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/list_sort.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-wbt.h"
 18
 19/*
 20 * Mark a hardware queue as needing a restart.
 21 */
 22void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx)
 23{
 24	if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
 25		return;
 26
 27	set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
 28}
 29EXPORT_SYMBOL_GPL(blk_mq_sched_mark_restart_hctx);
 30
 31void __blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx)
 32{
 33	clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
 34
 35	/*
 36	 * Order clearing SCHED_RESTART and list_empty_careful(&hctx->dispatch)
 37	 * in blk_mq_run_hw_queue(). Its pair is the barrier in
 38	 * blk_mq_dispatch_rq_list(). So dispatch code won't see SCHED_RESTART,
 39	 * meantime new request added to hctx->dispatch is missed to check in
 40	 * blk_mq_run_hw_queue().
 41	 */
 42	smp_mb();
 43
 44	blk_mq_run_hw_queue(hctx, true);
 45}
 46
 47static int sched_rq_cmp(void *priv, const struct list_head *a,
 48			const struct list_head *b)
 49{
 50	struct request *rqa = container_of(a, struct request, queuelist);
 51	struct request *rqb = container_of(b, struct request, queuelist);
 52
 53	return rqa->mq_hctx > rqb->mq_hctx;
 54}
 55
 56static bool blk_mq_dispatch_hctx_list(struct list_head *rq_list)
 57{
 58	struct blk_mq_hw_ctx *hctx =
 59		list_first_entry(rq_list, struct request, queuelist)->mq_hctx;
 60	struct request *rq;
 61	LIST_HEAD(hctx_list);
 62
 63	list_for_each_entry(rq, rq_list, queuelist) {
 64		if (rq->mq_hctx != hctx) {
 65			list_cut_before(&hctx_list, rq_list, &rq->queuelist);
 66			goto dispatch;
 67		}
 68	}
 69	list_splice_tail_init(rq_list, &hctx_list);
 70
 71dispatch:
 72	return blk_mq_dispatch_rq_list(hctx, &hctx_list, false);
 73}
 74
 75#define BLK_MQ_BUDGET_DELAY	3		/* ms units */
 76
 77/*
 78 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
 79 * its queue by itself in its completion handler, so we don't need to
 80 * restart queue if .get_budget() fails to get the budget.
 81 *
 82 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
 83 * be run again.  This is necessary to avoid starving flushes.
 84 */
 85static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
 86{
 87	struct request_queue *q = hctx->queue;
 88	struct elevator_queue *e = q->elevator;
 89	bool multi_hctxs = false, run_queue = false;
 90	bool dispatched = false, busy = false;
 91	unsigned int max_dispatch;
 92	LIST_HEAD(rq_list);
 93	int count = 0;
 94
 95	if (hctx->dispatch_busy)
 96		max_dispatch = 1;
 97	else
 98		max_dispatch = hctx->queue->nr_requests;
 99
100	do {
101		struct request *rq;
102		int budget_token;
103
104		if (e->type->ops.has_work && !e->type->ops.has_work(hctx))
105			break;
106
107		if (!list_empty_careful(&hctx->dispatch)) {
108			busy = true;
109			break;
110		}
111
112		budget_token = blk_mq_get_dispatch_budget(q);
113		if (budget_token < 0)
114			break;
115
116		rq = e->type->ops.dispatch_request(hctx);
117		if (!rq) {
118			blk_mq_put_dispatch_budget(q, budget_token);
119			/*
120			 * We're releasing without dispatching. Holding the
121			 * budget could have blocked any "hctx"s with the
122			 * same queue and if we didn't dispatch then there's
123			 * no guarantee anyone will kick the queue.  Kick it
124			 * ourselves.
125			 */
126			run_queue = true;
127			break;
128		}
129
130		blk_mq_set_rq_budget_token(rq, budget_token);
131
132		/*
133		 * Now this rq owns the budget which has to be released
134		 * if this rq won't be queued to driver via .queue_rq()
135		 * in blk_mq_dispatch_rq_list().
136		 */
137		list_add_tail(&rq->queuelist, &rq_list);
138		count++;
139		if (rq->mq_hctx != hctx)
140			multi_hctxs = true;
141
142		/*
143		 * If we cannot get tag for the request, stop dequeueing
144		 * requests from the IO scheduler. We are unlikely to be able
145		 * to submit them anyway and it creates false impression for
146		 * scheduling heuristics that the device can take more IO.
147		 */
148		if (!blk_mq_get_driver_tag(rq))
149			break;
150	} while (count < max_dispatch);
151
152	if (!count) {
153		if (run_queue)
154			blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
155	} else if (multi_hctxs) {
156		/*
157		 * Requests from different hctx may be dequeued from some
158		 * schedulers, such as bfq and deadline.
159		 *
160		 * Sort the requests in the list according to their hctx,
161		 * dispatch batching requests from same hctx at a time.
162		 */
163		list_sort(NULL, &rq_list, sched_rq_cmp);
164		do {
165			dispatched |= blk_mq_dispatch_hctx_list(&rq_list);
166		} while (!list_empty(&rq_list));
167	} else {
168		dispatched = blk_mq_dispatch_rq_list(hctx, &rq_list, false);
169	}
170
171	if (busy)
172		return -EAGAIN;
173	return !!dispatched;
174}
175
176static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
177{
178	unsigned long end = jiffies + HZ;
179	int ret;
180
181	do {
182		ret = __blk_mq_do_dispatch_sched(hctx);
183		if (ret != 1)
184			break;
185		if (need_resched() || time_is_before_jiffies(end)) {
186			blk_mq_delay_run_hw_queue(hctx, 0);
187			break;
188		}
189	} while (1);
190
191	return ret;
192}
193
194static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
195					  struct blk_mq_ctx *ctx)
196{
197	unsigned short idx = ctx->index_hw[hctx->type];
198
199	if (++idx == hctx->nr_ctx)
200		idx = 0;
201
202	return hctx->ctxs[idx];
203}
204
205/*
206 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
207 * its queue by itself in its completion handler, so we don't need to
208 * restart queue if .get_budget() fails to get the budget.
209 *
210 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
211 * be run again.  This is necessary to avoid starving flushes.
212 */
213static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
214{
215	struct request_queue *q = hctx->queue;
216	LIST_HEAD(rq_list);
217	struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from);
218	int ret = 0;
219	struct request *rq;
220
221	do {
222		int budget_token;
223
224		if (!list_empty_careful(&hctx->dispatch)) {
225			ret = -EAGAIN;
226			break;
227		}
228
229		if (!sbitmap_any_bit_set(&hctx->ctx_map))
230			break;
231
232		budget_token = blk_mq_get_dispatch_budget(q);
233		if (budget_token < 0)
234			break;
235
236		rq = blk_mq_dequeue_from_ctx(hctx, ctx);
237		if (!rq) {
238			blk_mq_put_dispatch_budget(q, budget_token);
239			/*
240			 * We're releasing without dispatching. Holding the
241			 * budget could have blocked any "hctx"s with the
242			 * same queue and if we didn't dispatch then there's
243			 * no guarantee anyone will kick the queue.  Kick it
244			 * ourselves.
245			 */
246			blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
247			break;
248		}
249
250		blk_mq_set_rq_budget_token(rq, budget_token);
251
252		/*
253		 * Now this rq owns the budget which has to be released
254		 * if this rq won't be queued to driver via .queue_rq()
255		 * in blk_mq_dispatch_rq_list().
256		 */
257		list_add(&rq->queuelist, &rq_list);
258
259		/* round robin for fair dispatch */
260		ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
261
262	} while (blk_mq_dispatch_rq_list(rq->mq_hctx, &rq_list, false));
263
264	WRITE_ONCE(hctx->dispatch_from, ctx);
265	return ret;
266}
267
268static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
269{
270	bool need_dispatch = false;
271	LIST_HEAD(rq_list);
272
273	/*
274	 * If we have previous entries on our dispatch list, grab them first for
275	 * more fair dispatch.
276	 */
277	if (!list_empty_careful(&hctx->dispatch)) {
278		spin_lock(&hctx->lock);
279		if (!list_empty(&hctx->dispatch))
280			list_splice_init(&hctx->dispatch, &rq_list);
281		spin_unlock(&hctx->lock);
282	}
283
284	/*
285	 * Only ask the scheduler for requests, if we didn't have residual
286	 * requests from the dispatch list. This is to avoid the case where
287	 * we only ever dispatch a fraction of the requests available because
288	 * of low device queue depth. Once we pull requests out of the IO
289	 * scheduler, we can no longer merge or sort them. So it's best to
290	 * leave them there for as long as we can. Mark the hw queue as
291	 * needing a restart in that case.
292	 *
293	 * We want to dispatch from the scheduler if there was nothing
294	 * on the dispatch list or we were able to dispatch from the
295	 * dispatch list.
296	 */
297	if (!list_empty(&rq_list)) {
298		blk_mq_sched_mark_restart_hctx(hctx);
299		if (!blk_mq_dispatch_rq_list(hctx, &rq_list, true))
300			return 0;
301		need_dispatch = true;
302	} else {
303		need_dispatch = hctx->dispatch_busy;
304	}
305
306	if (hctx->queue->elevator)
307		return blk_mq_do_dispatch_sched(hctx);
308
309	/* dequeue request one by one from sw queue if queue is busy */
310	if (need_dispatch)
311		return blk_mq_do_dispatch_ctx(hctx);
312	blk_mq_flush_busy_ctxs(hctx, &rq_list);
313	blk_mq_dispatch_rq_list(hctx, &rq_list, true);
314	return 0;
315}
316
317void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
318{
319	struct request_queue *q = hctx->queue;
320
321	/* RCU or SRCU read lock is needed before checking quiesced flag */
322	if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
323		return;
324
325	/*
326	 * A return of -EAGAIN is an indication that hctx->dispatch is not
327	 * empty and we must run again in order to avoid starving flushes.
328	 */
329	if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) {
330		if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN)
331			blk_mq_run_hw_queue(hctx, true);
332	}
333}
334
335bool blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio,
336		unsigned int nr_segs)
337{
338	struct elevator_queue *e = q->elevator;
339	struct blk_mq_ctx *ctx;
340	struct blk_mq_hw_ctx *hctx;
341	bool ret = false;
342	enum hctx_type type;
343
344	if (e && e->type->ops.bio_merge) {
345		ret = e->type->ops.bio_merge(q, bio, nr_segs);
346		goto out_put;
347	}
348
349	ctx = blk_mq_get_ctx(q);
350	hctx = blk_mq_map_queue(bio->bi_opf, ctx);
351	type = hctx->type;
352	if (list_empty_careful(&ctx->rq_lists[type]))
353		goto out_put;
354
355	/* default per sw-queue merge */
356	spin_lock(&ctx->lock);
357	/*
358	 * Reverse check our software queue for entries that we could
359	 * potentially merge with. Currently includes a hand-wavy stop
360	 * count of 8, to not spend too much time checking for merges.
361	 */
362	if (blk_bio_list_merge(q, &ctx->rq_lists[type], bio, nr_segs))
363		ret = true;
364
365	spin_unlock(&ctx->lock);
366out_put:
367	return ret;
368}
369
370bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq,
371				   struct list_head *free)
372{
373	return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq, free);
374}
375EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
376
377/* called in queue's release handler, tagset has gone away */
378static void blk_mq_sched_tags_teardown(struct request_queue *q, unsigned int flags)
379{
380	struct blk_mq_hw_ctx *hctx;
381	unsigned long i;
382
383	queue_for_each_hw_ctx(q, hctx, i)
384		hctx->sched_tags = NULL;
385
386	if (blk_mq_is_shared_tags(flags))
387		q->sched_shared_tags = NULL;
388}
389
390void blk_mq_sched_reg_debugfs(struct request_queue *q)
391{
392	struct blk_mq_hw_ctx *hctx;
393	unsigned long i;
394
395	mutex_lock(&q->debugfs_mutex);
396	blk_mq_debugfs_register_sched(q);
397	queue_for_each_hw_ctx(q, hctx, i)
398		blk_mq_debugfs_register_sched_hctx(q, hctx);
399	mutex_unlock(&q->debugfs_mutex);
400}
401
402void blk_mq_sched_unreg_debugfs(struct request_queue *q)
403{
404	struct blk_mq_hw_ctx *hctx;
405	unsigned long i;
406
407	mutex_lock(&q->debugfs_mutex);
408	queue_for_each_hw_ctx(q, hctx, i)
409		blk_mq_debugfs_unregister_sched_hctx(hctx);
410	blk_mq_debugfs_unregister_sched(q);
411	mutex_unlock(&q->debugfs_mutex);
412}
413
414void blk_mq_free_sched_tags(struct elevator_tags *et,
415		struct blk_mq_tag_set *set)
416{
417	unsigned long i;
418
419	/* Shared tags are stored at index 0 in @tags. */
420	if (blk_mq_is_shared_tags(set->flags))
421		blk_mq_free_map_and_rqs(set, et->tags[0], BLK_MQ_NO_HCTX_IDX);
422	else {
423		for (i = 0; i < et->nr_hw_queues; i++)
424			blk_mq_free_map_and_rqs(set, et->tags[i], i);
425	}
426
427	kfree(et);
428}
429
430void blk_mq_free_sched_res(struct elevator_resources *res,
431		struct elevator_type *type,
432		struct blk_mq_tag_set *set)
433{
434	if (res->et) {
435		blk_mq_free_sched_tags(res->et, set);
436		res->et = NULL;
437	}
438	if (res->data) {
439		blk_mq_free_sched_data(type, res->data);
440		res->data = NULL;
441	}
442}
443
444void blk_mq_free_sched_res_batch(struct xarray *elv_tbl,
445		struct blk_mq_tag_set *set)
446{
447	struct request_queue *q;
448	struct elv_change_ctx *ctx;
449
450	lockdep_assert_held_write(&set->update_nr_hwq_lock);
451
452	list_for_each_entry(q, &set->tag_list, tag_set_list) {
453		/*
454		 * Accessing q->elevator without holding q->elevator_lock is
455		 * safe because we're holding here set->update_nr_hwq_lock in
456		 * the writer context. So, scheduler update/switch code (which
457		 * acquires the same lock but in the reader context) can't run
458		 * concurrently.
459		 */
460		if (q->elevator) {
461			ctx = xa_load(elv_tbl, q->id);
462			if (!ctx) {
463				WARN_ON_ONCE(1);
464				continue;
465			}
466			blk_mq_free_sched_res(&ctx->res, ctx->type, set);
467		}
468	}
469}
470
471void blk_mq_free_sched_ctx_batch(struct xarray *elv_tbl)
472{
473	unsigned long i;
474	struct elv_change_ctx *ctx;
475
476	xa_for_each(elv_tbl, i, ctx) {
477		xa_erase(elv_tbl, i);
478		kfree(ctx);
479	}
480}
481
482int blk_mq_alloc_sched_ctx_batch(struct xarray *elv_tbl,
483		struct blk_mq_tag_set *set)
484{
485	struct request_queue *q;
486	struct elv_change_ctx *ctx;
487
488	lockdep_assert_held_write(&set->update_nr_hwq_lock);
489
490	list_for_each_entry(q, &set->tag_list, tag_set_list) {
491		ctx = kzalloc(sizeof(struct elv_change_ctx), GFP_KERNEL);
492		if (!ctx)
493			return -ENOMEM;
494
495		if (xa_insert(elv_tbl, q->id, ctx, GFP_KERNEL)) {
496			kfree(ctx);
497			return -ENOMEM;
498		}
499	}
500	return 0;
501}
502
503struct elevator_tags *blk_mq_alloc_sched_tags(struct blk_mq_tag_set *set,
504		unsigned int nr_hw_queues, unsigned int nr_requests)
505{
506	unsigned int nr_tags;
507	int i;
508	struct elevator_tags *et;
509	gfp_t gfp = GFP_NOIO | __GFP_ZERO | __GFP_NOWARN | __GFP_NORETRY;
510
511	if (blk_mq_is_shared_tags(set->flags))
512		nr_tags = 1;
513	else
514		nr_tags = nr_hw_queues;
515
516	et = kmalloc(struct_size(et, tags, nr_tags), gfp);
517	if (!et)
518		return NULL;
519
520	et->nr_requests = nr_requests;
521	et->nr_hw_queues = nr_hw_queues;
522
523	if (blk_mq_is_shared_tags(set->flags)) {
524		/* Shared tags are stored at index 0 in @tags. */
525		et->tags[0] = blk_mq_alloc_map_and_rqs(set, BLK_MQ_NO_HCTX_IDX,
526					MAX_SCHED_RQ);
527		if (!et->tags[0])
528			goto out;
529	} else {
530		for (i = 0; i < et->nr_hw_queues; i++) {
531			et->tags[i] = blk_mq_alloc_map_and_rqs(set, i,
532					et->nr_requests);
533			if (!et->tags[i])
534				goto out_unwind;
535		}
536	}
537
538	return et;
539out_unwind:
540	while (--i >= 0)
541		blk_mq_free_map_and_rqs(set, et->tags[i], i);
542out:
543	kfree(et);
544	return NULL;
545}
546
547int blk_mq_alloc_sched_res(struct request_queue *q,
548		struct elevator_type *type,
549		struct elevator_resources *res,
550		unsigned int nr_hw_queues)
551{
552	struct blk_mq_tag_set *set = q->tag_set;
553
554	res->et = blk_mq_alloc_sched_tags(set, nr_hw_queues,
555			blk_mq_default_nr_requests(set));
556	if (!res->et)
557		return -ENOMEM;
558
559	res->data = blk_mq_alloc_sched_data(q, type);
560	if (IS_ERR(res->data)) {
561		blk_mq_free_sched_tags(res->et, set);
562		return -ENOMEM;
563	}
564
565	return 0;
566}
567
568int blk_mq_alloc_sched_res_batch(struct xarray *elv_tbl,
569		struct blk_mq_tag_set *set, unsigned int nr_hw_queues)
570{
571	struct elv_change_ctx *ctx;
572	struct request_queue *q;
573	int ret = -ENOMEM;
574
575	lockdep_assert_held_write(&set->update_nr_hwq_lock);
576
577	list_for_each_entry(q, &set->tag_list, tag_set_list) {
578		/*
579		 * Accessing q->elevator without holding q->elevator_lock is
580		 * safe because we're holding here set->update_nr_hwq_lock in
581		 * the writer context. So, scheduler update/switch code (which
582		 * acquires the same lock but in the reader context) can't run
583		 * concurrently.
584		 */
585		if (q->elevator) {
586			ctx = xa_load(elv_tbl, q->id);
587			if (WARN_ON_ONCE(!ctx)) {
588				ret = -ENOENT;
589				goto out_unwind;
590			}
591
592			ret = blk_mq_alloc_sched_res(q, q->elevator->type,
593					&ctx->res, nr_hw_queues);
594			if (ret)
595				goto out_unwind;
596		}
597	}
598	return 0;
599
600out_unwind:
601	list_for_each_entry_continue_reverse(q, &set->tag_list, tag_set_list) {
602		if (q->elevator) {
603			ctx = xa_load(elv_tbl, q->id);
604			if (ctx)
605				blk_mq_free_sched_res(&ctx->res,
606						ctx->type, set);
607		}
608	}
609	return ret;
610}
611
612/* caller must have a reference to @e, will grab another one if successful */
613int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e,
614		struct elevator_resources *res)
615{
616	unsigned int flags = q->tag_set->flags;
617	struct elevator_tags *et = res->et;
618	struct blk_mq_hw_ctx *hctx;
619	struct elevator_queue *eq;
620	unsigned long i;
621	int ret;
622
623	eq = elevator_alloc(q, e, res);
624	if (!eq)
625		return -ENOMEM;
626
627	q->nr_requests = et->nr_requests;
628
629	if (blk_mq_is_shared_tags(flags)) {
630		/* Shared tags are stored at index 0 in @et->tags. */
631		q->sched_shared_tags = et->tags[0];
632		blk_mq_tag_update_sched_shared_tags(q, et->nr_requests);
633	}
634
635	queue_for_each_hw_ctx(q, hctx, i) {
636		if (blk_mq_is_shared_tags(flags))
637			hctx->sched_tags = q->sched_shared_tags;
638		else
639			hctx->sched_tags = et->tags[i];
640	}
641
642	ret = e->ops.init_sched(q, eq);
643	if (ret)
644		goto out;
645
646	queue_for_each_hw_ctx(q, hctx, i) {
647		if (e->ops.init_hctx) {
648			ret = e->ops.init_hctx(hctx, i);
649			if (ret) {
650				blk_mq_exit_sched(q, eq);
651				kobject_put(&eq->kobj);
652				return ret;
653			}
654		}
655	}
656	return 0;
657
658out:
659	blk_mq_sched_tags_teardown(q, flags);
660	kobject_put(&eq->kobj);
661	q->elevator = NULL;
662	return ret;
663}
664
665/*
666 * called in either blk_queue_cleanup or elevator_switch, tagset
667 * is required for freeing requests
668 */
669void blk_mq_sched_free_rqs(struct request_queue *q)
670{
671	struct blk_mq_hw_ctx *hctx;
672	unsigned long i;
673
674	if (blk_mq_is_shared_tags(q->tag_set->flags)) {
675		blk_mq_free_rqs(q->tag_set, q->sched_shared_tags,
676				BLK_MQ_NO_HCTX_IDX);
677	} else {
678		queue_for_each_hw_ctx(q, hctx, i) {
679			if (hctx->sched_tags)
680				blk_mq_free_rqs(q->tag_set,
681						hctx->sched_tags, i);
682		}
683	}
684}
685
686void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
687{
688	struct blk_mq_hw_ctx *hctx;
689	unsigned long i;
690	unsigned int flags = 0;
691
692	queue_for_each_hw_ctx(q, hctx, i) {
693		if (e->type->ops.exit_hctx && hctx->sched_data) {
694			e->type->ops.exit_hctx(hctx, i);
695			hctx->sched_data = NULL;
696		}
697		flags = hctx->flags;
698	}
699
700	if (e->type->ops.exit_sched)
701		e->type->ops.exit_sched(e);
702	blk_mq_sched_tags_teardown(q, flags);
703	set_bit(ELEVATOR_FLAG_DYING, &q->elevator->flags);
704	q->elevator = NULL;
705}