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)
36 struct request_queue *q = rq->q;
37 struct io_context *ioc;
41 * May not have an IO context if it's a passthrough request
43 ioc = current->io_context;
47 spin_lock_irq(&q->queue_lock);
48 icq = ioc_lookup_icq(ioc, q);
49 spin_unlock_irq(&q->queue_lock);
52 icq = ioc_create_icq(ioc, q, GFP_ATOMIC);
56 get_io_context(icq->ioc);
61 * Mark a hardware queue as needing a restart. For shared queues, maintain
62 * a count of how many hardware queues are marked for restart.
64 static void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx)
66 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
69 set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
72 void blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx)
74 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
76 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
78 blk_mq_run_hw_queue(hctx, true);
82 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
83 * its queue by itself in its completion handler, so we don't need to
84 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
86 static void blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
88 struct request_queue *q = hctx->queue;
89 struct elevator_queue *e = q->elevator;
95 if (e->type->ops.has_work && !e->type->ops.has_work(hctx))
98 if (!blk_mq_get_dispatch_budget(hctx))
101 rq = e->type->ops.dispatch_request(hctx);
103 blk_mq_put_dispatch_budget(hctx);
108 * Now this rq owns the budget which has to be released
109 * if this rq won't be queued to driver via .queue_rq()
110 * in blk_mq_dispatch_rq_list().
112 list_add(&rq->queuelist, &rq_list);
113 } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
116 static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
117 struct blk_mq_ctx *ctx)
119 unsigned short idx = ctx->index_hw[hctx->type];
121 if (++idx == hctx->nr_ctx)
124 return hctx->ctxs[idx];
128 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
129 * its queue by itself in its completion handler, so we don't need to
130 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
132 static void blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
134 struct request_queue *q = hctx->queue;
136 struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from);
141 if (!sbitmap_any_bit_set(&hctx->ctx_map))
144 if (!blk_mq_get_dispatch_budget(hctx))
147 rq = blk_mq_dequeue_from_ctx(hctx, ctx);
149 blk_mq_put_dispatch_budget(hctx);
154 * Now this rq owns the budget which has to be released
155 * if this rq won't be queued to driver via .queue_rq()
156 * in blk_mq_dispatch_rq_list().
158 list_add(&rq->queuelist, &rq_list);
160 /* round robin for fair dispatch */
161 ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
163 } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
165 WRITE_ONCE(hctx->dispatch_from, ctx);
168 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
170 struct request_queue *q = hctx->queue;
171 struct elevator_queue *e = q->elevator;
172 const bool has_sched_dispatch = e && e->type->ops.dispatch_request;
175 /* RCU or SRCU read lock is needed before checking quiesced flag */
176 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
182 * If we have previous entries on our dispatch list, grab them first for
183 * more fair dispatch.
185 if (!list_empty_careful(&hctx->dispatch)) {
186 spin_lock(&hctx->lock);
187 if (!list_empty(&hctx->dispatch))
188 list_splice_init(&hctx->dispatch, &rq_list);
189 spin_unlock(&hctx->lock);
193 * Only ask the scheduler for requests, if we didn't have residual
194 * requests from the dispatch list. This is to avoid the case where
195 * we only ever dispatch a fraction of the requests available because
196 * of low device queue depth. Once we pull requests out of the IO
197 * scheduler, we can no longer merge or sort them. So it's best to
198 * leave them there for as long as we can. Mark the hw queue as
199 * needing a restart in that case.
201 * We want to dispatch from the scheduler if there was nothing
202 * on the dispatch list or we were able to dispatch from the
205 if (!list_empty(&rq_list)) {
206 blk_mq_sched_mark_restart_hctx(hctx);
207 if (blk_mq_dispatch_rq_list(q, &rq_list, false)) {
208 if (has_sched_dispatch)
209 blk_mq_do_dispatch_sched(hctx);
211 blk_mq_do_dispatch_ctx(hctx);
213 } else if (has_sched_dispatch) {
214 blk_mq_do_dispatch_sched(hctx);
215 } else if (hctx->dispatch_busy) {
216 /* dequeue request one by one from sw queue if queue is busy */
217 blk_mq_do_dispatch_ctx(hctx);
219 blk_mq_flush_busy_ctxs(hctx, &rq_list);
220 blk_mq_dispatch_rq_list(q, &rq_list, false);
224 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
225 struct request **merged_request)
229 switch (elv_merge(q, &rq, bio)) {
230 case ELEVATOR_BACK_MERGE:
231 if (!blk_mq_sched_allow_merge(q, rq, bio))
233 if (!bio_attempt_back_merge(q, rq, bio))
235 *merged_request = attempt_back_merge(q, rq);
236 if (!*merged_request)
237 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
239 case ELEVATOR_FRONT_MERGE:
240 if (!blk_mq_sched_allow_merge(q, rq, bio))
242 if (!bio_attempt_front_merge(q, rq, bio))
244 *merged_request = attempt_front_merge(q, rq);
245 if (!*merged_request)
246 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
248 case ELEVATOR_DISCARD_MERGE:
249 return bio_attempt_discard_merge(q, rq, bio);
254 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
257 * Iterate list of requests and see if we can merge this bio with any
260 bool blk_mq_bio_list_merge(struct request_queue *q, struct list_head *list,
266 list_for_each_entry_reverse(rq, list, queuelist) {
272 if (!blk_rq_merge_ok(rq, bio))
275 switch (blk_try_merge(rq, bio)) {
276 case ELEVATOR_BACK_MERGE:
277 if (blk_mq_sched_allow_merge(q, rq, bio))
278 merged = bio_attempt_back_merge(q, rq, bio);
280 case ELEVATOR_FRONT_MERGE:
281 if (blk_mq_sched_allow_merge(q, rq, bio))
282 merged = bio_attempt_front_merge(q, rq, bio);
284 case ELEVATOR_DISCARD_MERGE:
285 merged = bio_attempt_discard_merge(q, rq, bio);
296 EXPORT_SYMBOL_GPL(blk_mq_bio_list_merge);
299 * Reverse check our software queue for entries that we could potentially
300 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
301 * too much time checking for merges.
303 static bool blk_mq_attempt_merge(struct request_queue *q,
304 struct blk_mq_ctx *ctx, struct bio *bio)
306 lockdep_assert_held(&ctx->lock);
308 if (blk_mq_bio_list_merge(q, &ctx->rq_list, bio)) {
316 bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio)
318 struct elevator_queue *e = q->elevator;
319 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
320 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, bio->bi_opf, ctx->cpu);
323 if (e && e->type->ops.bio_merge) {
325 return e->type->ops.bio_merge(hctx, bio);
328 if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
329 !list_empty_careful(&ctx->rq_list)) {
330 /* default per sw-queue merge */
331 spin_lock(&ctx->lock);
332 ret = blk_mq_attempt_merge(q, ctx, bio);
333 spin_unlock(&ctx->lock);
340 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
342 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
344 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
346 void blk_mq_sched_request_inserted(struct request *rq)
348 trace_block_rq_insert(rq->q, rq);
350 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
352 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
356 /* dispatch flush rq directly */
357 if (rq->rq_flags & RQF_FLUSH_SEQ) {
358 spin_lock(&hctx->lock);
359 list_add(&rq->queuelist, &hctx->dispatch);
360 spin_unlock(&hctx->lock);
365 rq->rq_flags |= RQF_SORTED;
370 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
371 bool run_queue, bool async)
373 struct request_queue *q = rq->q;
374 struct elevator_queue *e = q->elevator;
375 struct blk_mq_ctx *ctx = rq->mq_ctx;
376 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
378 /* flush rq in flush machinery need to be dispatched directly */
379 if (!(rq->rq_flags & RQF_FLUSH_SEQ) && op_is_flush(rq->cmd_flags)) {
380 blk_insert_flush(rq);
384 WARN_ON(e && (rq->tag != -1));
386 if (blk_mq_sched_bypass_insert(hctx, !!e, rq))
389 if (e && e->type->ops.insert_requests) {
392 list_add(&rq->queuelist, &list);
393 e->type->ops.insert_requests(hctx, &list, at_head);
395 spin_lock(&ctx->lock);
396 __blk_mq_insert_request(hctx, rq, at_head);
397 spin_unlock(&ctx->lock);
402 blk_mq_run_hw_queue(hctx, async);
405 void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx *hctx,
406 struct blk_mq_ctx *ctx,
407 struct list_head *list, bool run_queue_async)
409 struct elevator_queue *e;
411 e = hctx->queue->elevator;
412 if (e && e->type->ops.insert_requests)
413 e->type->ops.insert_requests(hctx, list, false);
416 * try to issue requests directly if the hw queue isn't
417 * busy in case of 'none' scheduler, and this way may save
418 * us one extra enqueue & dequeue to sw queue.
420 if (!hctx->dispatch_busy && !e && !run_queue_async) {
421 blk_mq_try_issue_list_directly(hctx, list);
422 if (list_empty(list))
425 blk_mq_insert_requests(hctx, ctx, list);
428 blk_mq_run_hw_queue(hctx, run_queue_async);
431 static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
432 struct blk_mq_hw_ctx *hctx,
433 unsigned int hctx_idx)
435 if (hctx->sched_tags) {
436 blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx);
437 blk_mq_free_rq_map(hctx->sched_tags);
438 hctx->sched_tags = NULL;
442 static int blk_mq_sched_alloc_tags(struct request_queue *q,
443 struct blk_mq_hw_ctx *hctx,
444 unsigned int hctx_idx)
446 struct blk_mq_tag_set *set = q->tag_set;
449 hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
451 if (!hctx->sched_tags)
454 ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests);
456 blk_mq_sched_free_tags(set, hctx, hctx_idx);
461 static void blk_mq_sched_tags_teardown(struct request_queue *q)
463 struct blk_mq_tag_set *set = q->tag_set;
464 struct blk_mq_hw_ctx *hctx;
467 queue_for_each_hw_ctx(q, hctx, i)
468 blk_mq_sched_free_tags(set, hctx, i);
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;
480 q->nr_requests = q->tag_set->queue_depth;
485 * Default to double of smaller one between hw queue_depth and 128,
486 * since we don't split into sync/async like the old code did.
487 * Additionally, this is a per-hw queue depth.
489 q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
492 queue_for_each_hw_ctx(q, hctx, i) {
493 ret = blk_mq_sched_alloc_tags(q, hctx, i);
498 ret = e->ops.init_sched(q, e);
502 blk_mq_debugfs_register_sched(q);
504 queue_for_each_hw_ctx(q, hctx, i) {
505 if (e->ops.init_hctx) {
506 ret = e->ops.init_hctx(hctx, i);
509 blk_mq_exit_sched(q, eq);
510 kobject_put(&eq->kobj);
514 blk_mq_debugfs_register_sched_hctx(q, hctx);
520 blk_mq_sched_tags_teardown(q);
525 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
527 struct blk_mq_hw_ctx *hctx;
530 queue_for_each_hw_ctx(q, hctx, i) {
531 blk_mq_debugfs_unregister_sched_hctx(hctx);
532 if (e->type->ops.exit_hctx && hctx->sched_data) {
533 e->type->ops.exit_hctx(hctx, i);
534 hctx->sched_data = NULL;
537 blk_mq_debugfs_unregister_sched(q);
538 if (e->type->ops.exit_sched)
539 e->type->ops.exit_sched(e);
540 blk_mq_sched_tags_teardown(q);