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> | |
6e6fcbc2 | 9 | #include <linux/list_sort.h> |
bd166ef1 JA |
10 | |
11 | #include <trace/events/block.h> | |
12 | ||
13 | #include "blk.h" | |
14 | #include "blk-mq.h" | |
d332ce09 | 15 | #include "blk-mq-debugfs.h" |
bd166ef1 | 16 | #include "blk-mq-sched.h" |
bd166ef1 JA |
17 | #include "blk-wbt.h" |
18 | ||
8e8320c9 | 19 | /* |
c31e76bc | 20 | * Mark a hardware queue as needing a restart. |
8e8320c9 | 21 | */ |
7211aef8 | 22 | void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx) |
8e8320c9 JA |
23 | { |
24 | if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
25 | return; | |
26 | ||
97889f9a | 27 | set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state); |
8e8320c9 | 28 | } |
7211aef8 | 29 | EXPORT_SYMBOL_GPL(blk_mq_sched_mark_restart_hctx); |
8e8320c9 | 30 | |
e9ea1596 | 31 | void __blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx) |
8e8320c9 | 32 | { |
97889f9a | 33 | clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state); |
8e8320c9 | 34 | |
d7d8535f ML |
35 | /* |
36 | * Order clearing SCHED_RESTART and list_empty_careful(&hctx->dispatch) | |
37 | * in blk_mq_run_hw_queue(). Its pair is the barrier in | |
38 | * blk_mq_dispatch_rq_list(). So dispatch code won't see SCHED_RESTART, | |
39 | * meantime new request added to hctx->dispatch is missed to check in | |
40 | * blk_mq_run_hw_queue(). | |
41 | */ | |
42 | smp_mb(); | |
43 | ||
97889f9a | 44 | blk_mq_run_hw_queue(hctx, true); |
8e8320c9 JA |
45 | } |
46 | ||
4f0f586b ST |
47 | static int sched_rq_cmp(void *priv, const struct list_head *a, |
48 | const struct list_head *b) | |
6e6fcbc2 ML |
49 | { |
50 | struct request *rqa = container_of(a, struct request, queuelist); | |
51 | struct request *rqb = container_of(b, struct request, queuelist); | |
52 | ||
53 | return rqa->mq_hctx > rqb->mq_hctx; | |
54 | } | |
55 | ||
56 | static bool blk_mq_dispatch_hctx_list(struct list_head *rq_list) | |
57 | { | |
58 | struct blk_mq_hw_ctx *hctx = | |
59 | list_first_entry(rq_list, struct request, queuelist)->mq_hctx; | |
60 | struct request *rq; | |
61 | LIST_HEAD(hctx_list); | |
62 | unsigned int count = 0; | |
6e6fcbc2 ML |
63 | |
64 | list_for_each_entry(rq, rq_list, queuelist) { | |
65 | if (rq->mq_hctx != hctx) { | |
66 | list_cut_before(&hctx_list, rq_list, &rq->queuelist); | |
67 | goto dispatch; | |
68 | } | |
69 | count++; | |
70 | } | |
71 | list_splice_tail_init(rq_list, &hctx_list); | |
72 | ||
73 | dispatch: | |
106e71c5 | 74 | return blk_mq_dispatch_rq_list(hctx, &hctx_list, count); |
6e6fcbc2 ML |
75 | } |
76 | ||
a0823421 DA |
77 | #define BLK_MQ_BUDGET_DELAY 3 /* ms units */ |
78 | ||
1f460b63 ML |
79 | /* |
80 | * Only SCSI implements .get_budget and .put_budget, and SCSI restarts | |
81 | * its queue by itself in its completion handler, so we don't need to | |
01542f65 | 82 | * restart queue if .get_budget() fails to get the budget. |
28d65729 SQ |
83 | * |
84 | * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to | |
85 | * be run again. This is necessary to avoid starving flushes. | |
1f460b63 | 86 | */ |
6e6fcbc2 | 87 | static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx) |
caf8eb0d ML |
88 | { |
89 | struct request_queue *q = hctx->queue; | |
90 | struct elevator_queue *e = q->elevator; | |
6e6fcbc2 ML |
91 | bool multi_hctxs = false, run_queue = false; |
92 | bool dispatched = false, busy = false; | |
93 | unsigned int max_dispatch; | |
caf8eb0d | 94 | LIST_HEAD(rq_list); |
6e6fcbc2 ML |
95 | int count = 0; |
96 | ||
97 | if (hctx->dispatch_busy) | |
98 | max_dispatch = 1; | |
99 | else | |
100 | max_dispatch = hctx->queue->nr_requests; | |
caf8eb0d ML |
101 | |
102 | do { | |
6e6fcbc2 | 103 | struct request *rq; |
2a5a24aa | 104 | int budget_token; |
6e6fcbc2 | 105 | |
f9cd4bfe | 106 | if (e->type->ops.has_work && !e->type->ops.has_work(hctx)) |
caf8eb0d | 107 | break; |
de148297 | 108 | |
28d65729 | 109 | if (!list_empty_careful(&hctx->dispatch)) { |
6e6fcbc2 | 110 | busy = true; |
28d65729 SQ |
111 | break; |
112 | } | |
113 | ||
2a5a24aa ML |
114 | budget_token = blk_mq_get_dispatch_budget(q); |
115 | if (budget_token < 0) | |
1f460b63 | 116 | break; |
de148297 | 117 | |
f9cd4bfe | 118 | rq = e->type->ops.dispatch_request(hctx); |
de148297 | 119 | if (!rq) { |
2a5a24aa | 120 | blk_mq_put_dispatch_budget(q, budget_token); |
a0823421 DA |
121 | /* |
122 | * We're releasing without dispatching. Holding the | |
123 | * budget could have blocked any "hctx"s with the | |
124 | * same queue and if we didn't dispatch then there's | |
125 | * no guarantee anyone will kick the queue. Kick it | |
126 | * ourselves. | |
127 | */ | |
6e6fcbc2 | 128 | run_queue = true; |
de148297 | 129 | break; |
de148297 ML |
130 | } |
131 | ||
2a5a24aa ML |
132 | blk_mq_set_rq_budget_token(rq, budget_token); |
133 | ||
de148297 ML |
134 | /* |
135 | * Now this rq owns the budget which has to be released | |
136 | * if this rq won't be queued to driver via .queue_rq() | |
137 | * in blk_mq_dispatch_rq_list(). | |
138 | */ | |
6e6fcbc2 | 139 | list_add_tail(&rq->queuelist, &rq_list); |
61347154 | 140 | count++; |
6e6fcbc2 ML |
141 | if (rq->mq_hctx != hctx) |
142 | multi_hctxs = true; | |
61347154 JK |
143 | |
144 | /* | |
145 | * If we cannot get tag for the request, stop dequeueing | |
146 | * requests from the IO scheduler. We are unlikely to be able | |
147 | * to submit them anyway and it creates false impression for | |
148 | * scheduling heuristics that the device can take more IO. | |
149 | */ | |
150 | if (!blk_mq_get_driver_tag(rq)) | |
151 | break; | |
152 | } while (count < max_dispatch); | |
6e6fcbc2 ML |
153 | |
154 | if (!count) { | |
155 | if (run_queue) | |
156 | blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY); | |
157 | } else if (multi_hctxs) { | |
158 | /* | |
159 | * Requests from different hctx may be dequeued from some | |
160 | * schedulers, such as bfq and deadline. | |
161 | * | |
162 | * Sort the requests in the list according to their hctx, | |
163 | * dispatch batching requests from same hctx at a time. | |
164 | */ | |
165 | list_sort(NULL, &rq_list, sched_rq_cmp); | |
166 | do { | |
167 | dispatched |= blk_mq_dispatch_hctx_list(&rq_list); | |
168 | } while (!list_empty(&rq_list)); | |
169 | } else { | |
170 | dispatched = blk_mq_dispatch_rq_list(hctx, &rq_list, count); | |
171 | } | |
172 | ||
173 | if (busy) | |
174 | return -EAGAIN; | |
175 | return !!dispatched; | |
176 | } | |
177 | ||
178 | static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx) | |
179 | { | |
572299f0 | 180 | unsigned long end = jiffies + HZ; |
6e6fcbc2 ML |
181 | int ret; |
182 | ||
183 | do { | |
184 | ret = __blk_mq_do_dispatch_sched(hctx); | |
572299f0 SK |
185 | if (ret != 1) |
186 | break; | |
187 | if (need_resched() || time_is_before_jiffies(end)) { | |
188 | blk_mq_delay_run_hw_queue(hctx, 0); | |
189 | break; | |
190 | } | |
191 | } while (1); | |
28d65729 SQ |
192 | |
193 | return ret; | |
caf8eb0d ML |
194 | } |
195 | ||
b347689f ML |
196 | static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx, |
197 | struct blk_mq_ctx *ctx) | |
198 | { | |
f31967f0 | 199 | unsigned short idx = ctx->index_hw[hctx->type]; |
b347689f ML |
200 | |
201 | if (++idx == hctx->nr_ctx) | |
202 | idx = 0; | |
203 | ||
204 | return hctx->ctxs[idx]; | |
205 | } | |
206 | ||
1f460b63 ML |
207 | /* |
208 | * Only SCSI implements .get_budget and .put_budget, and SCSI restarts | |
209 | * its queue by itself in its completion handler, so we don't need to | |
01542f65 | 210 | * restart queue if .get_budget() fails to get the budget. |
28d65729 SQ |
211 | * |
212 | * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to | |
c4aecaa2 | 213 | * be run again. This is necessary to avoid starving flushes. |
1f460b63 | 214 | */ |
28d65729 | 215 | static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx) |
b347689f ML |
216 | { |
217 | struct request_queue *q = hctx->queue; | |
218 | LIST_HEAD(rq_list); | |
219 | struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from); | |
28d65729 | 220 | int ret = 0; |
445874e8 | 221 | struct request *rq; |
b347689f ML |
222 | |
223 | do { | |
2a5a24aa ML |
224 | int budget_token; |
225 | ||
28d65729 SQ |
226 | if (!list_empty_careful(&hctx->dispatch)) { |
227 | ret = -EAGAIN; | |
228 | break; | |
229 | } | |
230 | ||
b347689f ML |
231 | if (!sbitmap_any_bit_set(&hctx->ctx_map)) |
232 | break; | |
233 | ||
2a5a24aa ML |
234 | budget_token = blk_mq_get_dispatch_budget(q); |
235 | if (budget_token < 0) | |
1f460b63 | 236 | break; |
b347689f ML |
237 | |
238 | rq = blk_mq_dequeue_from_ctx(hctx, ctx); | |
239 | if (!rq) { | |
2a5a24aa | 240 | blk_mq_put_dispatch_budget(q, budget_token); |
a0823421 DA |
241 | /* |
242 | * We're releasing without dispatching. Holding the | |
243 | * budget could have blocked any "hctx"s with the | |
244 | * same queue and if we didn't dispatch then there's | |
245 | * no guarantee anyone will kick the queue. Kick it | |
246 | * ourselves. | |
247 | */ | |
248 | blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY); | |
b347689f | 249 | break; |
b347689f ML |
250 | } |
251 | ||
2a5a24aa ML |
252 | blk_mq_set_rq_budget_token(rq, budget_token); |
253 | ||
b347689f ML |
254 | /* |
255 | * Now this rq owns the budget which has to be released | |
256 | * if this rq won't be queued to driver via .queue_rq() | |
257 | * in blk_mq_dispatch_rq_list(). | |
258 | */ | |
259 | list_add(&rq->queuelist, &rq_list); | |
260 | ||
261 | /* round robin for fair dispatch */ | |
262 | ctx = blk_mq_next_ctx(hctx, rq->mq_ctx); | |
263 | ||
1fd40b5e | 264 | } while (blk_mq_dispatch_rq_list(rq->mq_hctx, &rq_list, 1)); |
b347689f ML |
265 | |
266 | WRITE_ONCE(hctx->dispatch_from, ctx); | |
28d65729 | 267 | return ret; |
b347689f ML |
268 | } |
269 | ||
e1b586f2 | 270 | static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx) |
bd166ef1 | 271 | { |
89ea5ceb | 272 | bool need_dispatch = false; |
bd166ef1 JA |
273 | LIST_HEAD(rq_list); |
274 | ||
bd166ef1 JA |
275 | /* |
276 | * If we have previous entries on our dispatch list, grab them first for | |
277 | * more fair dispatch. | |
278 | */ | |
279 | if (!list_empty_careful(&hctx->dispatch)) { | |
280 | spin_lock(&hctx->lock); | |
281 | if (!list_empty(&hctx->dispatch)) | |
282 | list_splice_init(&hctx->dispatch, &rq_list); | |
283 | spin_unlock(&hctx->lock); | |
284 | } | |
285 | ||
286 | /* | |
287 | * Only ask the scheduler for requests, if we didn't have residual | |
288 | * requests from the dispatch list. This is to avoid the case where | |
289 | * we only ever dispatch a fraction of the requests available because | |
290 | * of low device queue depth. Once we pull requests out of the IO | |
291 | * scheduler, we can no longer merge or sort them. So it's best to | |
292 | * leave them there for as long as we can. Mark the hw queue as | |
293 | * needing a restart in that case. | |
caf8eb0d ML |
294 | * |
295 | * We want to dispatch from the scheduler if there was nothing | |
296 | * on the dispatch list or we were able to dispatch from the | |
297 | * dispatch list. | |
bd166ef1 | 298 | */ |
c13660a0 | 299 | if (!list_empty(&rq_list)) { |
d38d3515 | 300 | blk_mq_sched_mark_restart_hctx(hctx); |
89ea5ceb CH |
301 | if (!blk_mq_dispatch_rq_list(hctx, &rq_list, 0)) |
302 | return 0; | |
303 | need_dispatch = true; | |
caf8eb0d | 304 | } else { |
89ea5ceb | 305 | need_dispatch = hctx->dispatch_busy; |
64765a75 | 306 | } |
28d65729 | 307 | |
89ea5ceb CH |
308 | if (hctx->queue->elevator) |
309 | return blk_mq_do_dispatch_sched(hctx); | |
310 | ||
311 | /* dequeue request one by one from sw queue if queue is busy */ | |
312 | if (need_dispatch) | |
313 | return blk_mq_do_dispatch_ctx(hctx); | |
314 | blk_mq_flush_busy_ctxs(hctx, &rq_list); | |
315 | blk_mq_dispatch_rq_list(hctx, &rq_list, 0); | |
316 | return 0; | |
28d65729 SQ |
317 | } |
318 | ||
319 | void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx) | |
320 | { | |
321 | struct request_queue *q = hctx->queue; | |
322 | ||
323 | /* RCU or SRCU read lock is needed before checking quiesced flag */ | |
324 | if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q))) | |
325 | return; | |
326 | ||
28d65729 SQ |
327 | /* |
328 | * A return of -EAGAIN is an indication that hctx->dispatch is not | |
329 | * empty and we must run again in order to avoid starving flushes. | |
330 | */ | |
331 | if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) { | |
332 | if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) | |
333 | blk_mq_run_hw_queue(hctx, true); | |
334 | } | |
bd166ef1 JA |
335 | } |
336 | ||
179ae84f | 337 | bool blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio, |
14ccb66b | 338 | unsigned int nr_segs) |
bd166ef1 JA |
339 | { |
340 | struct elevator_queue *e = q->elevator; | |
efed9a33 OS |
341 | struct blk_mq_ctx *ctx; |
342 | struct blk_mq_hw_ctx *hctx; | |
9bddeb2a | 343 | bool ret = false; |
c16d6b5a | 344 | enum hctx_type type; |
bd166ef1 | 345 | |
900e0807 JA |
346 | if (e && e->type->ops.bio_merge) { |
347 | ret = e->type->ops.bio_merge(q, bio, nr_segs); | |
348 | goto out_put; | |
349 | } | |
bd166ef1 | 350 | |
efed9a33 OS |
351 | ctx = blk_mq_get_ctx(q); |
352 | hctx = blk_mq_map_queue(q, bio->bi_opf, ctx); | |
c16d6b5a | 353 | type = hctx->type; |
cdfcef9e BW |
354 | if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE) || |
355 | list_empty_careful(&ctx->rq_lists[type])) | |
900e0807 | 356 | goto out_put; |
cdfcef9e BW |
357 | |
358 | /* default per sw-queue merge */ | |
359 | spin_lock(&ctx->lock); | |
360 | /* | |
361 | * Reverse check our software queue for entries that we could | |
362 | * potentially merge with. Currently includes a hand-wavy stop | |
363 | * count of 8, to not spend too much time checking for merges. | |
364 | */ | |
9a14d6ce | 365 | if (blk_bio_list_merge(q, &ctx->rq_lists[type], bio, nr_segs)) |
cdfcef9e | 366 | ret = true; |
9bddeb2a | 367 | |
cdfcef9e | 368 | spin_unlock(&ctx->lock); |
900e0807 | 369 | out_put: |
9bddeb2a | 370 | return ret; |
bd166ef1 JA |
371 | } |
372 | ||
fd2ef39c JK |
373 | bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq, |
374 | struct list_head *free) | |
bd166ef1 | 375 | { |
fd2ef39c | 376 | return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq, free); |
bd166ef1 JA |
377 | } |
378 | EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge); | |
379 | ||
d99a6bb3 JG |
380 | static int blk_mq_sched_alloc_map_and_rqs(struct request_queue *q, |
381 | struct blk_mq_hw_ctx *hctx, | |
382 | unsigned int hctx_idx) | |
6917ff0b | 383 | { |
079a2e3e JG |
384 | if (blk_mq_is_shared_tags(q->tag_set->flags)) { |
385 | hctx->sched_tags = q->sched_shared_tags; | |
e155b0c2 JG |
386 | return 0; |
387 | } | |
388 | ||
63064be1 JG |
389 | hctx->sched_tags = blk_mq_alloc_map_and_rqs(q->tag_set, hctx_idx, |
390 | q->nr_requests); | |
6917ff0b | 391 | |
6917ff0b OS |
392 | if (!hctx->sched_tags) |
393 | return -ENOMEM; | |
63064be1 | 394 | return 0; |
6917ff0b OS |
395 | } |
396 | ||
079a2e3e | 397 | static void blk_mq_exit_sched_shared_tags(struct request_queue *queue) |
e155b0c2 | 398 | { |
079a2e3e JG |
399 | blk_mq_free_rq_map(queue->sched_shared_tags); |
400 | queue->sched_shared_tags = NULL; | |
e155b0c2 JG |
401 | } |
402 | ||
c3e22192 | 403 | /* called in queue's release handler, tagset has gone away */ |
e155b0c2 | 404 | static void blk_mq_sched_tags_teardown(struct request_queue *q, unsigned int flags) |
bd166ef1 | 405 | { |
bd166ef1 | 406 | struct blk_mq_hw_ctx *hctx; |
4f481208 | 407 | unsigned long i; |
6917ff0b | 408 | |
c3e22192 ML |
409 | queue_for_each_hw_ctx(q, hctx, i) { |
410 | if (hctx->sched_tags) { | |
8bdf7b3f | 411 | if (!blk_mq_is_shared_tags(flags)) |
e155b0c2 | 412 | blk_mq_free_rq_map(hctx->sched_tags); |
c3e22192 ML |
413 | hctx->sched_tags = NULL; |
414 | } | |
415 | } | |
e155b0c2 | 416 | |
079a2e3e JG |
417 | if (blk_mq_is_shared_tags(flags)) |
418 | blk_mq_exit_sched_shared_tags(q); | |
6917ff0b OS |
419 | } |
420 | ||
079a2e3e | 421 | static int blk_mq_init_sched_shared_tags(struct request_queue *queue) |
d97e594c JG |
422 | { |
423 | struct blk_mq_tag_set *set = queue->tag_set; | |
d97e594c JG |
424 | |
425 | /* | |
426 | * Set initial depth at max so that we don't need to reallocate for | |
427 | * updating nr_requests. | |
428 | */ | |
079a2e3e | 429 | queue->sched_shared_tags = blk_mq_alloc_map_and_rqs(set, |
e155b0c2 JG |
430 | BLK_MQ_NO_HCTX_IDX, |
431 | MAX_SCHED_RQ); | |
079a2e3e | 432 | if (!queue->sched_shared_tags) |
e155b0c2 | 433 | return -ENOMEM; |
d97e594c | 434 | |
079a2e3e | 435 | blk_mq_tag_update_sched_shared_tags(queue); |
d97e594c JG |
436 | |
437 | return 0; | |
438 | } | |
439 | ||
8ed40ee3 | 440 | /* caller must have a reference to @e, will grab another one if successful */ |
6917ff0b OS |
441 | int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e) |
442 | { | |
4f481208 | 443 | unsigned int flags = q->tag_set->flags; |
6917ff0b | 444 | struct blk_mq_hw_ctx *hctx; |
ee056f98 | 445 | struct elevator_queue *eq; |
4f481208 | 446 | unsigned long i; |
6917ff0b OS |
447 | int ret; |
448 | ||
bd166ef1 | 449 | /* |
32825c45 ML |
450 | * Default to double of smaller one between hw queue_depth and 128, |
451 | * since we don't split into sync/async like the old code did. | |
452 | * Additionally, this is a per-hw queue depth. | |
bd166ef1 | 453 | */ |
32825c45 | 454 | q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth, |
d2a27964 | 455 | BLKDEV_DEFAULT_RQ); |
bd166ef1 | 456 | |
079a2e3e JG |
457 | if (blk_mq_is_shared_tags(flags)) { |
458 | ret = blk_mq_init_sched_shared_tags(q); | |
bd166ef1 | 459 | if (ret) |
e155b0c2 | 460 | return ret; |
d97e594c JG |
461 | } |
462 | ||
e155b0c2 JG |
463 | queue_for_each_hw_ctx(q, hctx, i) { |
464 | ret = blk_mq_sched_alloc_map_and_rqs(q, hctx, i); | |
d97e594c | 465 | if (ret) |
d99a6bb3 | 466 | goto err_free_map_and_rqs; |
bd166ef1 JA |
467 | } |
468 | ||
f9cd4bfe | 469 | ret = e->ops.init_sched(q, e); |
6917ff0b | 470 | if (ret) |
e155b0c2 | 471 | goto err_free_map_and_rqs; |
bd166ef1 | 472 | |
5cf9c91b | 473 | mutex_lock(&q->debugfs_mutex); |
d332ce09 | 474 | blk_mq_debugfs_register_sched(q); |
5cf9c91b | 475 | mutex_unlock(&q->debugfs_mutex); |
d332ce09 OS |
476 | |
477 | queue_for_each_hw_ctx(q, hctx, i) { | |
f9cd4bfe JA |
478 | if (e->ops.init_hctx) { |
479 | ret = e->ops.init_hctx(hctx, i); | |
ee056f98 OS |
480 | if (ret) { |
481 | eq = q->elevator; | |
1820f4f0 | 482 | blk_mq_sched_free_rqs(q); |
ee056f98 OS |
483 | blk_mq_exit_sched(q, eq); |
484 | kobject_put(&eq->kobj); | |
485 | return ret; | |
486 | } | |
487 | } | |
5cf9c91b | 488 | mutex_lock(&q->debugfs_mutex); |
d332ce09 | 489 | blk_mq_debugfs_register_sched_hctx(q, hctx); |
5cf9c91b | 490 | mutex_unlock(&q->debugfs_mutex); |
ee056f98 OS |
491 | } |
492 | ||
bd166ef1 | 493 | return 0; |
bd166ef1 | 494 | |
d99a6bb3 | 495 | err_free_map_and_rqs: |
1820f4f0 | 496 | blk_mq_sched_free_rqs(q); |
e155b0c2 JG |
497 | blk_mq_sched_tags_teardown(q, flags); |
498 | ||
54d5329d | 499 | q->elevator = NULL; |
6917ff0b | 500 | return ret; |
bd166ef1 | 501 | } |
d3484991 | 502 | |
c3e22192 ML |
503 | /* |
504 | * called in either blk_queue_cleanup or elevator_switch, tagset | |
505 | * is required for freeing requests | |
506 | */ | |
1820f4f0 | 507 | void blk_mq_sched_free_rqs(struct request_queue *q) |
c3e22192 ML |
508 | { |
509 | struct blk_mq_hw_ctx *hctx; | |
4f481208 | 510 | unsigned long i; |
c3e22192 | 511 | |
079a2e3e JG |
512 | if (blk_mq_is_shared_tags(q->tag_set->flags)) { |
513 | blk_mq_free_rqs(q->tag_set, q->sched_shared_tags, | |
e155b0c2 JG |
514 | BLK_MQ_NO_HCTX_IDX); |
515 | } else { | |
516 | queue_for_each_hw_ctx(q, hctx, i) { | |
517 | if (hctx->sched_tags) | |
518 | blk_mq_free_rqs(q->tag_set, | |
519 | hctx->sched_tags, i); | |
520 | } | |
c3e22192 ML |
521 | } |
522 | } | |
523 | ||
54d5329d OS |
524 | void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e) |
525 | { | |
ee056f98 | 526 | struct blk_mq_hw_ctx *hctx; |
4f481208 | 527 | unsigned long i; |
f0c1c4d2 | 528 | unsigned int flags = 0; |
ee056f98 | 529 | |
d332ce09 | 530 | queue_for_each_hw_ctx(q, hctx, i) { |
5cf9c91b | 531 | mutex_lock(&q->debugfs_mutex); |
d332ce09 | 532 | blk_mq_debugfs_unregister_sched_hctx(hctx); |
5cf9c91b CH |
533 | mutex_unlock(&q->debugfs_mutex); |
534 | ||
f9cd4bfe JA |
535 | if (e->type->ops.exit_hctx && hctx->sched_data) { |
536 | e->type->ops.exit_hctx(hctx, i); | |
d332ce09 | 537 | hctx->sched_data = NULL; |
ee056f98 | 538 | } |
f0c1c4d2 | 539 | flags = hctx->flags; |
ee056f98 | 540 | } |
5cf9c91b CH |
541 | |
542 | mutex_lock(&q->debugfs_mutex); | |
d332ce09 | 543 | blk_mq_debugfs_unregister_sched(q); |
5cf9c91b CH |
544 | mutex_unlock(&q->debugfs_mutex); |
545 | ||
f9cd4bfe JA |
546 | if (e->type->ops.exit_sched) |
547 | e->type->ops.exit_sched(e); | |
e155b0c2 | 548 | blk_mq_sched_tags_teardown(q, flags); |
54d5329d OS |
549 | q->elevator = NULL; |
550 | } |