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 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);
64 EXPORT_SYMBOL_GPL(blk_mq_sched_mark_restart_hctx);
66 void blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx)
68 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
70 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
73 * Order clearing SCHED_RESTART and list_empty_careful(&hctx->dispatch)
74 * in blk_mq_run_hw_queue(). Its pair is the barrier in
75 * blk_mq_dispatch_rq_list(). So dispatch code won't see SCHED_RESTART,
76 * meantime new request added to hctx->dispatch is missed to check in
77 * blk_mq_run_hw_queue().
81 blk_mq_run_hw_queue(hctx, true);
85 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
86 * its queue by itself in its completion handler, so we don't need to
87 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
89 static void blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
91 struct request_queue *q = hctx->queue;
92 struct elevator_queue *e = q->elevator;
98 if (e->type->ops.mq.has_work &&
99 !e->type->ops.mq.has_work(hctx))
102 if (!blk_mq_get_dispatch_budget(hctx))
105 rq = e->type->ops.mq.dispatch_request(hctx);
107 blk_mq_put_dispatch_budget(hctx);
112 * Now this rq owns the budget which has to be released
113 * if this rq won't be queued to driver via .queue_rq()
114 * in blk_mq_dispatch_rq_list().
116 list_add(&rq->queuelist, &rq_list);
117 } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
120 static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
121 struct blk_mq_ctx *ctx)
123 unsigned idx = ctx->index_hw;
125 if (++idx == hctx->nr_ctx)
128 return hctx->ctxs[idx];
132 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
133 * its queue by itself in its completion handler, so we don't need to
134 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
136 static void blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
138 struct request_queue *q = hctx->queue;
140 struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from);
145 if (!sbitmap_any_bit_set(&hctx->ctx_map))
148 if (!blk_mq_get_dispatch_budget(hctx))
151 rq = blk_mq_dequeue_from_ctx(hctx, ctx);
153 blk_mq_put_dispatch_budget(hctx);
158 * Now this rq owns the budget which has to be released
159 * if this rq won't be queued to driver via .queue_rq()
160 * in blk_mq_dispatch_rq_list().
162 list_add(&rq->queuelist, &rq_list);
164 /* round robin for fair dispatch */
165 ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
167 } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
169 WRITE_ONCE(hctx->dispatch_from, ctx);
172 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
174 struct request_queue *q = hctx->queue;
175 struct elevator_queue *e = q->elevator;
176 const bool has_sched_dispatch = e && e->type->ops.mq.dispatch_request;
179 /* RCU or SRCU read lock is needed before checking quiesced flag */
180 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
186 * If we have previous entries on our dispatch list, grab them first for
187 * more fair dispatch.
189 if (!list_empty_careful(&hctx->dispatch)) {
190 spin_lock(&hctx->lock);
191 if (!list_empty(&hctx->dispatch))
192 list_splice_init(&hctx->dispatch, &rq_list);
193 spin_unlock(&hctx->lock);
197 * Only ask the scheduler for requests, if we didn't have residual
198 * requests from the dispatch list. This is to avoid the case where
199 * we only ever dispatch a fraction of the requests available because
200 * of low device queue depth. Once we pull requests out of the IO
201 * scheduler, we can no longer merge or sort them. So it's best to
202 * leave them there for as long as we can. Mark the hw queue as
203 * needing a restart in that case.
205 * We want to dispatch from the scheduler if there was nothing
206 * on the dispatch list or we were able to dispatch from the
209 if (!list_empty(&rq_list)) {
210 blk_mq_sched_mark_restart_hctx(hctx);
211 if (blk_mq_dispatch_rq_list(q, &rq_list, false)) {
212 if (has_sched_dispatch)
213 blk_mq_do_dispatch_sched(hctx);
215 blk_mq_do_dispatch_ctx(hctx);
217 } else if (has_sched_dispatch) {
218 blk_mq_do_dispatch_sched(hctx);
219 } else if (hctx->dispatch_busy) {
220 /* dequeue request one by one from sw queue if queue is busy */
221 blk_mq_do_dispatch_ctx(hctx);
223 blk_mq_flush_busy_ctxs(hctx, &rq_list);
224 blk_mq_dispatch_rq_list(q, &rq_list, false);
228 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
229 struct request **merged_request)
233 switch (elv_merge(q, &rq, bio)) {
234 case ELEVATOR_BACK_MERGE:
235 if (!blk_mq_sched_allow_merge(q, rq, bio))
237 if (!bio_attempt_back_merge(q, rq, bio))
239 *merged_request = attempt_back_merge(q, rq);
240 if (!*merged_request)
241 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
243 case ELEVATOR_FRONT_MERGE:
244 if (!blk_mq_sched_allow_merge(q, rq, bio))
246 if (!bio_attempt_front_merge(q, rq, bio))
248 *merged_request = attempt_front_merge(q, rq);
249 if (!*merged_request)
250 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
252 case ELEVATOR_DISCARD_MERGE:
253 return bio_attempt_discard_merge(q, rq, bio);
258 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
261 * Iterate list of requests and see if we can merge this bio with any
264 bool blk_mq_bio_list_merge(struct request_queue *q, struct list_head *list,
270 list_for_each_entry_reverse(rq, list, queuelist) {
276 if (!blk_rq_merge_ok(rq, bio))
279 switch (blk_try_merge(rq, bio)) {
280 case ELEVATOR_BACK_MERGE:
281 if (blk_mq_sched_allow_merge(q, rq, bio))
282 merged = bio_attempt_back_merge(q, rq, bio);
284 case ELEVATOR_FRONT_MERGE:
285 if (blk_mq_sched_allow_merge(q, rq, bio))
286 merged = bio_attempt_front_merge(q, rq, bio);
288 case ELEVATOR_DISCARD_MERGE:
289 merged = bio_attempt_discard_merge(q, rq, bio);
300 EXPORT_SYMBOL_GPL(blk_mq_bio_list_merge);
303 * Reverse check our software queue for entries that we could potentially
304 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
305 * too much time checking for merges.
307 static bool blk_mq_attempt_merge(struct request_queue *q,
308 struct blk_mq_ctx *ctx, struct bio *bio)
310 lockdep_assert_held(&ctx->lock);
312 if (blk_mq_bio_list_merge(q, &ctx->rq_list, bio)) {
320 bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio)
322 struct elevator_queue *e = q->elevator;
323 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
324 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
327 if (e && e->type->ops.mq.bio_merge) {
329 return e->type->ops.mq.bio_merge(hctx, bio);
332 if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
333 !list_empty_careful(&ctx->rq_list)) {
334 /* default per sw-queue merge */
335 spin_lock(&ctx->lock);
336 ret = blk_mq_attempt_merge(q, ctx, bio);
337 spin_unlock(&ctx->lock);
344 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
346 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
348 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
350 void blk_mq_sched_request_inserted(struct request *rq)
352 trace_block_rq_insert(rq->q, rq);
354 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
356 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
360 /* dispatch flush rq directly */
361 if (rq->rq_flags & RQF_FLUSH_SEQ) {
362 spin_lock(&hctx->lock);
363 list_add(&rq->queuelist, &hctx->dispatch);
364 spin_unlock(&hctx->lock);
369 rq->rq_flags |= RQF_SORTED;
374 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
375 bool run_queue, bool async)
377 struct request_queue *q = rq->q;
378 struct elevator_queue *e = q->elevator;
379 struct blk_mq_ctx *ctx = rq->mq_ctx;
380 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
382 /* flush rq in flush machinery need to be dispatched directly */
383 if (!(rq->rq_flags & RQF_FLUSH_SEQ) && op_is_flush(rq->cmd_flags)) {
384 blk_insert_flush(rq);
388 WARN_ON(e && (rq->tag != -1));
390 if (blk_mq_sched_bypass_insert(hctx, !!e, rq))
393 if (e && e->type->ops.mq.insert_requests) {
396 list_add(&rq->queuelist, &list);
397 e->type->ops.mq.insert_requests(hctx, &list, at_head);
399 spin_lock(&ctx->lock);
400 __blk_mq_insert_request(hctx, rq, at_head);
401 spin_unlock(&ctx->lock);
406 blk_mq_run_hw_queue(hctx, async);
409 void blk_mq_sched_insert_requests(struct request_queue *q,
410 struct blk_mq_ctx *ctx,
411 struct list_head *list, bool run_queue_async)
413 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
414 struct elevator_queue *e = hctx->queue->elevator;
416 if (e && e->type->ops.mq.insert_requests)
417 e->type->ops.mq.insert_requests(hctx, list, false);
420 * try to issue requests directly if the hw queue isn't
421 * busy in case of 'none' scheduler, and this way may save
422 * us one extra enqueue & dequeue to sw queue.
424 if (!hctx->dispatch_busy && !e && !run_queue_async) {
425 blk_mq_try_issue_list_directly(hctx, list);
426 if (list_empty(list))
429 blk_mq_insert_requests(hctx, ctx, list);
432 blk_mq_run_hw_queue(hctx, run_queue_async);
435 static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
436 struct blk_mq_hw_ctx *hctx,
437 unsigned int hctx_idx)
439 if (hctx->sched_tags) {
440 blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx);
441 blk_mq_free_rq_map(hctx->sched_tags);
442 hctx->sched_tags = NULL;
446 static int blk_mq_sched_alloc_tags(struct request_queue *q,
447 struct blk_mq_hw_ctx *hctx,
448 unsigned int hctx_idx)
450 struct blk_mq_tag_set *set = q->tag_set;
453 hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
455 if (!hctx->sched_tags)
458 ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests);
460 blk_mq_sched_free_tags(set, hctx, hctx_idx);
465 static void blk_mq_sched_tags_teardown(struct request_queue *q)
467 struct blk_mq_tag_set *set = q->tag_set;
468 struct blk_mq_hw_ctx *hctx;
471 queue_for_each_hw_ctx(q, hctx, i)
472 blk_mq_sched_free_tags(set, hctx, i);
475 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
477 struct blk_mq_hw_ctx *hctx;
478 struct elevator_queue *eq;
484 q->nr_requests = q->tag_set->queue_depth;
489 * Default to double of smaller one between hw queue_depth and 128,
490 * since we don't split into sync/async like the old code did.
491 * Additionally, this is a per-hw queue depth.
493 q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
496 queue_for_each_hw_ctx(q, hctx, i) {
497 ret = blk_mq_sched_alloc_tags(q, hctx, i);
502 ret = e->ops.mq.init_sched(q, e);
506 blk_mq_debugfs_register_sched(q);
508 queue_for_each_hw_ctx(q, hctx, i) {
509 if (e->ops.mq.init_hctx) {
510 ret = e->ops.mq.init_hctx(hctx, i);
513 blk_mq_exit_sched(q, eq);
514 kobject_put(&eq->kobj);
518 blk_mq_debugfs_register_sched_hctx(q, hctx);
524 blk_mq_sched_tags_teardown(q);
529 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
531 struct blk_mq_hw_ctx *hctx;
534 queue_for_each_hw_ctx(q, hctx, i) {
535 blk_mq_debugfs_unregister_sched_hctx(hctx);
536 if (e->type->ops.mq.exit_hctx && hctx->sched_data) {
537 e->type->ops.mq.exit_hctx(hctx, i);
538 hctx->sched_data = NULL;
541 blk_mq_debugfs_unregister_sched(q);
542 if (e->type->ops.mq.exit_sched)
543 e->type->ops.mq.exit_sched(e);
544 blk_mq_sched_tags_teardown(q);