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);
53 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
55 struct request_queue *q = hctx->queue;
56 struct elevator_queue *e = q->elevator;
57 const bool has_sched_dispatch = e && e->type->ops.mq.dispatch_request;
58 bool did_work = false;
61 /* RCU or SRCU read lock is needed before checking quiesced flag */
62 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
68 * If we have previous entries on our dispatch list, grab them first for
71 if (!list_empty_careful(&hctx->dispatch)) {
72 spin_lock(&hctx->lock);
73 if (!list_empty(&hctx->dispatch))
74 list_splice_init(&hctx->dispatch, &rq_list);
75 spin_unlock(&hctx->lock);
79 * Only ask the scheduler for requests, if we didn't have residual
80 * requests from the dispatch list. This is to avoid the case where
81 * we only ever dispatch a fraction of the requests available because
82 * of low device queue depth. Once we pull requests out of the IO
83 * scheduler, we can no longer merge or sort them. So it's best to
84 * leave them there for as long as we can. Mark the hw queue as
85 * needing a restart in that case.
87 if (!list_empty(&rq_list)) {
88 blk_mq_sched_mark_restart_hctx(hctx);
89 did_work = blk_mq_dispatch_rq_list(q, &rq_list);
90 } else if (!has_sched_dispatch) {
91 blk_mq_flush_busy_ctxs(hctx, &rq_list);
92 blk_mq_dispatch_rq_list(q, &rq_list);
96 * We want to dispatch from the scheduler if we had no work left
97 * on the dispatch list, OR if we did have work but weren't able
100 if (!did_work && has_sched_dispatch) {
104 rq = e->type->ops.mq.dispatch_request(hctx);
107 list_add(&rq->queuelist, &rq_list);
108 } while (blk_mq_dispatch_rq_list(q, &rq_list));
112 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
113 struct request **merged_request)
117 switch (elv_merge(q, &rq, bio)) {
118 case ELEVATOR_BACK_MERGE:
119 if (!blk_mq_sched_allow_merge(q, rq, bio))
121 if (!bio_attempt_back_merge(q, rq, bio))
123 *merged_request = attempt_back_merge(q, rq);
124 if (!*merged_request)
125 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
127 case ELEVATOR_FRONT_MERGE:
128 if (!blk_mq_sched_allow_merge(q, rq, bio))
130 if (!bio_attempt_front_merge(q, rq, bio))
132 *merged_request = attempt_front_merge(q, rq);
133 if (!*merged_request)
134 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
140 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
143 * Reverse check our software queue for entries that we could potentially
144 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
145 * too much time checking for merges.
147 static bool blk_mq_attempt_merge(struct request_queue *q,
148 struct blk_mq_ctx *ctx, struct bio *bio)
153 lockdep_assert_held(&ctx->lock);
155 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
161 if (!blk_rq_merge_ok(rq, bio))
164 switch (blk_try_merge(rq, bio)) {
165 case ELEVATOR_BACK_MERGE:
166 if (blk_mq_sched_allow_merge(q, rq, bio))
167 merged = bio_attempt_back_merge(q, rq, bio);
169 case ELEVATOR_FRONT_MERGE:
170 if (blk_mq_sched_allow_merge(q, rq, bio))
171 merged = bio_attempt_front_merge(q, rq, bio);
173 case ELEVATOR_DISCARD_MERGE:
174 merged = bio_attempt_discard_merge(q, rq, bio);
188 bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio)
190 struct elevator_queue *e = q->elevator;
191 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
192 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
195 if (e && e->type->ops.mq.bio_merge) {
197 return e->type->ops.mq.bio_merge(hctx, bio);
200 if (hctx->flags & BLK_MQ_F_SHOULD_MERGE) {
201 /* default per sw-queue merge */
202 spin_lock(&ctx->lock);
203 ret = blk_mq_attempt_merge(q, ctx, bio);
204 spin_unlock(&ctx->lock);
211 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
213 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
215 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
217 void blk_mq_sched_request_inserted(struct request *rq)
219 trace_block_rq_insert(rq->q, rq);
221 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
223 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
227 rq->rq_flags |= RQF_SORTED;
232 * If we already have a real request tag, send directly to
235 spin_lock(&hctx->lock);
236 list_add(&rq->queuelist, &hctx->dispatch);
237 spin_unlock(&hctx->lock);
241 static bool blk_mq_sched_restart_hctx(struct blk_mq_hw_ctx *hctx)
243 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) {
244 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
245 if (blk_mq_hctx_has_pending(hctx)) {
246 blk_mq_run_hw_queue(hctx, true);
254 * list_for_each_entry_rcu_rr - iterate in a round-robin fashion over rcu list
256 * @skip: the list element that will not be examined. Iteration starts at
258 * @head: head of the list to examine. This list must have at least one
259 * element, namely @skip.
260 * @member: name of the list_head structure within typeof(*pos).
262 #define list_for_each_entry_rcu_rr(pos, skip, head, member) \
263 for ((pos) = (skip); \
264 (pos = (pos)->member.next != (head) ? list_entry_rcu( \
265 (pos)->member.next, typeof(*pos), member) : \
266 list_entry_rcu((pos)->member.next->next, typeof(*pos), member)), \
270 * Called after a driver tag has been freed to check whether a hctx needs to
271 * be restarted. Restarts @hctx if its tag set is not shared. Restarts hardware
272 * queues in a round-robin fashion if the tag set of @hctx is shared with other
275 void blk_mq_sched_restart(struct blk_mq_hw_ctx *const hctx)
277 struct blk_mq_tags *const tags = hctx->tags;
278 struct blk_mq_tag_set *const set = hctx->queue->tag_set;
279 struct request_queue *const queue = hctx->queue, *q;
280 struct blk_mq_hw_ctx *hctx2;
283 if (set->flags & BLK_MQ_F_TAG_SHARED) {
285 list_for_each_entry_rcu_rr(q, queue, &set->tag_list,
287 queue_for_each_hw_ctx(q, hctx2, i)
288 if (hctx2->tags == tags &&
289 blk_mq_sched_restart_hctx(hctx2))
292 j = hctx->queue_num + 1;
293 for (i = 0; i < queue->nr_hw_queues; i++, j++) {
294 if (j == queue->nr_hw_queues)
296 hctx2 = queue->queue_hw_ctx[j];
297 if (hctx2->tags == tags &&
298 blk_mq_sched_restart_hctx(hctx2))
304 blk_mq_sched_restart_hctx(hctx);
309 * Add flush/fua to the queue. If we fail getting a driver tag, then
310 * punt to the requeue list. Requeue will re-invoke us from a context
311 * that's safe to block from.
313 static void blk_mq_sched_insert_flush(struct blk_mq_hw_ctx *hctx,
314 struct request *rq, bool can_block)
316 if (blk_mq_get_driver_tag(rq, &hctx, can_block)) {
317 blk_insert_flush(rq);
318 blk_mq_run_hw_queue(hctx, true);
320 blk_mq_add_to_requeue_list(rq, false, true);
323 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
324 bool run_queue, bool async, bool can_block)
326 struct request_queue *q = rq->q;
327 struct elevator_queue *e = q->elevator;
328 struct blk_mq_ctx *ctx = rq->mq_ctx;
329 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
331 if (rq->tag == -1 && op_is_flush(rq->cmd_flags)) {
332 blk_mq_sched_insert_flush(hctx, rq, can_block);
336 if (e && blk_mq_sched_bypass_insert(hctx, rq))
339 if (e && e->type->ops.mq.insert_requests) {
342 list_add(&rq->queuelist, &list);
343 e->type->ops.mq.insert_requests(hctx, &list, at_head);
345 spin_lock(&ctx->lock);
346 __blk_mq_insert_request(hctx, rq, at_head);
347 spin_unlock(&ctx->lock);
352 blk_mq_run_hw_queue(hctx, async);
355 void blk_mq_sched_insert_requests(struct request_queue *q,
356 struct blk_mq_ctx *ctx,
357 struct list_head *list, bool run_queue_async)
359 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
360 struct elevator_queue *e = hctx->queue->elevator;
363 struct request *rq, *next;
366 * We bypass requests that already have a driver tag assigned,
367 * which should only be flushes. Flushes are only ever inserted
368 * as single requests, so we shouldn't ever hit the
369 * WARN_ON_ONCE() below (but let's handle it just in case).
371 list_for_each_entry_safe(rq, next, list, queuelist) {
372 if (WARN_ON_ONCE(rq->tag != -1)) {
373 list_del_init(&rq->queuelist);
374 blk_mq_sched_bypass_insert(hctx, rq);
379 if (e && e->type->ops.mq.insert_requests)
380 e->type->ops.mq.insert_requests(hctx, list, false);
382 blk_mq_insert_requests(hctx, ctx, list);
384 blk_mq_run_hw_queue(hctx, run_queue_async);
387 static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
388 struct blk_mq_hw_ctx *hctx,
389 unsigned int hctx_idx)
391 if (hctx->sched_tags) {
392 blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx);
393 blk_mq_free_rq_map(hctx->sched_tags);
394 hctx->sched_tags = NULL;
398 static int blk_mq_sched_alloc_tags(struct request_queue *q,
399 struct blk_mq_hw_ctx *hctx,
400 unsigned int hctx_idx)
402 struct blk_mq_tag_set *set = q->tag_set;
405 hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
407 if (!hctx->sched_tags)
410 ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests);
412 blk_mq_sched_free_tags(set, hctx, hctx_idx);
417 static void blk_mq_sched_tags_teardown(struct request_queue *q)
419 struct blk_mq_tag_set *set = q->tag_set;
420 struct blk_mq_hw_ctx *hctx;
423 queue_for_each_hw_ctx(q, hctx, i)
424 blk_mq_sched_free_tags(set, hctx, i);
427 int blk_mq_sched_init_hctx(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
428 unsigned int hctx_idx)
430 struct elevator_queue *e = q->elevator;
436 ret = blk_mq_sched_alloc_tags(q, hctx, hctx_idx);
440 if (e->type->ops.mq.init_hctx) {
441 ret = e->type->ops.mq.init_hctx(hctx, hctx_idx);
443 blk_mq_sched_free_tags(q->tag_set, hctx, hctx_idx);
448 blk_mq_debugfs_register_sched_hctx(q, hctx);
453 void blk_mq_sched_exit_hctx(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
454 unsigned int hctx_idx)
456 struct elevator_queue *e = q->elevator;
461 blk_mq_debugfs_unregister_sched_hctx(hctx);
463 if (e->type->ops.mq.exit_hctx && hctx->sched_data) {
464 e->type->ops.mq.exit_hctx(hctx, hctx_idx);
465 hctx->sched_data = NULL;
468 blk_mq_sched_free_tags(q->tag_set, hctx, hctx_idx);
471 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
473 struct blk_mq_hw_ctx *hctx;
474 struct elevator_queue *eq;
484 * Default to 256, since we don't split into sync/async like the
485 * old code did. Additionally, this is a per-hw queue depth.
487 q->nr_requests = 2 * BLKDEV_MAX_RQ;
489 queue_for_each_hw_ctx(q, hctx, i) {
490 ret = blk_mq_sched_alloc_tags(q, hctx, i);
495 ret = e->ops.mq.init_sched(q, e);
499 blk_mq_debugfs_register_sched(q);
501 queue_for_each_hw_ctx(q, hctx, i) {
502 if (e->ops.mq.init_hctx) {
503 ret = e->ops.mq.init_hctx(hctx, i);
506 blk_mq_exit_sched(q, eq);
507 kobject_put(&eq->kobj);
511 blk_mq_debugfs_register_sched_hctx(q, hctx);
517 blk_mq_sched_tags_teardown(q);
522 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
524 struct blk_mq_hw_ctx *hctx;
527 queue_for_each_hw_ctx(q, hctx, i) {
528 blk_mq_debugfs_unregister_sched_hctx(hctx);
529 if (e->type->ops.mq.exit_hctx && hctx->sched_data) {
530 e->type->ops.mq.exit_hctx(hctx, i);
531 hctx->sched_data = NULL;
534 blk_mq_debugfs_unregister_sched(q);
535 if (e->type->ops.mq.exit_sched)
536 e->type->ops.mq.exit_sched(e);
537 blk_mq_sched_tags_teardown(q);
541 int blk_mq_sched_init(struct request_queue *q)
545 mutex_lock(&q->sysfs_lock);
546 ret = elevator_init(q, NULL);
547 mutex_unlock(&q->sysfs_lock);