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-sched.h"
15 #include "blk-mq-tag.h"
18 void blk_mq_sched_free_hctx_data(struct request_queue *q,
19 void (*exit)(struct blk_mq_hw_ctx *))
21 struct blk_mq_hw_ctx *hctx;
24 queue_for_each_hw_ctx(q, hctx, i) {
25 if (exit && hctx->sched_data)
27 kfree(hctx->sched_data);
28 hctx->sched_data = NULL;
31 EXPORT_SYMBOL_GPL(blk_mq_sched_free_hctx_data);
33 static void __blk_mq_sched_assign_ioc(struct request_queue *q,
36 struct io_context *ioc)
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);
51 if (!blk_mq_sched_get_rq_priv(q, rq, bio)) {
52 rq->rq_flags |= RQF_ELVPRIV;
53 get_io_context(icq->ioc);
60 static void blk_mq_sched_assign_ioc(struct request_queue *q,
61 struct request *rq, struct bio *bio)
63 struct io_context *ioc;
67 __blk_mq_sched_assign_ioc(q, rq, bio, ioc);
70 struct request *blk_mq_sched_get_request(struct request_queue *q,
73 struct blk_mq_alloc_data *data)
75 struct elevator_queue *e = q->elevator;
78 blk_queue_enter_live(q);
80 if (likely(!data->ctx))
81 data->ctx = blk_mq_get_ctx(q);
82 if (likely(!data->hctx))
83 data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
86 * For a reserved tag, allocate a normal request since we might
87 * have driver dependencies on the value of the internal tag.
89 if (e && !(data->flags & BLK_MQ_REQ_RESERVED)) {
90 data->flags |= BLK_MQ_REQ_INTERNAL;
93 * Flush requests are special and go directly to the
96 if (!op_is_flush(op) && e->type->ops.mq.get_request) {
97 rq = e->type->ops.mq.get_request(q, op, data);
99 rq->rq_flags |= RQF_QUEUED;
101 rq = __blk_mq_alloc_request(data, op);
103 rq = __blk_mq_alloc_request(data, op);
107 if (!op_is_flush(op)) {
109 if (e && e->type->icq_cache)
110 blk_mq_sched_assign_ioc(q, rq, bio);
112 data->hctx->queued++;
120 void blk_mq_sched_put_request(struct request *rq)
122 struct request_queue *q = rq->q;
123 struct elevator_queue *e = q->elevator;
125 if (rq->rq_flags & RQF_ELVPRIV) {
126 blk_mq_sched_put_rq_priv(rq->q, rq);
128 put_io_context(rq->elv.icq->ioc);
133 if ((rq->rq_flags & RQF_QUEUED) && e && e->type->ops.mq.put_request)
134 e->type->ops.mq.put_request(rq);
136 blk_mq_finish_request(rq);
139 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
141 struct request_queue *q = hctx->queue;
142 struct elevator_queue *e = q->elevator;
143 const bool has_sched_dispatch = e && e->type->ops.mq.dispatch_request;
144 bool did_work = false;
147 if (unlikely(blk_mq_hctx_stopped(hctx)))
153 * If we have previous entries on our dispatch list, grab them first for
154 * more fair dispatch.
156 if (!list_empty_careful(&hctx->dispatch)) {
157 spin_lock(&hctx->lock);
158 if (!list_empty(&hctx->dispatch))
159 list_splice_init(&hctx->dispatch, &rq_list);
160 spin_unlock(&hctx->lock);
164 * Only ask the scheduler for requests, if we didn't have residual
165 * requests from the dispatch list. This is to avoid the case where
166 * we only ever dispatch a fraction of the requests available because
167 * of low device queue depth. Once we pull requests out of the IO
168 * scheduler, we can no longer merge or sort them. So it's best to
169 * leave them there for as long as we can. Mark the hw queue as
170 * needing a restart in that case.
172 if (!list_empty(&rq_list)) {
173 blk_mq_sched_mark_restart_hctx(hctx);
174 did_work = blk_mq_dispatch_rq_list(q, &rq_list);
175 } else if (!has_sched_dispatch) {
176 blk_mq_flush_busy_ctxs(hctx, &rq_list);
177 blk_mq_dispatch_rq_list(q, &rq_list);
181 * We want to dispatch from the scheduler if we had no work left
182 * on the dispatch list, OR if we did have work but weren't able
185 if (!did_work && has_sched_dispatch) {
189 rq = e->type->ops.mq.dispatch_request(hctx);
192 list_add(&rq->queuelist, &rq_list);
193 } while (blk_mq_dispatch_rq_list(q, &rq_list));
197 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
198 struct request **merged_request)
202 switch (elv_merge(q, &rq, bio)) {
203 case ELEVATOR_BACK_MERGE:
204 if (!blk_mq_sched_allow_merge(q, rq, bio))
206 if (!bio_attempt_back_merge(q, rq, bio))
208 *merged_request = attempt_back_merge(q, rq);
209 if (!*merged_request)
210 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
212 case ELEVATOR_FRONT_MERGE:
213 if (!blk_mq_sched_allow_merge(q, rq, bio))
215 if (!bio_attempt_front_merge(q, rq, bio))
217 *merged_request = attempt_front_merge(q, rq);
218 if (!*merged_request)
219 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
225 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
227 bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio)
229 struct elevator_queue *e = q->elevator;
231 if (e->type->ops.mq.bio_merge) {
232 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
233 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
236 return e->type->ops.mq.bio_merge(hctx, bio);
242 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
244 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
246 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
248 void blk_mq_sched_request_inserted(struct request *rq)
250 trace_block_rq_insert(rq->q, rq);
252 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
254 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
258 rq->rq_flags |= RQF_SORTED;
263 * If we already have a real request tag, send directly to
266 spin_lock(&hctx->lock);
267 list_add(&rq->queuelist, &hctx->dispatch);
268 spin_unlock(&hctx->lock);
272 static bool blk_mq_sched_restart_hctx(struct blk_mq_hw_ctx *hctx)
274 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) {
275 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
276 if (blk_mq_hctx_has_pending(hctx)) {
277 blk_mq_run_hw_queue(hctx, true);
285 * list_for_each_entry_rcu_rr - iterate in a round-robin fashion over rcu list
287 * @skip: the list element that will not be examined. Iteration starts at
289 * @head: head of the list to examine. This list must have at least one
290 * element, namely @skip.
291 * @member: name of the list_head structure within typeof(*pos).
293 #define list_for_each_entry_rcu_rr(pos, skip, head, member) \
294 for ((pos) = (skip); \
295 (pos = (pos)->member.next != (head) ? list_entry_rcu( \
296 (pos)->member.next, typeof(*pos), member) : \
297 list_entry_rcu((pos)->member.next->next, typeof(*pos), member)), \
301 * Called after a driver tag has been freed to check whether a hctx needs to
302 * be restarted. Restarts @hctx if its tag set is not shared. Restarts hardware
303 * queues in a round-robin fashion if the tag set of @hctx is shared with other
306 void blk_mq_sched_restart(struct blk_mq_hw_ctx *const hctx)
308 struct blk_mq_tags *const tags = hctx->tags;
309 struct blk_mq_tag_set *const set = hctx->queue->tag_set;
310 struct request_queue *const queue = hctx->queue, *q;
311 struct blk_mq_hw_ctx *hctx2;
314 if (set->flags & BLK_MQ_F_TAG_SHARED) {
316 list_for_each_entry_rcu_rr(q, queue, &set->tag_list,
318 queue_for_each_hw_ctx(q, hctx2, i)
319 if (hctx2->tags == tags &&
320 blk_mq_sched_restart_hctx(hctx2))
323 j = hctx->queue_num + 1;
324 for (i = 0; i < queue->nr_hw_queues; i++, j++) {
325 if (j == queue->nr_hw_queues)
327 hctx2 = queue->queue_hw_ctx[j];
328 if (hctx2->tags == tags &&
329 blk_mq_sched_restart_hctx(hctx2))
335 blk_mq_sched_restart_hctx(hctx);
340 * Add flush/fua to the queue. If we fail getting a driver tag, then
341 * punt to the requeue list. Requeue will re-invoke us from a context
342 * that's safe to block from.
344 static void blk_mq_sched_insert_flush(struct blk_mq_hw_ctx *hctx,
345 struct request *rq, bool can_block)
347 if (blk_mq_get_driver_tag(rq, &hctx, can_block)) {
348 blk_insert_flush(rq);
349 blk_mq_run_hw_queue(hctx, true);
351 blk_mq_add_to_requeue_list(rq, false, true);
354 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
355 bool run_queue, bool async, bool can_block)
357 struct request_queue *q = rq->q;
358 struct elevator_queue *e = q->elevator;
359 struct blk_mq_ctx *ctx = rq->mq_ctx;
360 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
362 if (rq->tag == -1 && op_is_flush(rq->cmd_flags)) {
363 blk_mq_sched_insert_flush(hctx, rq, can_block);
367 if (e && blk_mq_sched_bypass_insert(hctx, rq))
370 if (e && e->type->ops.mq.insert_requests) {
373 list_add(&rq->queuelist, &list);
374 e->type->ops.mq.insert_requests(hctx, &list, at_head);
376 spin_lock(&ctx->lock);
377 __blk_mq_insert_request(hctx, rq, at_head);
378 spin_unlock(&ctx->lock);
383 blk_mq_run_hw_queue(hctx, async);
386 void blk_mq_sched_insert_requests(struct request_queue *q,
387 struct blk_mq_ctx *ctx,
388 struct list_head *list, bool run_queue_async)
390 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
391 struct elevator_queue *e = hctx->queue->elevator;
394 struct request *rq, *next;
397 * We bypass requests that already have a driver tag assigned,
398 * which should only be flushes. Flushes are only ever inserted
399 * as single requests, so we shouldn't ever hit the
400 * WARN_ON_ONCE() below (but let's handle it just in case).
402 list_for_each_entry_safe(rq, next, list, queuelist) {
403 if (WARN_ON_ONCE(rq->tag != -1)) {
404 list_del_init(&rq->queuelist);
405 blk_mq_sched_bypass_insert(hctx, rq);
410 if (e && e->type->ops.mq.insert_requests)
411 e->type->ops.mq.insert_requests(hctx, list, false);
413 blk_mq_insert_requests(hctx, ctx, list);
415 blk_mq_run_hw_queue(hctx, run_queue_async);
418 static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
419 struct blk_mq_hw_ctx *hctx,
420 unsigned int hctx_idx)
422 if (hctx->sched_tags) {
423 blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx);
424 blk_mq_free_rq_map(hctx->sched_tags);
425 hctx->sched_tags = NULL;
429 static int blk_mq_sched_alloc_tags(struct request_queue *q,
430 struct blk_mq_hw_ctx *hctx,
431 unsigned int hctx_idx)
433 struct blk_mq_tag_set *set = q->tag_set;
436 hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
438 if (!hctx->sched_tags)
441 ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests);
443 blk_mq_sched_free_tags(set, hctx, hctx_idx);
448 static void blk_mq_sched_tags_teardown(struct request_queue *q)
450 struct blk_mq_tag_set *set = q->tag_set;
451 struct blk_mq_hw_ctx *hctx;
454 queue_for_each_hw_ctx(q, hctx, i)
455 blk_mq_sched_free_tags(set, hctx, i);
458 int blk_mq_sched_init_hctx(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
459 unsigned int hctx_idx)
461 struct elevator_queue *e = q->elevator;
467 ret = blk_mq_sched_alloc_tags(q, hctx, hctx_idx);
471 if (e->type->ops.mq.init_hctx) {
472 ret = e->type->ops.mq.init_hctx(hctx, hctx_idx);
474 blk_mq_sched_free_tags(q->tag_set, hctx, hctx_idx);
482 void blk_mq_sched_exit_hctx(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
483 unsigned int hctx_idx)
485 struct elevator_queue *e = q->elevator;
490 if (e->type->ops.mq.exit_hctx && hctx->sched_data) {
491 e->type->ops.mq.exit_hctx(hctx, hctx_idx);
492 hctx->sched_data = NULL;
495 blk_mq_sched_free_tags(q->tag_set, hctx, hctx_idx);
498 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
500 struct blk_mq_hw_ctx *hctx;
501 struct elevator_queue *eq;
511 * Default to 256, since we don't split into sync/async like the
512 * old code did. Additionally, this is a per-hw queue depth.
514 q->nr_requests = 2 * BLKDEV_MAX_RQ;
516 queue_for_each_hw_ctx(q, hctx, i) {
517 ret = blk_mq_sched_alloc_tags(q, hctx, i);
522 ret = e->ops.mq.init_sched(q, e);
526 if (e->ops.mq.init_hctx) {
527 queue_for_each_hw_ctx(q, hctx, i) {
528 ret = e->ops.mq.init_hctx(hctx, i);
531 blk_mq_exit_sched(q, eq);
532 kobject_put(&eq->kobj);
541 blk_mq_sched_tags_teardown(q);
546 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
548 struct blk_mq_hw_ctx *hctx;
551 if (e->type->ops.mq.exit_hctx) {
552 queue_for_each_hw_ctx(q, hctx, i) {
553 if (hctx->sched_data) {
554 e->type->ops.mq.exit_hctx(hctx, i);
555 hctx->sched_data = NULL;
559 if (e->type->ops.mq.exit_sched)
560 e->type->ops.mq.exit_sched(e);
561 blk_mq_sched_tags_teardown(q);
565 int blk_mq_sched_init(struct request_queue *q)
569 mutex_lock(&q->sysfs_lock);
570 ret = elevator_init(q, NULL);
571 mutex_unlock(&q->sysfs_lock);