Commit | Line | Data |
---|---|---|
3dcf60bc | 1 | // SPDX-License-Identifier: GPL-2.0 |
bd166ef1 JA |
2 | /* |
3 | * blk-mq scheduling framework | |
4 | * | |
5 | * Copyright (C) 2016 Jens Axboe | |
6 | */ | |
7 | #include <linux/kernel.h> | |
8 | #include <linux/module.h> | |
9 | #include <linux/blk-mq.h> | |
6e6fcbc2 | 10 | #include <linux/list_sort.h> |
bd166ef1 JA |
11 | |
12 | #include <trace/events/block.h> | |
13 | ||
14 | #include "blk.h" | |
15 | #include "blk-mq.h" | |
d332ce09 | 16 | #include "blk-mq-debugfs.h" |
bd166ef1 JA |
17 | #include "blk-mq-sched.h" |
18 | #include "blk-mq-tag.h" | |
19 | #include "blk-wbt.h" | |
20 | ||
21 | void blk_mq_sched_free_hctx_data(struct request_queue *q, | |
22 | void (*exit)(struct blk_mq_hw_ctx *)) | |
23 | { | |
24 | struct blk_mq_hw_ctx *hctx; | |
25 | int i; | |
26 | ||
27 | queue_for_each_hw_ctx(q, hctx, i) { | |
28 | if (exit && hctx->sched_data) | |
29 | exit(hctx); | |
30 | kfree(hctx->sched_data); | |
31 | hctx->sched_data = NULL; | |
32 | } | |
33 | } | |
34 | EXPORT_SYMBOL_GPL(blk_mq_sched_free_hctx_data); | |
35 | ||
e2b3fa5a | 36 | void blk_mq_sched_assign_ioc(struct request *rq) |
bd166ef1 | 37 | { |
44e8c2bf | 38 | struct request_queue *q = rq->q; |
0c62bff1 | 39 | struct io_context *ioc; |
bd166ef1 JA |
40 | struct io_cq *icq; |
41 | ||
0c62bff1 JA |
42 | /* |
43 | * May not have an IO context if it's a passthrough request | |
44 | */ | |
45 | ioc = current->io_context; | |
46 | if (!ioc) | |
47 | return; | |
48 | ||
0d945c1f | 49 | spin_lock_irq(&q->queue_lock); |
bd166ef1 | 50 | icq = ioc_lookup_icq(ioc, q); |
0d945c1f | 51 | spin_unlock_irq(&q->queue_lock); |
bd166ef1 JA |
52 | |
53 | if (!icq) { | |
54 | icq = ioc_create_icq(ioc, q, GFP_ATOMIC); | |
55 | if (!icq) | |
56 | return; | |
57 | } | |
ea511e3c | 58 | get_io_context(icq->ioc); |
44e8c2bf | 59 | rq->elv.icq = icq; |
bd166ef1 JA |
60 | } |
61 | ||
8e8320c9 JA |
62 | /* |
63 | * Mark a hardware queue as needing a restart. For shared queues, maintain | |
64 | * a count of how many hardware queues are marked for restart. | |
65 | */ | |
7211aef8 | 66 | void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx) |
8e8320c9 JA |
67 | { |
68 | if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
69 | return; | |
70 | ||
97889f9a | 71 | set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state); |
8e8320c9 | 72 | } |
7211aef8 | 73 | EXPORT_SYMBOL_GPL(blk_mq_sched_mark_restart_hctx); |
8e8320c9 | 74 | |
97889f9a | 75 | void blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx) |
8e8320c9 JA |
76 | { |
77 | if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
97889f9a ML |
78 | return; |
79 | clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state); | |
8e8320c9 | 80 | |
d7d8535f ML |
81 | /* |
82 | * Order clearing SCHED_RESTART and list_empty_careful(&hctx->dispatch) | |
83 | * in blk_mq_run_hw_queue(). Its pair is the barrier in | |
84 | * blk_mq_dispatch_rq_list(). So dispatch code won't see SCHED_RESTART, | |
85 | * meantime new request added to hctx->dispatch is missed to check in | |
86 | * blk_mq_run_hw_queue(). | |
87 | */ | |
88 | smp_mb(); | |
89 | ||
97889f9a | 90 | blk_mq_run_hw_queue(hctx, true); |
8e8320c9 JA |
91 | } |
92 | ||
6e6fcbc2 ML |
93 | static int sched_rq_cmp(void *priv, struct list_head *a, struct list_head *b) |
94 | { | |
95 | struct request *rqa = container_of(a, struct request, queuelist); | |
96 | struct request *rqb = container_of(b, struct request, queuelist); | |
97 | ||
98 | return rqa->mq_hctx > rqb->mq_hctx; | |
99 | } | |
100 | ||
101 | static bool blk_mq_dispatch_hctx_list(struct list_head *rq_list) | |
102 | { | |
103 | struct blk_mq_hw_ctx *hctx = | |
104 | list_first_entry(rq_list, struct request, queuelist)->mq_hctx; | |
105 | struct request *rq; | |
106 | LIST_HEAD(hctx_list); | |
107 | unsigned int count = 0; | |
6e6fcbc2 ML |
108 | |
109 | list_for_each_entry(rq, rq_list, queuelist) { | |
110 | if (rq->mq_hctx != hctx) { | |
111 | list_cut_before(&hctx_list, rq_list, &rq->queuelist); | |
112 | goto dispatch; | |
113 | } | |
114 | count++; | |
115 | } | |
116 | list_splice_tail_init(rq_list, &hctx_list); | |
117 | ||
118 | dispatch: | |
106e71c5 | 119 | return blk_mq_dispatch_rq_list(hctx, &hctx_list, count); |
6e6fcbc2 ML |
120 | } |
121 | ||
a0823421 DA |
122 | #define BLK_MQ_BUDGET_DELAY 3 /* ms units */ |
123 | ||
1f460b63 ML |
124 | /* |
125 | * Only SCSI implements .get_budget and .put_budget, and SCSI restarts | |
126 | * its queue by itself in its completion handler, so we don't need to | |
127 | * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE. | |
28d65729 SQ |
128 | * |
129 | * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to | |
130 | * be run again. This is necessary to avoid starving flushes. | |
1f460b63 | 131 | */ |
6e6fcbc2 | 132 | static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx) |
caf8eb0d ML |
133 | { |
134 | struct request_queue *q = hctx->queue; | |
135 | struct elevator_queue *e = q->elevator; | |
6e6fcbc2 ML |
136 | bool multi_hctxs = false, run_queue = false; |
137 | bool dispatched = false, busy = false; | |
138 | unsigned int max_dispatch; | |
caf8eb0d | 139 | LIST_HEAD(rq_list); |
6e6fcbc2 ML |
140 | int count = 0; |
141 | ||
142 | if (hctx->dispatch_busy) | |
143 | max_dispatch = 1; | |
144 | else | |
145 | max_dispatch = hctx->queue->nr_requests; | |
caf8eb0d ML |
146 | |
147 | do { | |
6e6fcbc2 ML |
148 | struct request *rq; |
149 | ||
f9cd4bfe | 150 | if (e->type->ops.has_work && !e->type->ops.has_work(hctx)) |
caf8eb0d | 151 | break; |
de148297 | 152 | |
28d65729 | 153 | if (!list_empty_careful(&hctx->dispatch)) { |
6e6fcbc2 | 154 | busy = true; |
28d65729 SQ |
155 | break; |
156 | } | |
157 | ||
65c76369 | 158 | if (!blk_mq_get_dispatch_budget(q)) |
1f460b63 | 159 | break; |
de148297 | 160 | |
f9cd4bfe | 161 | rq = e->type->ops.dispatch_request(hctx); |
de148297 | 162 | if (!rq) { |
65c76369 | 163 | blk_mq_put_dispatch_budget(q); |
a0823421 DA |
164 | /* |
165 | * We're releasing without dispatching. Holding the | |
166 | * budget could have blocked any "hctx"s with the | |
167 | * same queue and if we didn't dispatch then there's | |
168 | * no guarantee anyone will kick the queue. Kick it | |
169 | * ourselves. | |
170 | */ | |
6e6fcbc2 | 171 | run_queue = true; |
de148297 | 172 | break; |
de148297 ML |
173 | } |
174 | ||
175 | /* | |
176 | * Now this rq owns the budget which has to be released | |
177 | * if this rq won't be queued to driver via .queue_rq() | |
178 | * in blk_mq_dispatch_rq_list(). | |
179 | */ | |
6e6fcbc2 ML |
180 | list_add_tail(&rq->queuelist, &rq_list); |
181 | if (rq->mq_hctx != hctx) | |
182 | multi_hctxs = true; | |
183 | } while (++count < max_dispatch); | |
184 | ||
185 | if (!count) { | |
186 | if (run_queue) | |
187 | blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY); | |
188 | } else if (multi_hctxs) { | |
189 | /* | |
190 | * Requests from different hctx may be dequeued from some | |
191 | * schedulers, such as bfq and deadline. | |
192 | * | |
193 | * Sort the requests in the list according to their hctx, | |
194 | * dispatch batching requests from same hctx at a time. | |
195 | */ | |
196 | list_sort(NULL, &rq_list, sched_rq_cmp); | |
197 | do { | |
198 | dispatched |= blk_mq_dispatch_hctx_list(&rq_list); | |
199 | } while (!list_empty(&rq_list)); | |
200 | } else { | |
201 | dispatched = blk_mq_dispatch_rq_list(hctx, &rq_list, count); | |
202 | } | |
203 | ||
204 | if (busy) | |
205 | return -EAGAIN; | |
206 | return !!dispatched; | |
207 | } | |
208 | ||
209 | static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx) | |
210 | { | |
211 | int ret; | |
212 | ||
213 | do { | |
214 | ret = __blk_mq_do_dispatch_sched(hctx); | |
215 | } while (ret == 1); | |
28d65729 SQ |
216 | |
217 | return ret; | |
caf8eb0d ML |
218 | } |
219 | ||
b347689f ML |
220 | static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx, |
221 | struct blk_mq_ctx *ctx) | |
222 | { | |
f31967f0 | 223 | unsigned short idx = ctx->index_hw[hctx->type]; |
b347689f ML |
224 | |
225 | if (++idx == hctx->nr_ctx) | |
226 | idx = 0; | |
227 | ||
228 | return hctx->ctxs[idx]; | |
229 | } | |
230 | ||
1f460b63 ML |
231 | /* |
232 | * Only SCSI implements .get_budget and .put_budget, and SCSI restarts | |
233 | * its queue by itself in its completion handler, so we don't need to | |
234 | * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE. | |
28d65729 SQ |
235 | * |
236 | * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to | |
c4aecaa2 | 237 | * be run again. This is necessary to avoid starving flushes. |
1f460b63 | 238 | */ |
28d65729 | 239 | static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx) |
b347689f ML |
240 | { |
241 | struct request_queue *q = hctx->queue; | |
242 | LIST_HEAD(rq_list); | |
243 | struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from); | |
28d65729 | 244 | int ret = 0; |
445874e8 | 245 | struct request *rq; |
b347689f ML |
246 | |
247 | do { | |
28d65729 SQ |
248 | if (!list_empty_careful(&hctx->dispatch)) { |
249 | ret = -EAGAIN; | |
250 | break; | |
251 | } | |
252 | ||
b347689f ML |
253 | if (!sbitmap_any_bit_set(&hctx->ctx_map)) |
254 | break; | |
255 | ||
65c76369 | 256 | if (!blk_mq_get_dispatch_budget(q)) |
1f460b63 | 257 | break; |
b347689f ML |
258 | |
259 | rq = blk_mq_dequeue_from_ctx(hctx, ctx); | |
260 | if (!rq) { | |
65c76369 | 261 | blk_mq_put_dispatch_budget(q); |
a0823421 DA |
262 | /* |
263 | * We're releasing without dispatching. Holding the | |
264 | * budget could have blocked any "hctx"s with the | |
265 | * same queue and if we didn't dispatch then there's | |
266 | * no guarantee anyone will kick the queue. Kick it | |
267 | * ourselves. | |
268 | */ | |
269 | blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY); | |
b347689f | 270 | break; |
b347689f ML |
271 | } |
272 | ||
273 | /* | |
274 | * Now this rq owns the budget which has to be released | |
275 | * if this rq won't be queued to driver via .queue_rq() | |
276 | * in blk_mq_dispatch_rq_list(). | |
277 | */ | |
278 | list_add(&rq->queuelist, &rq_list); | |
279 | ||
280 | /* round robin for fair dispatch */ | |
281 | ctx = blk_mq_next_ctx(hctx, rq->mq_ctx); | |
282 | ||
1fd40b5e | 283 | } while (blk_mq_dispatch_rq_list(rq->mq_hctx, &rq_list, 1)); |
b347689f ML |
284 | |
285 | WRITE_ONCE(hctx->dispatch_from, ctx); | |
28d65729 | 286 | return ret; |
b347689f ML |
287 | } |
288 | ||
e1b586f2 | 289 | static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx) |
bd166ef1 | 290 | { |
81380ca1 OS |
291 | struct request_queue *q = hctx->queue; |
292 | struct elevator_queue *e = q->elevator; | |
f9cd4bfe | 293 | const bool has_sched_dispatch = e && e->type->ops.dispatch_request; |
28d65729 | 294 | int ret = 0; |
bd166ef1 JA |
295 | LIST_HEAD(rq_list); |
296 | ||
bd166ef1 JA |
297 | /* |
298 | * If we have previous entries on our dispatch list, grab them first for | |
299 | * more fair dispatch. | |
300 | */ | |
301 | if (!list_empty_careful(&hctx->dispatch)) { | |
302 | spin_lock(&hctx->lock); | |
303 | if (!list_empty(&hctx->dispatch)) | |
304 | list_splice_init(&hctx->dispatch, &rq_list); | |
305 | spin_unlock(&hctx->lock); | |
306 | } | |
307 | ||
308 | /* | |
309 | * Only ask the scheduler for requests, if we didn't have residual | |
310 | * requests from the dispatch list. This is to avoid the case where | |
311 | * we only ever dispatch a fraction of the requests available because | |
312 | * of low device queue depth. Once we pull requests out of the IO | |
313 | * scheduler, we can no longer merge or sort them. So it's best to | |
314 | * leave them there for as long as we can. Mark the hw queue as | |
315 | * needing a restart in that case. | |
caf8eb0d ML |
316 | * |
317 | * We want to dispatch from the scheduler if there was nothing | |
318 | * on the dispatch list or we were able to dispatch from the | |
319 | * dispatch list. | |
bd166ef1 | 320 | */ |
c13660a0 | 321 | if (!list_empty(&rq_list)) { |
d38d3515 | 322 | blk_mq_sched_mark_restart_hctx(hctx); |
1fd40b5e | 323 | if (blk_mq_dispatch_rq_list(hctx, &rq_list, 0)) { |
b347689f | 324 | if (has_sched_dispatch) |
28d65729 | 325 | ret = blk_mq_do_dispatch_sched(hctx); |
b347689f | 326 | else |
28d65729 | 327 | ret = blk_mq_do_dispatch_ctx(hctx); |
b347689f | 328 | } |
caf8eb0d | 329 | } else if (has_sched_dispatch) { |
28d65729 | 330 | ret = blk_mq_do_dispatch_sched(hctx); |
6e768717 ML |
331 | } else if (hctx->dispatch_busy) { |
332 | /* dequeue request one by one from sw queue if queue is busy */ | |
28d65729 | 333 | ret = blk_mq_do_dispatch_ctx(hctx); |
caf8eb0d | 334 | } else { |
c13660a0 | 335 | blk_mq_flush_busy_ctxs(hctx, &rq_list); |
1fd40b5e | 336 | blk_mq_dispatch_rq_list(hctx, &rq_list, 0); |
64765a75 | 337 | } |
28d65729 SQ |
338 | |
339 | return ret; | |
340 | } | |
341 | ||
342 | void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx) | |
343 | { | |
344 | struct request_queue *q = hctx->queue; | |
345 | ||
346 | /* RCU or SRCU read lock is needed before checking quiesced flag */ | |
347 | if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q))) | |
348 | return; | |
349 | ||
350 | hctx->run++; | |
351 | ||
352 | /* | |
353 | * A return of -EAGAIN is an indication that hctx->dispatch is not | |
354 | * empty and we must run again in order to avoid starving flushes. | |
355 | */ | |
356 | if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) { | |
357 | if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) | |
358 | blk_mq_run_hw_queue(hctx, true); | |
359 | } | |
bd166ef1 JA |
360 | } |
361 | ||
e4d750c9 | 362 | bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio, |
14ccb66b | 363 | unsigned int nr_segs, struct request **merged_request) |
bd166ef1 JA |
364 | { |
365 | struct request *rq; | |
bd166ef1 | 366 | |
34fe7c05 CH |
367 | switch (elv_merge(q, &rq, bio)) { |
368 | case ELEVATOR_BACK_MERGE: | |
bd166ef1 JA |
369 | if (!blk_mq_sched_allow_merge(q, rq, bio)) |
370 | return false; | |
7d7ca7c5 | 371 | if (bio_attempt_back_merge(rq, bio, nr_segs) != BIO_MERGE_OK) |
34fe7c05 CH |
372 | return false; |
373 | *merged_request = attempt_back_merge(q, rq); | |
374 | if (!*merged_request) | |
375 | elv_merged_request(q, rq, ELEVATOR_BACK_MERGE); | |
376 | return true; | |
377 | case ELEVATOR_FRONT_MERGE: | |
bd166ef1 JA |
378 | if (!blk_mq_sched_allow_merge(q, rq, bio)) |
379 | return false; | |
7d7ca7c5 | 380 | if (bio_attempt_front_merge(rq, bio, nr_segs) != BIO_MERGE_OK) |
34fe7c05 CH |
381 | return false; |
382 | *merged_request = attempt_front_merge(q, rq); | |
383 | if (!*merged_request) | |
384 | elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE); | |
385 | return true; | |
bea99a50 | 386 | case ELEVATOR_DISCARD_MERGE: |
7d7ca7c5 | 387 | return bio_attempt_discard_merge(q, rq, bio) == BIO_MERGE_OK; |
34fe7c05 CH |
388 | default: |
389 | return false; | |
bd166ef1 | 390 | } |
bd166ef1 JA |
391 | } |
392 | EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge); | |
393 | ||
9c558734 JA |
394 | /* |
395 | * Reverse check our software queue for entries that we could potentially | |
396 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
397 | * too much time checking for merges. | |
398 | */ | |
399 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
c16d6b5a | 400 | struct blk_mq_hw_ctx *hctx, |
14ccb66b CH |
401 | struct blk_mq_ctx *ctx, struct bio *bio, |
402 | unsigned int nr_segs) | |
9c558734 | 403 | { |
c16d6b5a ML |
404 | enum hctx_type type = hctx->type; |
405 | ||
9c558734 JA |
406 | lockdep_assert_held(&ctx->lock); |
407 | ||
bdc6a287 | 408 | if (blk_bio_list_merge(q, &ctx->rq_lists[type], bio, nr_segs)) { |
9c558734 JA |
409 | ctx->rq_merged++; |
410 | return true; | |
411 | } | |
412 | ||
413 | return false; | |
414 | } | |
9bddeb2a | 415 | |
14ccb66b CH |
416 | bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio, |
417 | unsigned int nr_segs) | |
bd166ef1 JA |
418 | { |
419 | struct elevator_queue *e = q->elevator; | |
9bddeb2a | 420 | struct blk_mq_ctx *ctx = blk_mq_get_ctx(q); |
8ccdf4a3 | 421 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, bio->bi_opf, ctx); |
9bddeb2a | 422 | bool ret = false; |
c16d6b5a | 423 | enum hctx_type type; |
bd166ef1 | 424 | |
c05f4220 | 425 | if (e && e->type->ops.bio_merge) |
14ccb66b | 426 | return e->type->ops.bio_merge(hctx, bio, nr_segs); |
bd166ef1 | 427 | |
c16d6b5a | 428 | type = hctx->type; |
b04f50ab | 429 | if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) && |
c16d6b5a | 430 | !list_empty_careful(&ctx->rq_lists[type])) { |
9bddeb2a ML |
431 | /* default per sw-queue merge */ |
432 | spin_lock(&ctx->lock); | |
14ccb66b | 433 | ret = blk_mq_attempt_merge(q, hctx, ctx, bio, nr_segs); |
9bddeb2a ML |
434 | spin_unlock(&ctx->lock); |
435 | } | |
436 | ||
9bddeb2a | 437 | return ret; |
bd166ef1 JA |
438 | } |
439 | ||
440 | bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq) | |
441 | { | |
442 | return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq); | |
443 | } | |
444 | EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge); | |
445 | ||
446 | void blk_mq_sched_request_inserted(struct request *rq) | |
447 | { | |
448 | trace_block_rq_insert(rq->q, rq); | |
449 | } | |
450 | EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted); | |
451 | ||
0cacba6c | 452 | static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx, |
a6a252e6 | 453 | bool has_sched, |
0cacba6c | 454 | struct request *rq) |
bd166ef1 | 455 | { |
01e99aec ML |
456 | /* |
457 | * dispatch flush and passthrough rq directly | |
458 | * | |
459 | * passthrough request has to be added to hctx->dispatch directly. | |
460 | * For some reason, device may be in one situation which can't | |
461 | * handle FS request, so STS_RESOURCE is always returned and the | |
462 | * FS request will be added to hctx->dispatch. However passthrough | |
463 | * request may be required at that time for fixing the problem. If | |
464 | * passthrough request is added to scheduler queue, there isn't any | |
465 | * chance to dispatch it given we prioritize requests in hctx->dispatch. | |
466 | */ | |
467 | if ((rq->rq_flags & RQF_FLUSH_SEQ) || blk_rq_is_passthrough(rq)) | |
a6a252e6 | 468 | return true; |
a6a252e6 | 469 | |
923218f6 | 470 | if (has_sched) |
bd166ef1 | 471 | rq->rq_flags |= RQF_SORTED; |
bd166ef1 | 472 | |
a6a252e6 | 473 | return false; |
bd166ef1 | 474 | } |
bd166ef1 | 475 | |
bd6737f1 | 476 | void blk_mq_sched_insert_request(struct request *rq, bool at_head, |
9e97d295 | 477 | bool run_queue, bool async) |
bd6737f1 JA |
478 | { |
479 | struct request_queue *q = rq->q; | |
480 | struct elevator_queue *e = q->elevator; | |
481 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 482 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
bd6737f1 | 483 | |
a6a252e6 ML |
484 | /* flush rq in flush machinery need to be dispatched directly */ |
485 | if (!(rq->rq_flags & RQF_FLUSH_SEQ) && op_is_flush(rq->cmd_flags)) { | |
923218f6 ML |
486 | blk_insert_flush(rq); |
487 | goto run; | |
bd6737f1 JA |
488 | } |
489 | ||
923218f6 ML |
490 | WARN_ON(e && (rq->tag != -1)); |
491 | ||
01e99aec | 492 | if (blk_mq_sched_bypass_insert(hctx, !!e, rq)) { |
cc3200ea ML |
493 | /* |
494 | * Firstly normal IO request is inserted to scheduler queue or | |
495 | * sw queue, meantime we add flush request to dispatch queue( | |
496 | * hctx->dispatch) directly and there is at most one in-flight | |
497 | * flush request for each hw queue, so it doesn't matter to add | |
498 | * flush request to tail or front of the dispatch queue. | |
499 | * | |
500 | * Secondly in case of NCQ, flush request belongs to non-NCQ | |
501 | * command, and queueing it will fail when there is any | |
502 | * in-flight normal IO request(NCQ command). When adding flush | |
503 | * rq to the front of hctx->dispatch, it is easier to introduce | |
504 | * extra time to flush rq's latency because of S_SCHED_RESTART | |
505 | * compared with adding to the tail of dispatch queue, then | |
506 | * chance of flush merge is increased, and less flush requests | |
507 | * will be issued to controller. It is observed that ~10% time | |
508 | * is saved in blktests block/004 on disk attached to AHCI/NCQ | |
509 | * drive when adding flush rq to the front of hctx->dispatch. | |
510 | * | |
511 | * Simply queue flush rq to the front of hctx->dispatch so that | |
512 | * intensive flush workloads can benefit in case of NCQ HW. | |
513 | */ | |
514 | at_head = (rq->rq_flags & RQF_FLUSH_SEQ) ? true : at_head; | |
01e99aec | 515 | blk_mq_request_bypass_insert(rq, at_head, false); |
0cacba6c | 516 | goto run; |
01e99aec | 517 | } |
0cacba6c | 518 | |
f9cd4bfe | 519 | if (e && e->type->ops.insert_requests) { |
bd6737f1 JA |
520 | LIST_HEAD(list); |
521 | ||
522 | list_add(&rq->queuelist, &list); | |
f9cd4bfe | 523 | e->type->ops.insert_requests(hctx, &list, at_head); |
bd6737f1 JA |
524 | } else { |
525 | spin_lock(&ctx->lock); | |
526 | __blk_mq_insert_request(hctx, rq, at_head); | |
527 | spin_unlock(&ctx->lock); | |
528 | } | |
529 | ||
0cacba6c | 530 | run: |
bd6737f1 JA |
531 | if (run_queue) |
532 | blk_mq_run_hw_queue(hctx, async); | |
533 | } | |
534 | ||
67cae4c9 | 535 | void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx *hctx, |
bd6737f1 JA |
536 | struct blk_mq_ctx *ctx, |
537 | struct list_head *list, bool run_queue_async) | |
538 | { | |
f9afca4d | 539 | struct elevator_queue *e; |
e87eb301 ML |
540 | struct request_queue *q = hctx->queue; |
541 | ||
542 | /* | |
543 | * blk_mq_sched_insert_requests() is called from flush plug | |
544 | * context only, and hold one usage counter to prevent queue | |
545 | * from being released. | |
546 | */ | |
547 | percpu_ref_get(&q->q_usage_counter); | |
bd6737f1 | 548 | |
f9afca4d | 549 | e = hctx->queue->elevator; |
f9cd4bfe JA |
550 | if (e && e->type->ops.insert_requests) |
551 | e->type->ops.insert_requests(hctx, list, false); | |
6ce3dd6e ML |
552 | else { |
553 | /* | |
554 | * try to issue requests directly if the hw queue isn't | |
555 | * busy in case of 'none' scheduler, and this way may save | |
556 | * us one extra enqueue & dequeue to sw queue. | |
557 | */ | |
fd9c40f6 | 558 | if (!hctx->dispatch_busy && !e && !run_queue_async) { |
6ce3dd6e | 559 | blk_mq_try_issue_list_directly(hctx, list); |
fd9c40f6 | 560 | if (list_empty(list)) |
e87eb301 | 561 | goto out; |
fd9c40f6 BVA |
562 | } |
563 | blk_mq_insert_requests(hctx, ctx, list); | |
6ce3dd6e | 564 | } |
bd6737f1 JA |
565 | |
566 | blk_mq_run_hw_queue(hctx, run_queue_async); | |
e87eb301 ML |
567 | out: |
568 | percpu_ref_put(&q->q_usage_counter); | |
bd6737f1 JA |
569 | } |
570 | ||
bd166ef1 JA |
571 | static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set, |
572 | struct blk_mq_hw_ctx *hctx, | |
573 | unsigned int hctx_idx) | |
574 | { | |
575 | if (hctx->sched_tags) { | |
576 | blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx); | |
577 | blk_mq_free_rq_map(hctx->sched_tags); | |
578 | hctx->sched_tags = NULL; | |
579 | } | |
580 | } | |
581 | ||
6917ff0b OS |
582 | static int blk_mq_sched_alloc_tags(struct request_queue *q, |
583 | struct blk_mq_hw_ctx *hctx, | |
584 | unsigned int hctx_idx) | |
585 | { | |
586 | struct blk_mq_tag_set *set = q->tag_set; | |
587 | int ret; | |
588 | ||
589 | hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests, | |
590 | set->reserved_tags); | |
591 | if (!hctx->sched_tags) | |
592 | return -ENOMEM; | |
593 | ||
594 | ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests); | |
595 | if (ret) | |
596 | blk_mq_sched_free_tags(set, hctx, hctx_idx); | |
597 | ||
598 | return ret; | |
599 | } | |
600 | ||
c3e22192 | 601 | /* called in queue's release handler, tagset has gone away */ |
54d5329d | 602 | static void blk_mq_sched_tags_teardown(struct request_queue *q) |
bd166ef1 | 603 | { |
bd166ef1 | 604 | struct blk_mq_hw_ctx *hctx; |
6917ff0b OS |
605 | int i; |
606 | ||
c3e22192 ML |
607 | queue_for_each_hw_ctx(q, hctx, i) { |
608 | if (hctx->sched_tags) { | |
609 | blk_mq_free_rq_map(hctx->sched_tags); | |
610 | hctx->sched_tags = NULL; | |
611 | } | |
612 | } | |
6917ff0b OS |
613 | } |
614 | ||
615 | int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e) | |
616 | { | |
617 | struct blk_mq_hw_ctx *hctx; | |
ee056f98 | 618 | struct elevator_queue *eq; |
6917ff0b OS |
619 | unsigned int i; |
620 | int ret; | |
621 | ||
622 | if (!e) { | |
623 | q->elevator = NULL; | |
32a50fab | 624 | q->nr_requests = q->tag_set->queue_depth; |
6917ff0b OS |
625 | return 0; |
626 | } | |
bd166ef1 JA |
627 | |
628 | /* | |
32825c45 ML |
629 | * Default to double of smaller one between hw queue_depth and 128, |
630 | * since we don't split into sync/async like the old code did. | |
631 | * Additionally, this is a per-hw queue depth. | |
bd166ef1 | 632 | */ |
32825c45 ML |
633 | q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth, |
634 | BLKDEV_MAX_RQ); | |
bd166ef1 | 635 | |
bd166ef1 | 636 | queue_for_each_hw_ctx(q, hctx, i) { |
6917ff0b | 637 | ret = blk_mq_sched_alloc_tags(q, hctx, i); |
bd166ef1 | 638 | if (ret) |
6917ff0b | 639 | goto err; |
bd166ef1 JA |
640 | } |
641 | ||
f9cd4bfe | 642 | ret = e->ops.init_sched(q, e); |
6917ff0b OS |
643 | if (ret) |
644 | goto err; | |
bd166ef1 | 645 | |
d332ce09 OS |
646 | blk_mq_debugfs_register_sched(q); |
647 | ||
648 | queue_for_each_hw_ctx(q, hctx, i) { | |
f9cd4bfe JA |
649 | if (e->ops.init_hctx) { |
650 | ret = e->ops.init_hctx(hctx, i); | |
ee056f98 OS |
651 | if (ret) { |
652 | eq = q->elevator; | |
c3e22192 | 653 | blk_mq_sched_free_requests(q); |
ee056f98 OS |
654 | blk_mq_exit_sched(q, eq); |
655 | kobject_put(&eq->kobj); | |
656 | return ret; | |
657 | } | |
658 | } | |
d332ce09 | 659 | blk_mq_debugfs_register_sched_hctx(q, hctx); |
ee056f98 OS |
660 | } |
661 | ||
bd166ef1 | 662 | return 0; |
bd166ef1 | 663 | |
6917ff0b | 664 | err: |
c3e22192 | 665 | blk_mq_sched_free_requests(q); |
54d5329d OS |
666 | blk_mq_sched_tags_teardown(q); |
667 | q->elevator = NULL; | |
6917ff0b | 668 | return ret; |
bd166ef1 | 669 | } |
d3484991 | 670 | |
c3e22192 ML |
671 | /* |
672 | * called in either blk_queue_cleanup or elevator_switch, tagset | |
673 | * is required for freeing requests | |
674 | */ | |
675 | void blk_mq_sched_free_requests(struct request_queue *q) | |
676 | { | |
677 | struct blk_mq_hw_ctx *hctx; | |
678 | int i; | |
679 | ||
c3e22192 ML |
680 | queue_for_each_hw_ctx(q, hctx, i) { |
681 | if (hctx->sched_tags) | |
682 | blk_mq_free_rqs(q->tag_set, hctx->sched_tags, i); | |
683 | } | |
684 | } | |
685 | ||
54d5329d OS |
686 | void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e) |
687 | { | |
ee056f98 OS |
688 | struct blk_mq_hw_ctx *hctx; |
689 | unsigned int i; | |
690 | ||
d332ce09 OS |
691 | queue_for_each_hw_ctx(q, hctx, i) { |
692 | blk_mq_debugfs_unregister_sched_hctx(hctx); | |
f9cd4bfe JA |
693 | if (e->type->ops.exit_hctx && hctx->sched_data) { |
694 | e->type->ops.exit_hctx(hctx, i); | |
d332ce09 | 695 | hctx->sched_data = NULL; |
ee056f98 OS |
696 | } |
697 | } | |
d332ce09 | 698 | blk_mq_debugfs_unregister_sched(q); |
f9cd4bfe JA |
699 | if (e->type->ops.exit_sched) |
700 | e->type->ops.exit_sched(e); | |
54d5329d OS |
701 | blk_mq_sched_tags_teardown(q); |
702 | q->elevator = NULL; | |
703 | } |