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 | |
97889f9a | 81 | blk_mq_run_hw_queue(hctx, true); |
8e8320c9 JA |
82 | } |
83 | ||
6e6fcbc2 ML |
84 | static int sched_rq_cmp(void *priv, struct list_head *a, struct list_head *b) |
85 | { | |
86 | struct request *rqa = container_of(a, struct request, queuelist); | |
87 | struct request *rqb = container_of(b, struct request, queuelist); | |
88 | ||
89 | return rqa->mq_hctx > rqb->mq_hctx; | |
90 | } | |
91 | ||
92 | static bool blk_mq_dispatch_hctx_list(struct list_head *rq_list) | |
93 | { | |
94 | struct blk_mq_hw_ctx *hctx = | |
95 | list_first_entry(rq_list, struct request, queuelist)->mq_hctx; | |
96 | struct request *rq; | |
97 | LIST_HEAD(hctx_list); | |
98 | unsigned int count = 0; | |
99 | bool ret; | |
100 | ||
101 | list_for_each_entry(rq, rq_list, queuelist) { | |
102 | if (rq->mq_hctx != hctx) { | |
103 | list_cut_before(&hctx_list, rq_list, &rq->queuelist); | |
104 | goto dispatch; | |
105 | } | |
106 | count++; | |
107 | } | |
108 | list_splice_tail_init(rq_list, &hctx_list); | |
109 | ||
110 | dispatch: | |
111 | ret = blk_mq_dispatch_rq_list(hctx, &hctx_list, count); | |
112 | return ret; | |
113 | } | |
114 | ||
a0823421 DA |
115 | #define BLK_MQ_BUDGET_DELAY 3 /* ms units */ |
116 | ||
1f460b63 ML |
117 | /* |
118 | * Only SCSI implements .get_budget and .put_budget, and SCSI restarts | |
119 | * its queue by itself in its completion handler, so we don't need to | |
120 | * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE. | |
28d65729 SQ |
121 | * |
122 | * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to | |
123 | * be run again. This is necessary to avoid starving flushes. | |
1f460b63 | 124 | */ |
6e6fcbc2 | 125 | static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx) |
caf8eb0d ML |
126 | { |
127 | struct request_queue *q = hctx->queue; | |
128 | struct elevator_queue *e = q->elevator; | |
6e6fcbc2 ML |
129 | bool multi_hctxs = false, run_queue = false; |
130 | bool dispatched = false, busy = false; | |
131 | unsigned int max_dispatch; | |
caf8eb0d | 132 | LIST_HEAD(rq_list); |
6e6fcbc2 ML |
133 | int count = 0; |
134 | ||
135 | if (hctx->dispatch_busy) | |
136 | max_dispatch = 1; | |
137 | else | |
138 | max_dispatch = hctx->queue->nr_requests; | |
caf8eb0d ML |
139 | |
140 | do { | |
6e6fcbc2 ML |
141 | struct request *rq; |
142 | ||
f9cd4bfe | 143 | if (e->type->ops.has_work && !e->type->ops.has_work(hctx)) |
caf8eb0d | 144 | break; |
de148297 | 145 | |
28d65729 | 146 | if (!list_empty_careful(&hctx->dispatch)) { |
6e6fcbc2 | 147 | busy = true; |
28d65729 SQ |
148 | break; |
149 | } | |
150 | ||
65c76369 | 151 | if (!blk_mq_get_dispatch_budget(q)) |
1f460b63 | 152 | break; |
de148297 | 153 | |
f9cd4bfe | 154 | rq = e->type->ops.dispatch_request(hctx); |
de148297 | 155 | if (!rq) { |
65c76369 | 156 | blk_mq_put_dispatch_budget(q); |
a0823421 DA |
157 | /* |
158 | * We're releasing without dispatching. Holding the | |
159 | * budget could have blocked any "hctx"s with the | |
160 | * same queue and if we didn't dispatch then there's | |
161 | * no guarantee anyone will kick the queue. Kick it | |
162 | * ourselves. | |
163 | */ | |
6e6fcbc2 | 164 | run_queue = true; |
de148297 | 165 | break; |
de148297 ML |
166 | } |
167 | ||
168 | /* | |
169 | * Now this rq owns the budget which has to be released | |
170 | * if this rq won't be queued to driver via .queue_rq() | |
171 | * in blk_mq_dispatch_rq_list(). | |
172 | */ | |
6e6fcbc2 ML |
173 | list_add_tail(&rq->queuelist, &rq_list); |
174 | if (rq->mq_hctx != hctx) | |
175 | multi_hctxs = true; | |
176 | } while (++count < max_dispatch); | |
177 | ||
178 | if (!count) { | |
179 | if (run_queue) | |
180 | blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY); | |
181 | } else if (multi_hctxs) { | |
182 | /* | |
183 | * Requests from different hctx may be dequeued from some | |
184 | * schedulers, such as bfq and deadline. | |
185 | * | |
186 | * Sort the requests in the list according to their hctx, | |
187 | * dispatch batching requests from same hctx at a time. | |
188 | */ | |
189 | list_sort(NULL, &rq_list, sched_rq_cmp); | |
190 | do { | |
191 | dispatched |= blk_mq_dispatch_hctx_list(&rq_list); | |
192 | } while (!list_empty(&rq_list)); | |
193 | } else { | |
194 | dispatched = blk_mq_dispatch_rq_list(hctx, &rq_list, count); | |
195 | } | |
196 | ||
197 | if (busy) | |
198 | return -EAGAIN; | |
199 | return !!dispatched; | |
200 | } | |
201 | ||
202 | static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx) | |
203 | { | |
204 | int ret; | |
205 | ||
206 | do { | |
207 | ret = __blk_mq_do_dispatch_sched(hctx); | |
208 | } while (ret == 1); | |
28d65729 SQ |
209 | |
210 | return ret; | |
caf8eb0d ML |
211 | } |
212 | ||
b347689f ML |
213 | static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx, |
214 | struct blk_mq_ctx *ctx) | |
215 | { | |
f31967f0 | 216 | unsigned short idx = ctx->index_hw[hctx->type]; |
b347689f ML |
217 | |
218 | if (++idx == hctx->nr_ctx) | |
219 | idx = 0; | |
220 | ||
221 | return hctx->ctxs[idx]; | |
222 | } | |
223 | ||
1f460b63 ML |
224 | /* |
225 | * Only SCSI implements .get_budget and .put_budget, and SCSI restarts | |
226 | * its queue by itself in its completion handler, so we don't need to | |
227 | * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE. | |
28d65729 SQ |
228 | * |
229 | * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to | |
230 | * to be run again. This is necessary to avoid starving flushes. | |
1f460b63 | 231 | */ |
28d65729 | 232 | static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx) |
b347689f ML |
233 | { |
234 | struct request_queue *q = hctx->queue; | |
235 | LIST_HEAD(rq_list); | |
236 | struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from); | |
28d65729 | 237 | int ret = 0; |
445874e8 | 238 | struct request *rq; |
b347689f ML |
239 | |
240 | do { | |
28d65729 SQ |
241 | if (!list_empty_careful(&hctx->dispatch)) { |
242 | ret = -EAGAIN; | |
243 | break; | |
244 | } | |
245 | ||
b347689f ML |
246 | if (!sbitmap_any_bit_set(&hctx->ctx_map)) |
247 | break; | |
248 | ||
65c76369 | 249 | if (!blk_mq_get_dispatch_budget(q)) |
1f460b63 | 250 | break; |
b347689f ML |
251 | |
252 | rq = blk_mq_dequeue_from_ctx(hctx, ctx); | |
253 | if (!rq) { | |
65c76369 | 254 | blk_mq_put_dispatch_budget(q); |
a0823421 DA |
255 | /* |
256 | * We're releasing without dispatching. Holding the | |
257 | * budget could have blocked any "hctx"s with the | |
258 | * same queue and if we didn't dispatch then there's | |
259 | * no guarantee anyone will kick the queue. Kick it | |
260 | * ourselves. | |
261 | */ | |
262 | blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY); | |
b347689f | 263 | break; |
b347689f ML |
264 | } |
265 | ||
266 | /* | |
267 | * Now this rq owns the budget which has to be released | |
268 | * if this rq won't be queued to driver via .queue_rq() | |
269 | * in blk_mq_dispatch_rq_list(). | |
270 | */ | |
271 | list_add(&rq->queuelist, &rq_list); | |
272 | ||
273 | /* round robin for fair dispatch */ | |
274 | ctx = blk_mq_next_ctx(hctx, rq->mq_ctx); | |
275 | ||
1fd40b5e | 276 | } while (blk_mq_dispatch_rq_list(rq->mq_hctx, &rq_list, 1)); |
b347689f ML |
277 | |
278 | WRITE_ONCE(hctx->dispatch_from, ctx); | |
28d65729 | 279 | return ret; |
b347689f ML |
280 | } |
281 | ||
e1b586f2 | 282 | static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx) |
bd166ef1 | 283 | { |
81380ca1 OS |
284 | struct request_queue *q = hctx->queue; |
285 | struct elevator_queue *e = q->elevator; | |
f9cd4bfe | 286 | const bool has_sched_dispatch = e && e->type->ops.dispatch_request; |
28d65729 | 287 | int ret = 0; |
bd166ef1 JA |
288 | LIST_HEAD(rq_list); |
289 | ||
bd166ef1 JA |
290 | /* |
291 | * If we have previous entries on our dispatch list, grab them first for | |
292 | * more fair dispatch. | |
293 | */ | |
294 | if (!list_empty_careful(&hctx->dispatch)) { | |
295 | spin_lock(&hctx->lock); | |
296 | if (!list_empty(&hctx->dispatch)) | |
297 | list_splice_init(&hctx->dispatch, &rq_list); | |
298 | spin_unlock(&hctx->lock); | |
299 | } | |
300 | ||
301 | /* | |
302 | * Only ask the scheduler for requests, if we didn't have residual | |
303 | * requests from the dispatch list. This is to avoid the case where | |
304 | * we only ever dispatch a fraction of the requests available because | |
305 | * of low device queue depth. Once we pull requests out of the IO | |
306 | * scheduler, we can no longer merge or sort them. So it's best to | |
307 | * leave them there for as long as we can. Mark the hw queue as | |
308 | * needing a restart in that case. | |
caf8eb0d ML |
309 | * |
310 | * We want to dispatch from the scheduler if there was nothing | |
311 | * on the dispatch list or we were able to dispatch from the | |
312 | * dispatch list. | |
bd166ef1 | 313 | */ |
c13660a0 | 314 | if (!list_empty(&rq_list)) { |
d38d3515 | 315 | blk_mq_sched_mark_restart_hctx(hctx); |
1fd40b5e | 316 | if (blk_mq_dispatch_rq_list(hctx, &rq_list, 0)) { |
b347689f | 317 | if (has_sched_dispatch) |
28d65729 | 318 | ret = blk_mq_do_dispatch_sched(hctx); |
b347689f | 319 | else |
28d65729 | 320 | ret = blk_mq_do_dispatch_ctx(hctx); |
b347689f | 321 | } |
caf8eb0d | 322 | } else if (has_sched_dispatch) { |
28d65729 | 323 | ret = blk_mq_do_dispatch_sched(hctx); |
6e768717 ML |
324 | } else if (hctx->dispatch_busy) { |
325 | /* dequeue request one by one from sw queue if queue is busy */ | |
28d65729 | 326 | ret = blk_mq_do_dispatch_ctx(hctx); |
caf8eb0d | 327 | } else { |
c13660a0 | 328 | blk_mq_flush_busy_ctxs(hctx, &rq_list); |
1fd40b5e | 329 | blk_mq_dispatch_rq_list(hctx, &rq_list, 0); |
64765a75 | 330 | } |
28d65729 SQ |
331 | |
332 | return ret; | |
333 | } | |
334 | ||
335 | void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx) | |
336 | { | |
337 | struct request_queue *q = hctx->queue; | |
338 | ||
339 | /* RCU or SRCU read lock is needed before checking quiesced flag */ | |
340 | if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q))) | |
341 | return; | |
342 | ||
343 | hctx->run++; | |
344 | ||
345 | /* | |
346 | * A return of -EAGAIN is an indication that hctx->dispatch is not | |
347 | * empty and we must run again in order to avoid starving flushes. | |
348 | */ | |
349 | if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) { | |
350 | if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) | |
351 | blk_mq_run_hw_queue(hctx, true); | |
352 | } | |
bd166ef1 JA |
353 | } |
354 | ||
e4d750c9 | 355 | bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio, |
14ccb66b | 356 | unsigned int nr_segs, struct request **merged_request) |
bd166ef1 JA |
357 | { |
358 | struct request *rq; | |
bd166ef1 | 359 | |
34fe7c05 CH |
360 | switch (elv_merge(q, &rq, bio)) { |
361 | case ELEVATOR_BACK_MERGE: | |
bd166ef1 JA |
362 | if (!blk_mq_sched_allow_merge(q, rq, bio)) |
363 | return false; | |
14ccb66b | 364 | if (!bio_attempt_back_merge(rq, bio, nr_segs)) |
34fe7c05 CH |
365 | return false; |
366 | *merged_request = attempt_back_merge(q, rq); | |
367 | if (!*merged_request) | |
368 | elv_merged_request(q, rq, ELEVATOR_BACK_MERGE); | |
369 | return true; | |
370 | case ELEVATOR_FRONT_MERGE: | |
bd166ef1 JA |
371 | if (!blk_mq_sched_allow_merge(q, rq, bio)) |
372 | return false; | |
14ccb66b | 373 | if (!bio_attempt_front_merge(rq, bio, nr_segs)) |
34fe7c05 CH |
374 | return false; |
375 | *merged_request = attempt_front_merge(q, rq); | |
376 | if (!*merged_request) | |
377 | elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE); | |
378 | return true; | |
bea99a50 KB |
379 | case ELEVATOR_DISCARD_MERGE: |
380 | return bio_attempt_discard_merge(q, rq, bio); | |
34fe7c05 CH |
381 | default: |
382 | return false; | |
bd166ef1 | 383 | } |
bd166ef1 JA |
384 | } |
385 | EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge); | |
386 | ||
9bddeb2a | 387 | /* |
9c558734 JA |
388 | * Iterate list of requests and see if we can merge this bio with any |
389 | * of them. | |
9bddeb2a | 390 | */ |
9c558734 | 391 | bool blk_mq_bio_list_merge(struct request_queue *q, struct list_head *list, |
14ccb66b | 392 | struct bio *bio, unsigned int nr_segs) |
9bddeb2a ML |
393 | { |
394 | struct request *rq; | |
395 | int checked = 8; | |
396 | ||
9c558734 | 397 | list_for_each_entry_reverse(rq, list, queuelist) { |
9bddeb2a ML |
398 | bool merged = false; |
399 | ||
400 | if (!checked--) | |
401 | break; | |
402 | ||
403 | if (!blk_rq_merge_ok(rq, bio)) | |
404 | continue; | |
405 | ||
406 | switch (blk_try_merge(rq, bio)) { | |
407 | case ELEVATOR_BACK_MERGE: | |
408 | if (blk_mq_sched_allow_merge(q, rq, bio)) | |
14ccb66b CH |
409 | merged = bio_attempt_back_merge(rq, bio, |
410 | nr_segs); | |
9bddeb2a ML |
411 | break; |
412 | case ELEVATOR_FRONT_MERGE: | |
413 | if (blk_mq_sched_allow_merge(q, rq, bio)) | |
14ccb66b CH |
414 | merged = bio_attempt_front_merge(rq, bio, |
415 | nr_segs); | |
9bddeb2a ML |
416 | break; |
417 | case ELEVATOR_DISCARD_MERGE: | |
418 | merged = bio_attempt_discard_merge(q, rq, bio); | |
419 | break; | |
420 | default: | |
421 | continue; | |
422 | } | |
423 | ||
9bddeb2a ML |
424 | return merged; |
425 | } | |
426 | ||
427 | return false; | |
428 | } | |
9c558734 JA |
429 | EXPORT_SYMBOL_GPL(blk_mq_bio_list_merge); |
430 | ||
431 | /* | |
432 | * Reverse check our software queue for entries that we could potentially | |
433 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
434 | * too much time checking for merges. | |
435 | */ | |
436 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
c16d6b5a | 437 | struct blk_mq_hw_ctx *hctx, |
14ccb66b CH |
438 | struct blk_mq_ctx *ctx, struct bio *bio, |
439 | unsigned int nr_segs) | |
9c558734 | 440 | { |
c16d6b5a ML |
441 | enum hctx_type type = hctx->type; |
442 | ||
9c558734 JA |
443 | lockdep_assert_held(&ctx->lock); |
444 | ||
14ccb66b | 445 | if (blk_mq_bio_list_merge(q, &ctx->rq_lists[type], bio, nr_segs)) { |
9c558734 JA |
446 | ctx->rq_merged++; |
447 | return true; | |
448 | } | |
449 | ||
450 | return false; | |
451 | } | |
9bddeb2a | 452 | |
14ccb66b CH |
453 | bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio, |
454 | unsigned int nr_segs) | |
bd166ef1 JA |
455 | { |
456 | struct elevator_queue *e = q->elevator; | |
9bddeb2a | 457 | struct blk_mq_ctx *ctx = blk_mq_get_ctx(q); |
8ccdf4a3 | 458 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, bio->bi_opf, ctx); |
9bddeb2a | 459 | bool ret = false; |
c16d6b5a | 460 | enum hctx_type type; |
bd166ef1 | 461 | |
c05f4220 | 462 | if (e && e->type->ops.bio_merge) |
14ccb66b | 463 | return e->type->ops.bio_merge(hctx, bio, nr_segs); |
bd166ef1 | 464 | |
c16d6b5a | 465 | type = hctx->type; |
b04f50ab | 466 | if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) && |
c16d6b5a | 467 | !list_empty_careful(&ctx->rq_lists[type])) { |
9bddeb2a ML |
468 | /* default per sw-queue merge */ |
469 | spin_lock(&ctx->lock); | |
14ccb66b | 470 | ret = blk_mq_attempt_merge(q, hctx, ctx, bio, nr_segs); |
9bddeb2a ML |
471 | spin_unlock(&ctx->lock); |
472 | } | |
473 | ||
9bddeb2a | 474 | return ret; |
bd166ef1 JA |
475 | } |
476 | ||
477 | bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq) | |
478 | { | |
479 | return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq); | |
480 | } | |
481 | EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge); | |
482 | ||
483 | void blk_mq_sched_request_inserted(struct request *rq) | |
484 | { | |
485 | trace_block_rq_insert(rq->q, rq); | |
486 | } | |
487 | EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted); | |
488 | ||
0cacba6c | 489 | static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx, |
a6a252e6 | 490 | bool has_sched, |
0cacba6c | 491 | struct request *rq) |
bd166ef1 | 492 | { |
01e99aec ML |
493 | /* |
494 | * dispatch flush and passthrough rq directly | |
495 | * | |
496 | * passthrough request has to be added to hctx->dispatch directly. | |
497 | * For some reason, device may be in one situation which can't | |
498 | * handle FS request, so STS_RESOURCE is always returned and the | |
499 | * FS request will be added to hctx->dispatch. However passthrough | |
500 | * request may be required at that time for fixing the problem. If | |
501 | * passthrough request is added to scheduler queue, there isn't any | |
502 | * chance to dispatch it given we prioritize requests in hctx->dispatch. | |
503 | */ | |
504 | if ((rq->rq_flags & RQF_FLUSH_SEQ) || blk_rq_is_passthrough(rq)) | |
a6a252e6 | 505 | return true; |
a6a252e6 | 506 | |
923218f6 | 507 | if (has_sched) |
bd166ef1 | 508 | rq->rq_flags |= RQF_SORTED; |
bd166ef1 | 509 | |
a6a252e6 | 510 | return false; |
bd166ef1 | 511 | } |
bd166ef1 | 512 | |
bd6737f1 | 513 | void blk_mq_sched_insert_request(struct request *rq, bool at_head, |
9e97d295 | 514 | bool run_queue, bool async) |
bd6737f1 JA |
515 | { |
516 | struct request_queue *q = rq->q; | |
517 | struct elevator_queue *e = q->elevator; | |
518 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 519 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
bd6737f1 | 520 | |
a6a252e6 ML |
521 | /* flush rq in flush machinery need to be dispatched directly */ |
522 | if (!(rq->rq_flags & RQF_FLUSH_SEQ) && op_is_flush(rq->cmd_flags)) { | |
923218f6 ML |
523 | blk_insert_flush(rq); |
524 | goto run; | |
bd6737f1 JA |
525 | } |
526 | ||
923218f6 ML |
527 | WARN_ON(e && (rq->tag != -1)); |
528 | ||
01e99aec | 529 | if (blk_mq_sched_bypass_insert(hctx, !!e, rq)) { |
cc3200ea ML |
530 | /* |
531 | * Firstly normal IO request is inserted to scheduler queue or | |
532 | * sw queue, meantime we add flush request to dispatch queue( | |
533 | * hctx->dispatch) directly and there is at most one in-flight | |
534 | * flush request for each hw queue, so it doesn't matter to add | |
535 | * flush request to tail or front of the dispatch queue. | |
536 | * | |
537 | * Secondly in case of NCQ, flush request belongs to non-NCQ | |
538 | * command, and queueing it will fail when there is any | |
539 | * in-flight normal IO request(NCQ command). When adding flush | |
540 | * rq to the front of hctx->dispatch, it is easier to introduce | |
541 | * extra time to flush rq's latency because of S_SCHED_RESTART | |
542 | * compared with adding to the tail of dispatch queue, then | |
543 | * chance of flush merge is increased, and less flush requests | |
544 | * will be issued to controller. It is observed that ~10% time | |
545 | * is saved in blktests block/004 on disk attached to AHCI/NCQ | |
546 | * drive when adding flush rq to the front of hctx->dispatch. | |
547 | * | |
548 | * Simply queue flush rq to the front of hctx->dispatch so that | |
549 | * intensive flush workloads can benefit in case of NCQ HW. | |
550 | */ | |
551 | at_head = (rq->rq_flags & RQF_FLUSH_SEQ) ? true : at_head; | |
01e99aec | 552 | blk_mq_request_bypass_insert(rq, at_head, false); |
0cacba6c | 553 | goto run; |
01e99aec | 554 | } |
0cacba6c | 555 | |
f9cd4bfe | 556 | if (e && e->type->ops.insert_requests) { |
bd6737f1 JA |
557 | LIST_HEAD(list); |
558 | ||
559 | list_add(&rq->queuelist, &list); | |
f9cd4bfe | 560 | e->type->ops.insert_requests(hctx, &list, at_head); |
bd6737f1 JA |
561 | } else { |
562 | spin_lock(&ctx->lock); | |
563 | __blk_mq_insert_request(hctx, rq, at_head); | |
564 | spin_unlock(&ctx->lock); | |
565 | } | |
566 | ||
0cacba6c | 567 | run: |
bd6737f1 JA |
568 | if (run_queue) |
569 | blk_mq_run_hw_queue(hctx, async); | |
570 | } | |
571 | ||
67cae4c9 | 572 | void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx *hctx, |
bd6737f1 JA |
573 | struct blk_mq_ctx *ctx, |
574 | struct list_head *list, bool run_queue_async) | |
575 | { | |
f9afca4d | 576 | struct elevator_queue *e; |
e87eb301 ML |
577 | struct request_queue *q = hctx->queue; |
578 | ||
579 | /* | |
580 | * blk_mq_sched_insert_requests() is called from flush plug | |
581 | * context only, and hold one usage counter to prevent queue | |
582 | * from being released. | |
583 | */ | |
584 | percpu_ref_get(&q->q_usage_counter); | |
bd6737f1 | 585 | |
f9afca4d | 586 | e = hctx->queue->elevator; |
f9cd4bfe JA |
587 | if (e && e->type->ops.insert_requests) |
588 | e->type->ops.insert_requests(hctx, list, false); | |
6ce3dd6e ML |
589 | else { |
590 | /* | |
591 | * try to issue requests directly if the hw queue isn't | |
592 | * busy in case of 'none' scheduler, and this way may save | |
593 | * us one extra enqueue & dequeue to sw queue. | |
594 | */ | |
fd9c40f6 | 595 | if (!hctx->dispatch_busy && !e && !run_queue_async) { |
6ce3dd6e | 596 | blk_mq_try_issue_list_directly(hctx, list); |
fd9c40f6 | 597 | if (list_empty(list)) |
e87eb301 | 598 | goto out; |
fd9c40f6 BVA |
599 | } |
600 | blk_mq_insert_requests(hctx, ctx, list); | |
6ce3dd6e | 601 | } |
bd6737f1 JA |
602 | |
603 | blk_mq_run_hw_queue(hctx, run_queue_async); | |
e87eb301 ML |
604 | out: |
605 | percpu_ref_put(&q->q_usage_counter); | |
bd6737f1 JA |
606 | } |
607 | ||
bd166ef1 JA |
608 | static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set, |
609 | struct blk_mq_hw_ctx *hctx, | |
610 | unsigned int hctx_idx) | |
611 | { | |
612 | if (hctx->sched_tags) { | |
613 | blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx); | |
614 | blk_mq_free_rq_map(hctx->sched_tags); | |
615 | hctx->sched_tags = NULL; | |
616 | } | |
617 | } | |
618 | ||
6917ff0b OS |
619 | static int blk_mq_sched_alloc_tags(struct request_queue *q, |
620 | struct blk_mq_hw_ctx *hctx, | |
621 | unsigned int hctx_idx) | |
622 | { | |
623 | struct blk_mq_tag_set *set = q->tag_set; | |
624 | int ret; | |
625 | ||
626 | hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests, | |
627 | set->reserved_tags); | |
628 | if (!hctx->sched_tags) | |
629 | return -ENOMEM; | |
630 | ||
631 | ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests); | |
632 | if (ret) | |
633 | blk_mq_sched_free_tags(set, hctx, hctx_idx); | |
634 | ||
635 | return ret; | |
636 | } | |
637 | ||
c3e22192 | 638 | /* called in queue's release handler, tagset has gone away */ |
54d5329d | 639 | static void blk_mq_sched_tags_teardown(struct request_queue *q) |
bd166ef1 | 640 | { |
bd166ef1 | 641 | struct blk_mq_hw_ctx *hctx; |
6917ff0b OS |
642 | int i; |
643 | ||
c3e22192 ML |
644 | queue_for_each_hw_ctx(q, hctx, i) { |
645 | if (hctx->sched_tags) { | |
646 | blk_mq_free_rq_map(hctx->sched_tags); | |
647 | hctx->sched_tags = NULL; | |
648 | } | |
649 | } | |
6917ff0b OS |
650 | } |
651 | ||
652 | int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e) | |
653 | { | |
654 | struct blk_mq_hw_ctx *hctx; | |
ee056f98 | 655 | struct elevator_queue *eq; |
6917ff0b OS |
656 | unsigned int i; |
657 | int ret; | |
658 | ||
659 | if (!e) { | |
660 | q->elevator = NULL; | |
32a50fab | 661 | q->nr_requests = q->tag_set->queue_depth; |
6917ff0b OS |
662 | return 0; |
663 | } | |
bd166ef1 JA |
664 | |
665 | /* | |
32825c45 ML |
666 | * Default to double of smaller one between hw queue_depth and 128, |
667 | * since we don't split into sync/async like the old code did. | |
668 | * Additionally, this is a per-hw queue depth. | |
bd166ef1 | 669 | */ |
32825c45 ML |
670 | q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth, |
671 | BLKDEV_MAX_RQ); | |
bd166ef1 | 672 | |
bd166ef1 | 673 | queue_for_each_hw_ctx(q, hctx, i) { |
6917ff0b | 674 | ret = blk_mq_sched_alloc_tags(q, hctx, i); |
bd166ef1 | 675 | if (ret) |
6917ff0b | 676 | goto err; |
bd166ef1 JA |
677 | } |
678 | ||
f9cd4bfe | 679 | ret = e->ops.init_sched(q, e); |
6917ff0b OS |
680 | if (ret) |
681 | goto err; | |
bd166ef1 | 682 | |
d332ce09 OS |
683 | blk_mq_debugfs_register_sched(q); |
684 | ||
685 | queue_for_each_hw_ctx(q, hctx, i) { | |
f9cd4bfe JA |
686 | if (e->ops.init_hctx) { |
687 | ret = e->ops.init_hctx(hctx, i); | |
ee056f98 OS |
688 | if (ret) { |
689 | eq = q->elevator; | |
c3e22192 | 690 | blk_mq_sched_free_requests(q); |
ee056f98 OS |
691 | blk_mq_exit_sched(q, eq); |
692 | kobject_put(&eq->kobj); | |
693 | return ret; | |
694 | } | |
695 | } | |
d332ce09 | 696 | blk_mq_debugfs_register_sched_hctx(q, hctx); |
ee056f98 OS |
697 | } |
698 | ||
bd166ef1 | 699 | return 0; |
bd166ef1 | 700 | |
6917ff0b | 701 | err: |
c3e22192 | 702 | blk_mq_sched_free_requests(q); |
54d5329d OS |
703 | blk_mq_sched_tags_teardown(q); |
704 | q->elevator = NULL; | |
6917ff0b | 705 | return ret; |
bd166ef1 | 706 | } |
d3484991 | 707 | |
c3e22192 ML |
708 | /* |
709 | * called in either blk_queue_cleanup or elevator_switch, tagset | |
710 | * is required for freeing requests | |
711 | */ | |
712 | void blk_mq_sched_free_requests(struct request_queue *q) | |
713 | { | |
714 | struct blk_mq_hw_ctx *hctx; | |
715 | int i; | |
716 | ||
c3e22192 ML |
717 | queue_for_each_hw_ctx(q, hctx, i) { |
718 | if (hctx->sched_tags) | |
719 | blk_mq_free_rqs(q->tag_set, hctx->sched_tags, i); | |
720 | } | |
721 | } | |
722 | ||
54d5329d OS |
723 | void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e) |
724 | { | |
ee056f98 OS |
725 | struct blk_mq_hw_ctx *hctx; |
726 | unsigned int i; | |
727 | ||
d332ce09 OS |
728 | queue_for_each_hw_ctx(q, hctx, i) { |
729 | blk_mq_debugfs_unregister_sched_hctx(hctx); | |
f9cd4bfe JA |
730 | if (e->type->ops.exit_hctx && hctx->sched_data) { |
731 | e->type->ops.exit_hctx(hctx, i); | |
d332ce09 | 732 | hctx->sched_data = NULL; |
ee056f98 OS |
733 | } |
734 | } | |
d332ce09 | 735 | blk_mq_debugfs_unregister_sched(q); |
f9cd4bfe JA |
736 | if (e->type->ops.exit_sched) |
737 | e->type->ops.exit_sched(e); | |
54d5329d OS |
738 | blk_mq_sched_tags_teardown(q); |
739 | q->elevator = NULL; | |
740 | } |