1/*
2 * blk-mq scheduling framework
3 *
4 * Copyright (C) 2016 Jens Axboe
5 */
6#include <linux/kernel.h>
7#include <linux/module.h>
8#include <linux/blk-mq.h>
9
10#include <trace/events/block.h>
11
12#include "blk.h"
13#include "blk-mq.h"
14#include "blk-mq-debugfs.h"
15#include "blk-mq-sched.h"
16#include "blk-mq-tag.h"
17#include "blk-wbt.h"
18
19void blk_mq_sched_free_hctx_data(struct request_queue *q,
20 void (*exit)(struct blk_mq_hw_ctx *))
21{
22 struct blk_mq_hw_ctx *hctx;
23 int i;
24
25 queue_for_each_hw_ctx(q, hctx, i) {
26 if (exit && hctx->sched_data)
27 exit(hctx);
28 kfree(hctx->sched_data);
29 hctx->sched_data = NULL;
30 }
31}
32EXPORT_SYMBOL_GPL(blk_mq_sched_free_hctx_data);
33
34void blk_mq_sched_assign_ioc(struct request *rq)
35{
36 struct request_queue *q = rq->q;
37 struct io_context *ioc;
38 struct io_cq *icq;
39
40 /*
41 * May not have an IO context if it's a passthrough request
42 */
43 ioc = current->io_context;
44 if (!ioc)
45 return;
46
47 spin_lock_irq(&q->queue_lock);
48 icq = ioc_lookup_icq(ioc, q);
49 spin_unlock_irq(&q->queue_lock);
50
51 if (!icq) {
52 icq = ioc_create_icq(ioc, q, GFP_ATOMIC);
53 if (!icq)
54 return;
55 }
56 get_io_context(icq->ioc);
57 rq->elv.icq = icq;
58}
59
60/*
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.
63 */
64void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx)
65{
66 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
67 return;
68
69 set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
70}
71EXPORT_SYMBOL_GPL(blk_mq_sched_mark_restart_hctx);
72
73void blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx)
74{
75 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
76 return;
77 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
78
79 blk_mq_run_hw_queue(hctx, true);
80}
81
82/*
83 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
84 * its queue by itself in its completion handler, so we don't need to
85 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
86 */
87static void blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
88{
89 struct request_queue *q = hctx->queue;
90 struct elevator_queue *e = q->elevator;
91 LIST_HEAD(rq_list);
92
93 do {
94 struct request *rq;
95
96 if (e->type->ops.has_work && !e->type->ops.has_work(hctx))
97 break;
98
99 if (!blk_mq_get_dispatch_budget(hctx))
100 break;
101
102 rq = e->type->ops.dispatch_request(hctx);
103 if (!rq) {
104 blk_mq_put_dispatch_budget(hctx);
105 break;
106 }
107
108 /*
109 * Now this rq owns the budget which has to be released
110 * if this rq won't be queued to driver via .queue_rq()
111 * in blk_mq_dispatch_rq_list().
112 */
113 list_add(&rq->queuelist, &rq_list);
114 } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
115}
116
117static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
118 struct blk_mq_ctx *ctx)
119{
120 unsigned short idx = ctx->index_hw[hctx->type];
121
122 if (++idx == hctx->nr_ctx)
123 idx = 0;
124
125 return hctx->ctxs[idx];
126}
127
128/*
129 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
130 * its queue by itself in its completion handler, so we don't need to
131 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
132 */
133static void blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
134{
135 struct request_queue *q = hctx->queue;
136 LIST_HEAD(rq_list);
137 struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from);
138
139 do {
140 struct request *rq;
141
142 if (!sbitmap_any_bit_set(&hctx->ctx_map))
143 break;
144
145 if (!blk_mq_get_dispatch_budget(hctx))
146 break;
147
148 rq = blk_mq_dequeue_from_ctx(hctx, ctx);
149 if (!rq) {
150 blk_mq_put_dispatch_budget(hctx);
151 break;
152 }
153
154 /*
155 * Now this rq owns the budget which has to be released
156 * if this rq won't be queued to driver via .queue_rq()
157 * in blk_mq_dispatch_rq_list().
158 */
159 list_add(&rq->queuelist, &rq_list);
160
161 /* round robin for fair dispatch */
162 ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
163
164 } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
165
166 WRITE_ONCE(hctx->dispatch_from, ctx);
167}
168
169void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
170{
171 struct request_queue *q = hctx->queue;
172 struct elevator_queue *e = q->elevator;
173 const bool has_sched_dispatch = e && e->type->ops.dispatch_request;
174 LIST_HEAD(rq_list);
175
176 /* RCU or SRCU read lock is needed before checking quiesced flag */
177 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
178 return;
179
180 hctx->run++;
181
182 /*
183 * If we have previous entries on our dispatch list, grab them first for
184 * more fair dispatch.
185 */
186 if (!list_empty_careful(&hctx->dispatch)) {
187 spin_lock(&hctx->lock);
188 if (!list_empty(&hctx->dispatch))
189 list_splice_init(&hctx->dispatch, &rq_list);
190 spin_unlock(&hctx->lock);
191 }
192
193 /*
194 * Only ask the scheduler for requests, if we didn't have residual
195 * requests from the dispatch list. This is to avoid the case where
196 * we only ever dispatch a fraction of the requests available because
197 * of low device queue depth. Once we pull requests out of the IO
198 * scheduler, we can no longer merge or sort them. So it's best to
199 * leave them there for as long as we can. Mark the hw queue as
200 * needing a restart in that case.
201 *
202 * We want to dispatch from the scheduler if there was nothing
203 * on the dispatch list or we were able to dispatch from the
204 * dispatch list.
205 */
206 if (!list_empty(&rq_list)) {
207 blk_mq_sched_mark_restart_hctx(hctx);
208 if (blk_mq_dispatch_rq_list(q, &rq_list, false)) {
209 if (has_sched_dispatch)
210 blk_mq_do_dispatch_sched(hctx);
211 else
212 blk_mq_do_dispatch_ctx(hctx);
213 }
214 } else if (has_sched_dispatch) {
215 blk_mq_do_dispatch_sched(hctx);
216 } else if (hctx->dispatch_busy) {
217 /* dequeue request one by one from sw queue if queue is busy */
218 blk_mq_do_dispatch_ctx(hctx);
219 } else {
220 blk_mq_flush_busy_ctxs(hctx, &rq_list);
221 blk_mq_dispatch_rq_list(q, &rq_list, false);
222 }
223}
224
225bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
226 struct request **merged_request)
227{
228 struct request *rq;
229
230 switch (elv_merge(q, &rq, bio)) {
231 case ELEVATOR_BACK_MERGE:
232 if (!blk_mq_sched_allow_merge(q, rq, bio))
233 return false;
234 if (!bio_attempt_back_merge(q, rq, bio))
235 return false;
236 *merged_request = attempt_back_merge(q, rq);
237 if (!*merged_request)
238 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
239 return true;
240 case ELEVATOR_FRONT_MERGE:
241 if (!blk_mq_sched_allow_merge(q, rq, bio))
242 return false;
243 if (!bio_attempt_front_merge(q, rq, bio))
244 return false;
245 *merged_request = attempt_front_merge(q, rq);
246 if (!*merged_request)
247 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
248 return true;
249 case ELEVATOR_DISCARD_MERGE:
250 return bio_attempt_discard_merge(q, rq, bio);
251 default:
252 return false;
253 }
254}
255EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
256
257/*
258 * Iterate list of requests and see if we can merge this bio with any
259 * of them.
260 */
261bool blk_mq_bio_list_merge(struct request_queue *q, struct list_head *list,
262 struct bio *bio)
263{
264 struct request *rq;
265 int checked = 8;
266
267 list_for_each_entry_reverse(rq, list, queuelist) {
268 bool merged = false;
269
270 if (!checked--)
271 break;
272
273 if (!blk_rq_merge_ok(rq, bio))
274 continue;
275
276 switch (blk_try_merge(rq, bio)) {
277 case ELEVATOR_BACK_MERGE:
278 if (blk_mq_sched_allow_merge(q, rq, bio))
279 merged = bio_attempt_back_merge(q, rq, bio);
280 break;
281 case ELEVATOR_FRONT_MERGE:
282 if (blk_mq_sched_allow_merge(q, rq, bio))
283 merged = bio_attempt_front_merge(q, rq, bio);
284 break;
285 case ELEVATOR_DISCARD_MERGE:
286 merged = bio_attempt_discard_merge(q, rq, bio);
287 break;
288 default:
289 continue;
290 }
291
292 return merged;
293 }
294
295 return false;
296}
297EXPORT_SYMBOL_GPL(blk_mq_bio_list_merge);
298
299/*
300 * Reverse check our software queue for entries that we could potentially
301 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
302 * too much time checking for merges.
303 */
304static bool blk_mq_attempt_merge(struct request_queue *q,
305 struct blk_mq_hw_ctx *hctx,
306 struct blk_mq_ctx *ctx, struct bio *bio)
307{
308 enum hctx_type type = hctx->type;
309
310 lockdep_assert_held(&ctx->lock);
311
312 if (blk_mq_bio_list_merge(q, &ctx->rq_lists[type], bio)) {
313 ctx->rq_merged++;
314 return true;
315 }
316
317 return false;
318}
319
320bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio)
321{
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, bio->bi_opf, ctx);
325 bool ret = false;
326 enum hctx_type type;
327
328 if (e && e->type->ops.bio_merge) {
329 blk_mq_put_ctx(ctx);
330 return e->type->ops.bio_merge(hctx, bio);
331 }
332
333 type = hctx->type;
334 if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
335 !list_empty_careful(&ctx->rq_lists[type])) {
336 /* default per sw-queue merge */
337 spin_lock(&ctx->lock);
338 ret = blk_mq_attempt_merge(q, hctx, ctx, bio);
339 spin_unlock(&ctx->lock);
340 }
341
342 blk_mq_put_ctx(ctx);
343 return ret;
344}
345
346bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
347{
348 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
349}
350EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
351
352void blk_mq_sched_request_inserted(struct request *rq)
353{
354 trace_block_rq_insert(rq->q, rq);
355}
356EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
357
358static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
359 bool has_sched,
360 struct request *rq)
361{
362 /* dispatch flush rq directly */
363 if (rq->rq_flags & RQF_FLUSH_SEQ) {
364 spin_lock(&hctx->lock);
365 list_add(&rq->queuelist, &hctx->dispatch);
366 spin_unlock(&hctx->lock);
367 return true;
368 }
369
370 if (has_sched)
371 rq->rq_flags |= RQF_SORTED;
372
373 return false;
374}
375
376void blk_mq_sched_insert_request(struct request *rq, bool at_head,
377 bool run_queue, bool async)
378{
379 struct request_queue *q = rq->q;
380 struct elevator_queue *e = q->elevator;
381 struct blk_mq_ctx *ctx = rq->mq_ctx;
382 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
383
384 /* flush rq in flush machinery need to be dispatched directly */
385 if (!(rq->rq_flags & RQF_FLUSH_SEQ) && op_is_flush(rq->cmd_flags)) {
386 blk_insert_flush(rq);
387 goto run;
388 }
389
390 WARN_ON(e && (rq->tag != -1));
391
392 if (blk_mq_sched_bypass_insert(hctx, !!e, rq))
393 goto run;
394
395 if (e && e->type->ops.insert_requests) {
396 LIST_HEAD(list);
397
398 list_add(&rq->queuelist, &list);
399 e->type->ops.insert_requests(hctx, &list, at_head);
400 } else {
401 spin_lock(&ctx->lock);
402 __blk_mq_insert_request(hctx, rq, at_head);
403 spin_unlock(&ctx->lock);
404 }
405
406run:
407 if (run_queue)
408 blk_mq_run_hw_queue(hctx, async);
409}
410
411void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx *hctx,
412 struct blk_mq_ctx *ctx,
413 struct list_head *list, bool run_queue_async)
414{
415 struct elevator_queue *e;
416
417 e = hctx->queue->elevator;
418 if (e && e->type->ops.insert_requests)
419 e->type->ops.insert_requests(hctx, list, false);
420 else {
421 /*
422 * try to issue requests directly if the hw queue isn't
423 * busy in case of 'none' scheduler, and this way may save
424 * us one extra enqueue & dequeue to sw queue.
425 */
426 if (!hctx->dispatch_busy && !e && !run_queue_async)
427 blk_mq_try_issue_list_directly(hctx, list);
428 else
429 blk_mq_insert_requests(hctx, ctx, list);
430 }
431
432 blk_mq_run_hw_queue(hctx, run_queue_async);
433}
434
435static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
436 struct blk_mq_hw_ctx *hctx,
437 unsigned int hctx_idx)
438{
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;
443 }
444}
445
446static int blk_mq_sched_alloc_tags(struct request_queue *q,
447 struct blk_mq_hw_ctx *hctx,
448 unsigned int hctx_idx)
449{
450 struct blk_mq_tag_set *set = q->tag_set;
451 int ret;
452
453 hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
454 set->reserved_tags);
455 if (!hctx->sched_tags)
456 return -ENOMEM;
457
458 ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests);
459 if (ret)
460 blk_mq_sched_free_tags(set, hctx, hctx_idx);
461
462 return ret;
463}
464
465static void blk_mq_sched_tags_teardown(struct request_queue *q)
466{
467 struct blk_mq_tag_set *set = q->tag_set;
468 struct blk_mq_hw_ctx *hctx;
469 int i;
470
471 queue_for_each_hw_ctx(q, hctx, i)
472 blk_mq_sched_free_tags(set, hctx, i);
473}
474
475int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
476{
477 struct blk_mq_hw_ctx *hctx;
478 struct elevator_queue *eq;
479 unsigned int i;
480 int ret;
481
482 if (!e) {
483 q->elevator = NULL;
484 q->nr_requests = q->tag_set->queue_depth;
485 return 0;
486 }
487
488 /*
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.
492 */
493 q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
494 BLKDEV_MAX_RQ);
495
496 queue_for_each_hw_ctx(q, hctx, i) {
497 ret = blk_mq_sched_alloc_tags(q, hctx, i);
498 if (ret)
499 goto err;
500 }
501
502 ret = e->ops.init_sched(q, e);
503 if (ret)
504 goto err;
505
506 blk_mq_debugfs_register_sched(q);
507
508 queue_for_each_hw_ctx(q, hctx, i) {
509 if (e->ops.init_hctx) {
510 ret = e->ops.init_hctx(hctx, i);
511 if (ret) {
512 eq = q->elevator;
513 blk_mq_exit_sched(q, eq);
514 kobject_put(&eq->kobj);
515 return ret;
516 }
517 }
518 blk_mq_debugfs_register_sched_hctx(q, hctx);
519 }
520
521 return 0;
522
523err:
524 blk_mq_sched_tags_teardown(q);
525 q->elevator = NULL;
526 return ret;
527}
528
529void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
530{
531 struct blk_mq_hw_ctx *hctx;
532 unsigned int i;
533
534 queue_for_each_hw_ctx(q, hctx, i) {
535 blk_mq_debugfs_unregister_sched_hctx(hctx);
536 if (e->type->ops.exit_hctx && hctx->sched_data) {
537 e->type->ops.exit_hctx(hctx, i);
538 hctx->sched_data = NULL;
539 }
540 }
541 blk_mq_debugfs_unregister_sched(q);
542 if (e->type->ops.exit_sched)
543 e->type->ops.exit_sched(e);
544 blk_mq_sched_tags_teardown(q);
545 q->elevator = NULL;
546}
547