2 * blk-mq scheduling framework
4 * Copyright (C) 2016 Jens Axboe
6 #include <linux/kernel.h>
7 #include <linux/module.h>
8 #include <linux/blk-mq.h>
10 #include <trace/events/block.h>
14 #include "blk-mq-debugfs.h"
15 #include "blk-mq-sched.h"
16 #include "blk-mq-tag.h"
19 void blk_mq_sched_free_hctx_data(struct request_queue *q,
20 void (*exit)(struct blk_mq_hw_ctx *))
22 struct blk_mq_hw_ctx *hctx;
25 queue_for_each_hw_ctx(q, hctx, i) {
26 if (exit && hctx->sched_data)
28 kfree(hctx->sched_data);
29 hctx->sched_data = NULL;
32 EXPORT_SYMBOL_GPL(blk_mq_sched_free_hctx_data);
34 void blk_mq_sched_assign_ioc(struct request *rq, struct bio *bio)
36 struct request_queue *q = rq->q;
37 struct io_context *ioc = rq_ioc(bio);
40 spin_lock_irq(q->queue_lock);
41 icq = ioc_lookup_icq(ioc, q);
42 spin_unlock_irq(q->queue_lock);
45 icq = ioc_create_icq(ioc, q, GFP_ATOMIC);
49 get_io_context(icq->ioc);
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.
57 static void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx)
59 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
62 if (hctx->flags & BLK_MQ_F_TAG_SHARED) {
63 struct request_queue *q = hctx->queue;
65 if (!test_and_set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
66 atomic_inc(&q->shared_hctx_restart);
68 set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
71 static bool blk_mq_sched_restart_hctx(struct blk_mq_hw_ctx *hctx)
73 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
76 if (hctx->flags & BLK_MQ_F_TAG_SHARED) {
77 struct request_queue *q = hctx->queue;
79 if (test_and_clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
80 atomic_dec(&q->shared_hctx_restart);
82 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
84 if (blk_mq_hctx_has_pending(hctx)) {
85 blk_mq_run_hw_queue(hctx, true);
92 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
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;
100 /* RCU or SRCU read lock is needed before checking quiesced flag */
101 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
107 * If we have previous entries on our dispatch list, grab them first for
108 * more fair dispatch.
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);
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.
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);
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
139 if (!did_work && has_sched_dispatch) {
143 rq = e->type->ops.mq.dispatch_request(hctx);
146 list_add(&rq->queuelist, &rq_list);
147 } while (blk_mq_dispatch_rq_list(q, &rq_list));
151 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
152 struct request **merged_request)
156 switch (elv_merge(q, &rq, bio)) {
157 case ELEVATOR_BACK_MERGE:
158 if (!blk_mq_sched_allow_merge(q, rq, bio))
160 if (!bio_attempt_back_merge(q, rq, bio))
162 *merged_request = attempt_back_merge(q, rq);
163 if (!*merged_request)
164 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
166 case ELEVATOR_FRONT_MERGE:
167 if (!blk_mq_sched_allow_merge(q, rq, bio))
169 if (!bio_attempt_front_merge(q, rq, bio))
171 *merged_request = attempt_front_merge(q, rq);
172 if (!*merged_request)
173 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
179 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
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.
186 static bool blk_mq_attempt_merge(struct request_queue *q,
187 struct blk_mq_ctx *ctx, struct bio *bio)
192 lockdep_assert_held(&ctx->lock);
194 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
200 if (!blk_rq_merge_ok(rq, bio))
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);
208 case ELEVATOR_FRONT_MERGE:
209 if (blk_mq_sched_allow_merge(q, rq, bio))
210 merged = bio_attempt_front_merge(q, rq, bio);
212 case ELEVATOR_DISCARD_MERGE:
213 merged = bio_attempt_discard_merge(q, rq, bio);
227 bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio)
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);
234 if (e && e->type->ops.mq.bio_merge) {
236 return e->type->ops.mq.bio_merge(hctx, bio);
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);
250 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
252 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
254 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
256 void blk_mq_sched_request_inserted(struct request *rq)
258 trace_block_rq_insert(rq->q, rq);
260 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
262 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
266 rq->rq_flags |= RQF_SORTED;
271 * If we already have a real request tag, send directly to
274 spin_lock(&hctx->lock);
275 list_add(&rq->queuelist, &hctx->dispatch);
276 spin_unlock(&hctx->lock);
281 * list_for_each_entry_rcu_rr - iterate in a round-robin fashion over rcu list
283 * @skip: the list element that will not be examined. Iteration starts at
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).
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)), \
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
302 void blk_mq_sched_restart(struct blk_mq_hw_ctx *const hctx)
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;
310 if (set->flags & BLK_MQ_F_TAG_SHARED) {
312 * If this is 0, then we know that no hardware queues
313 * have RESTART marked. We're done.
315 if (!atomic_read(&queue->shared_hctx_restart))
319 list_for_each_entry_rcu_rr(q, queue, &set->tag_list,
321 queue_for_each_hw_ctx(q, hctx2, i)
322 if (hctx2->tags == tags &&
323 blk_mq_sched_restart_hctx(hctx2))
326 j = hctx->queue_num + 1;
327 for (i = 0; i < queue->nr_hw_queues; i++, j++) {
328 if (j == queue->nr_hw_queues)
330 hctx2 = queue->queue_hw_ctx[j];
331 if (hctx2->tags == tags &&
332 blk_mq_sched_restart_hctx(hctx2))
338 blk_mq_sched_restart_hctx(hctx);
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.
347 static void blk_mq_sched_insert_flush(struct blk_mq_hw_ctx *hctx,
348 struct request *rq, bool can_block)
350 if (blk_mq_get_driver_tag(rq, &hctx, can_block)) {
351 blk_insert_flush(rq);
352 blk_mq_run_hw_queue(hctx, true);
354 blk_mq_add_to_requeue_list(rq, false, true);
357 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
358 bool run_queue, bool async, bool can_block)
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);
365 if (rq->tag == -1 && op_is_flush(rq->cmd_flags)) {
366 blk_mq_sched_insert_flush(hctx, rq, can_block);
370 if (e && blk_mq_sched_bypass_insert(hctx, rq))
373 if (e && e->type->ops.mq.insert_requests) {
376 list_add(&rq->queuelist, &list);
377 e->type->ops.mq.insert_requests(hctx, &list, at_head);
379 spin_lock(&ctx->lock);
380 __blk_mq_insert_request(hctx, rq, at_head);
381 spin_unlock(&ctx->lock);
386 blk_mq_run_hw_queue(hctx, async);
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)
393 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
394 struct elevator_queue *e = hctx->queue->elevator;
397 struct request *rq, *next;
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).
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);
413 if (e && e->type->ops.mq.insert_requests)
414 e->type->ops.mq.insert_requests(hctx, list, false);
416 blk_mq_insert_requests(hctx, ctx, list);
418 blk_mq_run_hw_queue(hctx, run_queue_async);
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)
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;
432 static int blk_mq_sched_alloc_tags(struct request_queue *q,
433 struct blk_mq_hw_ctx *hctx,
434 unsigned int hctx_idx)
436 struct blk_mq_tag_set *set = q->tag_set;
439 hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
441 if (!hctx->sched_tags)
444 ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests);
446 blk_mq_sched_free_tags(set, hctx, hctx_idx);
451 static void blk_mq_sched_tags_teardown(struct request_queue *q)
453 struct blk_mq_tag_set *set = q->tag_set;
454 struct blk_mq_hw_ctx *hctx;
457 queue_for_each_hw_ctx(q, hctx, i)
458 blk_mq_sched_free_tags(set, hctx, i);
461 int blk_mq_sched_init_hctx(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
462 unsigned int hctx_idx)
464 struct elevator_queue *e = q->elevator;
470 ret = blk_mq_sched_alloc_tags(q, hctx, hctx_idx);
474 if (e->type->ops.mq.init_hctx) {
475 ret = e->type->ops.mq.init_hctx(hctx, hctx_idx);
477 blk_mq_sched_free_tags(q->tag_set, hctx, hctx_idx);
482 blk_mq_debugfs_register_sched_hctx(q, hctx);
487 void blk_mq_sched_exit_hctx(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
488 unsigned int hctx_idx)
490 struct elevator_queue *e = q->elevator;
495 blk_mq_debugfs_unregister_sched_hctx(hctx);
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;
502 blk_mq_sched_free_tags(q->tag_set, hctx, hctx_idx);
505 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
507 struct blk_mq_hw_ctx *hctx;
508 struct elevator_queue *eq;
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.
522 q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
525 queue_for_each_hw_ctx(q, hctx, i) {
526 ret = blk_mq_sched_alloc_tags(q, hctx, i);
531 ret = e->ops.mq.init_sched(q, e);
535 blk_mq_debugfs_register_sched(q);
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);
542 blk_mq_exit_sched(q, eq);
543 kobject_put(&eq->kobj);
547 blk_mq_debugfs_register_sched_hctx(q, hctx);
553 blk_mq_sched_tags_teardown(q);
558 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
560 struct blk_mq_hw_ctx *hctx;
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;
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);
577 int blk_mq_sched_init(struct request_queue *q)
581 mutex_lock(&q->sysfs_lock);
582 ret = elevator_init(q, NULL);
583 mutex_unlock(&q->sysfs_lock);