert_requests(struct request_queue *q,
390 				  struct blk_mq_ctx *ctx,
391 				  struct list_head *list, bool run_queue_async)
392 {
393 	struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
394 	struct elevator_queue *e = hctx->queue->elevator;
395 
396 	if (e) {
397 		struct request *rq, *next;
398 
399 		/*
400 		 * We bypass requests that already have a driver tag assigned,
401 		 * which should only be flushes. Flushes are only ever inserted
402 		 * as single requests, so we shouldn't ever hit the
403 		 * WARN_ON_ONCE() below (but let's handle it just in case).
404 		 */
405 		list_for_each_entry_safe(rq, next, list, queuelist) {
406 			if (WARN_ON_ONCE(rq->tag != -1)) {
407 				list_del_init(&rq->queuelist);
408 				blk_mq_sched_bypass_insert(hctx, rq);
409 			}
410 		}
411 	}
412 
413 	if (e && e->type->ops.mq.insert_requests)
414 		e->type->ops.mq.insert_requests(hctx, list, false);
415 	else
416 		blk_mq_insert_requests(hctx, ctx, list);
417 
418 	blk_mq_run_hw_queue(hctx, run_queue_async);
419 }
420 
421 static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
422 				   struct blk_mq_hw_ctx *hctx,
423 				   unsigned int hctx_idx)
424 {
425 	if (hctx->sched_tags) {
426 		blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx);
427 		blk_mq_free_rq_map(hctx->sched_tags);
428 		hctx->sched_tags = NULL;
429 	}
430 }
431 
432 static int blk_mq_sched_alloc_tags(struct request_queue *q,
433 				   struct blk_mq_hw_ctx *hctx,
434 				   unsigned int hctx_idx)
435 {
436 	struct blk_mq_tag_set *set = q->tag_set;
437 	int ret;
438 
439 	hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
440 					       set->reserved_tags);
441 	if (!hctx->sched_tags)
442 		return -ENOMEM;
443 
444 	ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests);
445 	if (ret)
446 		blk_mq_sched_free_tags(set, hctx, hctx_idx);
447 
448 	return ret;
449 }
450 
451 static void blk_mq_sched_tags_teardown(struct request_queue *q)
452 {
453 	struct blk_mq_tag_set *set = q->tag_set;
454 	struct blk_mq_hw_ctx *hctx;
455 	int i;
456 
457 	queue_for_each_hw_ctx(q, hctx, i)
458 		blk_mq_sched_free_tags(set, hctx, i);
459 }
460 
461 int blk_mq_sched_init_hctx(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
462 			   unsigned int hctx_idx)
463 {
464 	struct elevator_queue *e = q->elevator;
465 	int ret;
466 
467 	if (!e)
468 		return 0;
469 
470 	ret = blk_mq_sched_alloc_tags(q, hctx, hctx_idx);
471 	if (ret)
472 		return ret;
473 
474 	if (e->type->ops.mq.init_hctx) {
475 		ret = e->type->ops.mq.init_hctx(hctx, hctx_idx);
476 		if (ret) {
477 			blk_mq_sched_free_tags(q->tag_set, hctx, hctx_idx);
478 			return ret;
479 		}
480 	}
481 
482 	blk_mq_debugfs_register_sched_hctx(q, hctx);
483 
484 	return 0;
485 }
486 
487 void blk_mq_sched_exit_hctx(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
488 			    unsigned int hctx_idx)
489 {
490 	struct elevator_queue *e = q->elevator;
491 
492 	if (!e)
493 		return;
494 
495 	blk_mq_debugfs_unregister_sched_hctx(hctx);
496 
497 	if (e->type->ops.mq.exit_hctx && hctx->sched_data) {
498 		e->type->ops.mq.exit_hctx(hctx, hctx_idx);
499 		hctx->sched_data = NULL;
500 	}
501 
502 	blk_mq_sched_free_tags(q->tag_set, hctx, hctx_idx);
503 }
504 
505 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
506 {
507 	struct blk_mq_hw_ctx *hctx;
508 	struct elevator_queue *eq;
509 	unsigned int i;
510 	int ret;
511 
512 	if (!e) {
513 		q->elevator = NULL;
514 		return 0;
515 	}
516 
517 	/*
518 	 * Default to double of smaller one between hw queue_depth and 128,
519 	 * since we don't split into sync/async like the old code did.
520 	 * Additionally, this is a per-hw queue depth.
521 	 */
522 	q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
523 				   BLKDEV_MAX_RQ);
524 
525 	queue_for_each_hw_ctx(q, hctx, i) {
526 		ret = blk_mq_sched_alloc_tags(q, hctx, i);
527 		if (ret)
528 			goto err;
529 	}
530 
531 	ret = e->ops.mq.init_sched(q, e);
532 	if (ret)
533 		goto err;
534 
535 	blk_mq_debugfs_register_sched(q);
536 
537 	queue_for_each_hw_ctx(q, hctx, i) {
538 		if (e->ops.mq.init_hctx) {
539 			ret = e->ops.mq.init_hctx(hctx, i);
540 			if (ret) {
541 				eq = q->elevator;
542 				blk_mq_exit_sched(q, eq);
543 				kobject_put(&eq->kobj);
544 				return ret;
545 			}
546 		}
547 		blk_mq_debugfs_register_sched_hctx(q, hctx);
548 	}
549 
550 	return 0;
551 
552 err:
553 	blk_mq_sched_tags_teardown(q);
554 	q->elevator = NULL;
555 	return ret;
556 }
557 
558 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
559 {
560 	struct blk_mq_hw_ctx *hctx;
561 	unsigned int i;
562 
563 	queue_for_each_hw_ctx(q, hctx, i) {
564 		blk_mq_debugfs_unregister_sched_hctx(hctx);
565 		if (e->type->ops.mq.exit_hctx && hctx->sched_data) {
566 			e->type->ops.mq.exit_hctx(hctx, i);
567 			hctx->sched_data = NULL;
568 		}
569 	}
570 	blk_mq_debugfs_unregister_sched(q);
571 	if (e->type->ops.mq.exit_sched)
572 		e->type->ops.mq.exit_sched(e);
573 	blk_mq_sched_tags_teardown(q);
574 	q->elevator = NULL;
575 }
576 
577 int blk_mq_sched_init(struct request_queue *q)
578 {
579 	int ret;
580 
581 	mutex_lock(&q->sysfs_lock);
582 	ret = elevator_init(q, NULL);
583 	mutex_unlock(&q->sysfs_lock);
584 
585 	return ret;
586 }
587