blob: 9e8d6795a8c1be7eee1c727bb376f209da432ecf [file] [log] [blame]
Jens Axboebd166ef2017-01-17 06:03:22 -07001/*
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-sched.h"
15#include "blk-mq-tag.h"
16#include "blk-wbt.h"
17
18void blk_mq_sched_free_hctx_data(struct request_queue *q,
19 void (*exit)(struct blk_mq_hw_ctx *))
20{
21 struct blk_mq_hw_ctx *hctx;
22 int i;
23
24 queue_for_each_hw_ctx(q, hctx, i) {
25 if (exit && hctx->sched_data)
26 exit(hctx);
27 kfree(hctx->sched_data);
28 hctx->sched_data = NULL;
29 }
30}
31EXPORT_SYMBOL_GPL(blk_mq_sched_free_hctx_data);
32
33int blk_mq_sched_init_hctx_data(struct request_queue *q, size_t size,
34 int (*init)(struct blk_mq_hw_ctx *),
35 void (*exit)(struct blk_mq_hw_ctx *))
36{
37 struct blk_mq_hw_ctx *hctx;
38 int ret;
39 int i;
40
41 queue_for_each_hw_ctx(q, hctx, i) {
42 hctx->sched_data = kmalloc_node(size, GFP_KERNEL, hctx->numa_node);
43 if (!hctx->sched_data) {
44 ret = -ENOMEM;
45 goto error;
46 }
47
48 if (init) {
49 ret = init(hctx);
50 if (ret) {
51 /*
52 * We don't want to give exit() a partially
53 * initialized sched_data. init() must clean up
54 * if it fails.
55 */
56 kfree(hctx->sched_data);
57 hctx->sched_data = NULL;
58 goto error;
59 }
60 }
61 }
62
63 return 0;
64error:
65 blk_mq_sched_free_hctx_data(q, exit);
66 return ret;
67}
68EXPORT_SYMBOL_GPL(blk_mq_sched_init_hctx_data);
69
70static void __blk_mq_sched_assign_ioc(struct request_queue *q,
Paolo Valentef1ba8262017-02-07 18:24:43 +010071 struct request *rq,
72 struct bio *bio,
73 struct io_context *ioc)
Jens Axboebd166ef2017-01-17 06:03:22 -070074{
75 struct io_cq *icq;
76
77 spin_lock_irq(q->queue_lock);
78 icq = ioc_lookup_icq(ioc, q);
79 spin_unlock_irq(q->queue_lock);
80
81 if (!icq) {
82 icq = ioc_create_icq(ioc, q, GFP_ATOMIC);
83 if (!icq)
84 return;
85 }
86
87 rq->elv.icq = icq;
Paolo Valentef1ba8262017-02-07 18:24:43 +010088 if (!blk_mq_sched_get_rq_priv(q, rq, bio)) {
Jens Axboebd166ef2017-01-17 06:03:22 -070089 rq->rq_flags |= RQF_ELVPRIV;
90 get_io_context(icq->ioc);
91 return;
92 }
93
94 rq->elv.icq = NULL;
95}
96
97static void blk_mq_sched_assign_ioc(struct request_queue *q,
98 struct request *rq, struct bio *bio)
99{
100 struct io_context *ioc;
101
102 ioc = rq_ioc(bio);
103 if (ioc)
Paolo Valentef1ba8262017-02-07 18:24:43 +0100104 __blk_mq_sched_assign_ioc(q, rq, bio, ioc);
Jens Axboebd166ef2017-01-17 06:03:22 -0700105}
106
107struct request *blk_mq_sched_get_request(struct request_queue *q,
108 struct bio *bio,
109 unsigned int op,
110 struct blk_mq_alloc_data *data)
111{
112 struct elevator_queue *e = q->elevator;
113 struct blk_mq_hw_ctx *hctx;
114 struct blk_mq_ctx *ctx;
115 struct request *rq;
Jens Axboebd166ef2017-01-17 06:03:22 -0700116
117 blk_queue_enter_live(q);
118 ctx = blk_mq_get_ctx(q);
119 hctx = blk_mq_map_queue(q, ctx->cpu);
120
Jens Axboe5a797e02017-01-26 12:22:11 -0700121 blk_mq_set_alloc_data(data, q, data->flags, ctx, hctx);
Jens Axboebd166ef2017-01-17 06:03:22 -0700122
123 if (e) {
124 data->flags |= BLK_MQ_REQ_INTERNAL;
125
126 /*
127 * Flush requests are special and go directly to the
128 * dispatch list.
129 */
Christoph Hellwigf73f44e2017-01-27 08:30:47 -0700130 if (!op_is_flush(op) && e->type->ops.mq.get_request) {
Jens Axboebd166ef2017-01-17 06:03:22 -0700131 rq = e->type->ops.mq.get_request(q, op, data);
132 if (rq)
133 rq->rq_flags |= RQF_QUEUED;
134 } else
135 rq = __blk_mq_alloc_request(data, op);
136 } else {
137 rq = __blk_mq_alloc_request(data, op);
Jens Axboeb48fda02017-01-26 14:52:20 -0700138 if (rq)
139 data->hctx->tags->rqs[rq->tag] = rq;
Jens Axboebd166ef2017-01-17 06:03:22 -0700140 }
141
142 if (rq) {
Christoph Hellwigf73f44e2017-01-27 08:30:47 -0700143 if (!op_is_flush(op)) {
Jens Axboebd166ef2017-01-17 06:03:22 -0700144 rq->elv.icq = NULL;
145 if (e && e->type->icq_cache)
146 blk_mq_sched_assign_ioc(q, rq, bio);
147 }
148 data->hctx->queued++;
149 return rq;
150 }
151
152 blk_queue_exit(q);
153 return NULL;
154}
155
156void blk_mq_sched_put_request(struct request *rq)
157{
158 struct request_queue *q = rq->q;
159 struct elevator_queue *e = q->elevator;
160
161 if (rq->rq_flags & RQF_ELVPRIV) {
162 blk_mq_sched_put_rq_priv(rq->q, rq);
163 if (rq->elv.icq) {
164 put_io_context(rq->elv.icq->ioc);
165 rq->elv.icq = NULL;
166 }
167 }
168
169 if ((rq->rq_flags & RQF_QUEUED) && e && e->type->ops.mq.put_request)
170 e->type->ops.mq.put_request(rq);
171 else
172 blk_mq_finish_request(rq);
173}
174
175void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
176{
177 struct elevator_queue *e = hctx->queue->elevator;
Jens Axboe64765a72017-02-17 11:39:26 -0700178 const bool has_sched_dispatch = e && e->type->ops.mq.dispatch_request;
179 bool did_work = false;
Jens Axboebd166ef2017-01-17 06:03:22 -0700180 LIST_HEAD(rq_list);
181
182 if (unlikely(blk_mq_hctx_stopped(hctx)))
183 return;
184
185 hctx->run++;
186
187 /*
188 * If we have previous entries on our dispatch list, grab them first for
189 * more fair dispatch.
190 */
191 if (!list_empty_careful(&hctx->dispatch)) {
192 spin_lock(&hctx->lock);
193 if (!list_empty(&hctx->dispatch))
194 list_splice_init(&hctx->dispatch, &rq_list);
195 spin_unlock(&hctx->lock);
196 }
197
198 /*
199 * Only ask the scheduler for requests, if we didn't have residual
200 * requests from the dispatch list. This is to avoid the case where
201 * we only ever dispatch a fraction of the requests available because
202 * of low device queue depth. Once we pull requests out of the IO
203 * scheduler, we can no longer merge or sort them. So it's best to
204 * leave them there for as long as we can. Mark the hw queue as
205 * needing a restart in that case.
206 */
Jens Axboec13660a2017-01-26 12:40:07 -0700207 if (!list_empty(&rq_list)) {
Jens Axboebd166ef2017-01-17 06:03:22 -0700208 blk_mq_sched_mark_restart(hctx);
Jens Axboe64765a72017-02-17 11:39:26 -0700209 did_work = blk_mq_dispatch_rq_list(hctx, &rq_list);
210 } else if (!has_sched_dispatch) {
Jens Axboec13660a2017-01-26 12:40:07 -0700211 blk_mq_flush_busy_ctxs(hctx, &rq_list);
212 blk_mq_dispatch_rq_list(hctx, &rq_list);
Jens Axboe64765a72017-02-17 11:39:26 -0700213 }
214
215 /*
216 * We want to dispatch from the scheduler if we had no work left
217 * on the dispatch list, OR if we did have work but weren't able
218 * to make progress.
219 */
220 if (!did_work && has_sched_dispatch) {
Jens Axboec13660a2017-01-26 12:40:07 -0700221 do {
222 struct request *rq;
Jens Axboebd166ef2017-01-17 06:03:22 -0700223
Jens Axboec13660a2017-01-26 12:40:07 -0700224 rq = e->type->ops.mq.dispatch_request(hctx);
225 if (!rq)
226 break;
227 list_add(&rq->queuelist, &rq_list);
228 } while (blk_mq_dispatch_rq_list(hctx, &rq_list));
229 }
Jens Axboebd166ef2017-01-17 06:03:22 -0700230}
231
232void blk_mq_sched_move_to_dispatch(struct blk_mq_hw_ctx *hctx,
233 struct list_head *rq_list,
234 struct request *(*get_rq)(struct blk_mq_hw_ctx *))
235{
236 do {
237 struct request *rq;
238
239 rq = get_rq(hctx);
240 if (!rq)
241 break;
242
243 list_add_tail(&rq->queuelist, rq_list);
244 } while (1);
245}
246EXPORT_SYMBOL_GPL(blk_mq_sched_move_to_dispatch);
247
Jens Axboee4d750c2017-02-03 09:48:28 -0700248bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
249 struct request **merged_request)
Jens Axboebd166ef2017-01-17 06:03:22 -0700250{
251 struct request *rq;
Jens Axboebd166ef2017-01-17 06:03:22 -0700252
Christoph Hellwig34fe7c02017-02-08 14:46:48 +0100253 switch (elv_merge(q, &rq, bio)) {
254 case ELEVATOR_BACK_MERGE:
Jens Axboebd166ef2017-01-17 06:03:22 -0700255 if (!blk_mq_sched_allow_merge(q, rq, bio))
256 return false;
Christoph Hellwig34fe7c02017-02-08 14:46:48 +0100257 if (!bio_attempt_back_merge(q, rq, bio))
258 return false;
259 *merged_request = attempt_back_merge(q, rq);
260 if (!*merged_request)
261 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
262 return true;
263 case ELEVATOR_FRONT_MERGE:
Jens Axboebd166ef2017-01-17 06:03:22 -0700264 if (!blk_mq_sched_allow_merge(q, rq, bio))
265 return false;
Christoph Hellwig34fe7c02017-02-08 14:46:48 +0100266 if (!bio_attempt_front_merge(q, rq, bio))
267 return false;
268 *merged_request = attempt_front_merge(q, rq);
269 if (!*merged_request)
270 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
271 return true;
272 default:
273 return false;
Jens Axboebd166ef2017-01-17 06:03:22 -0700274 }
Jens Axboebd166ef2017-01-17 06:03:22 -0700275}
276EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
277
278bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio)
279{
280 struct elevator_queue *e = q->elevator;
281
282 if (e->type->ops.mq.bio_merge) {
283 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
284 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
285
286 blk_mq_put_ctx(ctx);
287 return e->type->ops.mq.bio_merge(hctx, bio);
288 }
289
290 return false;
291}
292
293bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
294{
295 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
296}
297EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
298
299void blk_mq_sched_request_inserted(struct request *rq)
300{
301 trace_block_rq_insert(rq->q, rq);
302}
303EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
304
Omar Sandoval0cacba62017-02-02 15:42:39 -0800305static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
306 struct request *rq)
Jens Axboebd166ef2017-01-17 06:03:22 -0700307{
308 if (rq->tag == -1) {
309 rq->rq_flags |= RQF_SORTED;
310 return false;
311 }
312
313 /*
314 * If we already have a real request tag, send directly to
315 * the dispatch list.
316 */
317 spin_lock(&hctx->lock);
318 list_add(&rq->queuelist, &hctx->dispatch);
319 spin_unlock(&hctx->lock);
320 return true;
321}
Jens Axboebd166ef2017-01-17 06:03:22 -0700322
Jens Axboe50e1dab2017-01-26 14:42:34 -0700323static void blk_mq_sched_restart_hctx(struct blk_mq_hw_ctx *hctx)
324{
325 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) {
326 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
327 if (blk_mq_hctx_has_pending(hctx))
328 blk_mq_run_hw_queue(hctx, true);
329 }
330}
331
332void blk_mq_sched_restart_queues(struct blk_mq_hw_ctx *hctx)
333{
334 unsigned int i;
335
336 if (!(hctx->flags & BLK_MQ_F_TAG_SHARED))
337 blk_mq_sched_restart_hctx(hctx);
338 else {
339 struct request_queue *q = hctx->queue;
340
341 if (!test_bit(QUEUE_FLAG_RESTART, &q->queue_flags))
342 return;
343
344 clear_bit(QUEUE_FLAG_RESTART, &q->queue_flags);
345
346 queue_for_each_hw_ctx(q, hctx, i)
347 blk_mq_sched_restart_hctx(hctx);
348 }
349}
350
Jens Axboebd6737f2017-01-27 01:00:47 -0700351/*
352 * Add flush/fua to the queue. If we fail getting a driver tag, then
353 * punt to the requeue list. Requeue will re-invoke us from a context
354 * that's safe to block from.
355 */
356static void blk_mq_sched_insert_flush(struct blk_mq_hw_ctx *hctx,
357 struct request *rq, bool can_block)
358{
359 if (blk_mq_get_driver_tag(rq, &hctx, can_block)) {
360 blk_insert_flush(rq);
361 blk_mq_run_hw_queue(hctx, true);
362 } else
Jens Axboec7a571b2017-02-17 11:37:14 -0700363 blk_mq_add_to_requeue_list(rq, false, true);
Jens Axboebd6737f2017-01-27 01:00:47 -0700364}
365
366void blk_mq_sched_insert_request(struct request *rq, bool at_head,
367 bool run_queue, bool async, bool can_block)
368{
369 struct request_queue *q = rq->q;
370 struct elevator_queue *e = q->elevator;
371 struct blk_mq_ctx *ctx = rq->mq_ctx;
372 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
373
Jens Axboef3a8ab72017-01-27 09:08:23 -0700374 if (rq->tag == -1 && op_is_flush(rq->cmd_flags)) {
Jens Axboebd6737f2017-01-27 01:00:47 -0700375 blk_mq_sched_insert_flush(hctx, rq, can_block);
376 return;
377 }
378
Omar Sandoval0cacba62017-02-02 15:42:39 -0800379 if (e && blk_mq_sched_bypass_insert(hctx, rq))
380 goto run;
381
Jens Axboebd6737f2017-01-27 01:00:47 -0700382 if (e && e->type->ops.mq.insert_requests) {
383 LIST_HEAD(list);
384
385 list_add(&rq->queuelist, &list);
386 e->type->ops.mq.insert_requests(hctx, &list, at_head);
387 } else {
388 spin_lock(&ctx->lock);
389 __blk_mq_insert_request(hctx, rq, at_head);
390 spin_unlock(&ctx->lock);
391 }
392
Omar Sandoval0cacba62017-02-02 15:42:39 -0800393run:
Jens Axboebd6737f2017-01-27 01:00:47 -0700394 if (run_queue)
395 blk_mq_run_hw_queue(hctx, async);
396}
397
398void blk_mq_sched_insert_requests(struct request_queue *q,
399 struct blk_mq_ctx *ctx,
400 struct list_head *list, bool run_queue_async)
401{
402 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
403 struct elevator_queue *e = hctx->queue->elevator;
404
Omar Sandoval0cacba62017-02-02 15:42:39 -0800405 if (e) {
406 struct request *rq, *next;
407
408 /*
409 * We bypass requests that already have a driver tag assigned,
410 * which should only be flushes. Flushes are only ever inserted
411 * as single requests, so we shouldn't ever hit the
412 * WARN_ON_ONCE() below (but let's handle it just in case).
413 */
414 list_for_each_entry_safe(rq, next, list, queuelist) {
415 if (WARN_ON_ONCE(rq->tag != -1)) {
416 list_del_init(&rq->queuelist);
417 blk_mq_sched_bypass_insert(hctx, rq);
418 }
419 }
420 }
421
Jens Axboebd6737f2017-01-27 01:00:47 -0700422 if (e && e->type->ops.mq.insert_requests)
423 e->type->ops.mq.insert_requests(hctx, list, false);
424 else
425 blk_mq_insert_requests(hctx, ctx, list);
426
427 blk_mq_run_hw_queue(hctx, run_queue_async);
428}
429
Jens Axboebd166ef2017-01-17 06:03:22 -0700430static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
431 struct blk_mq_hw_ctx *hctx,
432 unsigned int hctx_idx)
433{
434 if (hctx->sched_tags) {
435 blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx);
436 blk_mq_free_rq_map(hctx->sched_tags);
437 hctx->sched_tags = NULL;
438 }
439}
440
441int blk_mq_sched_setup(struct request_queue *q)
442{
443 struct blk_mq_tag_set *set = q->tag_set;
444 struct blk_mq_hw_ctx *hctx;
445 int ret, i;
446
447 /*
448 * Default to 256, since we don't split into sync/async like the
449 * old code did. Additionally, this is a per-hw queue depth.
450 */
451 q->nr_requests = 2 * BLKDEV_MAX_RQ;
452
453 /*
454 * We're switching to using an IO scheduler, so setup the hctx
455 * scheduler tags and switch the request map from the regular
456 * tags to scheduler tags. First allocate what we need, so we
457 * can safely fail and fallback, if needed.
458 */
459 ret = 0;
460 queue_for_each_hw_ctx(q, hctx, i) {
461 hctx->sched_tags = blk_mq_alloc_rq_map(set, i, q->nr_requests, 0);
462 if (!hctx->sched_tags) {
463 ret = -ENOMEM;
464 break;
465 }
466 ret = blk_mq_alloc_rqs(set, hctx->sched_tags, i, q->nr_requests);
467 if (ret)
468 break;
469 }
470
471 /*
472 * If we failed, free what we did allocate
473 */
474 if (ret) {
475 queue_for_each_hw_ctx(q, hctx, i) {
476 if (!hctx->sched_tags)
477 continue;
478 blk_mq_sched_free_tags(set, hctx, i);
479 }
480
481 return ret;
482 }
483
484 return 0;
485}
486
487void blk_mq_sched_teardown(struct request_queue *q)
488{
489 struct blk_mq_tag_set *set = q->tag_set;
490 struct blk_mq_hw_ctx *hctx;
491 int i;
492
493 queue_for_each_hw_ctx(q, hctx, i)
494 blk_mq_sched_free_tags(set, hctx, i);
495}
Jens Axboed3484992017-01-13 14:43:58 -0700496
497int blk_mq_sched_init(struct request_queue *q)
498{
499 int ret;
500
501#if defined(CONFIG_DEFAULT_SQ_NONE)
502 if (q->nr_hw_queues == 1)
503 return 0;
504#endif
505#if defined(CONFIG_DEFAULT_MQ_NONE)
506 if (q->nr_hw_queues > 1)
507 return 0;
508#endif
509
510 mutex_lock(&q->sysfs_lock);
511 ret = elevator_init(q, NULL);
512 mutex_unlock(&q->sysfs_lock);
513
514 return ret;
515}