]>
Commit | Line | Data |
---|---|---|
75bb4625 JA |
1 | /* |
2 | * Block multiqueue core code | |
3 | * | |
4 | * Copyright (C) 2013-2014 Jens Axboe | |
5 | * Copyright (C) 2013-2014 Christoph Hellwig | |
6 | */ | |
320ae51f JA |
7 | #include <linux/kernel.h> |
8 | #include <linux/module.h> | |
9 | #include <linux/backing-dev.h> | |
10 | #include <linux/bio.h> | |
11 | #include <linux/blkdev.h> | |
f75782e4 | 12 | #include <linux/kmemleak.h> |
320ae51f JA |
13 | #include <linux/mm.h> |
14 | #include <linux/init.h> | |
15 | #include <linux/slab.h> | |
16 | #include <linux/workqueue.h> | |
17 | #include <linux/smp.h> | |
18 | #include <linux/llist.h> | |
19 | #include <linux/list_sort.h> | |
20 | #include <linux/cpu.h> | |
21 | #include <linux/cache.h> | |
22 | #include <linux/sched/sysctl.h> | |
105ab3d8 | 23 | #include <linux/sched/topology.h> |
174cd4b1 | 24 | #include <linux/sched/signal.h> |
320ae51f | 25 | #include <linux/delay.h> |
aedcd72f | 26 | #include <linux/crash_dump.h> |
88c7b2b7 | 27 | #include <linux/prefetch.h> |
320ae51f JA |
28 | |
29 | #include <trace/events/block.h> | |
30 | ||
31 | #include <linux/blk-mq.h> | |
32 | #include "blk.h" | |
33 | #include "blk-mq.h" | |
9c1051aa | 34 | #include "blk-mq-debugfs.h" |
320ae51f | 35 | #include "blk-mq-tag.h" |
986d413b | 36 | #include "blk-pm.h" |
cf43e6be | 37 | #include "blk-stat.h" |
bd166ef1 | 38 | #include "blk-mq-sched.h" |
c1c80384 | 39 | #include "blk-rq-qos.h" |
320ae51f | 40 | |
ea435e1b | 41 | static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie); |
34dbad5d OS |
42 | static void blk_mq_poll_stats_start(struct request_queue *q); |
43 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb); | |
44 | ||
720b8ccc SB |
45 | static int blk_mq_poll_stats_bkt(const struct request *rq) |
46 | { | |
47 | int ddir, bytes, bucket; | |
48 | ||
99c749a4 | 49 | ddir = rq_data_dir(rq); |
720b8ccc SB |
50 | bytes = blk_rq_bytes(rq); |
51 | ||
52 | bucket = ddir + 2*(ilog2(bytes) - 9); | |
53 | ||
54 | if (bucket < 0) | |
55 | return -1; | |
56 | else if (bucket >= BLK_MQ_POLL_STATS_BKTS) | |
57 | return ddir + BLK_MQ_POLL_STATS_BKTS - 2; | |
58 | ||
59 | return bucket; | |
60 | } | |
61 | ||
320ae51f JA |
62 | /* |
63 | * Check if any of the ctx's have pending work in this hardware queue | |
64 | */ | |
79f720a7 | 65 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 66 | { |
79f720a7 JA |
67 | return !list_empty_careful(&hctx->dispatch) || |
68 | sbitmap_any_bit_set(&hctx->ctx_map) || | |
bd166ef1 | 69 | blk_mq_sched_has_work(hctx); |
1429d7c9 JA |
70 | } |
71 | ||
320ae51f JA |
72 | /* |
73 | * Mark this ctx as having pending work in this hardware queue | |
74 | */ | |
75 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
76 | struct blk_mq_ctx *ctx) | |
77 | { | |
88459642 OS |
78 | if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw)) |
79 | sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw); | |
1429d7c9 JA |
80 | } |
81 | ||
82 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
83 | struct blk_mq_ctx *ctx) | |
84 | { | |
88459642 | 85 | sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw); |
320ae51f JA |
86 | } |
87 | ||
f299b7c7 JA |
88 | struct mq_inflight { |
89 | struct hd_struct *part; | |
90 | unsigned int *inflight; | |
91 | }; | |
92 | ||
93 | static void blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx, | |
94 | struct request *rq, void *priv, | |
95 | bool reserved) | |
96 | { | |
97 | struct mq_inflight *mi = priv; | |
98 | ||
6131837b OS |
99 | /* |
100 | * index[0] counts the specific partition that was asked for. index[1] | |
101 | * counts the ones that are active on the whole device, so increment | |
102 | * that if mi->part is indeed a partition, and not a whole device. | |
103 | */ | |
104 | if (rq->part == mi->part) | |
105 | mi->inflight[0]++; | |
106 | if (mi->part->partno) | |
107 | mi->inflight[1]++; | |
f299b7c7 JA |
108 | } |
109 | ||
110 | void blk_mq_in_flight(struct request_queue *q, struct hd_struct *part, | |
111 | unsigned int inflight[2]) | |
112 | { | |
113 | struct mq_inflight mi = { .part = part, .inflight = inflight, }; | |
114 | ||
b8d62b3a | 115 | inflight[0] = inflight[1] = 0; |
f299b7c7 JA |
116 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
117 | } | |
118 | ||
bf0ddaba OS |
119 | static void blk_mq_check_inflight_rw(struct blk_mq_hw_ctx *hctx, |
120 | struct request *rq, void *priv, | |
121 | bool reserved) | |
122 | { | |
123 | struct mq_inflight *mi = priv; | |
124 | ||
125 | if (rq->part == mi->part) | |
126 | mi->inflight[rq_data_dir(rq)]++; | |
127 | } | |
128 | ||
129 | void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part, | |
130 | unsigned int inflight[2]) | |
131 | { | |
132 | struct mq_inflight mi = { .part = part, .inflight = inflight, }; | |
133 | ||
134 | inflight[0] = inflight[1] = 0; | |
135 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight_rw, &mi); | |
136 | } | |
137 | ||
1671d522 | 138 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 139 | { |
4ecd4fef | 140 | int freeze_depth; |
cddd5d17 | 141 | |
4ecd4fef CH |
142 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
143 | if (freeze_depth == 1) { | |
3ef28e83 | 144 | percpu_ref_kill(&q->q_usage_counter); |
055f6e18 ML |
145 | if (q->mq_ops) |
146 | blk_mq_run_hw_queues(q, false); | |
cddd5d17 | 147 | } |
f3af020b | 148 | } |
1671d522 | 149 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
f3af020b | 150 | |
6bae363e | 151 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 152 | { |
3ef28e83 | 153 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 154 | } |
6bae363e | 155 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 156 | |
f91328c4 KB |
157 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
158 | unsigned long timeout) | |
159 | { | |
160 | return wait_event_timeout(q->mq_freeze_wq, | |
161 | percpu_ref_is_zero(&q->q_usage_counter), | |
162 | timeout); | |
163 | } | |
164 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 165 | |
f3af020b TH |
166 | /* |
167 | * Guarantee no request is in use, so we can change any data structure of | |
168 | * the queue afterward. | |
169 | */ | |
3ef28e83 | 170 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 171 | { |
3ef28e83 DW |
172 | /* |
173 | * In the !blk_mq case we are only calling this to kill the | |
174 | * q_usage_counter, otherwise this increases the freeze depth | |
175 | * and waits for it to return to zero. For this reason there is | |
176 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
177 | * exported to drivers as the only user for unfreeze is blk_mq. | |
178 | */ | |
1671d522 | 179 | blk_freeze_queue_start(q); |
f3af020b TH |
180 | blk_mq_freeze_queue_wait(q); |
181 | } | |
3ef28e83 DW |
182 | |
183 | void blk_mq_freeze_queue(struct request_queue *q) | |
184 | { | |
185 | /* | |
186 | * ...just an alias to keep freeze and unfreeze actions balanced | |
187 | * in the blk_mq_* namespace | |
188 | */ | |
189 | blk_freeze_queue(q); | |
190 | } | |
c761d96b | 191 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 192 | |
b4c6a028 | 193 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 194 | { |
4ecd4fef | 195 | int freeze_depth; |
320ae51f | 196 | |
4ecd4fef CH |
197 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
198 | WARN_ON_ONCE(freeze_depth < 0); | |
199 | if (!freeze_depth) { | |
bdd63160 | 200 | percpu_ref_resurrect(&q->q_usage_counter); |
320ae51f | 201 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 202 | } |
320ae51f | 203 | } |
b4c6a028 | 204 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 205 | |
852ec809 BVA |
206 | /* |
207 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
208 | * mpt3sas driver such that this function can be removed. | |
209 | */ | |
210 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
211 | { | |
8814ce8a | 212 | blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q); |
852ec809 BVA |
213 | } |
214 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
215 | ||
6a83e74d | 216 | /** |
69e07c4a | 217 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished |
6a83e74d BVA |
218 | * @q: request queue. |
219 | * | |
220 | * Note: this function does not prevent that the struct request end_io() | |
69e07c4a ML |
221 | * callback function is invoked. Once this function is returned, we make |
222 | * sure no dispatch can happen until the queue is unquiesced via | |
223 | * blk_mq_unquiesce_queue(). | |
6a83e74d BVA |
224 | */ |
225 | void blk_mq_quiesce_queue(struct request_queue *q) | |
226 | { | |
227 | struct blk_mq_hw_ctx *hctx; | |
228 | unsigned int i; | |
229 | bool rcu = false; | |
230 | ||
1d9e9bc6 | 231 | blk_mq_quiesce_queue_nowait(q); |
f4560ffe | 232 | |
6a83e74d BVA |
233 | queue_for_each_hw_ctx(q, hctx, i) { |
234 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
05707b64 | 235 | synchronize_srcu(hctx->srcu); |
6a83e74d BVA |
236 | else |
237 | rcu = true; | |
238 | } | |
239 | if (rcu) | |
240 | synchronize_rcu(); | |
241 | } | |
242 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
243 | ||
e4e73913 ML |
244 | /* |
245 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
246 | * @q: request queue. | |
247 | * | |
248 | * This function recovers queue into the state before quiescing | |
249 | * which is done by blk_mq_quiesce_queue. | |
250 | */ | |
251 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
252 | { | |
8814ce8a | 253 | blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q); |
f4560ffe | 254 | |
1d9e9bc6 ML |
255 | /* dispatch requests which are inserted during quiescing */ |
256 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
257 | } |
258 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
259 | ||
aed3ea94 JA |
260 | void blk_mq_wake_waiters(struct request_queue *q) |
261 | { | |
262 | struct blk_mq_hw_ctx *hctx; | |
263 | unsigned int i; | |
264 | ||
265 | queue_for_each_hw_ctx(q, hctx, i) | |
266 | if (blk_mq_hw_queue_mapped(hctx)) | |
267 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
268 | } | |
269 | ||
320ae51f JA |
270 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
271 | { | |
272 | return blk_mq_has_free_tags(hctx->tags); | |
273 | } | |
274 | EXPORT_SYMBOL(blk_mq_can_queue); | |
275 | ||
e4cdf1a1 CH |
276 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
277 | unsigned int tag, unsigned int op) | |
320ae51f | 278 | { |
e4cdf1a1 CH |
279 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
280 | struct request *rq = tags->static_rqs[tag]; | |
bf9ae8c5 | 281 | req_flags_t rq_flags = 0; |
c3a148d2 | 282 | |
e4cdf1a1 CH |
283 | if (data->flags & BLK_MQ_REQ_INTERNAL) { |
284 | rq->tag = -1; | |
285 | rq->internal_tag = tag; | |
286 | } else { | |
d263ed99 | 287 | if (data->hctx->flags & BLK_MQ_F_TAG_SHARED) { |
bf9ae8c5 | 288 | rq_flags = RQF_MQ_INFLIGHT; |
e4cdf1a1 CH |
289 | atomic_inc(&data->hctx->nr_active); |
290 | } | |
291 | rq->tag = tag; | |
292 | rq->internal_tag = -1; | |
293 | data->hctx->tags->rqs[rq->tag] = rq; | |
294 | } | |
295 | ||
af76e555 | 296 | /* csd/requeue_work/fifo_time is initialized before use */ |
e4cdf1a1 CH |
297 | rq->q = data->q; |
298 | rq->mq_ctx = data->ctx; | |
bf9ae8c5 | 299 | rq->rq_flags = rq_flags; |
ef295ecf | 300 | rq->cmd_flags = op; |
1b6d65a0 BVA |
301 | if (data->flags & BLK_MQ_REQ_PREEMPT) |
302 | rq->rq_flags |= RQF_PREEMPT; | |
e4cdf1a1 | 303 | if (blk_queue_io_stat(data->q)) |
e8064021 | 304 | rq->rq_flags |= RQF_IO_STAT; |
7c3fb70f | 305 | INIT_LIST_HEAD(&rq->queuelist); |
af76e555 CH |
306 | INIT_HLIST_NODE(&rq->hash); |
307 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
308 | rq->rq_disk = NULL; |
309 | rq->part = NULL; | |
522a7775 | 310 | rq->start_time_ns = ktime_get_ns(); |
544ccc8d | 311 | rq->io_start_time_ns = 0; |
af76e555 CH |
312 | rq->nr_phys_segments = 0; |
313 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
314 | rq->nr_integrity_segments = 0; | |
315 | #endif | |
af76e555 CH |
316 | rq->special = NULL; |
317 | /* tag was already set */ | |
af76e555 | 318 | rq->extra_len = 0; |
e14575b3 | 319 | rq->__deadline = 0; |
af76e555 | 320 | |
af76e555 | 321 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
322 | rq->timeout = 0; |
323 | ||
af76e555 CH |
324 | rq->end_io = NULL; |
325 | rq->end_io_data = NULL; | |
326 | rq->next_rq = NULL; | |
327 | ||
e4cdf1a1 | 328 | data->ctx->rq_dispatched[op_is_sync(op)]++; |
12f5b931 | 329 | refcount_set(&rq->ref, 1); |
e4cdf1a1 | 330 | return rq; |
5dee8577 CH |
331 | } |
332 | ||
d2c0d383 | 333 | static struct request *blk_mq_get_request(struct request_queue *q, |
f9afca4d JA |
334 | struct bio *bio, |
335 | struct blk_mq_alloc_data *data) | |
d2c0d383 CH |
336 | { |
337 | struct elevator_queue *e = q->elevator; | |
338 | struct request *rq; | |
e4cdf1a1 | 339 | unsigned int tag; |
21e768b4 | 340 | bool put_ctx_on_error = false; |
d2c0d383 CH |
341 | |
342 | blk_queue_enter_live(q); | |
343 | data->q = q; | |
21e768b4 BVA |
344 | if (likely(!data->ctx)) { |
345 | data->ctx = blk_mq_get_ctx(q); | |
346 | put_ctx_on_error = true; | |
347 | } | |
d2c0d383 | 348 | if (likely(!data->hctx)) |
f9afca4d JA |
349 | data->hctx = blk_mq_map_queue(q, data->cmd_flags, |
350 | data->ctx->cpu); | |
351 | if (data->cmd_flags & REQ_NOWAIT) | |
03a07c92 | 352 | data->flags |= BLK_MQ_REQ_NOWAIT; |
d2c0d383 CH |
353 | |
354 | if (e) { | |
355 | data->flags |= BLK_MQ_REQ_INTERNAL; | |
356 | ||
357 | /* | |
358 | * Flush requests are special and go directly to the | |
17a51199 JA |
359 | * dispatch list. Don't include reserved tags in the |
360 | * limiting, as it isn't useful. | |
d2c0d383 | 361 | */ |
f9afca4d JA |
362 | if (!op_is_flush(data->cmd_flags) && |
363 | e->type->ops.limit_depth && | |
17a51199 | 364 | !(data->flags & BLK_MQ_REQ_RESERVED)) |
f9afca4d | 365 | e->type->ops.limit_depth(data->cmd_flags, data); |
d263ed99 JW |
366 | } else { |
367 | blk_mq_tag_busy(data->hctx); | |
d2c0d383 CH |
368 | } |
369 | ||
e4cdf1a1 CH |
370 | tag = blk_mq_get_tag(data); |
371 | if (tag == BLK_MQ_TAG_FAIL) { | |
21e768b4 BVA |
372 | if (put_ctx_on_error) { |
373 | blk_mq_put_ctx(data->ctx); | |
1ad43c00 ML |
374 | data->ctx = NULL; |
375 | } | |
037cebb8 CH |
376 | blk_queue_exit(q); |
377 | return NULL; | |
d2c0d383 CH |
378 | } |
379 | ||
f9afca4d JA |
380 | rq = blk_mq_rq_ctx_init(data, tag, data->cmd_flags); |
381 | if (!op_is_flush(data->cmd_flags)) { | |
037cebb8 | 382 | rq->elv.icq = NULL; |
f9cd4bfe | 383 | if (e && e->type->ops.prepare_request) { |
44e8c2bf CH |
384 | if (e->type->icq_cache && rq_ioc(bio)) |
385 | blk_mq_sched_assign_ioc(rq, bio); | |
386 | ||
f9cd4bfe | 387 | e->type->ops.prepare_request(rq, bio); |
5bbf4e5a | 388 | rq->rq_flags |= RQF_ELVPRIV; |
44e8c2bf | 389 | } |
037cebb8 CH |
390 | } |
391 | data->hctx->queued++; | |
392 | return rq; | |
d2c0d383 CH |
393 | } |
394 | ||
cd6ce148 | 395 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, |
9a95e4ef | 396 | blk_mq_req_flags_t flags) |
320ae51f | 397 | { |
f9afca4d | 398 | struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op }; |
bd166ef1 | 399 | struct request *rq; |
a492f075 | 400 | int ret; |
320ae51f | 401 | |
3a0a5299 | 402 | ret = blk_queue_enter(q, flags); |
a492f075 JL |
403 | if (ret) |
404 | return ERR_PTR(ret); | |
320ae51f | 405 | |
f9afca4d | 406 | rq = blk_mq_get_request(q, NULL, &alloc_data); |
3280d66a | 407 | blk_queue_exit(q); |
841bac2c | 408 | |
bd166ef1 | 409 | if (!rq) |
a492f075 | 410 | return ERR_PTR(-EWOULDBLOCK); |
0c4de0f3 | 411 | |
1ad43c00 | 412 | blk_mq_put_ctx(alloc_data.ctx); |
1ad43c00 | 413 | |
0c4de0f3 CH |
414 | rq->__data_len = 0; |
415 | rq->__sector = (sector_t) -1; | |
416 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
417 | return rq; |
418 | } | |
4bb659b1 | 419 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 420 | |
cd6ce148 | 421 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
9a95e4ef | 422 | unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 423 | { |
f9afca4d | 424 | struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op }; |
1f5bd336 | 425 | struct request *rq; |
6d2809d5 | 426 | unsigned int cpu; |
1f5bd336 ML |
427 | int ret; |
428 | ||
429 | /* | |
430 | * If the tag allocator sleeps we could get an allocation for a | |
431 | * different hardware context. No need to complicate the low level | |
432 | * allocator for this for the rare use case of a command tied to | |
433 | * a specific queue. | |
434 | */ | |
435 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
436 | return ERR_PTR(-EINVAL); | |
437 | ||
438 | if (hctx_idx >= q->nr_hw_queues) | |
439 | return ERR_PTR(-EIO); | |
440 | ||
3a0a5299 | 441 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
442 | if (ret) |
443 | return ERR_PTR(ret); | |
444 | ||
c8712c6a CH |
445 | /* |
446 | * Check if the hardware context is actually mapped to anything. | |
447 | * If not tell the caller that it should skip this queue. | |
448 | */ | |
6d2809d5 OS |
449 | alloc_data.hctx = q->queue_hw_ctx[hctx_idx]; |
450 | if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) { | |
451 | blk_queue_exit(q); | |
452 | return ERR_PTR(-EXDEV); | |
c8712c6a | 453 | } |
20e4d813 | 454 | cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask); |
6d2809d5 | 455 | alloc_data.ctx = __blk_mq_get_ctx(q, cpu); |
1f5bd336 | 456 | |
f9afca4d | 457 | rq = blk_mq_get_request(q, NULL, &alloc_data); |
3280d66a | 458 | blk_queue_exit(q); |
c8712c6a | 459 | |
6d2809d5 OS |
460 | if (!rq) |
461 | return ERR_PTR(-EWOULDBLOCK); | |
462 | ||
463 | return rq; | |
1f5bd336 ML |
464 | } |
465 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
466 | ||
12f5b931 KB |
467 | static void __blk_mq_free_request(struct request *rq) |
468 | { | |
469 | struct request_queue *q = rq->q; | |
470 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
f9afca4d | 471 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, rq->cmd_flags, ctx->cpu); |
12f5b931 KB |
472 | const int sched_tag = rq->internal_tag; |
473 | ||
986d413b | 474 | blk_pm_mark_last_busy(rq); |
12f5b931 KB |
475 | if (rq->tag != -1) |
476 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
477 | if (sched_tag != -1) | |
478 | blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag); | |
479 | blk_mq_sched_restart(hctx); | |
480 | blk_queue_exit(q); | |
481 | } | |
482 | ||
6af54051 | 483 | void blk_mq_free_request(struct request *rq) |
320ae51f | 484 | { |
320ae51f | 485 | struct request_queue *q = rq->q; |
6af54051 CH |
486 | struct elevator_queue *e = q->elevator; |
487 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
f9afca4d | 488 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, rq->cmd_flags, ctx->cpu); |
6af54051 | 489 | |
5bbf4e5a | 490 | if (rq->rq_flags & RQF_ELVPRIV) { |
f9cd4bfe JA |
491 | if (e && e->type->ops.finish_request) |
492 | e->type->ops.finish_request(rq); | |
6af54051 CH |
493 | if (rq->elv.icq) { |
494 | put_io_context(rq->elv.icq->ioc); | |
495 | rq->elv.icq = NULL; | |
496 | } | |
497 | } | |
320ae51f | 498 | |
6af54051 | 499 | ctx->rq_completed[rq_is_sync(rq)]++; |
e8064021 | 500 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 501 | atomic_dec(&hctx->nr_active); |
87760e5e | 502 | |
7beb2f84 JA |
503 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
504 | laptop_io_completion(q->backing_dev_info); | |
505 | ||
a7905043 | 506 | rq_qos_done(q, rq); |
0d2602ca | 507 | |
12f5b931 KB |
508 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
509 | if (refcount_dec_and_test(&rq->ref)) | |
510 | __blk_mq_free_request(rq); | |
320ae51f | 511 | } |
1a3b595a | 512 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 513 | |
2a842aca | 514 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
320ae51f | 515 | { |
522a7775 OS |
516 | u64 now = ktime_get_ns(); |
517 | ||
4bc6339a OS |
518 | if (rq->rq_flags & RQF_STATS) { |
519 | blk_mq_poll_stats_start(rq->q); | |
522a7775 | 520 | blk_stat_add(rq, now); |
4bc6339a OS |
521 | } |
522 | ||
ed88660a OS |
523 | if (rq->internal_tag != -1) |
524 | blk_mq_sched_completed_request(rq, now); | |
525 | ||
522a7775 | 526 | blk_account_io_done(rq, now); |
0d11e6ac | 527 | |
91b63639 | 528 | if (rq->end_io) { |
a7905043 | 529 | rq_qos_done(rq->q, rq); |
320ae51f | 530 | rq->end_io(rq, error); |
91b63639 CH |
531 | } else { |
532 | if (unlikely(blk_bidi_rq(rq))) | |
533 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 534 | blk_mq_free_request(rq); |
91b63639 | 535 | } |
320ae51f | 536 | } |
c8a446ad | 537 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 538 | |
2a842aca | 539 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
540 | { |
541 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
542 | BUG(); | |
c8a446ad | 543 | __blk_mq_end_request(rq, error); |
63151a44 | 544 | } |
c8a446ad | 545 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 546 | |
30a91cb4 | 547 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 548 | { |
3d6efbf6 | 549 | struct request *rq = data; |
c7bb9ad1 | 550 | struct request_queue *q = rq->q; |
320ae51f | 551 | |
c7bb9ad1 | 552 | q->mq_ops->complete(rq); |
320ae51f | 553 | } |
320ae51f | 554 | |
453f8341 | 555 | static void __blk_mq_complete_request(struct request *rq) |
320ae51f JA |
556 | { |
557 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
c7bb9ad1 | 558 | struct request_queue *q = rq->q; |
38535201 | 559 | bool shared = false; |
320ae51f JA |
560 | int cpu; |
561 | ||
0fc09f92 | 562 | if (!blk_mq_mark_complete(rq)) |
12f5b931 | 563 | return; |
453f8341 | 564 | |
36e76539 ML |
565 | /* |
566 | * Most of single queue controllers, there is only one irq vector | |
567 | * for handling IO completion, and the only irq's affinity is set | |
568 | * as all possible CPUs. On most of ARCHs, this affinity means the | |
569 | * irq is handled on one specific CPU. | |
570 | * | |
571 | * So complete IO reqeust in softirq context in case of single queue | |
572 | * for not degrading IO performance by irqsoff latency. | |
573 | */ | |
c7bb9ad1 | 574 | if (q->nr_hw_queues == 1) { |
36e76539 ML |
575 | __blk_complete_request(rq); |
576 | return; | |
577 | } | |
578 | ||
c7bb9ad1 JA |
579 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags)) { |
580 | q->mq_ops->complete(rq); | |
30a91cb4 CH |
581 | return; |
582 | } | |
320ae51f JA |
583 | |
584 | cpu = get_cpu(); | |
c7bb9ad1 | 585 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &q->queue_flags)) |
38535201 CH |
586 | shared = cpus_share_cache(cpu, ctx->cpu); |
587 | ||
588 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 589 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
590 | rq->csd.info = rq; |
591 | rq->csd.flags = 0; | |
c46fff2a | 592 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 593 | } else { |
c7bb9ad1 | 594 | q->mq_ops->complete(rq); |
3d6efbf6 | 595 | } |
320ae51f JA |
596 | put_cpu(); |
597 | } | |
30a91cb4 | 598 | |
04ced159 | 599 | static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx) |
b7435db8 | 600 | __releases(hctx->srcu) |
04ced159 JA |
601 | { |
602 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
603 | rcu_read_unlock(); | |
604 | else | |
05707b64 | 605 | srcu_read_unlock(hctx->srcu, srcu_idx); |
04ced159 JA |
606 | } |
607 | ||
608 | static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx) | |
b7435db8 | 609 | __acquires(hctx->srcu) |
04ced159 | 610 | { |
08b5a6e2 JA |
611 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
612 | /* shut up gcc false positive */ | |
613 | *srcu_idx = 0; | |
04ced159 | 614 | rcu_read_lock(); |
08b5a6e2 | 615 | } else |
05707b64 | 616 | *srcu_idx = srcu_read_lock(hctx->srcu); |
04ced159 JA |
617 | } |
618 | ||
30a91cb4 CH |
619 | /** |
620 | * blk_mq_complete_request - end I/O on a request | |
621 | * @rq: the request being processed | |
622 | * | |
623 | * Description: | |
624 | * Ends all I/O on a request. It does not handle partial completions. | |
625 | * The actual completion happens out-of-order, through a IPI handler. | |
626 | **/ | |
08e0029a | 627 | void blk_mq_complete_request(struct request *rq) |
30a91cb4 | 628 | { |
12f5b931 | 629 | if (unlikely(blk_should_fake_timeout(rq->q))) |
30a91cb4 | 630 | return; |
12f5b931 | 631 | __blk_mq_complete_request(rq); |
30a91cb4 CH |
632 | } |
633 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 634 | |
973c0191 KB |
635 | int blk_mq_request_started(struct request *rq) |
636 | { | |
5a61c363 | 637 | return blk_mq_rq_state(rq) != MQ_RQ_IDLE; |
973c0191 KB |
638 | } |
639 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
640 | ||
e2490073 | 641 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
642 | { |
643 | struct request_queue *q = rq->q; | |
644 | ||
bd166ef1 JA |
645 | blk_mq_sched_started_request(rq); |
646 | ||
320ae51f JA |
647 | trace_block_rq_issue(q, rq); |
648 | ||
cf43e6be | 649 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
544ccc8d OS |
650 | rq->io_start_time_ns = ktime_get_ns(); |
651 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW | |
652 | rq->throtl_size = blk_rq_sectors(rq); | |
653 | #endif | |
cf43e6be | 654 | rq->rq_flags |= RQF_STATS; |
a7905043 | 655 | rq_qos_issue(q, rq); |
cf43e6be JA |
656 | } |
657 | ||
1d9bd516 | 658 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 659 | |
1d9bd516 | 660 | blk_add_timer(rq); |
12f5b931 | 661 | WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); |
49f5baa5 CH |
662 | |
663 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
664 | /* | |
665 | * Make sure space for the drain appears. We know we can do | |
666 | * this because max_hw_segments has been adjusted to be one | |
667 | * fewer than the device can handle. | |
668 | */ | |
669 | rq->nr_phys_segments++; | |
670 | } | |
320ae51f | 671 | } |
e2490073 | 672 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 673 | |
ed0791b2 | 674 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
675 | { |
676 | struct request_queue *q = rq->q; | |
677 | ||
923218f6 ML |
678 | blk_mq_put_driver_tag(rq); |
679 | ||
320ae51f | 680 | trace_block_rq_requeue(q, rq); |
a7905043 | 681 | rq_qos_requeue(q, rq); |
49f5baa5 | 682 | |
12f5b931 KB |
683 | if (blk_mq_request_started(rq)) { |
684 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); | |
da661267 | 685 | rq->rq_flags &= ~RQF_TIMED_OUT; |
e2490073 CH |
686 | if (q->dma_drain_size && blk_rq_bytes(rq)) |
687 | rq->nr_phys_segments--; | |
688 | } | |
320ae51f JA |
689 | } |
690 | ||
2b053aca | 691 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 692 | { |
ed0791b2 | 693 | __blk_mq_requeue_request(rq); |
ed0791b2 | 694 | |
105976f5 ML |
695 | /* this request will be re-inserted to io scheduler queue */ |
696 | blk_mq_sched_requeue_request(rq); | |
697 | ||
7d692330 | 698 | BUG_ON(!list_empty(&rq->queuelist)); |
2b053aca | 699 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
700 | } |
701 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
702 | ||
6fca6a61 CH |
703 | static void blk_mq_requeue_work(struct work_struct *work) |
704 | { | |
705 | struct request_queue *q = | |
2849450a | 706 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
707 | LIST_HEAD(rq_list); |
708 | struct request *rq, *next; | |
6fca6a61 | 709 | |
18e9781d | 710 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 711 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 712 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
713 | |
714 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
e8064021 | 715 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
6fca6a61 CH |
716 | continue; |
717 | ||
e8064021 | 718 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 719 | list_del_init(&rq->queuelist); |
9e97d295 | 720 | blk_mq_sched_insert_request(rq, true, false, false); |
6fca6a61 CH |
721 | } |
722 | ||
723 | while (!list_empty(&rq_list)) { | |
724 | rq = list_entry(rq_list.next, struct request, queuelist); | |
725 | list_del_init(&rq->queuelist); | |
9e97d295 | 726 | blk_mq_sched_insert_request(rq, false, false, false); |
6fca6a61 CH |
727 | } |
728 | ||
52d7f1b5 | 729 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
730 | } |
731 | ||
2b053aca BVA |
732 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
733 | bool kick_requeue_list) | |
6fca6a61 CH |
734 | { |
735 | struct request_queue *q = rq->q; | |
736 | unsigned long flags; | |
737 | ||
738 | /* | |
739 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
ff821d27 | 740 | * request head insertion from the workqueue. |
6fca6a61 | 741 | */ |
e8064021 | 742 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
743 | |
744 | spin_lock_irqsave(&q->requeue_lock, flags); | |
745 | if (at_head) { | |
e8064021 | 746 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
747 | list_add(&rq->queuelist, &q->requeue_list); |
748 | } else { | |
749 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
750 | } | |
751 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
752 | |
753 | if (kick_requeue_list) | |
754 | blk_mq_kick_requeue_list(q); | |
6fca6a61 CH |
755 | } |
756 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
757 | ||
758 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
759 | { | |
ae943d20 | 760 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
761 | } |
762 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
763 | ||
2849450a MS |
764 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
765 | unsigned long msecs) | |
766 | { | |
d4acf365 BVA |
767 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
768 | msecs_to_jiffies(msecs)); | |
2849450a MS |
769 | } |
770 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
771 | ||
0e62f51f JA |
772 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
773 | { | |
88c7b2b7 JA |
774 | if (tag < tags->nr_tags) { |
775 | prefetch(tags->rqs[tag]); | |
4ee86bab | 776 | return tags->rqs[tag]; |
88c7b2b7 | 777 | } |
4ee86bab HR |
778 | |
779 | return NULL; | |
24d2f903 CH |
780 | } |
781 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
782 | ||
358f70da | 783 | static void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 784 | { |
da661267 | 785 | req->rq_flags |= RQF_TIMED_OUT; |
d1210d5a CH |
786 | if (req->q->mq_ops->timeout) { |
787 | enum blk_eh_timer_return ret; | |
788 | ||
789 | ret = req->q->mq_ops->timeout(req, reserved); | |
790 | if (ret == BLK_EH_DONE) | |
791 | return; | |
792 | WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); | |
46f92d42 | 793 | } |
d1210d5a CH |
794 | |
795 | blk_add_timer(req); | |
87ee7b11 | 796 | } |
5b3f25fc | 797 | |
12f5b931 | 798 | static bool blk_mq_req_expired(struct request *rq, unsigned long *next) |
81481eb4 | 799 | { |
12f5b931 | 800 | unsigned long deadline; |
87ee7b11 | 801 | |
12f5b931 KB |
802 | if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) |
803 | return false; | |
da661267 CH |
804 | if (rq->rq_flags & RQF_TIMED_OUT) |
805 | return false; | |
a7af0af3 | 806 | |
12f5b931 KB |
807 | deadline = blk_rq_deadline(rq); |
808 | if (time_after_eq(jiffies, deadline)) | |
809 | return true; | |
a7af0af3 | 810 | |
12f5b931 KB |
811 | if (*next == 0) |
812 | *next = deadline; | |
813 | else if (time_after(*next, deadline)) | |
814 | *next = deadline; | |
815 | return false; | |
87ee7b11 JA |
816 | } |
817 | ||
12f5b931 | 818 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
1d9bd516 TH |
819 | struct request *rq, void *priv, bool reserved) |
820 | { | |
12f5b931 KB |
821 | unsigned long *next = priv; |
822 | ||
823 | /* | |
824 | * Just do a quick check if it is expired before locking the request in | |
825 | * so we're not unnecessarilly synchronizing across CPUs. | |
826 | */ | |
827 | if (!blk_mq_req_expired(rq, next)) | |
828 | return; | |
829 | ||
830 | /* | |
831 | * We have reason to believe the request may be expired. Take a | |
832 | * reference on the request to lock this request lifetime into its | |
833 | * currently allocated context to prevent it from being reallocated in | |
834 | * the event the completion by-passes this timeout handler. | |
835 | * | |
836 | * If the reference was already released, then the driver beat the | |
837 | * timeout handler to posting a natural completion. | |
838 | */ | |
839 | if (!refcount_inc_not_zero(&rq->ref)) | |
840 | return; | |
841 | ||
1d9bd516 | 842 | /* |
12f5b931 KB |
843 | * The request is now locked and cannot be reallocated underneath the |
844 | * timeout handler's processing. Re-verify this exact request is truly | |
845 | * expired; if it is not expired, then the request was completed and | |
846 | * reallocated as a new request. | |
1d9bd516 | 847 | */ |
12f5b931 | 848 | if (blk_mq_req_expired(rq, next)) |
1d9bd516 | 849 | blk_mq_rq_timed_out(rq, reserved); |
12f5b931 KB |
850 | if (refcount_dec_and_test(&rq->ref)) |
851 | __blk_mq_free_request(rq); | |
1d9bd516 TH |
852 | } |
853 | ||
287922eb | 854 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 855 | { |
287922eb CH |
856 | struct request_queue *q = |
857 | container_of(work, struct request_queue, timeout_work); | |
12f5b931 | 858 | unsigned long next = 0; |
1d9bd516 | 859 | struct blk_mq_hw_ctx *hctx; |
81481eb4 | 860 | int i; |
320ae51f | 861 | |
71f79fb3 GKB |
862 | /* A deadlock might occur if a request is stuck requiring a |
863 | * timeout at the same time a queue freeze is waiting | |
864 | * completion, since the timeout code would not be able to | |
865 | * acquire the queue reference here. | |
866 | * | |
867 | * That's why we don't use blk_queue_enter here; instead, we use | |
868 | * percpu_ref_tryget directly, because we need to be able to | |
869 | * obtain a reference even in the short window between the queue | |
870 | * starting to freeze, by dropping the first reference in | |
1671d522 | 871 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
872 | * consumed, marked by the instant q_usage_counter reaches |
873 | * zero. | |
874 | */ | |
875 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
876 | return; |
877 | ||
12f5b931 | 878 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next); |
320ae51f | 879 | |
12f5b931 KB |
880 | if (next != 0) { |
881 | mod_timer(&q->timeout, next); | |
0d2602ca | 882 | } else { |
fcd36c36 BVA |
883 | /* |
884 | * Request timeouts are handled as a forward rolling timer. If | |
885 | * we end up here it means that no requests are pending and | |
886 | * also that no request has been pending for a while. Mark | |
887 | * each hctx as idle. | |
888 | */ | |
f054b56c ML |
889 | queue_for_each_hw_ctx(q, hctx, i) { |
890 | /* the hctx may be unmapped, so check it here */ | |
891 | if (blk_mq_hw_queue_mapped(hctx)) | |
892 | blk_mq_tag_idle(hctx); | |
893 | } | |
0d2602ca | 894 | } |
287922eb | 895 | blk_queue_exit(q); |
320ae51f JA |
896 | } |
897 | ||
88459642 OS |
898 | struct flush_busy_ctx_data { |
899 | struct blk_mq_hw_ctx *hctx; | |
900 | struct list_head *list; | |
901 | }; | |
902 | ||
903 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
904 | { | |
905 | struct flush_busy_ctx_data *flush_data = data; | |
906 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
907 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
908 | ||
88459642 OS |
909 | spin_lock(&ctx->lock); |
910 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
e9a99a63 | 911 | sbitmap_clear_bit(sb, bitnr); |
88459642 OS |
912 | spin_unlock(&ctx->lock); |
913 | return true; | |
914 | } | |
915 | ||
1429d7c9 JA |
916 | /* |
917 | * Process software queues that have been marked busy, splicing them | |
918 | * to the for-dispatch | |
919 | */ | |
2c3ad667 | 920 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 921 | { |
88459642 OS |
922 | struct flush_busy_ctx_data data = { |
923 | .hctx = hctx, | |
924 | .list = list, | |
925 | }; | |
1429d7c9 | 926 | |
88459642 | 927 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 928 | } |
2c3ad667 | 929 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 930 | |
b347689f ML |
931 | struct dispatch_rq_data { |
932 | struct blk_mq_hw_ctx *hctx; | |
933 | struct request *rq; | |
934 | }; | |
935 | ||
936 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
937 | void *data) | |
938 | { | |
939 | struct dispatch_rq_data *dispatch_data = data; | |
940 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
941 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
942 | ||
943 | spin_lock(&ctx->lock); | |
b4f6f38d | 944 | if (!list_empty(&ctx->rq_list)) { |
b347689f ML |
945 | dispatch_data->rq = list_entry_rq(ctx->rq_list.next); |
946 | list_del_init(&dispatch_data->rq->queuelist); | |
947 | if (list_empty(&ctx->rq_list)) | |
948 | sbitmap_clear_bit(sb, bitnr); | |
949 | } | |
950 | spin_unlock(&ctx->lock); | |
951 | ||
952 | return !dispatch_data->rq; | |
953 | } | |
954 | ||
955 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
956 | struct blk_mq_ctx *start) | |
957 | { | |
958 | unsigned off = start ? start->index_hw : 0; | |
959 | struct dispatch_rq_data data = { | |
960 | .hctx = hctx, | |
961 | .rq = NULL, | |
962 | }; | |
963 | ||
964 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
965 | dispatch_rq_from_ctx, &data); | |
966 | ||
967 | return data.rq; | |
968 | } | |
969 | ||
703fd1c0 JA |
970 | static inline unsigned int queued_to_index(unsigned int queued) |
971 | { | |
972 | if (!queued) | |
973 | return 0; | |
1429d7c9 | 974 | |
703fd1c0 | 975 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
976 | } |
977 | ||
8ab6bb9e | 978 | bool blk_mq_get_driver_tag(struct request *rq) |
bd166ef1 JA |
979 | { |
980 | struct blk_mq_alloc_data data = { | |
981 | .q = rq->q, | |
f9afca4d | 982 | .hctx = blk_mq_map_queue(rq->q, rq->cmd_flags, rq->mq_ctx->cpu), |
8ab6bb9e | 983 | .flags = BLK_MQ_REQ_NOWAIT, |
f9afca4d | 984 | .cmd_flags = rq->cmd_flags, |
bd166ef1 | 985 | }; |
d263ed99 | 986 | bool shared; |
5feeacdd | 987 | |
81380ca1 OS |
988 | if (rq->tag != -1) |
989 | goto done; | |
bd166ef1 | 990 | |
415b806d SG |
991 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) |
992 | data.flags |= BLK_MQ_REQ_RESERVED; | |
993 | ||
d263ed99 | 994 | shared = blk_mq_tag_busy(data.hctx); |
bd166ef1 JA |
995 | rq->tag = blk_mq_get_tag(&data); |
996 | if (rq->tag >= 0) { | |
d263ed99 | 997 | if (shared) { |
200e86b3 JA |
998 | rq->rq_flags |= RQF_MQ_INFLIGHT; |
999 | atomic_inc(&data.hctx->nr_active); | |
1000 | } | |
bd166ef1 | 1001 | data.hctx->tags->rqs[rq->tag] = rq; |
bd166ef1 JA |
1002 | } |
1003 | ||
81380ca1 | 1004 | done: |
81380ca1 | 1005 | return rq->tag != -1; |
bd166ef1 JA |
1006 | } |
1007 | ||
eb619fdb JA |
1008 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1009 | int flags, void *key) | |
da55f2cc OS |
1010 | { |
1011 | struct blk_mq_hw_ctx *hctx; | |
1012 | ||
1013 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1014 | ||
5815839b | 1015 | spin_lock(&hctx->dispatch_wait_lock); |
eb619fdb | 1016 | list_del_init(&wait->entry); |
5815839b ML |
1017 | spin_unlock(&hctx->dispatch_wait_lock); |
1018 | ||
da55f2cc OS |
1019 | blk_mq_run_hw_queue(hctx, true); |
1020 | return 1; | |
1021 | } | |
1022 | ||
f906a6a0 JA |
1023 | /* |
1024 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1025 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1026 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1027 | * marking us as waiting. |
1028 | */ | |
2278d69f | 1029 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, |
f906a6a0 | 1030 | struct request *rq) |
da55f2cc | 1031 | { |
5815839b | 1032 | struct wait_queue_head *wq; |
f906a6a0 JA |
1033 | wait_queue_entry_t *wait; |
1034 | bool ret; | |
da55f2cc | 1035 | |
2278d69f ML |
1036 | if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) { |
1037 | if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
1038 | set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state); | |
f906a6a0 | 1039 | |
c27d53fb BVA |
1040 | /* |
1041 | * It's possible that a tag was freed in the window between the | |
1042 | * allocation failure and adding the hardware queue to the wait | |
1043 | * queue. | |
1044 | * | |
1045 | * Don't clear RESTART here, someone else could have set it. | |
1046 | * At most this will cost an extra queue run. | |
1047 | */ | |
8ab6bb9e | 1048 | return blk_mq_get_driver_tag(rq); |
eb619fdb | 1049 | } |
eb619fdb | 1050 | |
2278d69f | 1051 | wait = &hctx->dispatch_wait; |
c27d53fb BVA |
1052 | if (!list_empty_careful(&wait->entry)) |
1053 | return false; | |
1054 | ||
5815839b ML |
1055 | wq = &bt_wait_ptr(&hctx->tags->bitmap_tags, hctx)->wait; |
1056 | ||
1057 | spin_lock_irq(&wq->lock); | |
1058 | spin_lock(&hctx->dispatch_wait_lock); | |
c27d53fb | 1059 | if (!list_empty(&wait->entry)) { |
5815839b ML |
1060 | spin_unlock(&hctx->dispatch_wait_lock); |
1061 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1062 | return false; |
eb619fdb JA |
1063 | } |
1064 | ||
5815839b ML |
1065 | wait->flags &= ~WQ_FLAG_EXCLUSIVE; |
1066 | __add_wait_queue(wq, wait); | |
c27d53fb | 1067 | |
da55f2cc | 1068 | /* |
eb619fdb JA |
1069 | * It's possible that a tag was freed in the window between the |
1070 | * allocation failure and adding the hardware queue to the wait | |
1071 | * queue. | |
da55f2cc | 1072 | */ |
8ab6bb9e | 1073 | ret = blk_mq_get_driver_tag(rq); |
c27d53fb | 1074 | if (!ret) { |
5815839b ML |
1075 | spin_unlock(&hctx->dispatch_wait_lock); |
1076 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1077 | return false; |
eb619fdb | 1078 | } |
c27d53fb BVA |
1079 | |
1080 | /* | |
1081 | * We got a tag, remove ourselves from the wait queue to ensure | |
1082 | * someone else gets the wakeup. | |
1083 | */ | |
c27d53fb | 1084 | list_del_init(&wait->entry); |
5815839b ML |
1085 | spin_unlock(&hctx->dispatch_wait_lock); |
1086 | spin_unlock_irq(&wq->lock); | |
c27d53fb BVA |
1087 | |
1088 | return true; | |
da55f2cc OS |
1089 | } |
1090 | ||
6e768717 ML |
1091 | #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 |
1092 | #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 | |
1093 | /* | |
1094 | * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): | |
1095 | * - EWMA is one simple way to compute running average value | |
1096 | * - weight(7/8 and 1/8) is applied so that it can decrease exponentially | |
1097 | * - take 4 as factor for avoiding to get too small(0) result, and this | |
1098 | * factor doesn't matter because EWMA decreases exponentially | |
1099 | */ | |
1100 | static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) | |
1101 | { | |
1102 | unsigned int ewma; | |
1103 | ||
1104 | if (hctx->queue->elevator) | |
1105 | return; | |
1106 | ||
1107 | ewma = hctx->dispatch_busy; | |
1108 | ||
1109 | if (!ewma && !busy) | |
1110 | return; | |
1111 | ||
1112 | ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; | |
1113 | if (busy) | |
1114 | ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; | |
1115 | ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; | |
1116 | ||
1117 | hctx->dispatch_busy = ewma; | |
1118 | } | |
1119 | ||
86ff7c2a ML |
1120 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
1121 | ||
1f57f8d4 JA |
1122 | /* |
1123 | * Returns true if we did some work AND can potentially do more. | |
1124 | */ | |
de148297 | 1125 | bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list, |
eb619fdb | 1126 | bool got_budget) |
320ae51f | 1127 | { |
81380ca1 | 1128 | struct blk_mq_hw_ctx *hctx; |
6d6f167c | 1129 | struct request *rq, *nxt; |
eb619fdb | 1130 | bool no_tag = false; |
fc17b653 | 1131 | int errors, queued; |
86ff7c2a | 1132 | blk_status_t ret = BLK_STS_OK; |
320ae51f | 1133 | |
81380ca1 OS |
1134 | if (list_empty(list)) |
1135 | return false; | |
1136 | ||
de148297 ML |
1137 | WARN_ON(!list_is_singular(list) && got_budget); |
1138 | ||
320ae51f JA |
1139 | /* |
1140 | * Now process all the entries, sending them to the driver. | |
1141 | */ | |
93efe981 | 1142 | errors = queued = 0; |
81380ca1 | 1143 | do { |
74c45052 | 1144 | struct blk_mq_queue_data bd; |
320ae51f | 1145 | |
f04c3df3 | 1146 | rq = list_first_entry(list, struct request, queuelist); |
0bca799b | 1147 | |
f9afca4d | 1148 | hctx = blk_mq_map_queue(rq->q, rq->cmd_flags, rq->mq_ctx->cpu); |
0bca799b ML |
1149 | if (!got_budget && !blk_mq_get_dispatch_budget(hctx)) |
1150 | break; | |
1151 | ||
8ab6bb9e | 1152 | if (!blk_mq_get_driver_tag(rq)) { |
3c782d67 | 1153 | /* |
da55f2cc | 1154 | * The initial allocation attempt failed, so we need to |
eb619fdb JA |
1155 | * rerun the hardware queue when a tag is freed. The |
1156 | * waitqueue takes care of that. If the queue is run | |
1157 | * before we add this entry back on the dispatch list, | |
1158 | * we'll re-run it below. | |
3c782d67 | 1159 | */ |
2278d69f | 1160 | if (!blk_mq_mark_tag_wait(hctx, rq)) { |
0bca799b | 1161 | blk_mq_put_dispatch_budget(hctx); |
f906a6a0 JA |
1162 | /* |
1163 | * For non-shared tags, the RESTART check | |
1164 | * will suffice. | |
1165 | */ | |
1166 | if (hctx->flags & BLK_MQ_F_TAG_SHARED) | |
1167 | no_tag = true; | |
de148297 ML |
1168 | break; |
1169 | } | |
1170 | } | |
1171 | ||
320ae51f | 1172 | list_del_init(&rq->queuelist); |
320ae51f | 1173 | |
74c45052 | 1174 | bd.rq = rq; |
113285b4 JA |
1175 | |
1176 | /* | |
1177 | * Flag last if we have no more requests, or if we have more | |
1178 | * but can't assign a driver tag to it. | |
1179 | */ | |
1180 | if (list_empty(list)) | |
1181 | bd.last = true; | |
1182 | else { | |
113285b4 | 1183 | nxt = list_first_entry(list, struct request, queuelist); |
8ab6bb9e | 1184 | bd.last = !blk_mq_get_driver_tag(nxt); |
113285b4 | 1185 | } |
74c45052 JA |
1186 | |
1187 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
86ff7c2a | 1188 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) { |
6d6f167c JW |
1189 | /* |
1190 | * If an I/O scheduler has been configured and we got a | |
ff821d27 JA |
1191 | * driver tag for the next request already, free it |
1192 | * again. | |
6d6f167c JW |
1193 | */ |
1194 | if (!list_empty(list)) { | |
1195 | nxt = list_first_entry(list, struct request, queuelist); | |
1196 | blk_mq_put_driver_tag(nxt); | |
1197 | } | |
f04c3df3 | 1198 | list_add(&rq->queuelist, list); |
ed0791b2 | 1199 | __blk_mq_requeue_request(rq); |
320ae51f | 1200 | break; |
fc17b653 CH |
1201 | } |
1202 | ||
1203 | if (unlikely(ret != BLK_STS_OK)) { | |
93efe981 | 1204 | errors++; |
2a842aca | 1205 | blk_mq_end_request(rq, BLK_STS_IOERR); |
fc17b653 | 1206 | continue; |
320ae51f JA |
1207 | } |
1208 | ||
fc17b653 | 1209 | queued++; |
81380ca1 | 1210 | } while (!list_empty(list)); |
320ae51f | 1211 | |
703fd1c0 | 1212 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
1213 | |
1214 | /* | |
1215 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1216 | * that is where we will continue on next queue run. | |
1217 | */ | |
f04c3df3 | 1218 | if (!list_empty(list)) { |
86ff7c2a ML |
1219 | bool needs_restart; |
1220 | ||
320ae51f | 1221 | spin_lock(&hctx->lock); |
c13660a0 | 1222 | list_splice_init(list, &hctx->dispatch); |
320ae51f | 1223 | spin_unlock(&hctx->lock); |
f04c3df3 | 1224 | |
9ba52e58 | 1225 | /* |
710c785f BVA |
1226 | * If SCHED_RESTART was set by the caller of this function and |
1227 | * it is no longer set that means that it was cleared by another | |
1228 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1229 | * |
eb619fdb JA |
1230 | * If 'no_tag' is set, that means that we failed getting |
1231 | * a driver tag with an I/O scheduler attached. If our dispatch | |
1232 | * waitqueue is no longer active, ensure that we run the queue | |
1233 | * AFTER adding our entries back to the list. | |
bd166ef1 | 1234 | * |
710c785f BVA |
1235 | * If no I/O scheduler has been configured it is possible that |
1236 | * the hardware queue got stopped and restarted before requests | |
1237 | * were pushed back onto the dispatch list. Rerun the queue to | |
1238 | * avoid starvation. Notes: | |
1239 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1240 | * been stopped before rerunning a queue. | |
1241 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1242 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1243 | * and dm-rq. |
86ff7c2a ML |
1244 | * |
1245 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
1246 | * bit is set, run queue after a delay to avoid IO stalls | |
1247 | * that could otherwise occur if the queue is idle. | |
bd166ef1 | 1248 | */ |
86ff7c2a ML |
1249 | needs_restart = blk_mq_sched_needs_restart(hctx); |
1250 | if (!needs_restart || | |
eb619fdb | 1251 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
bd166ef1 | 1252 | blk_mq_run_hw_queue(hctx, true); |
86ff7c2a ML |
1253 | else if (needs_restart && (ret == BLK_STS_RESOURCE)) |
1254 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); | |
1f57f8d4 | 1255 | |
6e768717 | 1256 | blk_mq_update_dispatch_busy(hctx, true); |
1f57f8d4 | 1257 | return false; |
6e768717 ML |
1258 | } else |
1259 | blk_mq_update_dispatch_busy(hctx, false); | |
f04c3df3 | 1260 | |
1f57f8d4 JA |
1261 | /* |
1262 | * If the host/device is unable to accept more work, inform the | |
1263 | * caller of that. | |
1264 | */ | |
1265 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) | |
1266 | return false; | |
1267 | ||
93efe981 | 1268 | return (queued + errors) != 0; |
f04c3df3 JA |
1269 | } |
1270 | ||
6a83e74d BVA |
1271 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1272 | { | |
1273 | int srcu_idx; | |
1274 | ||
b7a71e66 JA |
1275 | /* |
1276 | * We should be running this queue from one of the CPUs that | |
1277 | * are mapped to it. | |
7df938fb ML |
1278 | * |
1279 | * There are at least two related races now between setting | |
1280 | * hctx->next_cpu from blk_mq_hctx_next_cpu() and running | |
1281 | * __blk_mq_run_hw_queue(): | |
1282 | * | |
1283 | * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(), | |
1284 | * but later it becomes online, then this warning is harmless | |
1285 | * at all | |
1286 | * | |
1287 | * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(), | |
1288 | * but later it becomes offline, then the warning can't be | |
1289 | * triggered, and we depend on blk-mq timeout handler to | |
1290 | * handle dispatched requests to this hctx | |
b7a71e66 | 1291 | */ |
7df938fb ML |
1292 | if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && |
1293 | cpu_online(hctx->next_cpu)) { | |
1294 | printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n", | |
1295 | raw_smp_processor_id(), | |
1296 | cpumask_empty(hctx->cpumask) ? "inactive": "active"); | |
1297 | dump_stack(); | |
1298 | } | |
6a83e74d | 1299 | |
b7a71e66 JA |
1300 | /* |
1301 | * We can't run the queue inline with ints disabled. Ensure that | |
1302 | * we catch bad users of this early. | |
1303 | */ | |
1304 | WARN_ON_ONCE(in_interrupt()); | |
1305 | ||
04ced159 | 1306 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1307 | |
04ced159 JA |
1308 | hctx_lock(hctx, &srcu_idx); |
1309 | blk_mq_sched_dispatch_requests(hctx); | |
1310 | hctx_unlock(hctx, srcu_idx); | |
6a83e74d BVA |
1311 | } |
1312 | ||
f82ddf19 ML |
1313 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
1314 | { | |
1315 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
1316 | ||
1317 | if (cpu >= nr_cpu_ids) | |
1318 | cpu = cpumask_first(hctx->cpumask); | |
1319 | return cpu; | |
1320 | } | |
1321 | ||
506e931f JA |
1322 | /* |
1323 | * It'd be great if the workqueue API had a way to pass | |
1324 | * in a mask and had some smarts for more clever placement. | |
1325 | * For now we just round-robin here, switching for every | |
1326 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1327 | */ | |
1328 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1329 | { | |
7bed4595 | 1330 | bool tried = false; |
476f8c98 | 1331 | int next_cpu = hctx->next_cpu; |
7bed4595 | 1332 | |
b657d7e6 CH |
1333 | if (hctx->queue->nr_hw_queues == 1) |
1334 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1335 | |
1336 | if (--hctx->next_cpu_batch <= 0) { | |
7bed4595 | 1337 | select_cpu: |
476f8c98 | 1338 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
20e4d813 | 1339 | cpu_online_mask); |
506e931f | 1340 | if (next_cpu >= nr_cpu_ids) |
f82ddf19 | 1341 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
1342 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
1343 | } | |
1344 | ||
7bed4595 ML |
1345 | /* |
1346 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
1347 | * and it should only happen in the path of handling CPU DEAD. | |
1348 | */ | |
476f8c98 | 1349 | if (!cpu_online(next_cpu)) { |
7bed4595 ML |
1350 | if (!tried) { |
1351 | tried = true; | |
1352 | goto select_cpu; | |
1353 | } | |
1354 | ||
1355 | /* | |
1356 | * Make sure to re-select CPU next time once after CPUs | |
1357 | * in hctx->cpumask become online again. | |
1358 | */ | |
476f8c98 | 1359 | hctx->next_cpu = next_cpu; |
7bed4595 ML |
1360 | hctx->next_cpu_batch = 1; |
1361 | return WORK_CPU_UNBOUND; | |
1362 | } | |
476f8c98 ML |
1363 | |
1364 | hctx->next_cpu = next_cpu; | |
1365 | return next_cpu; | |
506e931f JA |
1366 | } |
1367 | ||
7587a5ae BVA |
1368 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1369 | unsigned long msecs) | |
320ae51f | 1370 | { |
5435c023 | 1371 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f JA |
1372 | return; |
1373 | ||
1b792f2f | 1374 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1375 | int cpu = get_cpu(); |
1376 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1377 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1378 | put_cpu(); |
398205b8 PB |
1379 | return; |
1380 | } | |
e4043dcf | 1381 | |
2a90d4aa | 1382 | put_cpu(); |
e4043dcf | 1383 | } |
398205b8 | 1384 | |
ae943d20 BVA |
1385 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
1386 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
1387 | } |
1388 | ||
1389 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1390 | { | |
1391 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1392 | } | |
1393 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1394 | ||
79f720a7 | 1395 | bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 1396 | { |
24f5a90f ML |
1397 | int srcu_idx; |
1398 | bool need_run; | |
1399 | ||
1400 | /* | |
1401 | * When queue is quiesced, we may be switching io scheduler, or | |
1402 | * updating nr_hw_queues, or other things, and we can't run queue | |
1403 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
1404 | * | |
1405 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
1406 | * quiesced. | |
1407 | */ | |
04ced159 JA |
1408 | hctx_lock(hctx, &srcu_idx); |
1409 | need_run = !blk_queue_quiesced(hctx->queue) && | |
1410 | blk_mq_hctx_has_pending(hctx); | |
1411 | hctx_unlock(hctx, srcu_idx); | |
24f5a90f ML |
1412 | |
1413 | if (need_run) { | |
79f720a7 JA |
1414 | __blk_mq_delay_run_hw_queue(hctx, async, 0); |
1415 | return true; | |
1416 | } | |
1417 | ||
1418 | return false; | |
320ae51f | 1419 | } |
5b727272 | 1420 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1421 | |
b94ec296 | 1422 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1423 | { |
1424 | struct blk_mq_hw_ctx *hctx; | |
1425 | int i; | |
1426 | ||
1427 | queue_for_each_hw_ctx(q, hctx, i) { | |
79f720a7 | 1428 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1429 | continue; |
1430 | ||
b94ec296 | 1431 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1432 | } |
1433 | } | |
b94ec296 | 1434 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1435 | |
fd001443 BVA |
1436 | /** |
1437 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1438 | * @q: request queue. | |
1439 | * | |
1440 | * The caller is responsible for serializing this function against | |
1441 | * blk_mq_{start,stop}_hw_queue(). | |
1442 | */ | |
1443 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1444 | { | |
1445 | struct blk_mq_hw_ctx *hctx; | |
1446 | int i; | |
1447 | ||
1448 | queue_for_each_hw_ctx(q, hctx, i) | |
1449 | if (blk_mq_hctx_stopped(hctx)) | |
1450 | return true; | |
1451 | ||
1452 | return false; | |
1453 | } | |
1454 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1455 | ||
39a70c76 ML |
1456 | /* |
1457 | * This function is often used for pausing .queue_rq() by driver when | |
1458 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1459 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1460 | * |
1461 | * We do not guarantee that dispatch can be drained or blocked | |
1462 | * after blk_mq_stop_hw_queue() returns. Please use | |
1463 | * blk_mq_quiesce_queue() for that requirement. | |
1464 | */ | |
2719aa21 JA |
1465 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1466 | { | |
641a9ed6 | 1467 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 1468 | |
641a9ed6 | 1469 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 1470 | } |
641a9ed6 | 1471 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 1472 | |
39a70c76 ML |
1473 | /* |
1474 | * This function is often used for pausing .queue_rq() by driver when | |
1475 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1476 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1477 | * |
1478 | * We do not guarantee that dispatch can be drained or blocked | |
1479 | * after blk_mq_stop_hw_queues() returns. Please use | |
1480 | * blk_mq_quiesce_queue() for that requirement. | |
1481 | */ | |
2719aa21 JA |
1482 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1483 | { | |
641a9ed6 ML |
1484 | struct blk_mq_hw_ctx *hctx; |
1485 | int i; | |
1486 | ||
1487 | queue_for_each_hw_ctx(q, hctx, i) | |
1488 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
1489 | } |
1490 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1491 | ||
320ae51f JA |
1492 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1493 | { | |
1494 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1495 | |
0ffbce80 | 1496 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1497 | } |
1498 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1499 | ||
2f268556 CH |
1500 | void blk_mq_start_hw_queues(struct request_queue *q) |
1501 | { | |
1502 | struct blk_mq_hw_ctx *hctx; | |
1503 | int i; | |
1504 | ||
1505 | queue_for_each_hw_ctx(q, hctx, i) | |
1506 | blk_mq_start_hw_queue(hctx); | |
1507 | } | |
1508 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1509 | ||
ae911c5e JA |
1510 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1511 | { | |
1512 | if (!blk_mq_hctx_stopped(hctx)) | |
1513 | return; | |
1514 | ||
1515 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1516 | blk_mq_run_hw_queue(hctx, async); | |
1517 | } | |
1518 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1519 | ||
1b4a3258 | 1520 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1521 | { |
1522 | struct blk_mq_hw_ctx *hctx; | |
1523 | int i; | |
1524 | ||
ae911c5e JA |
1525 | queue_for_each_hw_ctx(q, hctx, i) |
1526 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1527 | } |
1528 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1529 | ||
70f4db63 | 1530 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1531 | { |
1532 | struct blk_mq_hw_ctx *hctx; | |
1533 | ||
9f993737 | 1534 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 1535 | |
21c6e939 | 1536 | /* |
15fe8a90 | 1537 | * If we are stopped, don't run the queue. |
21c6e939 | 1538 | */ |
15fe8a90 | 1539 | if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
0196d6b4 | 1540 | return; |
7587a5ae BVA |
1541 | |
1542 | __blk_mq_run_hw_queue(hctx); | |
1543 | } | |
1544 | ||
cfd0c552 | 1545 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1546 | struct request *rq, |
1547 | bool at_head) | |
320ae51f | 1548 | { |
e57690fe JA |
1549 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1550 | ||
7b607814 BVA |
1551 | lockdep_assert_held(&ctx->lock); |
1552 | ||
01b983c9 JA |
1553 | trace_block_rq_insert(hctx->queue, rq); |
1554 | ||
72a0a36e CH |
1555 | if (at_head) |
1556 | list_add(&rq->queuelist, &ctx->rq_list); | |
1557 | else | |
1558 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1559 | } |
4bb659b1 | 1560 | |
2c3ad667 JA |
1561 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1562 | bool at_head) | |
cfd0c552 ML |
1563 | { |
1564 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1565 | ||
7b607814 BVA |
1566 | lockdep_assert_held(&ctx->lock); |
1567 | ||
e57690fe | 1568 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1569 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1570 | } |
1571 | ||
157f377b JA |
1572 | /* |
1573 | * Should only be used carefully, when the caller knows we want to | |
1574 | * bypass a potential IO scheduler on the target device. | |
1575 | */ | |
b0850297 | 1576 | void blk_mq_request_bypass_insert(struct request *rq, bool run_queue) |
157f377b JA |
1577 | { |
1578 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
f9afca4d JA |
1579 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, rq->cmd_flags, |
1580 | ctx->cpu); | |
157f377b JA |
1581 | |
1582 | spin_lock(&hctx->lock); | |
1583 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
1584 | spin_unlock(&hctx->lock); | |
1585 | ||
b0850297 ML |
1586 | if (run_queue) |
1587 | blk_mq_run_hw_queue(hctx, false); | |
157f377b JA |
1588 | } |
1589 | ||
bd166ef1 JA |
1590 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1591 | struct list_head *list) | |
320ae51f JA |
1592 | |
1593 | { | |
3f0cedc7 ML |
1594 | struct request *rq; |
1595 | ||
320ae51f JA |
1596 | /* |
1597 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1598 | * offline now | |
1599 | */ | |
3f0cedc7 | 1600 | list_for_each_entry(rq, list, queuelist) { |
e57690fe | 1601 | BUG_ON(rq->mq_ctx != ctx); |
3f0cedc7 | 1602 | trace_block_rq_insert(hctx->queue, rq); |
320ae51f | 1603 | } |
3f0cedc7 ML |
1604 | |
1605 | spin_lock(&ctx->lock); | |
1606 | list_splice_tail_init(list, &ctx->rq_list); | |
cfd0c552 | 1607 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1608 | spin_unlock(&ctx->lock); |
320ae51f JA |
1609 | } |
1610 | ||
1611 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1612 | { | |
1613 | struct request *rqa = container_of(a, struct request, queuelist); | |
1614 | struct request *rqb = container_of(b, struct request, queuelist); | |
1615 | ||
1616 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1617 | (rqa->mq_ctx == rqb->mq_ctx && | |
1618 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1619 | } | |
1620 | ||
1621 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1622 | { | |
1623 | struct blk_mq_ctx *this_ctx; | |
1624 | struct request_queue *this_q; | |
1625 | struct request *rq; | |
1626 | LIST_HEAD(list); | |
1627 | LIST_HEAD(ctx_list); | |
1628 | unsigned int depth; | |
1629 | ||
1630 | list_splice_init(&plug->mq_list, &list); | |
1631 | ||
1632 | list_sort(NULL, &list, plug_ctx_cmp); | |
1633 | ||
1634 | this_q = NULL; | |
1635 | this_ctx = NULL; | |
1636 | depth = 0; | |
1637 | ||
1638 | while (!list_empty(&list)) { | |
1639 | rq = list_entry_rq(list.next); | |
1640 | list_del_init(&rq->queuelist); | |
1641 | BUG_ON(!rq->q); | |
1642 | if (rq->mq_ctx != this_ctx) { | |
1643 | if (this_ctx) { | |
587562d0 | 1644 | trace_block_unplug(this_q, depth, !from_schedule); |
bd166ef1 JA |
1645 | blk_mq_sched_insert_requests(this_q, this_ctx, |
1646 | &ctx_list, | |
1647 | from_schedule); | |
320ae51f JA |
1648 | } |
1649 | ||
1650 | this_ctx = rq->mq_ctx; | |
1651 | this_q = rq->q; | |
1652 | depth = 0; | |
1653 | } | |
1654 | ||
1655 | depth++; | |
1656 | list_add_tail(&rq->queuelist, &ctx_list); | |
1657 | } | |
1658 | ||
1659 | /* | |
1660 | * If 'this_ctx' is set, we know we have entries to complete | |
1661 | * on 'ctx_list'. Do those. | |
1662 | */ | |
1663 | if (this_ctx) { | |
587562d0 | 1664 | trace_block_unplug(this_q, depth, !from_schedule); |
bd166ef1 JA |
1665 | blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list, |
1666 | from_schedule); | |
320ae51f JA |
1667 | } |
1668 | } | |
1669 | ||
1670 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1671 | { | |
da8d7f07 | 1672 | blk_init_request_from_bio(rq, bio); |
4b570521 | 1673 | |
6e85eaf3 | 1674 | blk_account_io_start(rq, true); |
320ae51f JA |
1675 | } |
1676 | ||
fd2d3326 JA |
1677 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) |
1678 | { | |
bd166ef1 JA |
1679 | if (rq->tag != -1) |
1680 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); | |
1681 | ||
1682 | return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true); | |
fd2d3326 JA |
1683 | } |
1684 | ||
0f95549c MS |
1685 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
1686 | struct request *rq, | |
1687 | blk_qc_t *cookie) | |
f984df1f | 1688 | { |
f984df1f | 1689 | struct request_queue *q = rq->q; |
f984df1f SL |
1690 | struct blk_mq_queue_data bd = { |
1691 | .rq = rq, | |
d945a365 | 1692 | .last = true, |
f984df1f | 1693 | }; |
bd166ef1 | 1694 | blk_qc_t new_cookie; |
f06345ad | 1695 | blk_status_t ret; |
0f95549c MS |
1696 | |
1697 | new_cookie = request_to_qc_t(hctx, rq); | |
1698 | ||
1699 | /* | |
1700 | * For OK queue, we are done. For error, caller may kill it. | |
1701 | * Any other error (busy), just add it to our list as we | |
1702 | * previously would have done. | |
1703 | */ | |
1704 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
1705 | switch (ret) { | |
1706 | case BLK_STS_OK: | |
6ce3dd6e | 1707 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
1708 | *cookie = new_cookie; |
1709 | break; | |
1710 | case BLK_STS_RESOURCE: | |
86ff7c2a | 1711 | case BLK_STS_DEV_RESOURCE: |
6ce3dd6e | 1712 | blk_mq_update_dispatch_busy(hctx, true); |
0f95549c MS |
1713 | __blk_mq_requeue_request(rq); |
1714 | break; | |
1715 | default: | |
6ce3dd6e | 1716 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
1717 | *cookie = BLK_QC_T_NONE; |
1718 | break; | |
1719 | } | |
1720 | ||
1721 | return ret; | |
1722 | } | |
1723 | ||
0f95549c MS |
1724 | static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1725 | struct request *rq, | |
396eaf21 ML |
1726 | blk_qc_t *cookie, |
1727 | bool bypass_insert) | |
0f95549c MS |
1728 | { |
1729 | struct request_queue *q = rq->q; | |
d964f04a ML |
1730 | bool run_queue = true; |
1731 | ||
23d4ee19 ML |
1732 | /* |
1733 | * RCU or SRCU read lock is needed before checking quiesced flag. | |
1734 | * | |
1735 | * When queue is stopped or quiesced, ignore 'bypass_insert' from | |
c77ff7fd | 1736 | * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller, |
23d4ee19 ML |
1737 | * and avoid driver to try to dispatch again. |
1738 | */ | |
f4560ffe | 1739 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { |
d964f04a | 1740 | run_queue = false; |
23d4ee19 | 1741 | bypass_insert = false; |
d964f04a ML |
1742 | goto insert; |
1743 | } | |
f984df1f | 1744 | |
396eaf21 | 1745 | if (q->elevator && !bypass_insert) |
2253efc8 BVA |
1746 | goto insert; |
1747 | ||
0bca799b | 1748 | if (!blk_mq_get_dispatch_budget(hctx)) |
bd166ef1 JA |
1749 | goto insert; |
1750 | ||
8ab6bb9e | 1751 | if (!blk_mq_get_driver_tag(rq)) { |
0bca799b | 1752 | blk_mq_put_dispatch_budget(hctx); |
de148297 | 1753 | goto insert; |
88022d72 | 1754 | } |
de148297 | 1755 | |
0f95549c | 1756 | return __blk_mq_issue_directly(hctx, rq, cookie); |
2253efc8 | 1757 | insert: |
396eaf21 ML |
1758 | if (bypass_insert) |
1759 | return BLK_STS_RESOURCE; | |
0f95549c | 1760 | |
23d4ee19 | 1761 | blk_mq_sched_insert_request(rq, false, run_queue, false); |
0f95549c | 1762 | return BLK_STS_OK; |
f984df1f SL |
1763 | } |
1764 | ||
5eb6126e CH |
1765 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1766 | struct request *rq, blk_qc_t *cookie) | |
1767 | { | |
0f95549c | 1768 | blk_status_t ret; |
04ced159 | 1769 | int srcu_idx; |
bf4907c0 | 1770 | |
04ced159 | 1771 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1772 | |
04ced159 | 1773 | hctx_lock(hctx, &srcu_idx); |
0f95549c | 1774 | |
396eaf21 | 1775 | ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false); |
86ff7c2a | 1776 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) |
23d4ee19 | 1777 | blk_mq_sched_insert_request(rq, false, true, false); |
0f95549c MS |
1778 | else if (ret != BLK_STS_OK) |
1779 | blk_mq_end_request(rq, ret); | |
1780 | ||
04ced159 | 1781 | hctx_unlock(hctx, srcu_idx); |
5eb6126e CH |
1782 | } |
1783 | ||
c77ff7fd | 1784 | blk_status_t blk_mq_request_issue_directly(struct request *rq) |
396eaf21 ML |
1785 | { |
1786 | blk_status_t ret; | |
1787 | int srcu_idx; | |
1788 | blk_qc_t unused_cookie; | |
1789 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
f9afca4d JA |
1790 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, rq->cmd_flags, |
1791 | ctx->cpu); | |
396eaf21 ML |
1792 | |
1793 | hctx_lock(hctx, &srcu_idx); | |
1794 | ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true); | |
1795 | hctx_unlock(hctx, srcu_idx); | |
1796 | ||
1797 | return ret; | |
5eb6126e CH |
1798 | } |
1799 | ||
6ce3dd6e ML |
1800 | void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
1801 | struct list_head *list) | |
1802 | { | |
1803 | while (!list_empty(list)) { | |
1804 | blk_status_t ret; | |
1805 | struct request *rq = list_first_entry(list, struct request, | |
1806 | queuelist); | |
1807 | ||
1808 | list_del_init(&rq->queuelist); | |
1809 | ret = blk_mq_request_issue_directly(rq); | |
1810 | if (ret != BLK_STS_OK) { | |
8824f622 ML |
1811 | if (ret == BLK_STS_RESOURCE || |
1812 | ret == BLK_STS_DEV_RESOURCE) { | |
1813 | list_add(&rq->queuelist, list); | |
1814 | break; | |
1815 | } | |
1816 | blk_mq_end_request(rq, ret); | |
6ce3dd6e ML |
1817 | } |
1818 | } | |
1819 | } | |
1820 | ||
dece1635 | 1821 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1822 | { |
ef295ecf | 1823 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1824 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
f9afca4d | 1825 | struct blk_mq_alloc_data data = { .flags = 0, .cmd_flags = bio->bi_opf }; |
07068d5b | 1826 | struct request *rq; |
5eb6126e | 1827 | unsigned int request_count = 0; |
f984df1f | 1828 | struct blk_plug *plug; |
5b3f341f | 1829 | struct request *same_queue_rq = NULL; |
7b371636 | 1830 | blk_qc_t cookie; |
07068d5b JA |
1831 | |
1832 | blk_queue_bounce(q, &bio); | |
1833 | ||
af67c31f | 1834 | blk_queue_split(q, &bio); |
f36ea50c | 1835 | |
e23947bd | 1836 | if (!bio_integrity_prep(bio)) |
dece1635 | 1837 | return BLK_QC_T_NONE; |
07068d5b | 1838 | |
87c279e6 OS |
1839 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1840 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1841 | return BLK_QC_T_NONE; | |
f984df1f | 1842 | |
bd166ef1 JA |
1843 | if (blk_mq_sched_bio_merge(q, bio)) |
1844 | return BLK_QC_T_NONE; | |
1845 | ||
c1c80384 | 1846 | rq_qos_throttle(q, bio, NULL); |
87760e5e | 1847 | |
f9afca4d | 1848 | rq = blk_mq_get_request(q, bio, &data); |
87760e5e | 1849 | if (unlikely(!rq)) { |
c1c80384 | 1850 | rq_qos_cleanup(q, bio); |
03a07c92 GR |
1851 | if (bio->bi_opf & REQ_NOWAIT) |
1852 | bio_wouldblock_error(bio); | |
dece1635 | 1853 | return BLK_QC_T_NONE; |
87760e5e JA |
1854 | } |
1855 | ||
d6f1dda2 XW |
1856 | trace_block_getrq(q, bio, bio->bi_opf); |
1857 | ||
c1c80384 | 1858 | rq_qos_track(q, rq, bio); |
07068d5b | 1859 | |
fd2d3326 | 1860 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b | 1861 | |
f984df1f | 1862 | plug = current->plug; |
07068d5b | 1863 | if (unlikely(is_flush_fua)) { |
f984df1f | 1864 | blk_mq_put_ctx(data.ctx); |
07068d5b | 1865 | blk_mq_bio_to_request(rq, bio); |
923218f6 ML |
1866 | |
1867 | /* bypass scheduler for flush rq */ | |
1868 | blk_insert_flush(rq); | |
1869 | blk_mq_run_hw_queue(data.hctx, true); | |
a4d907b6 | 1870 | } else if (plug && q->nr_hw_queues == 1) { |
600271d9 SL |
1871 | struct request *last = NULL; |
1872 | ||
b00c53e8 | 1873 | blk_mq_put_ctx(data.ctx); |
e6c4438b | 1874 | blk_mq_bio_to_request(rq, bio); |
0a6219a9 ML |
1875 | |
1876 | /* | |
1877 | * @request_count may become stale because of schedule | |
1878 | * out, so check the list again. | |
1879 | */ | |
1880 | if (list_empty(&plug->mq_list)) | |
1881 | request_count = 0; | |
254d259d CH |
1882 | else if (blk_queue_nomerges(q)) |
1883 | request_count = blk_plug_queued_count(q); | |
1884 | ||
676d0607 | 1885 | if (!request_count) |
e6c4438b | 1886 | trace_block_plug(q); |
600271d9 SL |
1887 | else |
1888 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c | 1889 | |
600271d9 SL |
1890 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1891 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1892 | blk_flush_plug_list(plug, false); |
1893 | trace_block_plug(q); | |
320ae51f | 1894 | } |
b094f89c | 1895 | |
e6c4438b | 1896 | list_add_tail(&rq->queuelist, &plug->mq_list); |
2299722c | 1897 | } else if (plug && !blk_queue_nomerges(q)) { |
bd166ef1 | 1898 | blk_mq_bio_to_request(rq, bio); |
07068d5b | 1899 | |
07068d5b | 1900 | /* |
6a83e74d | 1901 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1902 | * Otherwise the existing request in the plug list will be |
1903 | * issued. So the plug list will have one request at most | |
2299722c CH |
1904 | * The plug list might get flushed before this. If that happens, |
1905 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 1906 | */ |
2299722c CH |
1907 | if (list_empty(&plug->mq_list)) |
1908 | same_queue_rq = NULL; | |
1909 | if (same_queue_rq) | |
1910 | list_del_init(&same_queue_rq->queuelist); | |
1911 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1912 | ||
bf4907c0 JA |
1913 | blk_mq_put_ctx(data.ctx); |
1914 | ||
dad7a3be ML |
1915 | if (same_queue_rq) { |
1916 | data.hctx = blk_mq_map_queue(q, | |
f9afca4d | 1917 | same_queue_rq->cmd_flags, |
dad7a3be | 1918 | same_queue_rq->mq_ctx->cpu); |
2299722c CH |
1919 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, |
1920 | &cookie); | |
dad7a3be | 1921 | } |
6ce3dd6e ML |
1922 | } else if ((q->nr_hw_queues > 1 && is_sync) || (!q->elevator && |
1923 | !data.hctx->dispatch_busy)) { | |
bf4907c0 | 1924 | blk_mq_put_ctx(data.ctx); |
2299722c | 1925 | blk_mq_bio_to_request(rq, bio); |
2299722c | 1926 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); |
ab42f35d | 1927 | } else { |
b00c53e8 | 1928 | blk_mq_put_ctx(data.ctx); |
ab42f35d | 1929 | blk_mq_bio_to_request(rq, bio); |
8fa9f556 | 1930 | blk_mq_sched_insert_request(rq, false, true, true); |
ab42f35d | 1931 | } |
320ae51f | 1932 | |
7b371636 | 1933 | return cookie; |
320ae51f JA |
1934 | } |
1935 | ||
cc71a6f4 JA |
1936 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
1937 | unsigned int hctx_idx) | |
95363efd | 1938 | { |
e9b267d9 | 1939 | struct page *page; |
320ae51f | 1940 | |
24d2f903 | 1941 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1942 | int i; |
320ae51f | 1943 | |
24d2f903 | 1944 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
1945 | struct request *rq = tags->static_rqs[i]; |
1946 | ||
1947 | if (!rq) | |
e9b267d9 | 1948 | continue; |
d6296d39 | 1949 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 1950 | tags->static_rqs[i] = NULL; |
e9b267d9 | 1951 | } |
320ae51f | 1952 | } |
320ae51f | 1953 | |
24d2f903 CH |
1954 | while (!list_empty(&tags->page_list)) { |
1955 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1956 | list_del_init(&page->lru); |
f75782e4 CM |
1957 | /* |
1958 | * Remove kmemleak object previously allocated in | |
1959 | * blk_mq_init_rq_map(). | |
1960 | */ | |
1961 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1962 | __free_pages(page, page->private); |
1963 | } | |
cc71a6f4 | 1964 | } |
320ae51f | 1965 | |
cc71a6f4 JA |
1966 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
1967 | { | |
24d2f903 | 1968 | kfree(tags->rqs); |
cc71a6f4 | 1969 | tags->rqs = NULL; |
2af8cbe3 JA |
1970 | kfree(tags->static_rqs); |
1971 | tags->static_rqs = NULL; | |
320ae51f | 1972 | |
24d2f903 | 1973 | blk_mq_free_tags(tags); |
320ae51f JA |
1974 | } |
1975 | ||
cc71a6f4 JA |
1976 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
1977 | unsigned int hctx_idx, | |
1978 | unsigned int nr_tags, | |
1979 | unsigned int reserved_tags) | |
320ae51f | 1980 | { |
24d2f903 | 1981 | struct blk_mq_tags *tags; |
59f082e4 | 1982 | int node; |
320ae51f | 1983 | |
ed76e329 | 1984 | node = blk_mq_hw_queue_to_node(&set->map[0], hctx_idx); |
59f082e4 SL |
1985 | if (node == NUMA_NO_NODE) |
1986 | node = set->numa_node; | |
1987 | ||
1988 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, | |
24391c0d | 1989 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
24d2f903 CH |
1990 | if (!tags) |
1991 | return NULL; | |
320ae51f | 1992 | |
590b5b7d | 1993 | tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
36e1f3d1 | 1994 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 1995 | node); |
24d2f903 CH |
1996 | if (!tags->rqs) { |
1997 | blk_mq_free_tags(tags); | |
1998 | return NULL; | |
1999 | } | |
320ae51f | 2000 | |
590b5b7d KC |
2001 | tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
2002 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
2003 | node); | |
2af8cbe3 JA |
2004 | if (!tags->static_rqs) { |
2005 | kfree(tags->rqs); | |
2006 | blk_mq_free_tags(tags); | |
2007 | return NULL; | |
2008 | } | |
2009 | ||
cc71a6f4 JA |
2010 | return tags; |
2011 | } | |
2012 | ||
2013 | static size_t order_to_size(unsigned int order) | |
2014 | { | |
2015 | return (size_t)PAGE_SIZE << order; | |
2016 | } | |
2017 | ||
1d9bd516 TH |
2018 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
2019 | unsigned int hctx_idx, int node) | |
2020 | { | |
2021 | int ret; | |
2022 | ||
2023 | if (set->ops->init_request) { | |
2024 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
2025 | if (ret) | |
2026 | return ret; | |
2027 | } | |
2028 | ||
12f5b931 | 2029 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
1d9bd516 TH |
2030 | return 0; |
2031 | } | |
2032 | ||
cc71a6f4 JA |
2033 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2034 | unsigned int hctx_idx, unsigned int depth) | |
2035 | { | |
2036 | unsigned int i, j, entries_per_page, max_order = 4; | |
2037 | size_t rq_size, left; | |
59f082e4 SL |
2038 | int node; |
2039 | ||
ed76e329 | 2040 | node = blk_mq_hw_queue_to_node(&set->map[0], hctx_idx); |
59f082e4 SL |
2041 | if (node == NUMA_NO_NODE) |
2042 | node = set->numa_node; | |
cc71a6f4 JA |
2043 | |
2044 | INIT_LIST_HEAD(&tags->page_list); | |
2045 | ||
320ae51f JA |
2046 | /* |
2047 | * rq_size is the size of the request plus driver payload, rounded | |
2048 | * to the cacheline size | |
2049 | */ | |
24d2f903 | 2050 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 2051 | cache_line_size()); |
cc71a6f4 | 2052 | left = rq_size * depth; |
320ae51f | 2053 | |
cc71a6f4 | 2054 | for (i = 0; i < depth; ) { |
320ae51f JA |
2055 | int this_order = max_order; |
2056 | struct page *page; | |
2057 | int to_do; | |
2058 | void *p; | |
2059 | ||
b3a834b1 | 2060 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
2061 | this_order--; |
2062 | ||
2063 | do { | |
59f082e4 | 2064 | page = alloc_pages_node(node, |
36e1f3d1 | 2065 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 2066 | this_order); |
320ae51f JA |
2067 | if (page) |
2068 | break; | |
2069 | if (!this_order--) | |
2070 | break; | |
2071 | if (order_to_size(this_order) < rq_size) | |
2072 | break; | |
2073 | } while (1); | |
2074 | ||
2075 | if (!page) | |
24d2f903 | 2076 | goto fail; |
320ae51f JA |
2077 | |
2078 | page->private = this_order; | |
24d2f903 | 2079 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
2080 | |
2081 | p = page_address(page); | |
f75782e4 CM |
2082 | /* |
2083 | * Allow kmemleak to scan these pages as they contain pointers | |
2084 | * to additional allocations like via ops->init_request(). | |
2085 | */ | |
36e1f3d1 | 2086 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 2087 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 2088 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
2089 | left -= to_do * rq_size; |
2090 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
2091 | struct request *rq = p; |
2092 | ||
2093 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
2094 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
2095 | tags->static_rqs[i] = NULL; | |
2096 | goto fail; | |
e9b267d9 CH |
2097 | } |
2098 | ||
320ae51f JA |
2099 | p += rq_size; |
2100 | i++; | |
2101 | } | |
2102 | } | |
cc71a6f4 | 2103 | return 0; |
320ae51f | 2104 | |
24d2f903 | 2105 | fail: |
cc71a6f4 JA |
2106 | blk_mq_free_rqs(set, tags, hctx_idx); |
2107 | return -ENOMEM; | |
320ae51f JA |
2108 | } |
2109 | ||
e57690fe JA |
2110 | /* |
2111 | * 'cpu' is going away. splice any existing rq_list entries from this | |
2112 | * software queue to the hw queue dispatch list, and ensure that it | |
2113 | * gets run. | |
2114 | */ | |
9467f859 | 2115 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 2116 | { |
9467f859 | 2117 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
2118 | struct blk_mq_ctx *ctx; |
2119 | LIST_HEAD(tmp); | |
2120 | ||
9467f859 | 2121 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 2122 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
2123 | |
2124 | spin_lock(&ctx->lock); | |
2125 | if (!list_empty(&ctx->rq_list)) { | |
2126 | list_splice_init(&ctx->rq_list, &tmp); | |
2127 | blk_mq_hctx_clear_pending(hctx, ctx); | |
2128 | } | |
2129 | spin_unlock(&ctx->lock); | |
2130 | ||
2131 | if (list_empty(&tmp)) | |
9467f859 | 2132 | return 0; |
484b4061 | 2133 | |
e57690fe JA |
2134 | spin_lock(&hctx->lock); |
2135 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
2136 | spin_unlock(&hctx->lock); | |
484b4061 JA |
2137 | |
2138 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 2139 | return 0; |
484b4061 JA |
2140 | } |
2141 | ||
9467f859 | 2142 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 2143 | { |
9467f859 TG |
2144 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
2145 | &hctx->cpuhp_dead); | |
484b4061 JA |
2146 | } |
2147 | ||
c3b4afca | 2148 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
2149 | static void blk_mq_exit_hctx(struct request_queue *q, |
2150 | struct blk_mq_tag_set *set, | |
2151 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
2152 | { | |
8ab0b7dc ML |
2153 | if (blk_mq_hw_queue_mapped(hctx)) |
2154 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 2155 | |
f70ced09 | 2156 | if (set->ops->exit_request) |
d6296d39 | 2157 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); |
f70ced09 | 2158 | |
08e98fc6 ML |
2159 | if (set->ops->exit_hctx) |
2160 | set->ops->exit_hctx(hctx, hctx_idx); | |
2161 | ||
6a83e74d | 2162 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2163 | cleanup_srcu_struct(hctx->srcu); |
6a83e74d | 2164 | |
9467f859 | 2165 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 2166 | blk_free_flush_queue(hctx->fq); |
88459642 | 2167 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2168 | } |
2169 | ||
624dbe47 ML |
2170 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
2171 | struct blk_mq_tag_set *set, int nr_queue) | |
2172 | { | |
2173 | struct blk_mq_hw_ctx *hctx; | |
2174 | unsigned int i; | |
2175 | ||
2176 | queue_for_each_hw_ctx(q, hctx, i) { | |
2177 | if (i == nr_queue) | |
2178 | break; | |
477e19de | 2179 | blk_mq_debugfs_unregister_hctx(hctx); |
08e98fc6 | 2180 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 2181 | } |
624dbe47 ML |
2182 | } |
2183 | ||
08e98fc6 ML |
2184 | static int blk_mq_init_hctx(struct request_queue *q, |
2185 | struct blk_mq_tag_set *set, | |
2186 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 2187 | { |
08e98fc6 ML |
2188 | int node; |
2189 | ||
2190 | node = hctx->numa_node; | |
2191 | if (node == NUMA_NO_NODE) | |
2192 | node = hctx->numa_node = set->numa_node; | |
2193 | ||
9f993737 | 2194 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
2195 | spin_lock_init(&hctx->lock); |
2196 | INIT_LIST_HEAD(&hctx->dispatch); | |
2197 | hctx->queue = q; | |
2404e607 | 2198 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 2199 | |
9467f859 | 2200 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
2201 | |
2202 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
2203 | |
2204 | /* | |
08e98fc6 ML |
2205 | * Allocate space for all possible cpus to avoid allocation at |
2206 | * runtime | |
320ae51f | 2207 | */ |
d904bfa7 | 2208 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
5b202853 | 2209 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, node); |
08e98fc6 ML |
2210 | if (!hctx->ctxs) |
2211 | goto unregister_cpu_notifier; | |
320ae51f | 2212 | |
5b202853 JW |
2213 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), |
2214 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, node)) | |
08e98fc6 | 2215 | goto free_ctxs; |
320ae51f | 2216 | |
08e98fc6 | 2217 | hctx->nr_ctx = 0; |
320ae51f | 2218 | |
5815839b | 2219 | spin_lock_init(&hctx->dispatch_wait_lock); |
eb619fdb JA |
2220 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
2221 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
2222 | ||
08e98fc6 ML |
2223 | if (set->ops->init_hctx && |
2224 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
2225 | goto free_bitmap; | |
320ae51f | 2226 | |
5b202853 JW |
2227 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size, |
2228 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); | |
f70ced09 | 2229 | if (!hctx->fq) |
d48ece20 | 2230 | goto exit_hctx; |
320ae51f | 2231 | |
1d9bd516 | 2232 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node)) |
f70ced09 | 2233 | goto free_fq; |
320ae51f | 2234 | |
6a83e74d | 2235 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2236 | init_srcu_struct(hctx->srcu); |
6a83e74d | 2237 | |
08e98fc6 | 2238 | return 0; |
320ae51f | 2239 | |
f70ced09 ML |
2240 | free_fq: |
2241 | kfree(hctx->fq); | |
2242 | exit_hctx: | |
2243 | if (set->ops->exit_hctx) | |
2244 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 2245 | free_bitmap: |
88459642 | 2246 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2247 | free_ctxs: |
2248 | kfree(hctx->ctxs); | |
2249 | unregister_cpu_notifier: | |
9467f859 | 2250 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
2251 | return -1; |
2252 | } | |
320ae51f | 2253 | |
320ae51f JA |
2254 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
2255 | unsigned int nr_hw_queues) | |
2256 | { | |
2257 | unsigned int i; | |
2258 | ||
2259 | for_each_possible_cpu(i) { | |
2260 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2261 | struct blk_mq_hw_ctx *hctx; | |
2262 | ||
320ae51f JA |
2263 | __ctx->cpu = i; |
2264 | spin_lock_init(&__ctx->lock); | |
2265 | INIT_LIST_HEAD(&__ctx->rq_list); | |
2266 | __ctx->queue = q; | |
2267 | ||
320ae51f JA |
2268 | /* |
2269 | * Set local node, IFF we have more than one hw queue. If | |
2270 | * not, we remain on the home node of the device | |
2271 | */ | |
f9afca4d | 2272 | hctx = blk_mq_map_queue_type(q, 0, i); |
320ae51f | 2273 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) |
bffed457 | 2274 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
2275 | } |
2276 | } | |
2277 | ||
cc71a6f4 JA |
2278 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
2279 | { | |
2280 | int ret = 0; | |
2281 | ||
2282 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
2283 | set->queue_depth, set->reserved_tags); | |
2284 | if (!set->tags[hctx_idx]) | |
2285 | return false; | |
2286 | ||
2287 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2288 | set->queue_depth); | |
2289 | if (!ret) | |
2290 | return true; | |
2291 | ||
2292 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2293 | set->tags[hctx_idx] = NULL; | |
2294 | return false; | |
2295 | } | |
2296 | ||
2297 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2298 | unsigned int hctx_idx) | |
2299 | { | |
bd166ef1 JA |
2300 | if (set->tags[hctx_idx]) { |
2301 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); | |
2302 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2303 | set->tags[hctx_idx] = NULL; | |
2304 | } | |
cc71a6f4 JA |
2305 | } |
2306 | ||
4b855ad3 | 2307 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 2308 | { |
4412efec | 2309 | unsigned int i, hctx_idx; |
320ae51f JA |
2310 | struct blk_mq_hw_ctx *hctx; |
2311 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2312 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2313 | |
60de074b AM |
2314 | /* |
2315 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
2316 | */ | |
2317 | mutex_lock(&q->sysfs_lock); | |
2318 | ||
320ae51f | 2319 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 2320 | cpumask_clear(hctx->cpumask); |
320ae51f | 2321 | hctx->nr_ctx = 0; |
d416c92c | 2322 | hctx->dispatch_from = NULL; |
320ae51f JA |
2323 | } |
2324 | ||
2325 | /* | |
4b855ad3 | 2326 | * Map software to hardware queues. |
4412efec ML |
2327 | * |
2328 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 2329 | */ |
20e4d813 | 2330 | for_each_possible_cpu(i) { |
ed76e329 | 2331 | hctx_idx = set->map[0].mq_map[i]; |
4412efec ML |
2332 | /* unmapped hw queue can be remapped after CPU topo changed */ |
2333 | if (!set->tags[hctx_idx] && | |
2334 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
2335 | /* | |
2336 | * If tags initialization fail for some hctx, | |
2337 | * that hctx won't be brought online. In this | |
2338 | * case, remap the current ctx to hctx[0] which | |
2339 | * is guaranteed to always have tags allocated | |
2340 | */ | |
ed76e329 | 2341 | set->map[0].mq_map[i] = 0; |
4412efec ML |
2342 | } |
2343 | ||
897bb0c7 | 2344 | ctx = per_cpu_ptr(q->queue_ctx, i); |
f9afca4d | 2345 | hctx = blk_mq_map_queue_type(q, 0, i); |
868f2f0b | 2346 | |
e4043dcf | 2347 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
2348 | ctx->index_hw = hctx->nr_ctx; |
2349 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2350 | } | |
506e931f | 2351 | |
60de074b AM |
2352 | mutex_unlock(&q->sysfs_lock); |
2353 | ||
506e931f | 2354 | queue_for_each_hw_ctx(q, hctx, i) { |
4412efec ML |
2355 | /* |
2356 | * If no software queues are mapped to this hardware queue, | |
2357 | * disable it and free the request entries. | |
2358 | */ | |
2359 | if (!hctx->nr_ctx) { | |
2360 | /* Never unmap queue 0. We need it as a | |
2361 | * fallback in case of a new remap fails | |
2362 | * allocation | |
2363 | */ | |
2364 | if (i && set->tags[i]) | |
2365 | blk_mq_free_map_and_requests(set, i); | |
2366 | ||
2367 | hctx->tags = NULL; | |
2368 | continue; | |
2369 | } | |
484b4061 | 2370 | |
2a34c087 ML |
2371 | hctx->tags = set->tags[i]; |
2372 | WARN_ON(!hctx->tags); | |
2373 | ||
889fa31f CY |
2374 | /* |
2375 | * Set the map size to the number of mapped software queues. | |
2376 | * This is more accurate and more efficient than looping | |
2377 | * over all possibly mapped software queues. | |
2378 | */ | |
88459642 | 2379 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2380 | |
484b4061 JA |
2381 | /* |
2382 | * Initialize batch roundrobin counts | |
2383 | */ | |
f82ddf19 | 2384 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
2385 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2386 | } | |
320ae51f JA |
2387 | } |
2388 | ||
8e8320c9 JA |
2389 | /* |
2390 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
2391 | * the queue isn't live yet. | |
2392 | */ | |
2404e607 | 2393 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2394 | { |
2395 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2396 | int i; |
2397 | ||
2404e607 | 2398 | queue_for_each_hw_ctx(q, hctx, i) { |
97889f9a | 2399 | if (shared) |
2404e607 | 2400 | hctx->flags |= BLK_MQ_F_TAG_SHARED; |
97889f9a | 2401 | else |
2404e607 JM |
2402 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; |
2403 | } | |
2404 | } | |
2405 | ||
8e8320c9 JA |
2406 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, |
2407 | bool shared) | |
2404e607 JM |
2408 | { |
2409 | struct request_queue *q; | |
0d2602ca | 2410 | |
705cda97 BVA |
2411 | lockdep_assert_held(&set->tag_list_lock); |
2412 | ||
0d2602ca JA |
2413 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2414 | blk_mq_freeze_queue(q); | |
2404e607 | 2415 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2416 | blk_mq_unfreeze_queue(q); |
2417 | } | |
2418 | } | |
2419 | ||
2420 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2421 | { | |
2422 | struct blk_mq_tag_set *set = q->tag_set; | |
2423 | ||
0d2602ca | 2424 | mutex_lock(&set->tag_list_lock); |
705cda97 | 2425 | list_del_rcu(&q->tag_set_list); |
2404e607 JM |
2426 | if (list_is_singular(&set->tag_list)) { |
2427 | /* just transitioned to unshared */ | |
2428 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2429 | /* update existing queue */ | |
2430 | blk_mq_update_tag_set_depth(set, false); | |
2431 | } | |
0d2602ca | 2432 | mutex_unlock(&set->tag_list_lock); |
a347c7ad | 2433 | INIT_LIST_HEAD(&q->tag_set_list); |
0d2602ca JA |
2434 | } |
2435 | ||
2436 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2437 | struct request_queue *q) | |
2438 | { | |
0d2602ca | 2439 | mutex_lock(&set->tag_list_lock); |
2404e607 | 2440 | |
ff821d27 JA |
2441 | /* |
2442 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
2443 | */ | |
2444 | if (!list_empty(&set->tag_list) && | |
2445 | !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2404e607 JM |
2446 | set->flags |= BLK_MQ_F_TAG_SHARED; |
2447 | /* update existing queue */ | |
2448 | blk_mq_update_tag_set_depth(set, true); | |
2449 | } | |
2450 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2451 | queue_set_hctx_shared(q, true); | |
705cda97 | 2452 | list_add_tail_rcu(&q->tag_set_list, &set->tag_list); |
2404e607 | 2453 | |
0d2602ca JA |
2454 | mutex_unlock(&set->tag_list_lock); |
2455 | } | |
2456 | ||
e09aae7e ML |
2457 | /* |
2458 | * It is the actual release handler for mq, but we do it from | |
2459 | * request queue's release handler for avoiding use-after-free | |
2460 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2461 | * but we can't group ctx/kctx kobj without it. | |
2462 | */ | |
2463 | void blk_mq_release(struct request_queue *q) | |
2464 | { | |
2465 | struct blk_mq_hw_ctx *hctx; | |
2466 | unsigned int i; | |
2467 | ||
2468 | /* hctx kobj stays in hctx */ | |
c3b4afca ML |
2469 | queue_for_each_hw_ctx(q, hctx, i) { |
2470 | if (!hctx) | |
2471 | continue; | |
6c8b232e | 2472 | kobject_put(&hctx->kobj); |
c3b4afca | 2473 | } |
e09aae7e ML |
2474 | |
2475 | kfree(q->queue_hw_ctx); | |
2476 | ||
7ea5fe31 ML |
2477 | /* |
2478 | * release .mq_kobj and sw queue's kobject now because | |
2479 | * both share lifetime with request queue. | |
2480 | */ | |
2481 | blk_mq_sysfs_deinit(q); | |
2482 | ||
e09aae7e ML |
2483 | free_percpu(q->queue_ctx); |
2484 | } | |
2485 | ||
24d2f903 | 2486 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2487 | { |
2488 | struct request_queue *uninit_q, *q; | |
2489 | ||
5ee0524b | 2490 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node, NULL); |
b62c21b7 MS |
2491 | if (!uninit_q) |
2492 | return ERR_PTR(-ENOMEM); | |
2493 | ||
2494 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2495 | if (IS_ERR(q)) | |
2496 | blk_cleanup_queue(uninit_q); | |
2497 | ||
2498 | return q; | |
2499 | } | |
2500 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2501 | ||
9316a9ed JA |
2502 | /* |
2503 | * Helper for setting up a queue with mq ops, given queue depth, and | |
2504 | * the passed in mq ops flags. | |
2505 | */ | |
2506 | struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set, | |
2507 | const struct blk_mq_ops *ops, | |
2508 | unsigned int queue_depth, | |
2509 | unsigned int set_flags) | |
2510 | { | |
2511 | struct request_queue *q; | |
2512 | int ret; | |
2513 | ||
2514 | memset(set, 0, sizeof(*set)); | |
2515 | set->ops = ops; | |
2516 | set->nr_hw_queues = 1; | |
2517 | set->queue_depth = queue_depth; | |
2518 | set->numa_node = NUMA_NO_NODE; | |
2519 | set->flags = set_flags; | |
2520 | ||
2521 | ret = blk_mq_alloc_tag_set(set); | |
2522 | if (ret) | |
2523 | return ERR_PTR(ret); | |
2524 | ||
2525 | q = blk_mq_init_queue(set); | |
2526 | if (IS_ERR(q)) { | |
2527 | blk_mq_free_tag_set(set); | |
2528 | return q; | |
2529 | } | |
2530 | ||
2531 | return q; | |
2532 | } | |
2533 | EXPORT_SYMBOL(blk_mq_init_sq_queue); | |
2534 | ||
07319678 BVA |
2535 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) |
2536 | { | |
2537 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2538 | ||
05707b64 | 2539 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu), |
07319678 BVA |
2540 | __alignof__(struct blk_mq_hw_ctx)) != |
2541 | sizeof(struct blk_mq_hw_ctx)); | |
2542 | ||
2543 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2544 | hw_ctx_size += sizeof(struct srcu_struct); | |
2545 | ||
2546 | return hw_ctx_size; | |
2547 | } | |
2548 | ||
34d11ffa JW |
2549 | static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx( |
2550 | struct blk_mq_tag_set *set, struct request_queue *q, | |
2551 | int hctx_idx, int node) | |
2552 | { | |
2553 | struct blk_mq_hw_ctx *hctx; | |
2554 | ||
2555 | hctx = kzalloc_node(blk_mq_hw_ctx_size(set), | |
2556 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
2557 | node); | |
2558 | if (!hctx) | |
2559 | return NULL; | |
2560 | ||
2561 | if (!zalloc_cpumask_var_node(&hctx->cpumask, | |
2562 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
2563 | node)) { | |
2564 | kfree(hctx); | |
2565 | return NULL; | |
2566 | } | |
2567 | ||
2568 | atomic_set(&hctx->nr_active, 0); | |
2569 | hctx->numa_node = node; | |
2570 | hctx->queue_num = hctx_idx; | |
2571 | ||
2572 | if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) { | |
2573 | free_cpumask_var(hctx->cpumask); | |
2574 | kfree(hctx); | |
2575 | return NULL; | |
2576 | } | |
2577 | blk_mq_hctx_kobj_init(hctx); | |
2578 | ||
2579 | return hctx; | |
2580 | } | |
2581 | ||
868f2f0b KB |
2582 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2583 | struct request_queue *q) | |
320ae51f | 2584 | { |
e01ad46d | 2585 | int i, j, end; |
868f2f0b | 2586 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; |
f14bbe77 | 2587 | |
fb350e0a ML |
2588 | /* protect against switching io scheduler */ |
2589 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 2590 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2591 | int node; |
34d11ffa | 2592 | struct blk_mq_hw_ctx *hctx; |
868f2f0b | 2593 | |
ed76e329 | 2594 | node = blk_mq_hw_queue_to_node(&set->map[0], i); |
34d11ffa JW |
2595 | /* |
2596 | * If the hw queue has been mapped to another numa node, | |
2597 | * we need to realloc the hctx. If allocation fails, fallback | |
2598 | * to use the previous one. | |
2599 | */ | |
2600 | if (hctxs[i] && (hctxs[i]->numa_node == node)) | |
2601 | continue; | |
868f2f0b | 2602 | |
34d11ffa JW |
2603 | hctx = blk_mq_alloc_and_init_hctx(set, q, i, node); |
2604 | if (hctx) { | |
2605 | if (hctxs[i]) { | |
2606 | blk_mq_exit_hctx(q, set, hctxs[i], i); | |
2607 | kobject_put(&hctxs[i]->kobj); | |
2608 | } | |
2609 | hctxs[i] = hctx; | |
2610 | } else { | |
2611 | if (hctxs[i]) | |
2612 | pr_warn("Allocate new hctx on node %d fails,\ | |
2613 | fallback to previous one on node %d\n", | |
2614 | node, hctxs[i]->numa_node); | |
2615 | else | |
2616 | break; | |
868f2f0b | 2617 | } |
320ae51f | 2618 | } |
e01ad46d JW |
2619 | /* |
2620 | * Increasing nr_hw_queues fails. Free the newly allocated | |
2621 | * hctxs and keep the previous q->nr_hw_queues. | |
2622 | */ | |
2623 | if (i != set->nr_hw_queues) { | |
2624 | j = q->nr_hw_queues; | |
2625 | end = i; | |
2626 | } else { | |
2627 | j = i; | |
2628 | end = q->nr_hw_queues; | |
2629 | q->nr_hw_queues = set->nr_hw_queues; | |
2630 | } | |
34d11ffa | 2631 | |
e01ad46d | 2632 | for (; j < end; j++) { |
868f2f0b KB |
2633 | struct blk_mq_hw_ctx *hctx = hctxs[j]; |
2634 | ||
2635 | if (hctx) { | |
cc71a6f4 JA |
2636 | if (hctx->tags) |
2637 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 2638 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 2639 | kobject_put(&hctx->kobj); |
868f2f0b KB |
2640 | hctxs[j] = NULL; |
2641 | ||
2642 | } | |
2643 | } | |
fb350e0a | 2644 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
2645 | } |
2646 | ||
2647 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2648 | struct request_queue *q) | |
2649 | { | |
66841672 ML |
2650 | /* mark the queue as mq asap */ |
2651 | q->mq_ops = set->ops; | |
2652 | ||
34dbad5d | 2653 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
2654 | blk_mq_poll_stats_bkt, |
2655 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
2656 | if (!q->poll_cb) |
2657 | goto err_exit; | |
2658 | ||
868f2f0b KB |
2659 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2660 | if (!q->queue_ctx) | |
c7de5726 | 2661 | goto err_exit; |
868f2f0b | 2662 | |
737f98cf ML |
2663 | /* init q->mq_kobj and sw queues' kobjects */ |
2664 | blk_mq_sysfs_init(q); | |
2665 | ||
590b5b7d | 2666 | q->queue_hw_ctx = kcalloc_node(nr_cpu_ids, sizeof(*(q->queue_hw_ctx)), |
868f2f0b KB |
2667 | GFP_KERNEL, set->numa_node); |
2668 | if (!q->queue_hw_ctx) | |
2669 | goto err_percpu; | |
2670 | ||
868f2f0b KB |
2671 | blk_mq_realloc_hw_ctxs(set, q); |
2672 | if (!q->nr_hw_queues) | |
2673 | goto err_hctxs; | |
320ae51f | 2674 | |
287922eb | 2675 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2676 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2677 | |
2678 | q->nr_queues = nr_cpu_ids; | |
a8908939 | 2679 | q->tag_set = set; |
320ae51f | 2680 | |
94eddfbe | 2681 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2682 | |
05f1dd53 | 2683 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
f78bac2c | 2684 | queue_flag_set_unlocked(QUEUE_FLAG_NO_SG_MERGE, q); |
05f1dd53 | 2685 | |
1be036e9 CH |
2686 | q->sg_reserved_size = INT_MAX; |
2687 | ||
2849450a | 2688 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2689 | INIT_LIST_HEAD(&q->requeue_list); |
2690 | spin_lock_init(&q->requeue_lock); | |
2691 | ||
254d259d | 2692 | blk_queue_make_request(q, blk_mq_make_request); |
ea435e1b CH |
2693 | if (q->mq_ops->poll) |
2694 | q->poll_fn = blk_mq_poll; | |
07068d5b | 2695 | |
eba71768 JA |
2696 | /* |
2697 | * Do this after blk_queue_make_request() overrides it... | |
2698 | */ | |
2699 | q->nr_requests = set->queue_depth; | |
2700 | ||
64f1c21e JA |
2701 | /* |
2702 | * Default to classic polling | |
2703 | */ | |
2704 | q->poll_nsec = -1; | |
2705 | ||
24d2f903 | 2706 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 2707 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 2708 | blk_mq_map_swqueue(q); |
4593fdbe | 2709 | |
d3484991 JA |
2710 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { |
2711 | int ret; | |
2712 | ||
131d08e1 | 2713 | ret = elevator_init_mq(q); |
d3484991 JA |
2714 | if (ret) |
2715 | return ERR_PTR(ret); | |
2716 | } | |
2717 | ||
320ae51f | 2718 | return q; |
18741986 | 2719 | |
320ae51f | 2720 | err_hctxs: |
868f2f0b | 2721 | kfree(q->queue_hw_ctx); |
320ae51f | 2722 | err_percpu: |
868f2f0b | 2723 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2724 | err_exit: |
2725 | q->mq_ops = NULL; | |
320ae51f JA |
2726 | return ERR_PTR(-ENOMEM); |
2727 | } | |
b62c21b7 | 2728 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2729 | |
2730 | void blk_mq_free_queue(struct request_queue *q) | |
2731 | { | |
624dbe47 | 2732 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2733 | |
0d2602ca | 2734 | blk_mq_del_queue_tag_set(q); |
624dbe47 | 2735 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
320ae51f | 2736 | } |
320ae51f | 2737 | |
a5164405 JA |
2738 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2739 | { | |
2740 | int i; | |
2741 | ||
cc71a6f4 JA |
2742 | for (i = 0; i < set->nr_hw_queues; i++) |
2743 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 2744 | goto out_unwind; |
a5164405 JA |
2745 | |
2746 | return 0; | |
2747 | ||
2748 | out_unwind: | |
2749 | while (--i >= 0) | |
cc71a6f4 | 2750 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 2751 | |
a5164405 JA |
2752 | return -ENOMEM; |
2753 | } | |
2754 | ||
2755 | /* | |
2756 | * Allocate the request maps associated with this tag_set. Note that this | |
2757 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2758 | * will be updated to reflect the allocated depth. | |
2759 | */ | |
2760 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2761 | { | |
2762 | unsigned int depth; | |
2763 | int err; | |
2764 | ||
2765 | depth = set->queue_depth; | |
2766 | do { | |
2767 | err = __blk_mq_alloc_rq_maps(set); | |
2768 | if (!err) | |
2769 | break; | |
2770 | ||
2771 | set->queue_depth >>= 1; | |
2772 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2773 | err = -ENOMEM; | |
2774 | break; | |
2775 | } | |
2776 | } while (set->queue_depth); | |
2777 | ||
2778 | if (!set->queue_depth || err) { | |
2779 | pr_err("blk-mq: failed to allocate request map\n"); | |
2780 | return -ENOMEM; | |
2781 | } | |
2782 | ||
2783 | if (depth != set->queue_depth) | |
2784 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2785 | depth, set->queue_depth); | |
2786 | ||
2787 | return 0; | |
2788 | } | |
2789 | ||
ebe8bddb OS |
2790 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
2791 | { | |
7d4901a9 | 2792 | if (set->ops->map_queues) { |
7d4901a9 ML |
2793 | /* |
2794 | * transport .map_queues is usually done in the following | |
2795 | * way: | |
2796 | * | |
2797 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
2798 | * mask = get_cpu_mask(queue) | |
2799 | * for_each_cpu(cpu, mask) | |
ed76e329 | 2800 | * set->map.mq_map[cpu] = queue; |
7d4901a9 ML |
2801 | * } |
2802 | * | |
2803 | * When we need to remap, the table has to be cleared for | |
2804 | * killing stale mapping since one CPU may not be mapped | |
2805 | * to any hw queue. | |
2806 | */ | |
ed76e329 | 2807 | blk_mq_clear_mq_map(&set->map[0]); |
7d4901a9 | 2808 | |
ebe8bddb | 2809 | return set->ops->map_queues(set); |
7d4901a9 | 2810 | } else |
ed76e329 | 2811 | return blk_mq_map_queues(&set->map[0]); |
ebe8bddb OS |
2812 | } |
2813 | ||
a4391c64 JA |
2814 | /* |
2815 | * Alloc a tag set to be associated with one or more request queues. | |
2816 | * May fail with EINVAL for various error conditions. May adjust the | |
c018c84f | 2817 | * requested depth down, if it's too large. In that case, the set |
a4391c64 JA |
2818 | * value will be stored in set->queue_depth. |
2819 | */ | |
24d2f903 CH |
2820 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2821 | { | |
da695ba2 CH |
2822 | int ret; |
2823 | ||
205fb5f5 BVA |
2824 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2825 | ||
24d2f903 CH |
2826 | if (!set->nr_hw_queues) |
2827 | return -EINVAL; | |
a4391c64 | 2828 | if (!set->queue_depth) |
24d2f903 CH |
2829 | return -EINVAL; |
2830 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2831 | return -EINVAL; | |
2832 | ||
7d7e0f90 | 2833 | if (!set->ops->queue_rq) |
24d2f903 CH |
2834 | return -EINVAL; |
2835 | ||
de148297 ML |
2836 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
2837 | return -EINVAL; | |
2838 | ||
a4391c64 JA |
2839 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2840 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2841 | BLK_MQ_MAX_DEPTH); | |
2842 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2843 | } | |
24d2f903 | 2844 | |
6637fadf SL |
2845 | /* |
2846 | * If a crashdump is active, then we are potentially in a very | |
2847 | * memory constrained environment. Limit us to 1 queue and | |
2848 | * 64 tags to prevent using too much memory. | |
2849 | */ | |
2850 | if (is_kdump_kernel()) { | |
2851 | set->nr_hw_queues = 1; | |
2852 | set->queue_depth = min(64U, set->queue_depth); | |
2853 | } | |
868f2f0b KB |
2854 | /* |
2855 | * There is no use for more h/w queues than cpus. | |
2856 | */ | |
2857 | if (set->nr_hw_queues > nr_cpu_ids) | |
2858 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2859 | |
590b5b7d | 2860 | set->tags = kcalloc_node(nr_cpu_ids, sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2861 | GFP_KERNEL, set->numa_node); |
2862 | if (!set->tags) | |
a5164405 | 2863 | return -ENOMEM; |
24d2f903 | 2864 | |
da695ba2 | 2865 | ret = -ENOMEM; |
ed76e329 JA |
2866 | set->map[0].mq_map = kcalloc_node(nr_cpu_ids, |
2867 | sizeof(*set->map[0].mq_map), | |
2868 | GFP_KERNEL, set->numa_node); | |
2869 | if (!set->map[0].mq_map) | |
bdd17e75 | 2870 | goto out_free_tags; |
ed76e329 | 2871 | set->map[0].nr_queues = set->nr_hw_queues; |
bdd17e75 | 2872 | |
ebe8bddb | 2873 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
2874 | if (ret) |
2875 | goto out_free_mq_map; | |
2876 | ||
2877 | ret = blk_mq_alloc_rq_maps(set); | |
2878 | if (ret) | |
bdd17e75 | 2879 | goto out_free_mq_map; |
24d2f903 | 2880 | |
0d2602ca JA |
2881 | mutex_init(&set->tag_list_lock); |
2882 | INIT_LIST_HEAD(&set->tag_list); | |
2883 | ||
24d2f903 | 2884 | return 0; |
bdd17e75 CH |
2885 | |
2886 | out_free_mq_map: | |
ed76e329 JA |
2887 | kfree(set->map[0].mq_map); |
2888 | set->map[0].mq_map = NULL; | |
bdd17e75 | 2889 | out_free_tags: |
5676e7b6 RE |
2890 | kfree(set->tags); |
2891 | set->tags = NULL; | |
da695ba2 | 2892 | return ret; |
24d2f903 CH |
2893 | } |
2894 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2895 | ||
2896 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2897 | { | |
2898 | int i; | |
2899 | ||
cc71a6f4 JA |
2900 | for (i = 0; i < nr_cpu_ids; i++) |
2901 | blk_mq_free_map_and_requests(set, i); | |
484b4061 | 2902 | |
ed76e329 JA |
2903 | kfree(set->map[0].mq_map); |
2904 | set->map[0].mq_map = NULL; | |
bdd17e75 | 2905 | |
981bd189 | 2906 | kfree(set->tags); |
5676e7b6 | 2907 | set->tags = NULL; |
24d2f903 CH |
2908 | } |
2909 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2910 | ||
e3a2b3f9 JA |
2911 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2912 | { | |
2913 | struct blk_mq_tag_set *set = q->tag_set; | |
2914 | struct blk_mq_hw_ctx *hctx; | |
2915 | int i, ret; | |
2916 | ||
bd166ef1 | 2917 | if (!set) |
e3a2b3f9 JA |
2918 | return -EINVAL; |
2919 | ||
70f36b60 | 2920 | blk_mq_freeze_queue(q); |
24f5a90f | 2921 | blk_mq_quiesce_queue(q); |
70f36b60 | 2922 | |
e3a2b3f9 JA |
2923 | ret = 0; |
2924 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2925 | if (!hctx->tags) |
2926 | continue; | |
bd166ef1 JA |
2927 | /* |
2928 | * If we're using an MQ scheduler, just update the scheduler | |
2929 | * queue depth. This is similar to what the old code would do. | |
2930 | */ | |
70f36b60 | 2931 | if (!hctx->sched_tags) { |
c2e82a23 | 2932 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, |
70f36b60 JA |
2933 | false); |
2934 | } else { | |
2935 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
2936 | nr, true); | |
2937 | } | |
e3a2b3f9 JA |
2938 | if (ret) |
2939 | break; | |
2940 | } | |
2941 | ||
2942 | if (!ret) | |
2943 | q->nr_requests = nr; | |
2944 | ||
24f5a90f | 2945 | blk_mq_unquiesce_queue(q); |
70f36b60 | 2946 | blk_mq_unfreeze_queue(q); |
70f36b60 | 2947 | |
e3a2b3f9 JA |
2948 | return ret; |
2949 | } | |
2950 | ||
d48ece20 JW |
2951 | /* |
2952 | * request_queue and elevator_type pair. | |
2953 | * It is just used by __blk_mq_update_nr_hw_queues to cache | |
2954 | * the elevator_type associated with a request_queue. | |
2955 | */ | |
2956 | struct blk_mq_qe_pair { | |
2957 | struct list_head node; | |
2958 | struct request_queue *q; | |
2959 | struct elevator_type *type; | |
2960 | }; | |
2961 | ||
2962 | /* | |
2963 | * Cache the elevator_type in qe pair list and switch the | |
2964 | * io scheduler to 'none' | |
2965 | */ | |
2966 | static bool blk_mq_elv_switch_none(struct list_head *head, | |
2967 | struct request_queue *q) | |
2968 | { | |
2969 | struct blk_mq_qe_pair *qe; | |
2970 | ||
2971 | if (!q->elevator) | |
2972 | return true; | |
2973 | ||
2974 | qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); | |
2975 | if (!qe) | |
2976 | return false; | |
2977 | ||
2978 | INIT_LIST_HEAD(&qe->node); | |
2979 | qe->q = q; | |
2980 | qe->type = q->elevator->type; | |
2981 | list_add(&qe->node, head); | |
2982 | ||
2983 | mutex_lock(&q->sysfs_lock); | |
2984 | /* | |
2985 | * After elevator_switch_mq, the previous elevator_queue will be | |
2986 | * released by elevator_release. The reference of the io scheduler | |
2987 | * module get by elevator_get will also be put. So we need to get | |
2988 | * a reference of the io scheduler module here to prevent it to be | |
2989 | * removed. | |
2990 | */ | |
2991 | __module_get(qe->type->elevator_owner); | |
2992 | elevator_switch_mq(q, NULL); | |
2993 | mutex_unlock(&q->sysfs_lock); | |
2994 | ||
2995 | return true; | |
2996 | } | |
2997 | ||
2998 | static void blk_mq_elv_switch_back(struct list_head *head, | |
2999 | struct request_queue *q) | |
3000 | { | |
3001 | struct blk_mq_qe_pair *qe; | |
3002 | struct elevator_type *t = NULL; | |
3003 | ||
3004 | list_for_each_entry(qe, head, node) | |
3005 | if (qe->q == q) { | |
3006 | t = qe->type; | |
3007 | break; | |
3008 | } | |
3009 | ||
3010 | if (!t) | |
3011 | return; | |
3012 | ||
3013 | list_del(&qe->node); | |
3014 | kfree(qe); | |
3015 | ||
3016 | mutex_lock(&q->sysfs_lock); | |
3017 | elevator_switch_mq(q, t); | |
3018 | mutex_unlock(&q->sysfs_lock); | |
3019 | } | |
3020 | ||
e4dc2b32 KB |
3021 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
3022 | int nr_hw_queues) | |
868f2f0b KB |
3023 | { |
3024 | struct request_queue *q; | |
d48ece20 | 3025 | LIST_HEAD(head); |
e01ad46d | 3026 | int prev_nr_hw_queues; |
868f2f0b | 3027 | |
705cda97 BVA |
3028 | lockdep_assert_held(&set->tag_list_lock); |
3029 | ||
868f2f0b KB |
3030 | if (nr_hw_queues > nr_cpu_ids) |
3031 | nr_hw_queues = nr_cpu_ids; | |
3032 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
3033 | return; | |
3034 | ||
3035 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3036 | blk_mq_freeze_queue(q); | |
f5bbbbe4 JW |
3037 | /* |
3038 | * Sync with blk_mq_queue_tag_busy_iter. | |
3039 | */ | |
3040 | synchronize_rcu(); | |
d48ece20 JW |
3041 | /* |
3042 | * Switch IO scheduler to 'none', cleaning up the data associated | |
3043 | * with the previous scheduler. We will switch back once we are done | |
3044 | * updating the new sw to hw queue mappings. | |
3045 | */ | |
3046 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3047 | if (!blk_mq_elv_switch_none(&head, q)) | |
3048 | goto switch_back; | |
868f2f0b | 3049 | |
477e19de JW |
3050 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3051 | blk_mq_debugfs_unregister_hctxs(q); | |
3052 | blk_mq_sysfs_unregister(q); | |
3053 | } | |
3054 | ||
e01ad46d | 3055 | prev_nr_hw_queues = set->nr_hw_queues; |
868f2f0b | 3056 | set->nr_hw_queues = nr_hw_queues; |
ebe8bddb | 3057 | blk_mq_update_queue_map(set); |
e01ad46d | 3058 | fallback: |
868f2f0b KB |
3059 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3060 | blk_mq_realloc_hw_ctxs(set, q); | |
e01ad46d JW |
3061 | if (q->nr_hw_queues != set->nr_hw_queues) { |
3062 | pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n", | |
3063 | nr_hw_queues, prev_nr_hw_queues); | |
3064 | set->nr_hw_queues = prev_nr_hw_queues; | |
ed76e329 | 3065 | blk_mq_map_queues(&set->map[0]); |
e01ad46d JW |
3066 | goto fallback; |
3067 | } | |
477e19de JW |
3068 | blk_mq_map_swqueue(q); |
3069 | } | |
3070 | ||
3071 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
3072 | blk_mq_sysfs_register(q); | |
3073 | blk_mq_debugfs_register_hctxs(q); | |
868f2f0b KB |
3074 | } |
3075 | ||
d48ece20 JW |
3076 | switch_back: |
3077 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3078 | blk_mq_elv_switch_back(&head, q); | |
3079 | ||
868f2f0b KB |
3080 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
3081 | blk_mq_unfreeze_queue(q); | |
3082 | } | |
e4dc2b32 KB |
3083 | |
3084 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
3085 | { | |
3086 | mutex_lock(&set->tag_list_lock); | |
3087 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
3088 | mutex_unlock(&set->tag_list_lock); | |
3089 | } | |
868f2f0b KB |
3090 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
3091 | ||
34dbad5d OS |
3092 | /* Enable polling stats and return whether they were already enabled. */ |
3093 | static bool blk_poll_stats_enable(struct request_queue *q) | |
3094 | { | |
3095 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
7dfdbc73 | 3096 | blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q)) |
34dbad5d OS |
3097 | return true; |
3098 | blk_stat_add_callback(q, q->poll_cb); | |
3099 | return false; | |
3100 | } | |
3101 | ||
3102 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
3103 | { | |
3104 | /* | |
3105 | * We don't arm the callback if polling stats are not enabled or the | |
3106 | * callback is already active. | |
3107 | */ | |
3108 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
3109 | blk_stat_is_active(q->poll_cb)) | |
3110 | return; | |
3111 | ||
3112 | blk_stat_activate_msecs(q->poll_cb, 100); | |
3113 | } | |
3114 | ||
3115 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
3116 | { | |
3117 | struct request_queue *q = cb->data; | |
720b8ccc | 3118 | int bucket; |
34dbad5d | 3119 | |
720b8ccc SB |
3120 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
3121 | if (cb->stat[bucket].nr_samples) | |
3122 | q->poll_stat[bucket] = cb->stat[bucket]; | |
3123 | } | |
34dbad5d OS |
3124 | } |
3125 | ||
64f1c21e JA |
3126 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
3127 | struct blk_mq_hw_ctx *hctx, | |
3128 | struct request *rq) | |
3129 | { | |
64f1c21e | 3130 | unsigned long ret = 0; |
720b8ccc | 3131 | int bucket; |
64f1c21e JA |
3132 | |
3133 | /* | |
3134 | * If stats collection isn't on, don't sleep but turn it on for | |
3135 | * future users | |
3136 | */ | |
34dbad5d | 3137 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
3138 | return 0; |
3139 | ||
64f1c21e JA |
3140 | /* |
3141 | * As an optimistic guess, use half of the mean service time | |
3142 | * for this type of request. We can (and should) make this smarter. | |
3143 | * For instance, if the completion latencies are tight, we can | |
3144 | * get closer than just half the mean. This is especially | |
3145 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
3146 | * than ~10 usec. We do use the stats for the relevant IO size |
3147 | * if available which does lead to better estimates. | |
64f1c21e | 3148 | */ |
720b8ccc SB |
3149 | bucket = blk_mq_poll_stats_bkt(rq); |
3150 | if (bucket < 0) | |
3151 | return ret; | |
3152 | ||
3153 | if (q->poll_stat[bucket].nr_samples) | |
3154 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
3155 | |
3156 | return ret; | |
3157 | } | |
3158 | ||
06426adf | 3159 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 3160 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
3161 | struct request *rq) |
3162 | { | |
3163 | struct hrtimer_sleeper hs; | |
3164 | enum hrtimer_mode mode; | |
64f1c21e | 3165 | unsigned int nsecs; |
06426adf JA |
3166 | ktime_t kt; |
3167 | ||
76a86f9d | 3168 | if (rq->rq_flags & RQF_MQ_POLL_SLEPT) |
64f1c21e JA |
3169 | return false; |
3170 | ||
3171 | /* | |
3172 | * poll_nsec can be: | |
3173 | * | |
3174 | * -1: don't ever hybrid sleep | |
3175 | * 0: use half of prev avg | |
3176 | * >0: use this specific value | |
3177 | */ | |
3178 | if (q->poll_nsec == -1) | |
3179 | return false; | |
3180 | else if (q->poll_nsec > 0) | |
3181 | nsecs = q->poll_nsec; | |
3182 | else | |
3183 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
3184 | ||
3185 | if (!nsecs) | |
06426adf JA |
3186 | return false; |
3187 | ||
76a86f9d | 3188 | rq->rq_flags |= RQF_MQ_POLL_SLEPT; |
06426adf JA |
3189 | |
3190 | /* | |
3191 | * This will be replaced with the stats tracking code, using | |
3192 | * 'avg_completion_time / 2' as the pre-sleep target. | |
3193 | */ | |
8b0e1953 | 3194 | kt = nsecs; |
06426adf JA |
3195 | |
3196 | mode = HRTIMER_MODE_REL; | |
3197 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
3198 | hrtimer_set_expires(&hs.timer, kt); | |
3199 | ||
3200 | hrtimer_init_sleeper(&hs, current); | |
3201 | do { | |
5a61c363 | 3202 | if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE) |
06426adf JA |
3203 | break; |
3204 | set_current_state(TASK_UNINTERRUPTIBLE); | |
3205 | hrtimer_start_expires(&hs.timer, mode); | |
3206 | if (hs.task) | |
3207 | io_schedule(); | |
3208 | hrtimer_cancel(&hs.timer); | |
3209 | mode = HRTIMER_MODE_ABS; | |
3210 | } while (hs.task && !signal_pending(current)); | |
3211 | ||
3212 | __set_current_state(TASK_RUNNING); | |
3213 | destroy_hrtimer_on_stack(&hs.timer); | |
3214 | return true; | |
3215 | } | |
3216 | ||
bbd7bb70 JA |
3217 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) |
3218 | { | |
3219 | struct request_queue *q = hctx->queue; | |
3220 | long state; | |
3221 | ||
06426adf JA |
3222 | /* |
3223 | * If we sleep, have the caller restart the poll loop to reset | |
3224 | * the state. Like for the other success return cases, the | |
3225 | * caller is responsible for checking if the IO completed. If | |
3226 | * the IO isn't complete, we'll get called again and will go | |
3227 | * straight to the busy poll loop. | |
3228 | */ | |
64f1c21e | 3229 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) |
06426adf JA |
3230 | return true; |
3231 | ||
bbd7bb70 JA |
3232 | hctx->poll_considered++; |
3233 | ||
3234 | state = current->state; | |
3235 | while (!need_resched()) { | |
3236 | int ret; | |
3237 | ||
3238 | hctx->poll_invoked++; | |
3239 | ||
3240 | ret = q->mq_ops->poll(hctx, rq->tag); | |
3241 | if (ret > 0) { | |
3242 | hctx->poll_success++; | |
3243 | set_current_state(TASK_RUNNING); | |
3244 | return true; | |
3245 | } | |
3246 | ||
3247 | if (signal_pending_state(state, current)) | |
3248 | set_current_state(TASK_RUNNING); | |
3249 | ||
3250 | if (current->state == TASK_RUNNING) | |
3251 | return true; | |
3252 | if (ret < 0) | |
3253 | break; | |
3254 | cpu_relax(); | |
3255 | } | |
3256 | ||
67b4110f | 3257 | __set_current_state(TASK_RUNNING); |
bbd7bb70 JA |
3258 | return false; |
3259 | } | |
3260 | ||
ea435e1b | 3261 | static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) |
bbd7bb70 JA |
3262 | { |
3263 | struct blk_mq_hw_ctx *hctx; | |
bbd7bb70 JA |
3264 | struct request *rq; |
3265 | ||
ea435e1b | 3266 | if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) |
bbd7bb70 JA |
3267 | return false; |
3268 | ||
bbd7bb70 | 3269 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; |
bd166ef1 JA |
3270 | if (!blk_qc_t_is_internal(cookie)) |
3271 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
3a07bb1d | 3272 | else { |
bd166ef1 | 3273 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); |
3a07bb1d JA |
3274 | /* |
3275 | * With scheduling, if the request has completed, we'll | |
3276 | * get a NULL return here, as we clear the sched tag when | |
3277 | * that happens. The request still remains valid, like always, | |
3278 | * so we should be safe with just the NULL check. | |
3279 | */ | |
3280 | if (!rq) | |
3281 | return false; | |
3282 | } | |
bbd7bb70 JA |
3283 | |
3284 | return __blk_mq_poll(hctx, rq); | |
3285 | } | |
bbd7bb70 | 3286 | |
9cf2bab6 JA |
3287 | unsigned int blk_mq_rq_cpu(struct request *rq) |
3288 | { | |
3289 | return rq->mq_ctx->cpu; | |
3290 | } | |
3291 | EXPORT_SYMBOL(blk_mq_rq_cpu); | |
3292 | ||
320ae51f JA |
3293 | static int __init blk_mq_init(void) |
3294 | { | |
9467f859 TG |
3295 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
3296 | blk_mq_hctx_notify_dead); | |
320ae51f JA |
3297 | return 0; |
3298 | } | |
3299 | subsys_initcall(blk_mq_init); |