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 set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
65 void blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx)
67 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
69 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
71 blk_mq_run_hw_queue(hctx, true);
75 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
76 * its queue by itself in its completion handler, so we don't need to
77 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
79 static void blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
81 struct request_queue *q = hctx->queue;
82 struct elevator_queue *e = q->elevator;
88 if (e->type->ops.mq.has_work &&
89 !e->type->ops.mq.has_work(hctx))
92 if (!blk_mq_get_dispatch_budget(hctx))
95 rq = e->type->ops.mq.dispatch_request(hctx);
97 blk_mq_put_dispatch_budget(hctx);
102 * Now this rq owns the budget which has to be released
103 * if this rq won't be queued to driver via .queue_rq()
104 * in blk_mq_dispatch_rq_list().
106 list_add(&rq->queuelist, &rq_list);
107 } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
110 static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
111 struct blk_mq_ctx *ctx)
113 unsigned idx = ctx->index_hw;
115 if (++idx == hctx->nr_ctx)
118 return hctx->ctxs[idx];
122 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
123 * its queue by itself in its completion handler, so we don't need to
124 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
126 static void blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
128 struct request_queue *q = hctx->queue;
130 struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from);
135 if (!sbitmap_any_bit_set(&hctx->ctx_map))
138 if (!blk_mq_get_dispatch_budget(hctx))
141 rq = blk_mq_dequeue_from_ctx(hctx, ctx);
143 blk_mq_put_dispatch_budget(hctx);
148 * Now this rq owns the budget which has to be released
149 * if this rq won't be queued to driver via .queue_rq()
150 * in blk_mq_dispatch_rq_list().
152 list_add(&rq->queuelist, &rq_list);
154 /* round robin for fair dispatch */
155 ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
157 } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
159 WRITE_ONCE(hctx->dispatch_from, ctx);
162 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
164 struct request_queue *q = hctx->queue;
165 struct elevator_queue *e = q->elevator;
166 const bool has_sched_dispatch = e && e->type->ops.mq.dispatch_request;
169 /* RCU or SRCU read lock is needed before checking quiesced flag */
170 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
176 * If we have previous entries on our dispatch list, grab them first for
177 * more fair dispatch.
179 if (!list_empty_careful(&hctx->dispatch)) {
180 spin_lock(&hctx->lock);
181 if (!list_empty(&hctx->dispatch))
182 list_splice_init(&hctx->dispatch, &rq_list);
183 spin_unlock(&hctx->lock);
187 * Only ask the scheduler for requests, if we didn't have residual
188 * requests from the dispatch list. This is to avoid the case where
189 * we only ever dispatch a fraction of the requests available because
190 * of low device queue depth. Once we pull requests out of the IO
191 * scheduler, we can no longer merge or sort them. So it's best to
192 * leave them there for as long as we can. Mark the hw queue as
193 * needing a restart in that case.
195 * We want to dispatch from the scheduler if there was nothing
196 * on the dispatch list or we were able to dispatch from the
199 if (!list_empty(&rq_list)) {
200 blk_mq_sched_mark_restart_hctx(hctx);
201 if (blk_mq_dispatch_rq_list(q, &rq_list, false)) {
202 if (has_sched_dispatch)
203 blk_mq_do_dispatch_sched(hctx);
205 blk_mq_do_dispatch_ctx(hctx);
207 } else if (has_sched_dispatch) {
208 blk_mq_do_dispatch_sched(hctx);
209 } else if (hctx->dispatch_busy) {
210 /* dequeue request one by one from sw queue if queue is busy */
211 blk_mq_do_dispatch_ctx(hctx);
213 blk_mq_flush_busy_ctxs(hctx, &rq_list);
214 blk_mq_dispatch_rq_list(q, &rq_list, false);
218 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
219 struct request **merged_request)
223 switch (elv_merge(q, &rq, bio)) {
224 case ELEVATOR_BACK_MERGE:
225 if (!blk_mq_sched_allow_merge(q, rq, bio))
227 if (!bio_attempt_back_merge(q, rq, bio))
229 *merged_request = attempt_back_merge(q, rq);
230 if (!*merged_request)
231 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
233 case ELEVATOR_FRONT_MERGE:
234 if (!blk_mq_sched_allow_merge(q, rq, bio))
236 if (!bio_attempt_front_merge(q, rq, bio))
238 *merged_request = attempt_front_merge(q, rq);
239 if (!*merged_request)
240 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
242 case ELEVATOR_DISCARD_MERGE:
243 return bio_attempt_discard_merge(q, rq, bio);
248 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
251 * Iterate list of requests and see if we can merge this bio with any
254 bool blk_mq_bio_list_merge(struct request_queue *q, struct list_head *list,
260 list_for_each_entry_reverse(rq, list, queuelist) {
266 if (!blk_rq_merge_ok(rq, bio))
269 switch (blk_try_merge(rq, bio)) {
270 case ELEVATOR_BACK_MERGE:
271 if (blk_mq_sched_allow_merge(q, rq, bio))
272 merged = bio_attempt_back_merge(q, rq, bio);
274 case ELEVATOR_FRONT_MERGE:
275 if (blk_mq_sched_allow_merge(q, rq, bio))
276 merged = bio_attempt_front_merge(q, rq, bio);
278 case ELEVATOR_DISCARD_MERGE:
279 merged = bio_attempt_discard_merge(q, rq, bio);
290 EXPORT_SYMBOL_GPL(blk_mq_bio_list_merge);
293 * Reverse check our software queue for entries that we could potentially
294 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
295 * too much time checking for merges.
297 static bool blk_mq_attempt_merge(struct request_queue *q,
298 struct blk_mq_ctx *ctx, struct bio *bio)
300 lockdep_assert_held(&ctx->lock);
302 if (blk_mq_bio_list_merge(q, &ctx->rq_list, bio)) {
310 bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio)
312 struct elevator_queue *e = q->elevator;
313 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
314 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
317 if (e && e->type->ops.mq.bio_merge) {
319 return e->type->ops.mq.bio_merge(hctx, bio);
322 if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
323 !list_empty_careful(&ctx->rq_list)) {
324 /* default per sw-queue merge */
325 spin_lock(&ctx->lock);
326 ret = blk_mq_attempt_merge(q, ctx, bio);
327 spin_unlock(&ctx->lock);
334 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
336 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
338 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
340 void blk_mq_sched_request_inserted(struct request *rq)
342 trace_block_rq_insert(rq->q, rq);
344 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
346 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
350 /* dispatch flush rq directly */
351 if (rq->rq_flags & RQF_FLUSH_SEQ) {
352 spin_lock(&hctx->lock);
353 list_add(&rq->queuelist, &hctx->dispatch);
354 spin_unlock(&hctx->lock);
359 rq->rq_flags |= RQF_SORTED;
364 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
365 bool run_queue, bool async)
367 struct request_queue *q = rq->q;
368 struct elevator_queue *e = q->elevator;
369 struct blk_mq_ctx *ctx = rq->mq_ctx;
370 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
372 /* flush rq in flush machinery need to be dispatched directly */
373 if (!(rq->rq_flags & RQF_FLUSH_SEQ) && op_is_flush(rq->cmd_flags)) {
374 blk_insert_flush(rq);
378 WARN_ON(e && (rq->tag != -1));
380 if (blk_mq_sched_bypass_insert(hctx, !!e, rq))
383 if (e && e->type->ops.mq.insert_requests) {
386 list_add(&rq->queuelist, &list);
387 e->type->ops.mq.insert_requests(hctx, &list, at_head);
389 spin_lock(&ctx->lock);
390 __blk_mq_insert_request(hctx, rq, at_head);
391 spin_unlock(&ctx->lock);
396 blk_mq_run_hw_queue(hctx, async);
399 void blk_mq_sched_insert_requests(struct request_queue *q,
400 struct blk_mq_ctx *ctx,
401 struct list_head *list, bool run_queue_async)
403 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
404 struct elevator_queue *e = hctx->queue->elevator;
406 if (e && e->type->ops.mq.insert_requests)
407 e->type->ops.mq.insert_requests(hctx, list, false);
410 * try to issue requests directly if the hw queue isn't
411 * busy in case of 'none' scheduler, and this way may save
412 * us one extra enqueue & dequeue to sw queue.
414 if (!hctx->dispatch_busy && !e && !run_queue_async) {
415 blk_mq_try_issue_list_directly(hctx, list);
416 if (list_empty(list))
419 blk_mq_insert_requests(hctx, ctx, list);
422 blk_mq_run_hw_queue(hctx, run_queue_async);
425 static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
426 struct blk_mq_hw_ctx *hctx,
427 unsigned int hctx_idx)
429 if (hctx->sched_tags) {
430 blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx);
431 blk_mq_free_rq_map(hctx->sched_tags);
432 hctx->sched_tags = NULL;
436 static int blk_mq_sched_alloc_tags(struct request_queue *q,
437 struct blk_mq_hw_ctx *hctx,
438 unsigned int hctx_idx)
440 struct blk_mq_tag_set *set = q->tag_set;
443 hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
445 if (!hctx->sched_tags)
448 ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests);
450 blk_mq_sched_free_tags(set, hctx, hctx_idx);
455 static void blk_mq_sched_tags_teardown(struct request_queue *q)
457 struct blk_mq_tag_set *set = q->tag_set;
458 struct blk_mq_hw_ctx *hctx;
461 queue_for_each_hw_ctx(q, hctx, i)
462 blk_mq_sched_free_tags(set, hctx, i);
465 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
467 struct blk_mq_hw_ctx *hctx;
468 struct elevator_queue *eq;
474 q->nr_requests = q->tag_set->queue_depth;
479 * Default to double of smaller one between hw queue_depth and 128,
480 * since we don't split into sync/async like the old code did.
481 * Additionally, this is a per-hw queue depth.
483 q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
486 queue_for_each_hw_ctx(q, hctx, i) {
487 ret = blk_mq_sched_alloc_tags(q, hctx, i);
492 ret = e->ops.mq.init_sched(q, e);
496 blk_mq_debugfs_register_sched(q);
498 queue_for_each_hw_ctx(q, hctx, i) {
499 if (e->ops.mq.init_hctx) {
500 ret = e->ops.mq.init_hctx(hctx, i);
503 blk_mq_exit_sched(q, eq);
504 kobject_put(&eq->kobj);
508 blk_mq_debugfs_register_sched_hctx(q, hctx);
514 blk_mq_sched_tags_teardown(q);
519 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
521 struct blk_mq_hw_ctx *hctx;
524 queue_for_each_hw_ctx(q, hctx, i) {
525 blk_mq_debugfs_unregister_sched_hctx(hctx);
526 if (e->type->ops.mq.exit_hctx && hctx->sched_data) {
527 e->type->ops.mq.exit_hctx(hctx, i);
528 hctx->sched_data = NULL;
531 blk_mq_debugfs_unregister_sched(q);
532 if (e->type->ops.mq.exit_sched)
533 e->type->ops.mq.exit_sched(e);
534 blk_mq_sched_tags_teardown(q);