Commit | Line | Data |
---|---|---|
3dcf60bc | 1 | // SPDX-License-Identifier: GPL-2.0 |
75bb4625 JA |
2 | /* |
3 | * Block multiqueue core code | |
4 | * | |
5 | * Copyright (C) 2013-2014 Jens Axboe | |
6 | * Copyright (C) 2013-2014 Christoph Hellwig | |
7 | */ | |
320ae51f JA |
8 | #include <linux/kernel.h> |
9 | #include <linux/module.h> | |
10 | #include <linux/backing-dev.h> | |
11 | #include <linux/bio.h> | |
12 | #include <linux/blkdev.h> | |
f75782e4 | 13 | #include <linux/kmemleak.h> |
320ae51f JA |
14 | #include <linux/mm.h> |
15 | #include <linux/init.h> | |
16 | #include <linux/slab.h> | |
17 | #include <linux/workqueue.h> | |
18 | #include <linux/smp.h> | |
19 | #include <linux/llist.h> | |
20 | #include <linux/list_sort.h> | |
21 | #include <linux/cpu.h> | |
22 | #include <linux/cache.h> | |
23 | #include <linux/sched/sysctl.h> | |
105ab3d8 | 24 | #include <linux/sched/topology.h> |
174cd4b1 | 25 | #include <linux/sched/signal.h> |
320ae51f | 26 | #include <linux/delay.h> |
aedcd72f | 27 | #include <linux/crash_dump.h> |
88c7b2b7 | 28 | #include <linux/prefetch.h> |
320ae51f JA |
29 | |
30 | #include <trace/events/block.h> | |
31 | ||
32 | #include <linux/blk-mq.h> | |
54d4e6ab | 33 | #include <linux/t10-pi.h> |
320ae51f JA |
34 | #include "blk.h" |
35 | #include "blk-mq.h" | |
9c1051aa | 36 | #include "blk-mq-debugfs.h" |
320ae51f | 37 | #include "blk-mq-tag.h" |
986d413b | 38 | #include "blk-pm.h" |
cf43e6be | 39 | #include "blk-stat.h" |
bd166ef1 | 40 | #include "blk-mq-sched.h" |
c1c80384 | 41 | #include "blk-rq-qos.h" |
320ae51f | 42 | |
34dbad5d OS |
43 | static void blk_mq_poll_stats_start(struct request_queue *q); |
44 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb); | |
45 | ||
720b8ccc SB |
46 | static int blk_mq_poll_stats_bkt(const struct request *rq) |
47 | { | |
3d244306 | 48 | int ddir, sectors, bucket; |
720b8ccc | 49 | |
99c749a4 | 50 | ddir = rq_data_dir(rq); |
3d244306 | 51 | sectors = blk_rq_stats_sectors(rq); |
720b8ccc | 52 | |
3d244306 | 53 | bucket = ddir + 2 * ilog2(sectors); |
720b8ccc SB |
54 | |
55 | if (bucket < 0) | |
56 | return -1; | |
57 | else if (bucket >= BLK_MQ_POLL_STATS_BKTS) | |
58 | return ddir + BLK_MQ_POLL_STATS_BKTS - 2; | |
59 | ||
60 | return bucket; | |
61 | } | |
62 | ||
320ae51f | 63 | /* |
85fae294 YY |
64 | * Check if any of the ctx, dispatch list or elevator |
65 | * have pending work in this hardware queue. | |
320ae51f | 66 | */ |
79f720a7 | 67 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 68 | { |
79f720a7 JA |
69 | return !list_empty_careful(&hctx->dispatch) || |
70 | sbitmap_any_bit_set(&hctx->ctx_map) || | |
bd166ef1 | 71 | blk_mq_sched_has_work(hctx); |
1429d7c9 JA |
72 | } |
73 | ||
320ae51f JA |
74 | /* |
75 | * Mark this ctx as having pending work in this hardware queue | |
76 | */ | |
77 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
78 | struct blk_mq_ctx *ctx) | |
79 | { | |
f31967f0 JA |
80 | const int bit = ctx->index_hw[hctx->type]; |
81 | ||
82 | if (!sbitmap_test_bit(&hctx->ctx_map, bit)) | |
83 | sbitmap_set_bit(&hctx->ctx_map, bit); | |
1429d7c9 JA |
84 | } |
85 | ||
86 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
87 | struct blk_mq_ctx *ctx) | |
88 | { | |
f31967f0 JA |
89 | const int bit = ctx->index_hw[hctx->type]; |
90 | ||
91 | sbitmap_clear_bit(&hctx->ctx_map, bit); | |
320ae51f JA |
92 | } |
93 | ||
f299b7c7 JA |
94 | struct mq_inflight { |
95 | struct hd_struct *part; | |
a2e80f6f | 96 | unsigned int inflight[2]; |
f299b7c7 JA |
97 | }; |
98 | ||
7baa8572 | 99 | static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx, |
f299b7c7 JA |
100 | struct request *rq, void *priv, |
101 | bool reserved) | |
102 | { | |
103 | struct mq_inflight *mi = priv; | |
104 | ||
6131837b | 105 | if (rq->part == mi->part) |
bb4e6b14 | 106 | mi->inflight[rq_data_dir(rq)]++; |
7baa8572 JA |
107 | |
108 | return true; | |
f299b7c7 JA |
109 | } |
110 | ||
e016b782 | 111 | unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part) |
f299b7c7 | 112 | { |
a2e80f6f | 113 | struct mq_inflight mi = { .part = part }; |
f299b7c7 | 114 | |
f299b7c7 | 115 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
e016b782 | 116 | |
a2e80f6f | 117 | return mi.inflight[0] + mi.inflight[1]; |
bf0ddaba OS |
118 | } |
119 | ||
120 | void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part, | |
121 | unsigned int inflight[2]) | |
122 | { | |
a2e80f6f | 123 | struct mq_inflight mi = { .part = part }; |
bf0ddaba | 124 | |
bb4e6b14 | 125 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
a2e80f6f PB |
126 | inflight[0] = mi.inflight[0]; |
127 | inflight[1] = mi.inflight[1]; | |
bf0ddaba OS |
128 | } |
129 | ||
1671d522 | 130 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 131 | { |
7996a8b5 BL |
132 | mutex_lock(&q->mq_freeze_lock); |
133 | if (++q->mq_freeze_depth == 1) { | |
3ef28e83 | 134 | percpu_ref_kill(&q->q_usage_counter); |
7996a8b5 | 135 | mutex_unlock(&q->mq_freeze_lock); |
344e9ffc | 136 | if (queue_is_mq(q)) |
055f6e18 | 137 | blk_mq_run_hw_queues(q, false); |
7996a8b5 BL |
138 | } else { |
139 | mutex_unlock(&q->mq_freeze_lock); | |
cddd5d17 | 140 | } |
f3af020b | 141 | } |
1671d522 | 142 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
f3af020b | 143 | |
6bae363e | 144 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 145 | { |
3ef28e83 | 146 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 147 | } |
6bae363e | 148 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 149 | |
f91328c4 KB |
150 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
151 | unsigned long timeout) | |
152 | { | |
153 | return wait_event_timeout(q->mq_freeze_wq, | |
154 | percpu_ref_is_zero(&q->q_usage_counter), | |
155 | timeout); | |
156 | } | |
157 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 158 | |
f3af020b TH |
159 | /* |
160 | * Guarantee no request is in use, so we can change any data structure of | |
161 | * the queue afterward. | |
162 | */ | |
3ef28e83 | 163 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 164 | { |
3ef28e83 DW |
165 | /* |
166 | * In the !blk_mq case we are only calling this to kill the | |
167 | * q_usage_counter, otherwise this increases the freeze depth | |
168 | * and waits for it to return to zero. For this reason there is | |
169 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
170 | * exported to drivers as the only user for unfreeze is blk_mq. | |
171 | */ | |
1671d522 | 172 | blk_freeze_queue_start(q); |
f3af020b TH |
173 | blk_mq_freeze_queue_wait(q); |
174 | } | |
3ef28e83 DW |
175 | |
176 | void blk_mq_freeze_queue(struct request_queue *q) | |
177 | { | |
178 | /* | |
179 | * ...just an alias to keep freeze and unfreeze actions balanced | |
180 | * in the blk_mq_* namespace | |
181 | */ | |
182 | blk_freeze_queue(q); | |
183 | } | |
c761d96b | 184 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 185 | |
b4c6a028 | 186 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 187 | { |
7996a8b5 BL |
188 | mutex_lock(&q->mq_freeze_lock); |
189 | q->mq_freeze_depth--; | |
190 | WARN_ON_ONCE(q->mq_freeze_depth < 0); | |
191 | if (!q->mq_freeze_depth) { | |
bdd63160 | 192 | percpu_ref_resurrect(&q->q_usage_counter); |
320ae51f | 193 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 194 | } |
7996a8b5 | 195 | mutex_unlock(&q->mq_freeze_lock); |
320ae51f | 196 | } |
b4c6a028 | 197 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 198 | |
852ec809 BVA |
199 | /* |
200 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
201 | * mpt3sas driver such that this function can be removed. | |
202 | */ | |
203 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
204 | { | |
8814ce8a | 205 | blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q); |
852ec809 BVA |
206 | } |
207 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
208 | ||
6a83e74d | 209 | /** |
69e07c4a | 210 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished |
6a83e74d BVA |
211 | * @q: request queue. |
212 | * | |
213 | * Note: this function does not prevent that the struct request end_io() | |
69e07c4a ML |
214 | * callback function is invoked. Once this function is returned, we make |
215 | * sure no dispatch can happen until the queue is unquiesced via | |
216 | * blk_mq_unquiesce_queue(). | |
6a83e74d BVA |
217 | */ |
218 | void blk_mq_quiesce_queue(struct request_queue *q) | |
219 | { | |
220 | struct blk_mq_hw_ctx *hctx; | |
221 | unsigned int i; | |
222 | bool rcu = false; | |
223 | ||
1d9e9bc6 | 224 | blk_mq_quiesce_queue_nowait(q); |
f4560ffe | 225 | |
6a83e74d BVA |
226 | queue_for_each_hw_ctx(q, hctx, i) { |
227 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
05707b64 | 228 | synchronize_srcu(hctx->srcu); |
6a83e74d BVA |
229 | else |
230 | rcu = true; | |
231 | } | |
232 | if (rcu) | |
233 | synchronize_rcu(); | |
234 | } | |
235 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
236 | ||
e4e73913 ML |
237 | /* |
238 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
239 | * @q: request queue. | |
240 | * | |
241 | * This function recovers queue into the state before quiescing | |
242 | * which is done by blk_mq_quiesce_queue. | |
243 | */ | |
244 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
245 | { | |
8814ce8a | 246 | blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q); |
f4560ffe | 247 | |
1d9e9bc6 ML |
248 | /* dispatch requests which are inserted during quiescing */ |
249 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
250 | } |
251 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
252 | ||
aed3ea94 JA |
253 | void blk_mq_wake_waiters(struct request_queue *q) |
254 | { | |
255 | struct blk_mq_hw_ctx *hctx; | |
256 | unsigned int i; | |
257 | ||
258 | queue_for_each_hw_ctx(q, hctx, i) | |
259 | if (blk_mq_hw_queue_mapped(hctx)) | |
260 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
261 | } | |
262 | ||
fe1f4526 | 263 | /* |
9a91b05b HT |
264 | * Only need start/end time stamping if we have iostat or |
265 | * blk stats enabled, or using an IO scheduler. | |
fe1f4526 JA |
266 | */ |
267 | static inline bool blk_mq_need_time_stamp(struct request *rq) | |
268 | { | |
9a91b05b | 269 | return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator; |
fe1f4526 JA |
270 | } |
271 | ||
e4cdf1a1 | 272 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
6f816b4b | 273 | unsigned int tag, unsigned int op, u64 alloc_time_ns) |
320ae51f | 274 | { |
e4cdf1a1 CH |
275 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
276 | struct request *rq = tags->static_rqs[tag]; | |
bf9ae8c5 | 277 | req_flags_t rq_flags = 0; |
c3a148d2 | 278 | |
e4cdf1a1 CH |
279 | if (data->flags & BLK_MQ_REQ_INTERNAL) { |
280 | rq->tag = -1; | |
281 | rq->internal_tag = tag; | |
282 | } else { | |
d263ed99 | 283 | if (data->hctx->flags & BLK_MQ_F_TAG_SHARED) { |
bf9ae8c5 | 284 | rq_flags = RQF_MQ_INFLIGHT; |
e4cdf1a1 CH |
285 | atomic_inc(&data->hctx->nr_active); |
286 | } | |
287 | rq->tag = tag; | |
288 | rq->internal_tag = -1; | |
289 | data->hctx->tags->rqs[rq->tag] = rq; | |
290 | } | |
291 | ||
af76e555 | 292 | /* csd/requeue_work/fifo_time is initialized before use */ |
e4cdf1a1 CH |
293 | rq->q = data->q; |
294 | rq->mq_ctx = data->ctx; | |
ea4f995e | 295 | rq->mq_hctx = data->hctx; |
bf9ae8c5 | 296 | rq->rq_flags = rq_flags; |
ef295ecf | 297 | rq->cmd_flags = op; |
1b6d65a0 BVA |
298 | if (data->flags & BLK_MQ_REQ_PREEMPT) |
299 | rq->rq_flags |= RQF_PREEMPT; | |
e4cdf1a1 | 300 | if (blk_queue_io_stat(data->q)) |
e8064021 | 301 | rq->rq_flags |= RQF_IO_STAT; |
7c3fb70f | 302 | INIT_LIST_HEAD(&rq->queuelist); |
af76e555 CH |
303 | INIT_HLIST_NODE(&rq->hash); |
304 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
305 | rq->rq_disk = NULL; |
306 | rq->part = NULL; | |
6f816b4b TH |
307 | #ifdef CONFIG_BLK_RQ_ALLOC_TIME |
308 | rq->alloc_time_ns = alloc_time_ns; | |
309 | #endif | |
fe1f4526 JA |
310 | if (blk_mq_need_time_stamp(rq)) |
311 | rq->start_time_ns = ktime_get_ns(); | |
312 | else | |
313 | rq->start_time_ns = 0; | |
544ccc8d | 314 | rq->io_start_time_ns = 0; |
3d244306 | 315 | rq->stats_sectors = 0; |
af76e555 CH |
316 | rq->nr_phys_segments = 0; |
317 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
318 | rq->nr_integrity_segments = 0; | |
319 | #endif | |
af76e555 | 320 | /* tag was already set */ |
079076b3 | 321 | WRITE_ONCE(rq->deadline, 0); |
af76e555 | 322 | |
f6be4fb4 JA |
323 | rq->timeout = 0; |
324 | ||
af76e555 CH |
325 | rq->end_io = NULL; |
326 | rq->end_io_data = NULL; | |
af76e555 | 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; |
c05f4220 | 340 | bool clear_ctx_on_error = false; |
6f816b4b | 341 | u64 alloc_time_ns = 0; |
d2c0d383 CH |
342 | |
343 | blk_queue_enter_live(q); | |
6f816b4b TH |
344 | |
345 | /* alloc_time includes depth and tag waits */ | |
346 | if (blk_queue_rq_alloc_time(q)) | |
347 | alloc_time_ns = ktime_get_ns(); | |
348 | ||
d2c0d383 | 349 | data->q = q; |
21e768b4 BVA |
350 | if (likely(!data->ctx)) { |
351 | data->ctx = blk_mq_get_ctx(q); | |
c05f4220 | 352 | clear_ctx_on_error = true; |
21e768b4 | 353 | } |
d2c0d383 | 354 | if (likely(!data->hctx)) |
f9afca4d | 355 | data->hctx = blk_mq_map_queue(q, data->cmd_flags, |
8ccdf4a3 | 356 | data->ctx); |
f9afca4d | 357 | if (data->cmd_flags & REQ_NOWAIT) |
03a07c92 | 358 | data->flags |= BLK_MQ_REQ_NOWAIT; |
d2c0d383 CH |
359 | |
360 | if (e) { | |
361 | data->flags |= BLK_MQ_REQ_INTERNAL; | |
362 | ||
363 | /* | |
364 | * Flush requests are special and go directly to the | |
17a51199 JA |
365 | * dispatch list. Don't include reserved tags in the |
366 | * limiting, as it isn't useful. | |
d2c0d383 | 367 | */ |
f9afca4d JA |
368 | if (!op_is_flush(data->cmd_flags) && |
369 | e->type->ops.limit_depth && | |
17a51199 | 370 | !(data->flags & BLK_MQ_REQ_RESERVED)) |
f9afca4d | 371 | e->type->ops.limit_depth(data->cmd_flags, data); |
d263ed99 JW |
372 | } else { |
373 | blk_mq_tag_busy(data->hctx); | |
d2c0d383 CH |
374 | } |
375 | ||
e4cdf1a1 CH |
376 | tag = blk_mq_get_tag(data); |
377 | if (tag == BLK_MQ_TAG_FAIL) { | |
c05f4220 | 378 | if (clear_ctx_on_error) |
1ad43c00 | 379 | data->ctx = NULL; |
037cebb8 CH |
380 | blk_queue_exit(q); |
381 | return NULL; | |
d2c0d383 CH |
382 | } |
383 | ||
6f816b4b | 384 | rq = blk_mq_rq_ctx_init(data, tag, data->cmd_flags, alloc_time_ns); |
f9afca4d | 385 | if (!op_is_flush(data->cmd_flags)) { |
037cebb8 | 386 | rq->elv.icq = NULL; |
f9cd4bfe | 387 | if (e && e->type->ops.prepare_request) { |
e2b3fa5a DLM |
388 | if (e->type->icq_cache) |
389 | blk_mq_sched_assign_ioc(rq); | |
44e8c2bf | 390 | |
f9cd4bfe | 391 | e->type->ops.prepare_request(rq, bio); |
5bbf4e5a | 392 | rq->rq_flags |= RQF_ELVPRIV; |
44e8c2bf | 393 | } |
037cebb8 CH |
394 | } |
395 | data->hctx->queued++; | |
396 | return rq; | |
d2c0d383 CH |
397 | } |
398 | ||
cd6ce148 | 399 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, |
9a95e4ef | 400 | blk_mq_req_flags_t flags) |
320ae51f | 401 | { |
f9afca4d | 402 | struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op }; |
bd166ef1 | 403 | struct request *rq; |
a492f075 | 404 | int ret; |
320ae51f | 405 | |
3a0a5299 | 406 | ret = blk_queue_enter(q, flags); |
a492f075 JL |
407 | if (ret) |
408 | return ERR_PTR(ret); | |
320ae51f | 409 | |
f9afca4d | 410 | rq = blk_mq_get_request(q, NULL, &alloc_data); |
3280d66a | 411 | blk_queue_exit(q); |
841bac2c | 412 | |
bd166ef1 | 413 | if (!rq) |
a492f075 | 414 | return ERR_PTR(-EWOULDBLOCK); |
0c4de0f3 CH |
415 | |
416 | rq->__data_len = 0; | |
417 | rq->__sector = (sector_t) -1; | |
418 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
419 | return rq; |
420 | } | |
4bb659b1 | 421 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 422 | |
cd6ce148 | 423 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
9a95e4ef | 424 | unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 425 | { |
f9afca4d | 426 | struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op }; |
1f5bd336 | 427 | struct request *rq; |
6d2809d5 | 428 | unsigned int cpu; |
1f5bd336 ML |
429 | int ret; |
430 | ||
431 | /* | |
432 | * If the tag allocator sleeps we could get an allocation for a | |
433 | * different hardware context. No need to complicate the low level | |
434 | * allocator for this for the rare use case of a command tied to | |
435 | * a specific queue. | |
436 | */ | |
437 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
438 | return ERR_PTR(-EINVAL); | |
439 | ||
440 | if (hctx_idx >= q->nr_hw_queues) | |
441 | return ERR_PTR(-EIO); | |
442 | ||
3a0a5299 | 443 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
444 | if (ret) |
445 | return ERR_PTR(ret); | |
446 | ||
c8712c6a CH |
447 | /* |
448 | * Check if the hardware context is actually mapped to anything. | |
449 | * If not tell the caller that it should skip this queue. | |
450 | */ | |
6d2809d5 OS |
451 | alloc_data.hctx = q->queue_hw_ctx[hctx_idx]; |
452 | if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) { | |
453 | blk_queue_exit(q); | |
454 | return ERR_PTR(-EXDEV); | |
c8712c6a | 455 | } |
20e4d813 | 456 | cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask); |
6d2809d5 | 457 | alloc_data.ctx = __blk_mq_get_ctx(q, cpu); |
1f5bd336 | 458 | |
f9afca4d | 459 | rq = blk_mq_get_request(q, NULL, &alloc_data); |
3280d66a | 460 | blk_queue_exit(q); |
c8712c6a | 461 | |
6d2809d5 OS |
462 | if (!rq) |
463 | return ERR_PTR(-EWOULDBLOCK); | |
464 | ||
465 | return rq; | |
1f5bd336 ML |
466 | } |
467 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
468 | ||
12f5b931 KB |
469 | static void __blk_mq_free_request(struct request *rq) |
470 | { | |
471 | struct request_queue *q = rq->q; | |
472 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 473 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
12f5b931 KB |
474 | const int sched_tag = rq->internal_tag; |
475 | ||
986d413b | 476 | blk_pm_mark_last_busy(rq); |
ea4f995e | 477 | rq->mq_hctx = NULL; |
12f5b931 | 478 | if (rq->tag != -1) |
cae740a0 | 479 | blk_mq_put_tag(hctx->tags, ctx, rq->tag); |
12f5b931 | 480 | if (sched_tag != -1) |
cae740a0 | 481 | blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag); |
12f5b931 KB |
482 | blk_mq_sched_restart(hctx); |
483 | blk_queue_exit(q); | |
484 | } | |
485 | ||
6af54051 | 486 | void blk_mq_free_request(struct request *rq) |
320ae51f | 487 | { |
320ae51f | 488 | struct request_queue *q = rq->q; |
6af54051 CH |
489 | struct elevator_queue *e = q->elevator; |
490 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 491 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
6af54051 | 492 | |
5bbf4e5a | 493 | if (rq->rq_flags & RQF_ELVPRIV) { |
f9cd4bfe JA |
494 | if (e && e->type->ops.finish_request) |
495 | e->type->ops.finish_request(rq); | |
6af54051 CH |
496 | if (rq->elv.icq) { |
497 | put_io_context(rq->elv.icq->ioc); | |
498 | rq->elv.icq = NULL; | |
499 | } | |
500 | } | |
320ae51f | 501 | |
6af54051 | 502 | ctx->rq_completed[rq_is_sync(rq)]++; |
e8064021 | 503 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 504 | atomic_dec(&hctx->nr_active); |
87760e5e | 505 | |
7beb2f84 JA |
506 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
507 | laptop_io_completion(q->backing_dev_info); | |
508 | ||
a7905043 | 509 | rq_qos_done(q, rq); |
0d2602ca | 510 | |
12f5b931 KB |
511 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
512 | if (refcount_dec_and_test(&rq->ref)) | |
513 | __blk_mq_free_request(rq); | |
320ae51f | 514 | } |
1a3b595a | 515 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 516 | |
2a842aca | 517 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
320ae51f | 518 | { |
fe1f4526 JA |
519 | u64 now = 0; |
520 | ||
521 | if (blk_mq_need_time_stamp(rq)) | |
522 | now = ktime_get_ns(); | |
522a7775 | 523 | |
4bc6339a OS |
524 | if (rq->rq_flags & RQF_STATS) { |
525 | blk_mq_poll_stats_start(rq->q); | |
522a7775 | 526 | blk_stat_add(rq, now); |
4bc6339a OS |
527 | } |
528 | ||
ed88660a OS |
529 | if (rq->internal_tag != -1) |
530 | blk_mq_sched_completed_request(rq, now); | |
531 | ||
522a7775 | 532 | blk_account_io_done(rq, now); |
0d11e6ac | 533 | |
91b63639 | 534 | if (rq->end_io) { |
a7905043 | 535 | rq_qos_done(rq->q, rq); |
320ae51f | 536 | rq->end_io(rq, error); |
91b63639 | 537 | } else { |
320ae51f | 538 | blk_mq_free_request(rq); |
91b63639 | 539 | } |
320ae51f | 540 | } |
c8a446ad | 541 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 542 | |
2a842aca | 543 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
544 | { |
545 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
546 | BUG(); | |
c8a446ad | 547 | __blk_mq_end_request(rq, error); |
63151a44 | 548 | } |
c8a446ad | 549 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 550 | |
30a91cb4 | 551 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 552 | { |
3d6efbf6 | 553 | struct request *rq = data; |
c7bb9ad1 | 554 | struct request_queue *q = rq->q; |
320ae51f | 555 | |
c7bb9ad1 | 556 | q->mq_ops->complete(rq); |
320ae51f | 557 | } |
320ae51f | 558 | |
453f8341 | 559 | static void __blk_mq_complete_request(struct request *rq) |
320ae51f JA |
560 | { |
561 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
c7bb9ad1 | 562 | struct request_queue *q = rq->q; |
38535201 | 563 | bool shared = false; |
320ae51f JA |
564 | int cpu; |
565 | ||
af78ff7c | 566 | WRITE_ONCE(rq->state, MQ_RQ_COMPLETE); |
36e76539 ML |
567 | /* |
568 | * Most of single queue controllers, there is only one irq vector | |
569 | * for handling IO completion, and the only irq's affinity is set | |
570 | * as all possible CPUs. On most of ARCHs, this affinity means the | |
571 | * irq is handled on one specific CPU. | |
572 | * | |
573 | * So complete IO reqeust in softirq context in case of single queue | |
574 | * for not degrading IO performance by irqsoff latency. | |
575 | */ | |
c7bb9ad1 | 576 | if (q->nr_hw_queues == 1) { |
36e76539 ML |
577 | __blk_complete_request(rq); |
578 | return; | |
579 | } | |
580 | ||
4ab32bf3 JA |
581 | /* |
582 | * For a polled request, always complete locallly, it's pointless | |
583 | * to redirect the completion. | |
584 | */ | |
585 | if ((rq->cmd_flags & REQ_HIPRI) || | |
586 | !test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags)) { | |
c7bb9ad1 | 587 | q->mq_ops->complete(rq); |
30a91cb4 CH |
588 | return; |
589 | } | |
320ae51f JA |
590 | |
591 | cpu = get_cpu(); | |
c7bb9ad1 | 592 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &q->queue_flags)) |
38535201 CH |
593 | shared = cpus_share_cache(cpu, ctx->cpu); |
594 | ||
595 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 596 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
597 | rq->csd.info = rq; |
598 | rq->csd.flags = 0; | |
c46fff2a | 599 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 600 | } else { |
c7bb9ad1 | 601 | q->mq_ops->complete(rq); |
3d6efbf6 | 602 | } |
320ae51f JA |
603 | put_cpu(); |
604 | } | |
30a91cb4 | 605 | |
04ced159 | 606 | static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx) |
b7435db8 | 607 | __releases(hctx->srcu) |
04ced159 JA |
608 | { |
609 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
610 | rcu_read_unlock(); | |
611 | else | |
05707b64 | 612 | srcu_read_unlock(hctx->srcu, srcu_idx); |
04ced159 JA |
613 | } |
614 | ||
615 | static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx) | |
b7435db8 | 616 | __acquires(hctx->srcu) |
04ced159 | 617 | { |
08b5a6e2 JA |
618 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
619 | /* shut up gcc false positive */ | |
620 | *srcu_idx = 0; | |
04ced159 | 621 | rcu_read_lock(); |
08b5a6e2 | 622 | } else |
05707b64 | 623 | *srcu_idx = srcu_read_lock(hctx->srcu); |
04ced159 JA |
624 | } |
625 | ||
30a91cb4 CH |
626 | /** |
627 | * blk_mq_complete_request - end I/O on a request | |
628 | * @rq: the request being processed | |
629 | * | |
630 | * Description: | |
631 | * Ends all I/O on a request. It does not handle partial completions. | |
632 | * The actual completion happens out-of-order, through a IPI handler. | |
633 | **/ | |
16c15eb1 | 634 | bool blk_mq_complete_request(struct request *rq) |
30a91cb4 | 635 | { |
12f5b931 | 636 | if (unlikely(blk_should_fake_timeout(rq->q))) |
16c15eb1 | 637 | return false; |
12f5b931 | 638 | __blk_mq_complete_request(rq); |
16c15eb1 | 639 | return true; |
30a91cb4 CH |
640 | } |
641 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 642 | |
105663f7 AA |
643 | /** |
644 | * blk_mq_start_request - Start processing a request | |
645 | * @rq: Pointer to request to be started | |
646 | * | |
647 | * Function used by device drivers to notify the block layer that a request | |
648 | * is going to be processed now, so blk layer can do proper initializations | |
649 | * such as starting the timeout timer. | |
650 | */ | |
e2490073 | 651 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
652 | { |
653 | struct request_queue *q = rq->q; | |
654 | ||
655 | trace_block_rq_issue(q, rq); | |
656 | ||
cf43e6be | 657 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
544ccc8d | 658 | rq->io_start_time_ns = ktime_get_ns(); |
3d244306 | 659 | rq->stats_sectors = blk_rq_sectors(rq); |
cf43e6be | 660 | rq->rq_flags |= RQF_STATS; |
a7905043 | 661 | rq_qos_issue(q, rq); |
cf43e6be JA |
662 | } |
663 | ||
1d9bd516 | 664 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 665 | |
1d9bd516 | 666 | blk_add_timer(rq); |
12f5b931 | 667 | WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); |
49f5baa5 | 668 | |
54d4e6ab MG |
669 | #ifdef CONFIG_BLK_DEV_INTEGRITY |
670 | if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE) | |
671 | q->integrity.profile->prepare_fn(rq); | |
672 | #endif | |
320ae51f | 673 | } |
e2490073 | 674 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 675 | |
ed0791b2 | 676 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
677 | { |
678 | struct request_queue *q = rq->q; | |
679 | ||
923218f6 ML |
680 | blk_mq_put_driver_tag(rq); |
681 | ||
320ae51f | 682 | trace_block_rq_requeue(q, rq); |
a7905043 | 683 | rq_qos_requeue(q, rq); |
49f5baa5 | 684 | |
12f5b931 KB |
685 | if (blk_mq_request_started(rq)) { |
686 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); | |
da661267 | 687 | rq->rq_flags &= ~RQF_TIMED_OUT; |
e2490073 | 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) { | |
aef1897c | 715 | if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP))) |
6fca6a61 CH |
716 | continue; |
717 | ||
e8064021 | 718 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 719 | list_del_init(&rq->queuelist); |
aef1897c JW |
720 | /* |
721 | * If RQF_DONTPREP, rq has contained some driver specific | |
722 | * data, so insert it to hctx dispatch list to avoid any | |
723 | * merge. | |
724 | */ | |
725 | if (rq->rq_flags & RQF_DONTPREP) | |
01e99aec | 726 | blk_mq_request_bypass_insert(rq, false, false); |
aef1897c JW |
727 | else |
728 | blk_mq_sched_insert_request(rq, true, false, false); | |
6fca6a61 CH |
729 | } |
730 | ||
731 | while (!list_empty(&rq_list)) { | |
732 | rq = list_entry(rq_list.next, struct request, queuelist); | |
733 | list_del_init(&rq->queuelist); | |
9e97d295 | 734 | blk_mq_sched_insert_request(rq, false, false, false); |
6fca6a61 CH |
735 | } |
736 | ||
52d7f1b5 | 737 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
738 | } |
739 | ||
2b053aca BVA |
740 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
741 | bool kick_requeue_list) | |
6fca6a61 CH |
742 | { |
743 | struct request_queue *q = rq->q; | |
744 | unsigned long flags; | |
745 | ||
746 | /* | |
747 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
ff821d27 | 748 | * request head insertion from the workqueue. |
6fca6a61 | 749 | */ |
e8064021 | 750 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
751 | |
752 | spin_lock_irqsave(&q->requeue_lock, flags); | |
753 | if (at_head) { | |
e8064021 | 754 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
755 | list_add(&rq->queuelist, &q->requeue_list); |
756 | } else { | |
757 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
758 | } | |
759 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
760 | |
761 | if (kick_requeue_list) | |
762 | blk_mq_kick_requeue_list(q); | |
6fca6a61 | 763 | } |
6fca6a61 CH |
764 | |
765 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
766 | { | |
ae943d20 | 767 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
768 | } |
769 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
770 | ||
2849450a MS |
771 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
772 | unsigned long msecs) | |
773 | { | |
d4acf365 BVA |
774 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
775 | msecs_to_jiffies(msecs)); | |
2849450a MS |
776 | } |
777 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
778 | ||
0e62f51f JA |
779 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
780 | { | |
88c7b2b7 JA |
781 | if (tag < tags->nr_tags) { |
782 | prefetch(tags->rqs[tag]); | |
4ee86bab | 783 | return tags->rqs[tag]; |
88c7b2b7 | 784 | } |
4ee86bab HR |
785 | |
786 | return NULL; | |
24d2f903 CH |
787 | } |
788 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
789 | ||
3c94d83c JA |
790 | static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq, |
791 | void *priv, bool reserved) | |
ae879912 JA |
792 | { |
793 | /* | |
3c94d83c JA |
794 | * If we find a request that is inflight and the queue matches, |
795 | * we know the queue is busy. Return false to stop the iteration. | |
ae879912 | 796 | */ |
3c94d83c | 797 | if (rq->state == MQ_RQ_IN_FLIGHT && rq->q == hctx->queue) { |
ae879912 JA |
798 | bool *busy = priv; |
799 | ||
800 | *busy = true; | |
801 | return false; | |
802 | } | |
803 | ||
804 | return true; | |
805 | } | |
806 | ||
3c94d83c | 807 | bool blk_mq_queue_inflight(struct request_queue *q) |
ae879912 JA |
808 | { |
809 | bool busy = false; | |
810 | ||
3c94d83c | 811 | blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy); |
ae879912 JA |
812 | return busy; |
813 | } | |
3c94d83c | 814 | EXPORT_SYMBOL_GPL(blk_mq_queue_inflight); |
ae879912 | 815 | |
358f70da | 816 | static void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 817 | { |
da661267 | 818 | req->rq_flags |= RQF_TIMED_OUT; |
d1210d5a CH |
819 | if (req->q->mq_ops->timeout) { |
820 | enum blk_eh_timer_return ret; | |
821 | ||
822 | ret = req->q->mq_ops->timeout(req, reserved); | |
823 | if (ret == BLK_EH_DONE) | |
824 | return; | |
825 | WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); | |
46f92d42 | 826 | } |
d1210d5a CH |
827 | |
828 | blk_add_timer(req); | |
87ee7b11 | 829 | } |
5b3f25fc | 830 | |
12f5b931 | 831 | static bool blk_mq_req_expired(struct request *rq, unsigned long *next) |
81481eb4 | 832 | { |
12f5b931 | 833 | unsigned long deadline; |
87ee7b11 | 834 | |
12f5b931 KB |
835 | if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) |
836 | return false; | |
da661267 CH |
837 | if (rq->rq_flags & RQF_TIMED_OUT) |
838 | return false; | |
a7af0af3 | 839 | |
079076b3 | 840 | deadline = READ_ONCE(rq->deadline); |
12f5b931 KB |
841 | if (time_after_eq(jiffies, deadline)) |
842 | return true; | |
a7af0af3 | 843 | |
12f5b931 KB |
844 | if (*next == 0) |
845 | *next = deadline; | |
846 | else if (time_after(*next, deadline)) | |
847 | *next = deadline; | |
848 | return false; | |
87ee7b11 JA |
849 | } |
850 | ||
7baa8572 | 851 | static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
1d9bd516 TH |
852 | struct request *rq, void *priv, bool reserved) |
853 | { | |
12f5b931 KB |
854 | unsigned long *next = priv; |
855 | ||
856 | /* | |
857 | * Just do a quick check if it is expired before locking the request in | |
858 | * so we're not unnecessarilly synchronizing across CPUs. | |
859 | */ | |
860 | if (!blk_mq_req_expired(rq, next)) | |
7baa8572 | 861 | return true; |
12f5b931 KB |
862 | |
863 | /* | |
864 | * We have reason to believe the request may be expired. Take a | |
865 | * reference on the request to lock this request lifetime into its | |
866 | * currently allocated context to prevent it from being reallocated in | |
867 | * the event the completion by-passes this timeout handler. | |
868 | * | |
869 | * If the reference was already released, then the driver beat the | |
870 | * timeout handler to posting a natural completion. | |
871 | */ | |
872 | if (!refcount_inc_not_zero(&rq->ref)) | |
7baa8572 | 873 | return true; |
12f5b931 | 874 | |
1d9bd516 | 875 | /* |
12f5b931 KB |
876 | * The request is now locked and cannot be reallocated underneath the |
877 | * timeout handler's processing. Re-verify this exact request is truly | |
878 | * expired; if it is not expired, then the request was completed and | |
879 | * reallocated as a new request. | |
1d9bd516 | 880 | */ |
12f5b931 | 881 | if (blk_mq_req_expired(rq, next)) |
1d9bd516 | 882 | blk_mq_rq_timed_out(rq, reserved); |
8d699663 YY |
883 | |
884 | if (is_flush_rq(rq, hctx)) | |
885 | rq->end_io(rq, 0); | |
886 | else if (refcount_dec_and_test(&rq->ref)) | |
12f5b931 | 887 | __blk_mq_free_request(rq); |
7baa8572 JA |
888 | |
889 | return true; | |
1d9bd516 TH |
890 | } |
891 | ||
287922eb | 892 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 893 | { |
287922eb CH |
894 | struct request_queue *q = |
895 | container_of(work, struct request_queue, timeout_work); | |
12f5b931 | 896 | unsigned long next = 0; |
1d9bd516 | 897 | struct blk_mq_hw_ctx *hctx; |
81481eb4 | 898 | int i; |
320ae51f | 899 | |
71f79fb3 GKB |
900 | /* A deadlock might occur if a request is stuck requiring a |
901 | * timeout at the same time a queue freeze is waiting | |
902 | * completion, since the timeout code would not be able to | |
903 | * acquire the queue reference here. | |
904 | * | |
905 | * That's why we don't use blk_queue_enter here; instead, we use | |
906 | * percpu_ref_tryget directly, because we need to be able to | |
907 | * obtain a reference even in the short window between the queue | |
908 | * starting to freeze, by dropping the first reference in | |
1671d522 | 909 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
910 | * consumed, marked by the instant q_usage_counter reaches |
911 | * zero. | |
912 | */ | |
913 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
914 | return; |
915 | ||
12f5b931 | 916 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next); |
320ae51f | 917 | |
12f5b931 KB |
918 | if (next != 0) { |
919 | mod_timer(&q->timeout, next); | |
0d2602ca | 920 | } else { |
fcd36c36 BVA |
921 | /* |
922 | * Request timeouts are handled as a forward rolling timer. If | |
923 | * we end up here it means that no requests are pending and | |
924 | * also that no request has been pending for a while. Mark | |
925 | * each hctx as idle. | |
926 | */ | |
f054b56c ML |
927 | queue_for_each_hw_ctx(q, hctx, i) { |
928 | /* the hctx may be unmapped, so check it here */ | |
929 | if (blk_mq_hw_queue_mapped(hctx)) | |
930 | blk_mq_tag_idle(hctx); | |
931 | } | |
0d2602ca | 932 | } |
287922eb | 933 | blk_queue_exit(q); |
320ae51f JA |
934 | } |
935 | ||
88459642 OS |
936 | struct flush_busy_ctx_data { |
937 | struct blk_mq_hw_ctx *hctx; | |
938 | struct list_head *list; | |
939 | }; | |
940 | ||
941 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
942 | { | |
943 | struct flush_busy_ctx_data *flush_data = data; | |
944 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
945 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 946 | enum hctx_type type = hctx->type; |
88459642 | 947 | |
88459642 | 948 | spin_lock(&ctx->lock); |
c16d6b5a | 949 | list_splice_tail_init(&ctx->rq_lists[type], flush_data->list); |
e9a99a63 | 950 | sbitmap_clear_bit(sb, bitnr); |
88459642 OS |
951 | spin_unlock(&ctx->lock); |
952 | return true; | |
953 | } | |
954 | ||
1429d7c9 JA |
955 | /* |
956 | * Process software queues that have been marked busy, splicing them | |
957 | * to the for-dispatch | |
958 | */ | |
2c3ad667 | 959 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 960 | { |
88459642 OS |
961 | struct flush_busy_ctx_data data = { |
962 | .hctx = hctx, | |
963 | .list = list, | |
964 | }; | |
1429d7c9 | 965 | |
88459642 | 966 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 967 | } |
2c3ad667 | 968 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 969 | |
b347689f ML |
970 | struct dispatch_rq_data { |
971 | struct blk_mq_hw_ctx *hctx; | |
972 | struct request *rq; | |
973 | }; | |
974 | ||
975 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
976 | void *data) | |
977 | { | |
978 | struct dispatch_rq_data *dispatch_data = data; | |
979 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
980 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 981 | enum hctx_type type = hctx->type; |
b347689f ML |
982 | |
983 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
984 | if (!list_empty(&ctx->rq_lists[type])) { |
985 | dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next); | |
b347689f | 986 | list_del_init(&dispatch_data->rq->queuelist); |
c16d6b5a | 987 | if (list_empty(&ctx->rq_lists[type])) |
b347689f ML |
988 | sbitmap_clear_bit(sb, bitnr); |
989 | } | |
990 | spin_unlock(&ctx->lock); | |
991 | ||
992 | return !dispatch_data->rq; | |
993 | } | |
994 | ||
995 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
996 | struct blk_mq_ctx *start) | |
997 | { | |
f31967f0 | 998 | unsigned off = start ? start->index_hw[hctx->type] : 0; |
b347689f ML |
999 | struct dispatch_rq_data data = { |
1000 | .hctx = hctx, | |
1001 | .rq = NULL, | |
1002 | }; | |
1003 | ||
1004 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1005 | dispatch_rq_from_ctx, &data); | |
1006 | ||
1007 | return data.rq; | |
1008 | } | |
1009 | ||
703fd1c0 JA |
1010 | static inline unsigned int queued_to_index(unsigned int queued) |
1011 | { | |
1012 | if (!queued) | |
1013 | return 0; | |
1429d7c9 | 1014 | |
703fd1c0 | 1015 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
1016 | } |
1017 | ||
8ab6bb9e | 1018 | bool blk_mq_get_driver_tag(struct request *rq) |
bd166ef1 JA |
1019 | { |
1020 | struct blk_mq_alloc_data data = { | |
1021 | .q = rq->q, | |
ea4f995e | 1022 | .hctx = rq->mq_hctx, |
8ab6bb9e | 1023 | .flags = BLK_MQ_REQ_NOWAIT, |
f9afca4d | 1024 | .cmd_flags = rq->cmd_flags, |
bd166ef1 | 1025 | }; |
d263ed99 | 1026 | bool shared; |
5feeacdd | 1027 | |
81380ca1 | 1028 | if (rq->tag != -1) |
1fead718 | 1029 | return true; |
bd166ef1 | 1030 | |
415b806d SG |
1031 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) |
1032 | data.flags |= BLK_MQ_REQ_RESERVED; | |
1033 | ||
d263ed99 | 1034 | shared = blk_mq_tag_busy(data.hctx); |
bd166ef1 JA |
1035 | rq->tag = blk_mq_get_tag(&data); |
1036 | if (rq->tag >= 0) { | |
d263ed99 | 1037 | if (shared) { |
200e86b3 JA |
1038 | rq->rq_flags |= RQF_MQ_INFLIGHT; |
1039 | atomic_inc(&data.hctx->nr_active); | |
1040 | } | |
bd166ef1 | 1041 | data.hctx->tags->rqs[rq->tag] = rq; |
bd166ef1 JA |
1042 | } |
1043 | ||
81380ca1 | 1044 | return rq->tag != -1; |
bd166ef1 JA |
1045 | } |
1046 | ||
eb619fdb JA |
1047 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1048 | int flags, void *key) | |
da55f2cc OS |
1049 | { |
1050 | struct blk_mq_hw_ctx *hctx; | |
1051 | ||
1052 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1053 | ||
5815839b | 1054 | spin_lock(&hctx->dispatch_wait_lock); |
e8618575 JA |
1055 | if (!list_empty(&wait->entry)) { |
1056 | struct sbitmap_queue *sbq; | |
1057 | ||
1058 | list_del_init(&wait->entry); | |
1059 | sbq = &hctx->tags->bitmap_tags; | |
1060 | atomic_dec(&sbq->ws_active); | |
1061 | } | |
5815839b ML |
1062 | spin_unlock(&hctx->dispatch_wait_lock); |
1063 | ||
da55f2cc OS |
1064 | blk_mq_run_hw_queue(hctx, true); |
1065 | return 1; | |
1066 | } | |
1067 | ||
f906a6a0 JA |
1068 | /* |
1069 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1070 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1071 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1072 | * marking us as waiting. |
1073 | */ | |
2278d69f | 1074 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, |
f906a6a0 | 1075 | struct request *rq) |
da55f2cc | 1076 | { |
e8618575 | 1077 | struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags; |
5815839b | 1078 | struct wait_queue_head *wq; |
f906a6a0 JA |
1079 | wait_queue_entry_t *wait; |
1080 | bool ret; | |
da55f2cc | 1081 | |
2278d69f | 1082 | if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) { |
684b7324 | 1083 | blk_mq_sched_mark_restart_hctx(hctx); |
f906a6a0 | 1084 | |
c27d53fb BVA |
1085 | /* |
1086 | * It's possible that a tag was freed in the window between the | |
1087 | * allocation failure and adding the hardware queue to the wait | |
1088 | * queue. | |
1089 | * | |
1090 | * Don't clear RESTART here, someone else could have set it. | |
1091 | * At most this will cost an extra queue run. | |
1092 | */ | |
8ab6bb9e | 1093 | return blk_mq_get_driver_tag(rq); |
eb619fdb | 1094 | } |
eb619fdb | 1095 | |
2278d69f | 1096 | wait = &hctx->dispatch_wait; |
c27d53fb BVA |
1097 | if (!list_empty_careful(&wait->entry)) |
1098 | return false; | |
1099 | ||
e8618575 | 1100 | wq = &bt_wait_ptr(sbq, hctx)->wait; |
5815839b ML |
1101 | |
1102 | spin_lock_irq(&wq->lock); | |
1103 | spin_lock(&hctx->dispatch_wait_lock); | |
c27d53fb | 1104 | if (!list_empty(&wait->entry)) { |
5815839b ML |
1105 | spin_unlock(&hctx->dispatch_wait_lock); |
1106 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1107 | return false; |
eb619fdb JA |
1108 | } |
1109 | ||
e8618575 | 1110 | atomic_inc(&sbq->ws_active); |
5815839b ML |
1111 | wait->flags &= ~WQ_FLAG_EXCLUSIVE; |
1112 | __add_wait_queue(wq, wait); | |
c27d53fb | 1113 | |
da55f2cc | 1114 | /* |
eb619fdb JA |
1115 | * It's possible that a tag was freed in the window between the |
1116 | * allocation failure and adding the hardware queue to the wait | |
1117 | * queue. | |
da55f2cc | 1118 | */ |
8ab6bb9e | 1119 | ret = blk_mq_get_driver_tag(rq); |
c27d53fb | 1120 | if (!ret) { |
5815839b ML |
1121 | spin_unlock(&hctx->dispatch_wait_lock); |
1122 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1123 | return false; |
eb619fdb | 1124 | } |
c27d53fb BVA |
1125 | |
1126 | /* | |
1127 | * We got a tag, remove ourselves from the wait queue to ensure | |
1128 | * someone else gets the wakeup. | |
1129 | */ | |
c27d53fb | 1130 | list_del_init(&wait->entry); |
e8618575 | 1131 | atomic_dec(&sbq->ws_active); |
5815839b ML |
1132 | spin_unlock(&hctx->dispatch_wait_lock); |
1133 | spin_unlock_irq(&wq->lock); | |
c27d53fb BVA |
1134 | |
1135 | return true; | |
da55f2cc OS |
1136 | } |
1137 | ||
6e768717 ML |
1138 | #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 |
1139 | #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 | |
1140 | /* | |
1141 | * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): | |
1142 | * - EWMA is one simple way to compute running average value | |
1143 | * - weight(7/8 and 1/8) is applied so that it can decrease exponentially | |
1144 | * - take 4 as factor for avoiding to get too small(0) result, and this | |
1145 | * factor doesn't matter because EWMA decreases exponentially | |
1146 | */ | |
1147 | static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) | |
1148 | { | |
1149 | unsigned int ewma; | |
1150 | ||
1151 | if (hctx->queue->elevator) | |
1152 | return; | |
1153 | ||
1154 | ewma = hctx->dispatch_busy; | |
1155 | ||
1156 | if (!ewma && !busy) | |
1157 | return; | |
1158 | ||
1159 | ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; | |
1160 | if (busy) | |
1161 | ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; | |
1162 | ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; | |
1163 | ||
1164 | hctx->dispatch_busy = ewma; | |
1165 | } | |
1166 | ||
86ff7c2a ML |
1167 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
1168 | ||
c92a4103 JT |
1169 | static void blk_mq_handle_dev_resource(struct request *rq, |
1170 | struct list_head *list) | |
1171 | { | |
1172 | struct request *next = | |
1173 | list_first_entry_or_null(list, struct request, queuelist); | |
1174 | ||
1175 | /* | |
1176 | * If an I/O scheduler has been configured and we got a driver tag for | |
1177 | * the next request already, free it. | |
1178 | */ | |
1179 | if (next) | |
1180 | blk_mq_put_driver_tag(next); | |
1181 | ||
1182 | list_add(&rq->queuelist, list); | |
1183 | __blk_mq_requeue_request(rq); | |
1184 | } | |
1185 | ||
1f57f8d4 JA |
1186 | /* |
1187 | * Returns true if we did some work AND can potentially do more. | |
1188 | */ | |
de148297 | 1189 | bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list, |
eb619fdb | 1190 | bool got_budget) |
320ae51f | 1191 | { |
81380ca1 | 1192 | struct blk_mq_hw_ctx *hctx; |
6d6f167c | 1193 | struct request *rq, *nxt; |
eb619fdb | 1194 | bool no_tag = false; |
fc17b653 | 1195 | int errors, queued; |
86ff7c2a | 1196 | blk_status_t ret = BLK_STS_OK; |
ab3cee37 | 1197 | bool no_budget_avail = false; |
320ae51f | 1198 | |
81380ca1 OS |
1199 | if (list_empty(list)) |
1200 | return false; | |
1201 | ||
de148297 ML |
1202 | WARN_ON(!list_is_singular(list) && got_budget); |
1203 | ||
320ae51f JA |
1204 | /* |
1205 | * Now process all the entries, sending them to the driver. | |
1206 | */ | |
93efe981 | 1207 | errors = queued = 0; |
81380ca1 | 1208 | do { |
74c45052 | 1209 | struct blk_mq_queue_data bd; |
320ae51f | 1210 | |
f04c3df3 | 1211 | rq = list_first_entry(list, struct request, queuelist); |
0bca799b | 1212 | |
ea4f995e | 1213 | hctx = rq->mq_hctx; |
5fe56de7 JG |
1214 | if (!got_budget && !blk_mq_get_dispatch_budget(hctx)) { |
1215 | blk_mq_put_driver_tag(rq); | |
ab3cee37 | 1216 | no_budget_avail = true; |
0bca799b | 1217 | break; |
5fe56de7 | 1218 | } |
0bca799b | 1219 | |
8ab6bb9e | 1220 | if (!blk_mq_get_driver_tag(rq)) { |
3c782d67 | 1221 | /* |
da55f2cc | 1222 | * The initial allocation attempt failed, so we need to |
eb619fdb JA |
1223 | * rerun the hardware queue when a tag is freed. The |
1224 | * waitqueue takes care of that. If the queue is run | |
1225 | * before we add this entry back on the dispatch list, | |
1226 | * we'll re-run it below. | |
3c782d67 | 1227 | */ |
2278d69f | 1228 | if (!blk_mq_mark_tag_wait(hctx, rq)) { |
0bca799b | 1229 | blk_mq_put_dispatch_budget(hctx); |
f906a6a0 JA |
1230 | /* |
1231 | * For non-shared tags, the RESTART check | |
1232 | * will suffice. | |
1233 | */ | |
1234 | if (hctx->flags & BLK_MQ_F_TAG_SHARED) | |
1235 | no_tag = true; | |
de148297 ML |
1236 | break; |
1237 | } | |
1238 | } | |
1239 | ||
320ae51f | 1240 | list_del_init(&rq->queuelist); |
320ae51f | 1241 | |
74c45052 | 1242 | bd.rq = rq; |
113285b4 JA |
1243 | |
1244 | /* | |
1245 | * Flag last if we have no more requests, or if we have more | |
1246 | * but can't assign a driver tag to it. | |
1247 | */ | |
1248 | if (list_empty(list)) | |
1249 | bd.last = true; | |
1250 | else { | |
113285b4 | 1251 | nxt = list_first_entry(list, struct request, queuelist); |
8ab6bb9e | 1252 | bd.last = !blk_mq_get_driver_tag(nxt); |
113285b4 | 1253 | } |
74c45052 JA |
1254 | |
1255 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
86ff7c2a | 1256 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) { |
c92a4103 | 1257 | blk_mq_handle_dev_resource(rq, list); |
320ae51f | 1258 | break; |
fc17b653 CH |
1259 | } |
1260 | ||
1261 | if (unlikely(ret != BLK_STS_OK)) { | |
93efe981 | 1262 | errors++; |
2a842aca | 1263 | blk_mq_end_request(rq, BLK_STS_IOERR); |
fc17b653 | 1264 | continue; |
320ae51f JA |
1265 | } |
1266 | ||
fc17b653 | 1267 | queued++; |
81380ca1 | 1268 | } while (!list_empty(list)); |
320ae51f | 1269 | |
703fd1c0 | 1270 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
1271 | |
1272 | /* | |
1273 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1274 | * that is where we will continue on next queue run. | |
1275 | */ | |
f04c3df3 | 1276 | if (!list_empty(list)) { |
86ff7c2a ML |
1277 | bool needs_restart; |
1278 | ||
d666ba98 JA |
1279 | /* |
1280 | * If we didn't flush the entire list, we could have told | |
1281 | * the driver there was more coming, but that turned out to | |
1282 | * be a lie. | |
1283 | */ | |
536167d4 | 1284 | if (q->mq_ops->commit_rqs && queued) |
d666ba98 JA |
1285 | q->mq_ops->commit_rqs(hctx); |
1286 | ||
320ae51f | 1287 | spin_lock(&hctx->lock); |
01e99aec | 1288 | list_splice_tail_init(list, &hctx->dispatch); |
320ae51f | 1289 | spin_unlock(&hctx->lock); |
f04c3df3 | 1290 | |
9ba52e58 | 1291 | /* |
710c785f BVA |
1292 | * If SCHED_RESTART was set by the caller of this function and |
1293 | * it is no longer set that means that it was cleared by another | |
1294 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1295 | * |
eb619fdb JA |
1296 | * If 'no_tag' is set, that means that we failed getting |
1297 | * a driver tag with an I/O scheduler attached. If our dispatch | |
1298 | * waitqueue is no longer active, ensure that we run the queue | |
1299 | * AFTER adding our entries back to the list. | |
bd166ef1 | 1300 | * |
710c785f BVA |
1301 | * If no I/O scheduler has been configured it is possible that |
1302 | * the hardware queue got stopped and restarted before requests | |
1303 | * were pushed back onto the dispatch list. Rerun the queue to | |
1304 | * avoid starvation. Notes: | |
1305 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1306 | * been stopped before rerunning a queue. | |
1307 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1308 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1309 | * and dm-rq. |
86ff7c2a ML |
1310 | * |
1311 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
1312 | * bit is set, run queue after a delay to avoid IO stalls | |
ab3cee37 DA |
1313 | * that could otherwise occur if the queue is idle. We'll do |
1314 | * similar if we couldn't get budget and SCHED_RESTART is set. | |
bd166ef1 | 1315 | */ |
86ff7c2a ML |
1316 | needs_restart = blk_mq_sched_needs_restart(hctx); |
1317 | if (!needs_restart || | |
eb619fdb | 1318 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
bd166ef1 | 1319 | blk_mq_run_hw_queue(hctx, true); |
ab3cee37 DA |
1320 | else if (needs_restart && (ret == BLK_STS_RESOURCE || |
1321 | no_budget_avail)) | |
86ff7c2a | 1322 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); |
1f57f8d4 | 1323 | |
6e768717 | 1324 | blk_mq_update_dispatch_busy(hctx, true); |
1f57f8d4 | 1325 | return false; |
6e768717 ML |
1326 | } else |
1327 | blk_mq_update_dispatch_busy(hctx, false); | |
f04c3df3 | 1328 | |
1f57f8d4 JA |
1329 | /* |
1330 | * If the host/device is unable to accept more work, inform the | |
1331 | * caller of that. | |
1332 | */ | |
1333 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) | |
1334 | return false; | |
1335 | ||
93efe981 | 1336 | return (queued + errors) != 0; |
f04c3df3 JA |
1337 | } |
1338 | ||
105663f7 AA |
1339 | /** |
1340 | * __blk_mq_run_hw_queue - Run a hardware queue. | |
1341 | * @hctx: Pointer to the hardware queue to run. | |
1342 | * | |
1343 | * Send pending requests to the hardware. | |
1344 | */ | |
6a83e74d BVA |
1345 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1346 | { | |
1347 | int srcu_idx; | |
1348 | ||
b7a71e66 JA |
1349 | /* |
1350 | * We should be running this queue from one of the CPUs that | |
1351 | * are mapped to it. | |
7df938fb ML |
1352 | * |
1353 | * There are at least two related races now between setting | |
1354 | * hctx->next_cpu from blk_mq_hctx_next_cpu() and running | |
1355 | * __blk_mq_run_hw_queue(): | |
1356 | * | |
1357 | * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(), | |
1358 | * but later it becomes online, then this warning is harmless | |
1359 | * at all | |
1360 | * | |
1361 | * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(), | |
1362 | * but later it becomes offline, then the warning can't be | |
1363 | * triggered, and we depend on blk-mq timeout handler to | |
1364 | * handle dispatched requests to this hctx | |
b7a71e66 | 1365 | */ |
7df938fb ML |
1366 | if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && |
1367 | cpu_online(hctx->next_cpu)) { | |
1368 | printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n", | |
1369 | raw_smp_processor_id(), | |
1370 | cpumask_empty(hctx->cpumask) ? "inactive": "active"); | |
1371 | dump_stack(); | |
1372 | } | |
6a83e74d | 1373 | |
b7a71e66 JA |
1374 | /* |
1375 | * We can't run the queue inline with ints disabled. Ensure that | |
1376 | * we catch bad users of this early. | |
1377 | */ | |
1378 | WARN_ON_ONCE(in_interrupt()); | |
1379 | ||
04ced159 | 1380 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1381 | |
04ced159 JA |
1382 | hctx_lock(hctx, &srcu_idx); |
1383 | blk_mq_sched_dispatch_requests(hctx); | |
1384 | hctx_unlock(hctx, srcu_idx); | |
6a83e74d BVA |
1385 | } |
1386 | ||
f82ddf19 ML |
1387 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
1388 | { | |
1389 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
1390 | ||
1391 | if (cpu >= nr_cpu_ids) | |
1392 | cpu = cpumask_first(hctx->cpumask); | |
1393 | return cpu; | |
1394 | } | |
1395 | ||
506e931f JA |
1396 | /* |
1397 | * It'd be great if the workqueue API had a way to pass | |
1398 | * in a mask and had some smarts for more clever placement. | |
1399 | * For now we just round-robin here, switching for every | |
1400 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1401 | */ | |
1402 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1403 | { | |
7bed4595 | 1404 | bool tried = false; |
476f8c98 | 1405 | int next_cpu = hctx->next_cpu; |
7bed4595 | 1406 | |
b657d7e6 CH |
1407 | if (hctx->queue->nr_hw_queues == 1) |
1408 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1409 | |
1410 | if (--hctx->next_cpu_batch <= 0) { | |
7bed4595 | 1411 | select_cpu: |
476f8c98 | 1412 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
20e4d813 | 1413 | cpu_online_mask); |
506e931f | 1414 | if (next_cpu >= nr_cpu_ids) |
f82ddf19 | 1415 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
1416 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
1417 | } | |
1418 | ||
7bed4595 ML |
1419 | /* |
1420 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
1421 | * and it should only happen in the path of handling CPU DEAD. | |
1422 | */ | |
476f8c98 | 1423 | if (!cpu_online(next_cpu)) { |
7bed4595 ML |
1424 | if (!tried) { |
1425 | tried = true; | |
1426 | goto select_cpu; | |
1427 | } | |
1428 | ||
1429 | /* | |
1430 | * Make sure to re-select CPU next time once after CPUs | |
1431 | * in hctx->cpumask become online again. | |
1432 | */ | |
476f8c98 | 1433 | hctx->next_cpu = next_cpu; |
7bed4595 ML |
1434 | hctx->next_cpu_batch = 1; |
1435 | return WORK_CPU_UNBOUND; | |
1436 | } | |
476f8c98 ML |
1437 | |
1438 | hctx->next_cpu = next_cpu; | |
1439 | return next_cpu; | |
506e931f JA |
1440 | } |
1441 | ||
105663f7 AA |
1442 | /** |
1443 | * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue. | |
1444 | * @hctx: Pointer to the hardware queue to run. | |
1445 | * @async: If we want to run the queue asynchronously. | |
1446 | * @msecs: Microseconds of delay to wait before running the queue. | |
1447 | * | |
1448 | * If !@async, try to run the queue now. Else, run the queue asynchronously and | |
1449 | * with a delay of @msecs. | |
1450 | */ | |
7587a5ae BVA |
1451 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1452 | unsigned long msecs) | |
320ae51f | 1453 | { |
5435c023 | 1454 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f JA |
1455 | return; |
1456 | ||
1b792f2f | 1457 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1458 | int cpu = get_cpu(); |
1459 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1460 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1461 | put_cpu(); |
398205b8 PB |
1462 | return; |
1463 | } | |
e4043dcf | 1464 | |
2a90d4aa | 1465 | put_cpu(); |
e4043dcf | 1466 | } |
398205b8 | 1467 | |
ae943d20 BVA |
1468 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
1469 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
1470 | } |
1471 | ||
105663f7 AA |
1472 | /** |
1473 | * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously. | |
1474 | * @hctx: Pointer to the hardware queue to run. | |
1475 | * @msecs: Microseconds of delay to wait before running the queue. | |
1476 | * | |
1477 | * Run a hardware queue asynchronously with a delay of @msecs. | |
1478 | */ | |
7587a5ae BVA |
1479 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) |
1480 | { | |
1481 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1482 | } | |
1483 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1484 | ||
105663f7 AA |
1485 | /** |
1486 | * blk_mq_run_hw_queue - Start to run a hardware queue. | |
1487 | * @hctx: Pointer to the hardware queue to run. | |
1488 | * @async: If we want to run the queue asynchronously. | |
1489 | * | |
1490 | * Check if the request queue is not in a quiesced state and if there are | |
1491 | * pending requests to be sent. If this is true, run the queue to send requests | |
1492 | * to hardware. | |
1493 | */ | |
626fb735 | 1494 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 1495 | { |
24f5a90f ML |
1496 | int srcu_idx; |
1497 | bool need_run; | |
1498 | ||
1499 | /* | |
1500 | * When queue is quiesced, we may be switching io scheduler, or | |
1501 | * updating nr_hw_queues, or other things, and we can't run queue | |
1502 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
1503 | * | |
1504 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
1505 | * quiesced. | |
1506 | */ | |
04ced159 JA |
1507 | hctx_lock(hctx, &srcu_idx); |
1508 | need_run = !blk_queue_quiesced(hctx->queue) && | |
1509 | blk_mq_hctx_has_pending(hctx); | |
1510 | hctx_unlock(hctx, srcu_idx); | |
24f5a90f | 1511 | |
626fb735 | 1512 | if (need_run) |
79f720a7 | 1513 | __blk_mq_delay_run_hw_queue(hctx, async, 0); |
320ae51f | 1514 | } |
5b727272 | 1515 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1516 | |
105663f7 AA |
1517 | /** |
1518 | * blk_mq_run_hw_queue - Run all hardware queues in a request queue. | |
1519 | * @q: Pointer to the request queue to run. | |
1520 | * @async: If we want to run the queue asynchronously. | |
1521 | */ | |
b94ec296 | 1522 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1523 | { |
1524 | struct blk_mq_hw_ctx *hctx; | |
1525 | int i; | |
1526 | ||
1527 | queue_for_each_hw_ctx(q, hctx, i) { | |
79f720a7 | 1528 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1529 | continue; |
1530 | ||
b94ec296 | 1531 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1532 | } |
1533 | } | |
b94ec296 | 1534 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1535 | |
b9151e7b DA |
1536 | /** |
1537 | * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously. | |
1538 | * @q: Pointer to the request queue to run. | |
1539 | * @msecs: Microseconds of delay to wait before running the queues. | |
1540 | */ | |
1541 | void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs) | |
1542 | { | |
1543 | struct blk_mq_hw_ctx *hctx; | |
1544 | int i; | |
1545 | ||
1546 | queue_for_each_hw_ctx(q, hctx, i) { | |
1547 | if (blk_mq_hctx_stopped(hctx)) | |
1548 | continue; | |
1549 | ||
1550 | blk_mq_delay_run_hw_queue(hctx, msecs); | |
1551 | } | |
1552 | } | |
1553 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queues); | |
1554 | ||
fd001443 BVA |
1555 | /** |
1556 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1557 | * @q: request queue. | |
1558 | * | |
1559 | * The caller is responsible for serializing this function against | |
1560 | * blk_mq_{start,stop}_hw_queue(). | |
1561 | */ | |
1562 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1563 | { | |
1564 | struct blk_mq_hw_ctx *hctx; | |
1565 | int i; | |
1566 | ||
1567 | queue_for_each_hw_ctx(q, hctx, i) | |
1568 | if (blk_mq_hctx_stopped(hctx)) | |
1569 | return true; | |
1570 | ||
1571 | return false; | |
1572 | } | |
1573 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1574 | ||
39a70c76 ML |
1575 | /* |
1576 | * This function is often used for pausing .queue_rq() by driver when | |
1577 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1578 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1579 | * |
1580 | * We do not guarantee that dispatch can be drained or blocked | |
1581 | * after blk_mq_stop_hw_queue() returns. Please use | |
1582 | * blk_mq_quiesce_queue() for that requirement. | |
1583 | */ | |
2719aa21 JA |
1584 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1585 | { | |
641a9ed6 | 1586 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 1587 | |
641a9ed6 | 1588 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 1589 | } |
641a9ed6 | 1590 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 1591 | |
39a70c76 ML |
1592 | /* |
1593 | * This function is often used for pausing .queue_rq() by driver when | |
1594 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1595 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1596 | * |
1597 | * We do not guarantee that dispatch can be drained or blocked | |
1598 | * after blk_mq_stop_hw_queues() returns. Please use | |
1599 | * blk_mq_quiesce_queue() for that requirement. | |
1600 | */ | |
2719aa21 JA |
1601 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1602 | { | |
641a9ed6 ML |
1603 | struct blk_mq_hw_ctx *hctx; |
1604 | int i; | |
1605 | ||
1606 | queue_for_each_hw_ctx(q, hctx, i) | |
1607 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
1608 | } |
1609 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1610 | ||
320ae51f JA |
1611 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1612 | { | |
1613 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1614 | |
0ffbce80 | 1615 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1616 | } |
1617 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1618 | ||
2f268556 CH |
1619 | void blk_mq_start_hw_queues(struct request_queue *q) |
1620 | { | |
1621 | struct blk_mq_hw_ctx *hctx; | |
1622 | int i; | |
1623 | ||
1624 | queue_for_each_hw_ctx(q, hctx, i) | |
1625 | blk_mq_start_hw_queue(hctx); | |
1626 | } | |
1627 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1628 | ||
ae911c5e JA |
1629 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1630 | { | |
1631 | if (!blk_mq_hctx_stopped(hctx)) | |
1632 | return; | |
1633 | ||
1634 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1635 | blk_mq_run_hw_queue(hctx, async); | |
1636 | } | |
1637 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1638 | ||
1b4a3258 | 1639 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1640 | { |
1641 | struct blk_mq_hw_ctx *hctx; | |
1642 | int i; | |
1643 | ||
ae911c5e JA |
1644 | queue_for_each_hw_ctx(q, hctx, i) |
1645 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1646 | } |
1647 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1648 | ||
70f4db63 | 1649 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1650 | { |
1651 | struct blk_mq_hw_ctx *hctx; | |
1652 | ||
9f993737 | 1653 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 1654 | |
21c6e939 | 1655 | /* |
15fe8a90 | 1656 | * If we are stopped, don't run the queue. |
21c6e939 | 1657 | */ |
15fe8a90 | 1658 | if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
0196d6b4 | 1659 | return; |
7587a5ae BVA |
1660 | |
1661 | __blk_mq_run_hw_queue(hctx); | |
1662 | } | |
1663 | ||
cfd0c552 | 1664 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1665 | struct request *rq, |
1666 | bool at_head) | |
320ae51f | 1667 | { |
e57690fe | 1668 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
c16d6b5a | 1669 | enum hctx_type type = hctx->type; |
e57690fe | 1670 | |
7b607814 BVA |
1671 | lockdep_assert_held(&ctx->lock); |
1672 | ||
01b983c9 JA |
1673 | trace_block_rq_insert(hctx->queue, rq); |
1674 | ||
72a0a36e | 1675 | if (at_head) |
c16d6b5a | 1676 | list_add(&rq->queuelist, &ctx->rq_lists[type]); |
72a0a36e | 1677 | else |
c16d6b5a | 1678 | list_add_tail(&rq->queuelist, &ctx->rq_lists[type]); |
cfd0c552 | 1679 | } |
4bb659b1 | 1680 | |
2c3ad667 JA |
1681 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1682 | bool at_head) | |
cfd0c552 ML |
1683 | { |
1684 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1685 | ||
7b607814 BVA |
1686 | lockdep_assert_held(&ctx->lock); |
1687 | ||
e57690fe | 1688 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1689 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1690 | } |
1691 | ||
105663f7 AA |
1692 | /** |
1693 | * blk_mq_request_bypass_insert - Insert a request at dispatch list. | |
1694 | * @rq: Pointer to request to be inserted. | |
1695 | * @run_queue: If we should run the hardware queue after inserting the request. | |
1696 | * | |
157f377b JA |
1697 | * Should only be used carefully, when the caller knows we want to |
1698 | * bypass a potential IO scheduler on the target device. | |
1699 | */ | |
01e99aec ML |
1700 | void blk_mq_request_bypass_insert(struct request *rq, bool at_head, |
1701 | bool run_queue) | |
157f377b | 1702 | { |
ea4f995e | 1703 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
157f377b JA |
1704 | |
1705 | spin_lock(&hctx->lock); | |
01e99aec ML |
1706 | if (at_head) |
1707 | list_add(&rq->queuelist, &hctx->dispatch); | |
1708 | else | |
1709 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
157f377b JA |
1710 | spin_unlock(&hctx->lock); |
1711 | ||
b0850297 ML |
1712 | if (run_queue) |
1713 | blk_mq_run_hw_queue(hctx, false); | |
157f377b JA |
1714 | } |
1715 | ||
bd166ef1 JA |
1716 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1717 | struct list_head *list) | |
320ae51f JA |
1718 | |
1719 | { | |
3f0cedc7 | 1720 | struct request *rq; |
c16d6b5a | 1721 | enum hctx_type type = hctx->type; |
3f0cedc7 | 1722 | |
320ae51f JA |
1723 | /* |
1724 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1725 | * offline now | |
1726 | */ | |
3f0cedc7 | 1727 | list_for_each_entry(rq, list, queuelist) { |
e57690fe | 1728 | BUG_ON(rq->mq_ctx != ctx); |
3f0cedc7 | 1729 | trace_block_rq_insert(hctx->queue, rq); |
320ae51f | 1730 | } |
3f0cedc7 ML |
1731 | |
1732 | spin_lock(&ctx->lock); | |
c16d6b5a | 1733 | list_splice_tail_init(list, &ctx->rq_lists[type]); |
cfd0c552 | 1734 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1735 | spin_unlock(&ctx->lock); |
320ae51f JA |
1736 | } |
1737 | ||
3110fc79 | 1738 | static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b) |
320ae51f JA |
1739 | { |
1740 | struct request *rqa = container_of(a, struct request, queuelist); | |
1741 | struct request *rqb = container_of(b, struct request, queuelist); | |
1742 | ||
7d30a621 PB |
1743 | if (rqa->mq_ctx != rqb->mq_ctx) |
1744 | return rqa->mq_ctx > rqb->mq_ctx; | |
1745 | if (rqa->mq_hctx != rqb->mq_hctx) | |
1746 | return rqa->mq_hctx > rqb->mq_hctx; | |
3110fc79 JA |
1747 | |
1748 | return blk_rq_pos(rqa) > blk_rq_pos(rqb); | |
320ae51f JA |
1749 | } |
1750 | ||
1751 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1752 | { | |
320ae51f | 1753 | LIST_HEAD(list); |
320ae51f | 1754 | |
95ed0c5b PB |
1755 | if (list_empty(&plug->mq_list)) |
1756 | return; | |
320ae51f JA |
1757 | list_splice_init(&plug->mq_list, &list); |
1758 | ||
ce5b009c JA |
1759 | if (plug->rq_count > 2 && plug->multiple_queues) |
1760 | list_sort(NULL, &list, plug_rq_cmp); | |
320ae51f | 1761 | |
bcc816df DZ |
1762 | plug->rq_count = 0; |
1763 | ||
95ed0c5b PB |
1764 | do { |
1765 | struct list_head rq_list; | |
1766 | struct request *rq, *head_rq = list_entry_rq(list.next); | |
1767 | struct list_head *pos = &head_rq->queuelist; /* skip first */ | |
1768 | struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx; | |
1769 | struct blk_mq_ctx *this_ctx = head_rq->mq_ctx; | |
1770 | unsigned int depth = 1; | |
1771 | ||
1772 | list_for_each_continue(pos, &list) { | |
1773 | rq = list_entry_rq(pos); | |
1774 | BUG_ON(!rq->q); | |
1775 | if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) | |
1776 | break; | |
1777 | depth++; | |
320ae51f JA |
1778 | } |
1779 | ||
95ed0c5b PB |
1780 | list_cut_before(&rq_list, &list, pos); |
1781 | trace_block_unplug(head_rq->q, depth, !from_schedule); | |
67cae4c9 | 1782 | blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list, |
bd166ef1 | 1783 | from_schedule); |
95ed0c5b | 1784 | } while(!list_empty(&list)); |
320ae51f JA |
1785 | } |
1786 | ||
14ccb66b CH |
1787 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio, |
1788 | unsigned int nr_segs) | |
320ae51f | 1789 | { |
f924cdde CH |
1790 | if (bio->bi_opf & REQ_RAHEAD) |
1791 | rq->cmd_flags |= REQ_FAILFAST_MASK; | |
1792 | ||
1793 | rq->__sector = bio->bi_iter.bi_sector; | |
1794 | rq->write_hint = bio->bi_write_hint; | |
14ccb66b | 1795 | blk_rq_bio_prep(rq, bio, nr_segs); |
4b570521 | 1796 | |
6e85eaf3 | 1797 | blk_account_io_start(rq, true); |
320ae51f JA |
1798 | } |
1799 | ||
0f95549c MS |
1800 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
1801 | struct request *rq, | |
be94f058 | 1802 | blk_qc_t *cookie, bool last) |
f984df1f | 1803 | { |
f984df1f | 1804 | struct request_queue *q = rq->q; |
f984df1f SL |
1805 | struct blk_mq_queue_data bd = { |
1806 | .rq = rq, | |
be94f058 | 1807 | .last = last, |
f984df1f | 1808 | }; |
bd166ef1 | 1809 | blk_qc_t new_cookie; |
f06345ad | 1810 | blk_status_t ret; |
0f95549c MS |
1811 | |
1812 | new_cookie = request_to_qc_t(hctx, rq); | |
1813 | ||
1814 | /* | |
1815 | * For OK queue, we are done. For error, caller may kill it. | |
1816 | * Any other error (busy), just add it to our list as we | |
1817 | * previously would have done. | |
1818 | */ | |
1819 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
1820 | switch (ret) { | |
1821 | case BLK_STS_OK: | |
6ce3dd6e | 1822 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
1823 | *cookie = new_cookie; |
1824 | break; | |
1825 | case BLK_STS_RESOURCE: | |
86ff7c2a | 1826 | case BLK_STS_DEV_RESOURCE: |
6ce3dd6e | 1827 | blk_mq_update_dispatch_busy(hctx, true); |
0f95549c MS |
1828 | __blk_mq_requeue_request(rq); |
1829 | break; | |
1830 | default: | |
6ce3dd6e | 1831 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
1832 | *cookie = BLK_QC_T_NONE; |
1833 | break; | |
1834 | } | |
1835 | ||
1836 | return ret; | |
1837 | } | |
1838 | ||
fd9c40f6 | 1839 | static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
0f95549c | 1840 | struct request *rq, |
396eaf21 | 1841 | blk_qc_t *cookie, |
fd9c40f6 | 1842 | bool bypass_insert, bool last) |
0f95549c MS |
1843 | { |
1844 | struct request_queue *q = rq->q; | |
d964f04a ML |
1845 | bool run_queue = true; |
1846 | ||
23d4ee19 | 1847 | /* |
fd9c40f6 | 1848 | * RCU or SRCU read lock is needed before checking quiesced flag. |
23d4ee19 | 1849 | * |
fd9c40f6 BVA |
1850 | * When queue is stopped or quiesced, ignore 'bypass_insert' from |
1851 | * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller, | |
1852 | * and avoid driver to try to dispatch again. | |
23d4ee19 | 1853 | */ |
fd9c40f6 | 1854 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { |
d964f04a | 1855 | run_queue = false; |
fd9c40f6 BVA |
1856 | bypass_insert = false; |
1857 | goto insert; | |
d964f04a | 1858 | } |
f984df1f | 1859 | |
fd9c40f6 BVA |
1860 | if (q->elevator && !bypass_insert) |
1861 | goto insert; | |
2253efc8 | 1862 | |
0bca799b | 1863 | if (!blk_mq_get_dispatch_budget(hctx)) |
fd9c40f6 | 1864 | goto insert; |
bd166ef1 | 1865 | |
8ab6bb9e | 1866 | if (!blk_mq_get_driver_tag(rq)) { |
0bca799b | 1867 | blk_mq_put_dispatch_budget(hctx); |
fd9c40f6 | 1868 | goto insert; |
88022d72 | 1869 | } |
de148297 | 1870 | |
fd9c40f6 BVA |
1871 | return __blk_mq_issue_directly(hctx, rq, cookie, last); |
1872 | insert: | |
1873 | if (bypass_insert) | |
1874 | return BLK_STS_RESOURCE; | |
1875 | ||
01e99aec | 1876 | blk_mq_request_bypass_insert(rq, false, run_queue); |
fd9c40f6 BVA |
1877 | return BLK_STS_OK; |
1878 | } | |
1879 | ||
105663f7 AA |
1880 | /** |
1881 | * blk_mq_try_issue_directly - Try to send a request directly to device driver. | |
1882 | * @hctx: Pointer of the associated hardware queue. | |
1883 | * @rq: Pointer to request to be sent. | |
1884 | * @cookie: Request queue cookie. | |
1885 | * | |
1886 | * If the device has enough resources to accept a new request now, send the | |
1887 | * request directly to device driver. Else, insert at hctx->dispatch queue, so | |
1888 | * we can try send it another time in the future. Requests inserted at this | |
1889 | * queue have higher priority. | |
1890 | */ | |
fd9c40f6 BVA |
1891 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1892 | struct request *rq, blk_qc_t *cookie) | |
1893 | { | |
1894 | blk_status_t ret; | |
1895 | int srcu_idx; | |
1896 | ||
1897 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); | |
1898 | ||
1899 | hctx_lock(hctx, &srcu_idx); | |
1900 | ||
1901 | ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true); | |
1902 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) | |
01e99aec | 1903 | blk_mq_request_bypass_insert(rq, false, true); |
fd9c40f6 BVA |
1904 | else if (ret != BLK_STS_OK) |
1905 | blk_mq_end_request(rq, ret); | |
1906 | ||
1907 | hctx_unlock(hctx, srcu_idx); | |
1908 | } | |
1909 | ||
1910 | blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last) | |
1911 | { | |
1912 | blk_status_t ret; | |
1913 | int srcu_idx; | |
1914 | blk_qc_t unused_cookie; | |
1915 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
1916 | ||
1917 | hctx_lock(hctx, &srcu_idx); | |
1918 | ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last); | |
04ced159 | 1919 | hctx_unlock(hctx, srcu_idx); |
7f556a44 JW |
1920 | |
1921 | return ret; | |
5eb6126e CH |
1922 | } |
1923 | ||
6ce3dd6e ML |
1924 | void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
1925 | struct list_head *list) | |
1926 | { | |
536167d4 KB |
1927 | int queued = 0; |
1928 | ||
6ce3dd6e | 1929 | while (!list_empty(list)) { |
fd9c40f6 | 1930 | blk_status_t ret; |
6ce3dd6e ML |
1931 | struct request *rq = list_first_entry(list, struct request, |
1932 | queuelist); | |
1933 | ||
1934 | list_del_init(&rq->queuelist); | |
fd9c40f6 BVA |
1935 | ret = blk_mq_request_issue_directly(rq, list_empty(list)); |
1936 | if (ret != BLK_STS_OK) { | |
1937 | if (ret == BLK_STS_RESOURCE || | |
1938 | ret == BLK_STS_DEV_RESOURCE) { | |
01e99aec | 1939 | blk_mq_request_bypass_insert(rq, false, |
c616cbee | 1940 | list_empty(list)); |
fd9c40f6 BVA |
1941 | break; |
1942 | } | |
1943 | blk_mq_end_request(rq, ret); | |
536167d4 KB |
1944 | } else |
1945 | queued++; | |
6ce3dd6e | 1946 | } |
d666ba98 JA |
1947 | |
1948 | /* | |
1949 | * If we didn't flush the entire list, we could have told | |
1950 | * the driver there was more coming, but that turned out to | |
1951 | * be a lie. | |
1952 | */ | |
536167d4 | 1953 | if (!list_empty(list) && hctx->queue->mq_ops->commit_rqs && queued) |
d666ba98 | 1954 | hctx->queue->mq_ops->commit_rqs(hctx); |
6ce3dd6e ML |
1955 | } |
1956 | ||
ce5b009c JA |
1957 | static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq) |
1958 | { | |
1959 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1960 | plug->rq_count++; | |
1961 | if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) { | |
1962 | struct request *tmp; | |
1963 | ||
1964 | tmp = list_first_entry(&plug->mq_list, struct request, | |
1965 | queuelist); | |
1966 | if (tmp->q != rq->q) | |
1967 | plug->multiple_queues = true; | |
1968 | } | |
1969 | } | |
1970 | ||
105663f7 AA |
1971 | /** |
1972 | * blk_mq_make_request - Create and send a request to block device. | |
1973 | * @q: Request queue pointer. | |
1974 | * @bio: Bio pointer. | |
1975 | * | |
1976 | * Builds up a request structure from @q and @bio and send to the device. The | |
1977 | * request may not be queued directly to hardware if: | |
1978 | * * This request can be merged with another one | |
1979 | * * We want to place request at plug queue for possible future merging | |
1980 | * * There is an IO scheduler active at this queue | |
1981 | * | |
1982 | * It will not queue the request if there is an error with the bio, or at the | |
1983 | * request creation. | |
1984 | * | |
1985 | * Returns: Request queue cookie. | |
1986 | */ | |
8cf7961d | 1987 | blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1988 | { |
ef295ecf | 1989 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1990 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
7809167d | 1991 | struct blk_mq_alloc_data data = { .flags = 0}; |
07068d5b | 1992 | struct request *rq; |
f984df1f | 1993 | struct blk_plug *plug; |
5b3f341f | 1994 | struct request *same_queue_rq = NULL; |
14ccb66b | 1995 | unsigned int nr_segs; |
7b371636 | 1996 | blk_qc_t cookie; |
07068d5b JA |
1997 | |
1998 | blk_queue_bounce(q, &bio); | |
14ccb66b | 1999 | __blk_queue_split(q, &bio, &nr_segs); |
f36ea50c | 2000 | |
e23947bd | 2001 | if (!bio_integrity_prep(bio)) |
dece1635 | 2002 | return BLK_QC_T_NONE; |
07068d5b | 2003 | |
87c279e6 | 2004 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
14ccb66b | 2005 | blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq)) |
87c279e6 | 2006 | return BLK_QC_T_NONE; |
f984df1f | 2007 | |
14ccb66b | 2008 | if (blk_mq_sched_bio_merge(q, bio, nr_segs)) |
bd166ef1 JA |
2009 | return BLK_QC_T_NONE; |
2010 | ||
d5337560 | 2011 | rq_qos_throttle(q, bio); |
87760e5e | 2012 | |
7809167d | 2013 | data.cmd_flags = bio->bi_opf; |
f9afca4d | 2014 | rq = blk_mq_get_request(q, bio, &data); |
87760e5e | 2015 | if (unlikely(!rq)) { |
c1c80384 | 2016 | rq_qos_cleanup(q, bio); |
7b6620d7 | 2017 | if (bio->bi_opf & REQ_NOWAIT) |
03a07c92 | 2018 | bio_wouldblock_error(bio); |
7b6620d7 | 2019 | return BLK_QC_T_NONE; |
87760e5e JA |
2020 | } |
2021 | ||
d6f1dda2 XW |
2022 | trace_block_getrq(q, bio, bio->bi_opf); |
2023 | ||
c1c80384 | 2024 | rq_qos_track(q, rq, bio); |
07068d5b | 2025 | |
fd2d3326 | 2026 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b | 2027 | |
970d168d BVA |
2028 | blk_mq_bio_to_request(rq, bio, nr_segs); |
2029 | ||
b49773e7 | 2030 | plug = blk_mq_plug(q, bio); |
07068d5b | 2031 | if (unlikely(is_flush_fua)) { |
105663f7 | 2032 | /* Bypass scheduler for flush requests */ |
923218f6 ML |
2033 | blk_insert_flush(rq); |
2034 | blk_mq_run_hw_queue(data.hctx, true); | |
3154df26 ML |
2035 | } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs || |
2036 | !blk_queue_nonrot(q))) { | |
b2c5d16b JA |
2037 | /* |
2038 | * Use plugging if we have a ->commit_rqs() hook as well, as | |
2039 | * we know the driver uses bd->last in a smart fashion. | |
3154df26 ML |
2040 | * |
2041 | * Use normal plugging if this disk is slow HDD, as sequential | |
2042 | * IO may benefit a lot from plug merging. | |
b2c5d16b | 2043 | */ |
5f0ed774 | 2044 | unsigned int request_count = plug->rq_count; |
600271d9 SL |
2045 | struct request *last = NULL; |
2046 | ||
676d0607 | 2047 | if (!request_count) |
e6c4438b | 2048 | trace_block_plug(q); |
600271d9 SL |
2049 | else |
2050 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c | 2051 | |
600271d9 SL |
2052 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
2053 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
2054 | blk_flush_plug_list(plug, false); |
2055 | trace_block_plug(q); | |
320ae51f | 2056 | } |
b094f89c | 2057 | |
ce5b009c | 2058 | blk_add_rq_to_plug(plug, rq); |
a12de1d4 | 2059 | } else if (q->elevator) { |
105663f7 | 2060 | /* Insert the request at the IO scheduler queue */ |
a12de1d4 | 2061 | blk_mq_sched_insert_request(rq, false, true, true); |
2299722c | 2062 | } else if (plug && !blk_queue_nomerges(q)) { |
07068d5b | 2063 | /* |
6a83e74d | 2064 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
2065 | * Otherwise the existing request in the plug list will be |
2066 | * issued. So the plug list will have one request at most | |
2299722c CH |
2067 | * The plug list might get flushed before this. If that happens, |
2068 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 2069 | */ |
2299722c CH |
2070 | if (list_empty(&plug->mq_list)) |
2071 | same_queue_rq = NULL; | |
4711b573 | 2072 | if (same_queue_rq) { |
2299722c | 2073 | list_del_init(&same_queue_rq->queuelist); |
4711b573 JA |
2074 | plug->rq_count--; |
2075 | } | |
ce5b009c | 2076 | blk_add_rq_to_plug(plug, rq); |
ff3b74b8 | 2077 | trace_block_plug(q); |
2299722c | 2078 | |
dad7a3be | 2079 | if (same_queue_rq) { |
ea4f995e | 2080 | data.hctx = same_queue_rq->mq_hctx; |
ff3b74b8 | 2081 | trace_block_unplug(q, 1, true); |
2299722c | 2082 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, |
fd9c40f6 | 2083 | &cookie); |
dad7a3be | 2084 | } |
a12de1d4 ML |
2085 | } else if ((q->nr_hw_queues > 1 && is_sync) || |
2086 | !data.hctx->dispatch_busy) { | |
105663f7 AA |
2087 | /* |
2088 | * There is no scheduler and we can try to send directly | |
2089 | * to the hardware. | |
2090 | */ | |
fd9c40f6 | 2091 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); |
ab42f35d | 2092 | } else { |
105663f7 | 2093 | /* Default case. */ |
8fa9f556 | 2094 | blk_mq_sched_insert_request(rq, false, true, true); |
ab42f35d | 2095 | } |
320ae51f | 2096 | |
7b371636 | 2097 | return cookie; |
320ae51f | 2098 | } |
8cf7961d | 2099 | EXPORT_SYMBOL_GPL(blk_mq_make_request); /* only for request based dm */ |
320ae51f | 2100 | |
cc71a6f4 JA |
2101 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2102 | unsigned int hctx_idx) | |
95363efd | 2103 | { |
e9b267d9 | 2104 | struct page *page; |
320ae51f | 2105 | |
24d2f903 | 2106 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 2107 | int i; |
320ae51f | 2108 | |
24d2f903 | 2109 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
2110 | struct request *rq = tags->static_rqs[i]; |
2111 | ||
2112 | if (!rq) | |
e9b267d9 | 2113 | continue; |
d6296d39 | 2114 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 2115 | tags->static_rqs[i] = NULL; |
e9b267d9 | 2116 | } |
320ae51f | 2117 | } |
320ae51f | 2118 | |
24d2f903 CH |
2119 | while (!list_empty(&tags->page_list)) { |
2120 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 2121 | list_del_init(&page->lru); |
f75782e4 CM |
2122 | /* |
2123 | * Remove kmemleak object previously allocated in | |
273938bf | 2124 | * blk_mq_alloc_rqs(). |
f75782e4 CM |
2125 | */ |
2126 | kmemleak_free(page_address(page)); | |
320ae51f JA |
2127 | __free_pages(page, page->private); |
2128 | } | |
cc71a6f4 | 2129 | } |
320ae51f | 2130 | |
cc71a6f4 JA |
2131 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
2132 | { | |
24d2f903 | 2133 | kfree(tags->rqs); |
cc71a6f4 | 2134 | tags->rqs = NULL; |
2af8cbe3 JA |
2135 | kfree(tags->static_rqs); |
2136 | tags->static_rqs = NULL; | |
320ae51f | 2137 | |
24d2f903 | 2138 | blk_mq_free_tags(tags); |
320ae51f JA |
2139 | } |
2140 | ||
cc71a6f4 JA |
2141 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
2142 | unsigned int hctx_idx, | |
2143 | unsigned int nr_tags, | |
2144 | unsigned int reserved_tags) | |
320ae51f | 2145 | { |
24d2f903 | 2146 | struct blk_mq_tags *tags; |
59f082e4 | 2147 | int node; |
320ae51f | 2148 | |
7d76f856 | 2149 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx); |
59f082e4 SL |
2150 | if (node == NUMA_NO_NODE) |
2151 | node = set->numa_node; | |
2152 | ||
2153 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, | |
24391c0d | 2154 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
24d2f903 CH |
2155 | if (!tags) |
2156 | return NULL; | |
320ae51f | 2157 | |
590b5b7d | 2158 | tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
36e1f3d1 | 2159 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 2160 | node); |
24d2f903 CH |
2161 | if (!tags->rqs) { |
2162 | blk_mq_free_tags(tags); | |
2163 | return NULL; | |
2164 | } | |
320ae51f | 2165 | |
590b5b7d KC |
2166 | tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
2167 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
2168 | node); | |
2af8cbe3 JA |
2169 | if (!tags->static_rqs) { |
2170 | kfree(tags->rqs); | |
2171 | blk_mq_free_tags(tags); | |
2172 | return NULL; | |
2173 | } | |
2174 | ||
cc71a6f4 JA |
2175 | return tags; |
2176 | } | |
2177 | ||
2178 | static size_t order_to_size(unsigned int order) | |
2179 | { | |
2180 | return (size_t)PAGE_SIZE << order; | |
2181 | } | |
2182 | ||
1d9bd516 TH |
2183 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
2184 | unsigned int hctx_idx, int node) | |
2185 | { | |
2186 | int ret; | |
2187 | ||
2188 | if (set->ops->init_request) { | |
2189 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
2190 | if (ret) | |
2191 | return ret; | |
2192 | } | |
2193 | ||
12f5b931 | 2194 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
1d9bd516 TH |
2195 | return 0; |
2196 | } | |
2197 | ||
cc71a6f4 JA |
2198 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2199 | unsigned int hctx_idx, unsigned int depth) | |
2200 | { | |
2201 | unsigned int i, j, entries_per_page, max_order = 4; | |
2202 | size_t rq_size, left; | |
59f082e4 SL |
2203 | int node; |
2204 | ||
7d76f856 | 2205 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx); |
59f082e4 SL |
2206 | if (node == NUMA_NO_NODE) |
2207 | node = set->numa_node; | |
cc71a6f4 JA |
2208 | |
2209 | INIT_LIST_HEAD(&tags->page_list); | |
2210 | ||
320ae51f JA |
2211 | /* |
2212 | * rq_size is the size of the request plus driver payload, rounded | |
2213 | * to the cacheline size | |
2214 | */ | |
24d2f903 | 2215 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 2216 | cache_line_size()); |
cc71a6f4 | 2217 | left = rq_size * depth; |
320ae51f | 2218 | |
cc71a6f4 | 2219 | for (i = 0; i < depth; ) { |
320ae51f JA |
2220 | int this_order = max_order; |
2221 | struct page *page; | |
2222 | int to_do; | |
2223 | void *p; | |
2224 | ||
b3a834b1 | 2225 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
2226 | this_order--; |
2227 | ||
2228 | do { | |
59f082e4 | 2229 | page = alloc_pages_node(node, |
36e1f3d1 | 2230 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 2231 | this_order); |
320ae51f JA |
2232 | if (page) |
2233 | break; | |
2234 | if (!this_order--) | |
2235 | break; | |
2236 | if (order_to_size(this_order) < rq_size) | |
2237 | break; | |
2238 | } while (1); | |
2239 | ||
2240 | if (!page) | |
24d2f903 | 2241 | goto fail; |
320ae51f JA |
2242 | |
2243 | page->private = this_order; | |
24d2f903 | 2244 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
2245 | |
2246 | p = page_address(page); | |
f75782e4 CM |
2247 | /* |
2248 | * Allow kmemleak to scan these pages as they contain pointers | |
2249 | * to additional allocations like via ops->init_request(). | |
2250 | */ | |
36e1f3d1 | 2251 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 2252 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 2253 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
2254 | left -= to_do * rq_size; |
2255 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
2256 | struct request *rq = p; |
2257 | ||
2258 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
2259 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
2260 | tags->static_rqs[i] = NULL; | |
2261 | goto fail; | |
e9b267d9 CH |
2262 | } |
2263 | ||
320ae51f JA |
2264 | p += rq_size; |
2265 | i++; | |
2266 | } | |
2267 | } | |
cc71a6f4 | 2268 | return 0; |
320ae51f | 2269 | |
24d2f903 | 2270 | fail: |
cc71a6f4 JA |
2271 | blk_mq_free_rqs(set, tags, hctx_idx); |
2272 | return -ENOMEM; | |
320ae51f JA |
2273 | } |
2274 | ||
e57690fe JA |
2275 | /* |
2276 | * 'cpu' is going away. splice any existing rq_list entries from this | |
2277 | * software queue to the hw queue dispatch list, and ensure that it | |
2278 | * gets run. | |
2279 | */ | |
9467f859 | 2280 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 2281 | { |
9467f859 | 2282 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
2283 | struct blk_mq_ctx *ctx; |
2284 | LIST_HEAD(tmp); | |
c16d6b5a | 2285 | enum hctx_type type; |
484b4061 | 2286 | |
9467f859 | 2287 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 2288 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
c16d6b5a | 2289 | type = hctx->type; |
484b4061 JA |
2290 | |
2291 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
2292 | if (!list_empty(&ctx->rq_lists[type])) { |
2293 | list_splice_init(&ctx->rq_lists[type], &tmp); | |
484b4061 JA |
2294 | blk_mq_hctx_clear_pending(hctx, ctx); |
2295 | } | |
2296 | spin_unlock(&ctx->lock); | |
2297 | ||
2298 | if (list_empty(&tmp)) | |
9467f859 | 2299 | return 0; |
484b4061 | 2300 | |
e57690fe JA |
2301 | spin_lock(&hctx->lock); |
2302 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
2303 | spin_unlock(&hctx->lock); | |
484b4061 JA |
2304 | |
2305 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 2306 | return 0; |
484b4061 JA |
2307 | } |
2308 | ||
9467f859 | 2309 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 2310 | { |
9467f859 TG |
2311 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
2312 | &hctx->cpuhp_dead); | |
484b4061 JA |
2313 | } |
2314 | ||
c3b4afca | 2315 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
2316 | static void blk_mq_exit_hctx(struct request_queue *q, |
2317 | struct blk_mq_tag_set *set, | |
2318 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
2319 | { | |
8ab0b7dc ML |
2320 | if (blk_mq_hw_queue_mapped(hctx)) |
2321 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 2322 | |
f70ced09 | 2323 | if (set->ops->exit_request) |
d6296d39 | 2324 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); |
f70ced09 | 2325 | |
08e98fc6 ML |
2326 | if (set->ops->exit_hctx) |
2327 | set->ops->exit_hctx(hctx, hctx_idx); | |
2328 | ||
9467f859 | 2329 | blk_mq_remove_cpuhp(hctx); |
2f8f1336 ML |
2330 | |
2331 | spin_lock(&q->unused_hctx_lock); | |
2332 | list_add(&hctx->hctx_list, &q->unused_hctx_list); | |
2333 | spin_unlock(&q->unused_hctx_lock); | |
08e98fc6 ML |
2334 | } |
2335 | ||
624dbe47 ML |
2336 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
2337 | struct blk_mq_tag_set *set, int nr_queue) | |
2338 | { | |
2339 | struct blk_mq_hw_ctx *hctx; | |
2340 | unsigned int i; | |
2341 | ||
2342 | queue_for_each_hw_ctx(q, hctx, i) { | |
2343 | if (i == nr_queue) | |
2344 | break; | |
477e19de | 2345 | blk_mq_debugfs_unregister_hctx(hctx); |
08e98fc6 | 2346 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 2347 | } |
624dbe47 ML |
2348 | } |
2349 | ||
7c6c5b7c ML |
2350 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) |
2351 | { | |
2352 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2353 | ||
2354 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu), | |
2355 | __alignof__(struct blk_mq_hw_ctx)) != | |
2356 | sizeof(struct blk_mq_hw_ctx)); | |
2357 | ||
2358 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2359 | hw_ctx_size += sizeof(struct srcu_struct); | |
2360 | ||
2361 | return hw_ctx_size; | |
2362 | } | |
2363 | ||
08e98fc6 ML |
2364 | static int blk_mq_init_hctx(struct request_queue *q, |
2365 | struct blk_mq_tag_set *set, | |
2366 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 2367 | { |
7c6c5b7c ML |
2368 | hctx->queue_num = hctx_idx; |
2369 | ||
2370 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); | |
2371 | ||
2372 | hctx->tags = set->tags[hctx_idx]; | |
2373 | ||
2374 | if (set->ops->init_hctx && | |
2375 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
2376 | goto unregister_cpu_notifier; | |
08e98fc6 | 2377 | |
7c6c5b7c ML |
2378 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, |
2379 | hctx->numa_node)) | |
2380 | goto exit_hctx; | |
2381 | return 0; | |
2382 | ||
2383 | exit_hctx: | |
2384 | if (set->ops->exit_hctx) | |
2385 | set->ops->exit_hctx(hctx, hctx_idx); | |
2386 | unregister_cpu_notifier: | |
2387 | blk_mq_remove_cpuhp(hctx); | |
2388 | return -1; | |
2389 | } | |
2390 | ||
2391 | static struct blk_mq_hw_ctx * | |
2392 | blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set, | |
2393 | int node) | |
2394 | { | |
2395 | struct blk_mq_hw_ctx *hctx; | |
2396 | gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY; | |
2397 | ||
2398 | hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node); | |
2399 | if (!hctx) | |
2400 | goto fail_alloc_hctx; | |
2401 | ||
2402 | if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node)) | |
2403 | goto free_hctx; | |
2404 | ||
2405 | atomic_set(&hctx->nr_active, 0); | |
08e98fc6 | 2406 | if (node == NUMA_NO_NODE) |
7c6c5b7c ML |
2407 | node = set->numa_node; |
2408 | hctx->numa_node = node; | |
08e98fc6 | 2409 | |
9f993737 | 2410 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
2411 | spin_lock_init(&hctx->lock); |
2412 | INIT_LIST_HEAD(&hctx->dispatch); | |
2413 | hctx->queue = q; | |
2404e607 | 2414 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 2415 | |
2f8f1336 ML |
2416 | INIT_LIST_HEAD(&hctx->hctx_list); |
2417 | ||
320ae51f | 2418 | /* |
08e98fc6 ML |
2419 | * Allocate space for all possible cpus to avoid allocation at |
2420 | * runtime | |
320ae51f | 2421 | */ |
d904bfa7 | 2422 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
7c6c5b7c | 2423 | gfp, node); |
08e98fc6 | 2424 | if (!hctx->ctxs) |
7c6c5b7c | 2425 | goto free_cpumask; |
320ae51f | 2426 | |
5b202853 | 2427 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), |
7c6c5b7c | 2428 | gfp, node)) |
08e98fc6 | 2429 | goto free_ctxs; |
08e98fc6 | 2430 | hctx->nr_ctx = 0; |
320ae51f | 2431 | |
5815839b | 2432 | spin_lock_init(&hctx->dispatch_wait_lock); |
eb619fdb JA |
2433 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
2434 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
2435 | ||
754a1572 | 2436 | hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp); |
f70ced09 | 2437 | if (!hctx->fq) |
7c6c5b7c | 2438 | goto free_bitmap; |
320ae51f | 2439 | |
6a83e74d | 2440 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2441 | init_srcu_struct(hctx->srcu); |
7c6c5b7c | 2442 | blk_mq_hctx_kobj_init(hctx); |
6a83e74d | 2443 | |
7c6c5b7c | 2444 | return hctx; |
320ae51f | 2445 | |
08e98fc6 | 2446 | free_bitmap: |
88459642 | 2447 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2448 | free_ctxs: |
2449 | kfree(hctx->ctxs); | |
7c6c5b7c ML |
2450 | free_cpumask: |
2451 | free_cpumask_var(hctx->cpumask); | |
2452 | free_hctx: | |
2453 | kfree(hctx); | |
2454 | fail_alloc_hctx: | |
2455 | return NULL; | |
08e98fc6 | 2456 | } |
320ae51f | 2457 | |
320ae51f JA |
2458 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
2459 | unsigned int nr_hw_queues) | |
2460 | { | |
b3c661b1 JA |
2461 | struct blk_mq_tag_set *set = q->tag_set; |
2462 | unsigned int i, j; | |
320ae51f JA |
2463 | |
2464 | for_each_possible_cpu(i) { | |
2465 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2466 | struct blk_mq_hw_ctx *hctx; | |
c16d6b5a | 2467 | int k; |
320ae51f | 2468 | |
320ae51f JA |
2469 | __ctx->cpu = i; |
2470 | spin_lock_init(&__ctx->lock); | |
c16d6b5a ML |
2471 | for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++) |
2472 | INIT_LIST_HEAD(&__ctx->rq_lists[k]); | |
2473 | ||
320ae51f JA |
2474 | __ctx->queue = q; |
2475 | ||
320ae51f JA |
2476 | /* |
2477 | * Set local node, IFF we have more than one hw queue. If | |
2478 | * not, we remain on the home node of the device | |
2479 | */ | |
b3c661b1 JA |
2480 | for (j = 0; j < set->nr_maps; j++) { |
2481 | hctx = blk_mq_map_queue_type(q, j, i); | |
2482 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
2483 | hctx->numa_node = local_memory_node(cpu_to_node(i)); | |
2484 | } | |
320ae51f JA |
2485 | } |
2486 | } | |
2487 | ||
cc71a6f4 JA |
2488 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
2489 | { | |
2490 | int ret = 0; | |
2491 | ||
2492 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
2493 | set->queue_depth, set->reserved_tags); | |
2494 | if (!set->tags[hctx_idx]) | |
2495 | return false; | |
2496 | ||
2497 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2498 | set->queue_depth); | |
2499 | if (!ret) | |
2500 | return true; | |
2501 | ||
2502 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2503 | set->tags[hctx_idx] = NULL; | |
2504 | return false; | |
2505 | } | |
2506 | ||
2507 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2508 | unsigned int hctx_idx) | |
2509 | { | |
4e6db0f2 | 2510 | if (set->tags && set->tags[hctx_idx]) { |
bd166ef1 JA |
2511 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); |
2512 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2513 | set->tags[hctx_idx] = NULL; | |
2514 | } | |
cc71a6f4 JA |
2515 | } |
2516 | ||
4b855ad3 | 2517 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 2518 | { |
b3c661b1 | 2519 | unsigned int i, j, hctx_idx; |
320ae51f JA |
2520 | struct blk_mq_hw_ctx *hctx; |
2521 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2522 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f JA |
2523 | |
2524 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 2525 | cpumask_clear(hctx->cpumask); |
320ae51f | 2526 | hctx->nr_ctx = 0; |
d416c92c | 2527 | hctx->dispatch_from = NULL; |
320ae51f JA |
2528 | } |
2529 | ||
2530 | /* | |
4b855ad3 | 2531 | * Map software to hardware queues. |
4412efec ML |
2532 | * |
2533 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 2534 | */ |
20e4d813 | 2535 | for_each_possible_cpu(i) { |
7d76f856 | 2536 | hctx_idx = set->map[HCTX_TYPE_DEFAULT].mq_map[i]; |
4412efec ML |
2537 | /* unmapped hw queue can be remapped after CPU topo changed */ |
2538 | if (!set->tags[hctx_idx] && | |
2539 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
2540 | /* | |
2541 | * If tags initialization fail for some hctx, | |
2542 | * that hctx won't be brought online. In this | |
2543 | * case, remap the current ctx to hctx[0] which | |
2544 | * is guaranteed to always have tags allocated | |
2545 | */ | |
7d76f856 | 2546 | set->map[HCTX_TYPE_DEFAULT].mq_map[i] = 0; |
4412efec ML |
2547 | } |
2548 | ||
897bb0c7 | 2549 | ctx = per_cpu_ptr(q->queue_ctx, i); |
b3c661b1 | 2550 | for (j = 0; j < set->nr_maps; j++) { |
bb94aea1 JW |
2551 | if (!set->map[j].nr_queues) { |
2552 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
2553 | HCTX_TYPE_DEFAULT, i); | |
e5edd5f2 | 2554 | continue; |
bb94aea1 | 2555 | } |
e5edd5f2 | 2556 | |
b3c661b1 | 2557 | hctx = blk_mq_map_queue_type(q, j, i); |
8ccdf4a3 | 2558 | ctx->hctxs[j] = hctx; |
b3c661b1 JA |
2559 | /* |
2560 | * If the CPU is already set in the mask, then we've | |
2561 | * mapped this one already. This can happen if | |
2562 | * devices share queues across queue maps. | |
2563 | */ | |
2564 | if (cpumask_test_cpu(i, hctx->cpumask)) | |
2565 | continue; | |
2566 | ||
2567 | cpumask_set_cpu(i, hctx->cpumask); | |
2568 | hctx->type = j; | |
2569 | ctx->index_hw[hctx->type] = hctx->nr_ctx; | |
2570 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2571 | ||
2572 | /* | |
2573 | * If the nr_ctx type overflows, we have exceeded the | |
2574 | * amount of sw queues we can support. | |
2575 | */ | |
2576 | BUG_ON(!hctx->nr_ctx); | |
2577 | } | |
bb94aea1 JW |
2578 | |
2579 | for (; j < HCTX_MAX_TYPES; j++) | |
2580 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
2581 | HCTX_TYPE_DEFAULT, i); | |
320ae51f | 2582 | } |
506e931f JA |
2583 | |
2584 | queue_for_each_hw_ctx(q, hctx, i) { | |
4412efec ML |
2585 | /* |
2586 | * If no software queues are mapped to this hardware queue, | |
2587 | * disable it and free the request entries. | |
2588 | */ | |
2589 | if (!hctx->nr_ctx) { | |
2590 | /* Never unmap queue 0. We need it as a | |
2591 | * fallback in case of a new remap fails | |
2592 | * allocation | |
2593 | */ | |
2594 | if (i && set->tags[i]) | |
2595 | blk_mq_free_map_and_requests(set, i); | |
2596 | ||
2597 | hctx->tags = NULL; | |
2598 | continue; | |
2599 | } | |
484b4061 | 2600 | |
2a34c087 ML |
2601 | hctx->tags = set->tags[i]; |
2602 | WARN_ON(!hctx->tags); | |
2603 | ||
889fa31f CY |
2604 | /* |
2605 | * Set the map size to the number of mapped software queues. | |
2606 | * This is more accurate and more efficient than looping | |
2607 | * over all possibly mapped software queues. | |
2608 | */ | |
88459642 | 2609 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2610 | |
484b4061 JA |
2611 | /* |
2612 | * Initialize batch roundrobin counts | |
2613 | */ | |
f82ddf19 | 2614 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
2615 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2616 | } | |
320ae51f JA |
2617 | } |
2618 | ||
8e8320c9 JA |
2619 | /* |
2620 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
2621 | * the queue isn't live yet. | |
2622 | */ | |
2404e607 | 2623 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2624 | { |
2625 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2626 | int i; |
2627 | ||
2404e607 | 2628 | queue_for_each_hw_ctx(q, hctx, i) { |
97889f9a | 2629 | if (shared) |
2404e607 | 2630 | hctx->flags |= BLK_MQ_F_TAG_SHARED; |
97889f9a | 2631 | else |
2404e607 JM |
2632 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; |
2633 | } | |
2634 | } | |
2635 | ||
8e8320c9 JA |
2636 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, |
2637 | bool shared) | |
2404e607 JM |
2638 | { |
2639 | struct request_queue *q; | |
0d2602ca | 2640 | |
705cda97 BVA |
2641 | lockdep_assert_held(&set->tag_list_lock); |
2642 | ||
0d2602ca JA |
2643 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2644 | blk_mq_freeze_queue(q); | |
2404e607 | 2645 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2646 | blk_mq_unfreeze_queue(q); |
2647 | } | |
2648 | } | |
2649 | ||
2650 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2651 | { | |
2652 | struct blk_mq_tag_set *set = q->tag_set; | |
2653 | ||
0d2602ca | 2654 | mutex_lock(&set->tag_list_lock); |
705cda97 | 2655 | list_del_rcu(&q->tag_set_list); |
2404e607 JM |
2656 | if (list_is_singular(&set->tag_list)) { |
2657 | /* just transitioned to unshared */ | |
2658 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2659 | /* update existing queue */ | |
2660 | blk_mq_update_tag_set_depth(set, false); | |
2661 | } | |
0d2602ca | 2662 | mutex_unlock(&set->tag_list_lock); |
a347c7ad | 2663 | INIT_LIST_HEAD(&q->tag_set_list); |
0d2602ca JA |
2664 | } |
2665 | ||
2666 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2667 | struct request_queue *q) | |
2668 | { | |
0d2602ca | 2669 | mutex_lock(&set->tag_list_lock); |
2404e607 | 2670 | |
ff821d27 JA |
2671 | /* |
2672 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
2673 | */ | |
2674 | if (!list_empty(&set->tag_list) && | |
2675 | !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2404e607 JM |
2676 | set->flags |= BLK_MQ_F_TAG_SHARED; |
2677 | /* update existing queue */ | |
2678 | blk_mq_update_tag_set_depth(set, true); | |
2679 | } | |
2680 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2681 | queue_set_hctx_shared(q, true); | |
705cda97 | 2682 | list_add_tail_rcu(&q->tag_set_list, &set->tag_list); |
2404e607 | 2683 | |
0d2602ca JA |
2684 | mutex_unlock(&set->tag_list_lock); |
2685 | } | |
2686 | ||
1db4909e ML |
2687 | /* All allocations will be freed in release handler of q->mq_kobj */ |
2688 | static int blk_mq_alloc_ctxs(struct request_queue *q) | |
2689 | { | |
2690 | struct blk_mq_ctxs *ctxs; | |
2691 | int cpu; | |
2692 | ||
2693 | ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL); | |
2694 | if (!ctxs) | |
2695 | return -ENOMEM; | |
2696 | ||
2697 | ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx); | |
2698 | if (!ctxs->queue_ctx) | |
2699 | goto fail; | |
2700 | ||
2701 | for_each_possible_cpu(cpu) { | |
2702 | struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu); | |
2703 | ctx->ctxs = ctxs; | |
2704 | } | |
2705 | ||
2706 | q->mq_kobj = &ctxs->kobj; | |
2707 | q->queue_ctx = ctxs->queue_ctx; | |
2708 | ||
2709 | return 0; | |
2710 | fail: | |
2711 | kfree(ctxs); | |
2712 | return -ENOMEM; | |
2713 | } | |
2714 | ||
e09aae7e ML |
2715 | /* |
2716 | * It is the actual release handler for mq, but we do it from | |
2717 | * request queue's release handler for avoiding use-after-free | |
2718 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2719 | * but we can't group ctx/kctx kobj without it. | |
2720 | */ | |
2721 | void blk_mq_release(struct request_queue *q) | |
2722 | { | |
2f8f1336 ML |
2723 | struct blk_mq_hw_ctx *hctx, *next; |
2724 | int i; | |
e09aae7e | 2725 | |
2f8f1336 ML |
2726 | queue_for_each_hw_ctx(q, hctx, i) |
2727 | WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list)); | |
2728 | ||
2729 | /* all hctx are in .unused_hctx_list now */ | |
2730 | list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) { | |
2731 | list_del_init(&hctx->hctx_list); | |
6c8b232e | 2732 | kobject_put(&hctx->kobj); |
c3b4afca | 2733 | } |
e09aae7e ML |
2734 | |
2735 | kfree(q->queue_hw_ctx); | |
2736 | ||
7ea5fe31 ML |
2737 | /* |
2738 | * release .mq_kobj and sw queue's kobject now because | |
2739 | * both share lifetime with request queue. | |
2740 | */ | |
2741 | blk_mq_sysfs_deinit(q); | |
e09aae7e ML |
2742 | } |
2743 | ||
2f227bb9 CH |
2744 | struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set, |
2745 | void *queuedata) | |
b62c21b7 MS |
2746 | { |
2747 | struct request_queue *uninit_q, *q; | |
2748 | ||
3d745ea5 | 2749 | uninit_q = __blk_alloc_queue(set->numa_node); |
b62c21b7 MS |
2750 | if (!uninit_q) |
2751 | return ERR_PTR(-ENOMEM); | |
2f227bb9 | 2752 | uninit_q->queuedata = queuedata; |
b62c21b7 | 2753 | |
737eb78e DLM |
2754 | /* |
2755 | * Initialize the queue without an elevator. device_add_disk() will do | |
2756 | * the initialization. | |
2757 | */ | |
2758 | q = blk_mq_init_allocated_queue(set, uninit_q, false); | |
b62c21b7 MS |
2759 | if (IS_ERR(q)) |
2760 | blk_cleanup_queue(uninit_q); | |
2761 | ||
2762 | return q; | |
2763 | } | |
2f227bb9 CH |
2764 | EXPORT_SYMBOL_GPL(blk_mq_init_queue_data); |
2765 | ||
2766 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) | |
2767 | { | |
2768 | return blk_mq_init_queue_data(set, NULL); | |
2769 | } | |
b62c21b7 MS |
2770 | EXPORT_SYMBOL(blk_mq_init_queue); |
2771 | ||
9316a9ed JA |
2772 | /* |
2773 | * Helper for setting up a queue with mq ops, given queue depth, and | |
2774 | * the passed in mq ops flags. | |
2775 | */ | |
2776 | struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set, | |
2777 | const struct blk_mq_ops *ops, | |
2778 | unsigned int queue_depth, | |
2779 | unsigned int set_flags) | |
2780 | { | |
2781 | struct request_queue *q; | |
2782 | int ret; | |
2783 | ||
2784 | memset(set, 0, sizeof(*set)); | |
2785 | set->ops = ops; | |
2786 | set->nr_hw_queues = 1; | |
b3c661b1 | 2787 | set->nr_maps = 1; |
9316a9ed JA |
2788 | set->queue_depth = queue_depth; |
2789 | set->numa_node = NUMA_NO_NODE; | |
2790 | set->flags = set_flags; | |
2791 | ||
2792 | ret = blk_mq_alloc_tag_set(set); | |
2793 | if (ret) | |
2794 | return ERR_PTR(ret); | |
2795 | ||
2796 | q = blk_mq_init_queue(set); | |
2797 | if (IS_ERR(q)) { | |
2798 | blk_mq_free_tag_set(set); | |
2799 | return q; | |
2800 | } | |
2801 | ||
2802 | return q; | |
2803 | } | |
2804 | EXPORT_SYMBOL(blk_mq_init_sq_queue); | |
2805 | ||
34d11ffa JW |
2806 | static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx( |
2807 | struct blk_mq_tag_set *set, struct request_queue *q, | |
2808 | int hctx_idx, int node) | |
2809 | { | |
2f8f1336 | 2810 | struct blk_mq_hw_ctx *hctx = NULL, *tmp; |
34d11ffa | 2811 | |
2f8f1336 ML |
2812 | /* reuse dead hctx first */ |
2813 | spin_lock(&q->unused_hctx_lock); | |
2814 | list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) { | |
2815 | if (tmp->numa_node == node) { | |
2816 | hctx = tmp; | |
2817 | break; | |
2818 | } | |
2819 | } | |
2820 | if (hctx) | |
2821 | list_del_init(&hctx->hctx_list); | |
2822 | spin_unlock(&q->unused_hctx_lock); | |
2823 | ||
2824 | if (!hctx) | |
2825 | hctx = blk_mq_alloc_hctx(q, set, node); | |
34d11ffa | 2826 | if (!hctx) |
7c6c5b7c | 2827 | goto fail; |
34d11ffa | 2828 | |
7c6c5b7c ML |
2829 | if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) |
2830 | goto free_hctx; | |
34d11ffa JW |
2831 | |
2832 | return hctx; | |
7c6c5b7c ML |
2833 | |
2834 | free_hctx: | |
2835 | kobject_put(&hctx->kobj); | |
2836 | fail: | |
2837 | return NULL; | |
34d11ffa JW |
2838 | } |
2839 | ||
868f2f0b KB |
2840 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2841 | struct request_queue *q) | |
320ae51f | 2842 | { |
e01ad46d | 2843 | int i, j, end; |
868f2f0b | 2844 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; |
f14bbe77 | 2845 | |
ac0d6b92 BVA |
2846 | if (q->nr_hw_queues < set->nr_hw_queues) { |
2847 | struct blk_mq_hw_ctx **new_hctxs; | |
2848 | ||
2849 | new_hctxs = kcalloc_node(set->nr_hw_queues, | |
2850 | sizeof(*new_hctxs), GFP_KERNEL, | |
2851 | set->numa_node); | |
2852 | if (!new_hctxs) | |
2853 | return; | |
2854 | if (hctxs) | |
2855 | memcpy(new_hctxs, hctxs, q->nr_hw_queues * | |
2856 | sizeof(*hctxs)); | |
2857 | q->queue_hw_ctx = new_hctxs; | |
ac0d6b92 BVA |
2858 | kfree(hctxs); |
2859 | hctxs = new_hctxs; | |
2860 | } | |
2861 | ||
fb350e0a ML |
2862 | /* protect against switching io scheduler */ |
2863 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 2864 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2865 | int node; |
34d11ffa | 2866 | struct blk_mq_hw_ctx *hctx; |
868f2f0b | 2867 | |
7d76f856 | 2868 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i); |
34d11ffa JW |
2869 | /* |
2870 | * If the hw queue has been mapped to another numa node, | |
2871 | * we need to realloc the hctx. If allocation fails, fallback | |
2872 | * to use the previous one. | |
2873 | */ | |
2874 | if (hctxs[i] && (hctxs[i]->numa_node == node)) | |
2875 | continue; | |
868f2f0b | 2876 | |
34d11ffa JW |
2877 | hctx = blk_mq_alloc_and_init_hctx(set, q, i, node); |
2878 | if (hctx) { | |
2f8f1336 | 2879 | if (hctxs[i]) |
34d11ffa | 2880 | blk_mq_exit_hctx(q, set, hctxs[i], i); |
34d11ffa JW |
2881 | hctxs[i] = hctx; |
2882 | } else { | |
2883 | if (hctxs[i]) | |
2884 | pr_warn("Allocate new hctx on node %d fails,\ | |
2885 | fallback to previous one on node %d\n", | |
2886 | node, hctxs[i]->numa_node); | |
2887 | else | |
2888 | break; | |
868f2f0b | 2889 | } |
320ae51f | 2890 | } |
e01ad46d JW |
2891 | /* |
2892 | * Increasing nr_hw_queues fails. Free the newly allocated | |
2893 | * hctxs and keep the previous q->nr_hw_queues. | |
2894 | */ | |
2895 | if (i != set->nr_hw_queues) { | |
2896 | j = q->nr_hw_queues; | |
2897 | end = i; | |
2898 | } else { | |
2899 | j = i; | |
2900 | end = q->nr_hw_queues; | |
2901 | q->nr_hw_queues = set->nr_hw_queues; | |
2902 | } | |
34d11ffa | 2903 | |
e01ad46d | 2904 | for (; j < end; j++) { |
868f2f0b KB |
2905 | struct blk_mq_hw_ctx *hctx = hctxs[j]; |
2906 | ||
2907 | if (hctx) { | |
cc71a6f4 JA |
2908 | if (hctx->tags) |
2909 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 2910 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 2911 | hctxs[j] = NULL; |
868f2f0b KB |
2912 | } |
2913 | } | |
fb350e0a | 2914 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
2915 | } |
2916 | ||
2917 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
737eb78e DLM |
2918 | struct request_queue *q, |
2919 | bool elevator_init) | |
868f2f0b | 2920 | { |
66841672 ML |
2921 | /* mark the queue as mq asap */ |
2922 | q->mq_ops = set->ops; | |
2923 | ||
34dbad5d | 2924 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
2925 | blk_mq_poll_stats_bkt, |
2926 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
2927 | if (!q->poll_cb) |
2928 | goto err_exit; | |
2929 | ||
1db4909e | 2930 | if (blk_mq_alloc_ctxs(q)) |
41de54c6 | 2931 | goto err_poll; |
868f2f0b | 2932 | |
737f98cf ML |
2933 | /* init q->mq_kobj and sw queues' kobjects */ |
2934 | blk_mq_sysfs_init(q); | |
2935 | ||
2f8f1336 ML |
2936 | INIT_LIST_HEAD(&q->unused_hctx_list); |
2937 | spin_lock_init(&q->unused_hctx_lock); | |
2938 | ||
868f2f0b KB |
2939 | blk_mq_realloc_hw_ctxs(set, q); |
2940 | if (!q->nr_hw_queues) | |
2941 | goto err_hctxs; | |
320ae51f | 2942 | |
287922eb | 2943 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2944 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f | 2945 | |
a8908939 | 2946 | q->tag_set = set; |
320ae51f | 2947 | |
94eddfbe | 2948 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
cd19181b ML |
2949 | if (set->nr_maps > HCTX_TYPE_POLL && |
2950 | set->map[HCTX_TYPE_POLL].nr_queues) | |
6544d229 | 2951 | blk_queue_flag_set(QUEUE_FLAG_POLL, q); |
320ae51f | 2952 | |
1be036e9 CH |
2953 | q->sg_reserved_size = INT_MAX; |
2954 | ||
2849450a | 2955 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2956 | INIT_LIST_HEAD(&q->requeue_list); |
2957 | spin_lock_init(&q->requeue_lock); | |
2958 | ||
eba71768 JA |
2959 | q->nr_requests = set->queue_depth; |
2960 | ||
64f1c21e JA |
2961 | /* |
2962 | * Default to classic polling | |
2963 | */ | |
29ece8b4 | 2964 | q->poll_nsec = BLK_MQ_POLL_CLASSIC; |
64f1c21e | 2965 | |
24d2f903 | 2966 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 2967 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 2968 | blk_mq_map_swqueue(q); |
4593fdbe | 2969 | |
737eb78e DLM |
2970 | if (elevator_init) |
2971 | elevator_init_mq(q); | |
d3484991 | 2972 | |
320ae51f | 2973 | return q; |
18741986 | 2974 | |
320ae51f | 2975 | err_hctxs: |
868f2f0b | 2976 | kfree(q->queue_hw_ctx); |
73d9c8d4 | 2977 | q->nr_hw_queues = 0; |
1db4909e | 2978 | blk_mq_sysfs_deinit(q); |
41de54c6 JS |
2979 | err_poll: |
2980 | blk_stat_free_callback(q->poll_cb); | |
2981 | q->poll_cb = NULL; | |
c7de5726 ML |
2982 | err_exit: |
2983 | q->mq_ops = NULL; | |
320ae51f JA |
2984 | return ERR_PTR(-ENOMEM); |
2985 | } | |
b62c21b7 | 2986 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f | 2987 | |
c7e2d94b ML |
2988 | /* tags can _not_ be used after returning from blk_mq_exit_queue */ |
2989 | void blk_mq_exit_queue(struct request_queue *q) | |
320ae51f | 2990 | { |
624dbe47 | 2991 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2992 | |
0d2602ca | 2993 | blk_mq_del_queue_tag_set(q); |
624dbe47 | 2994 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
320ae51f | 2995 | } |
320ae51f | 2996 | |
a5164405 JA |
2997 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2998 | { | |
2999 | int i; | |
3000 | ||
cc71a6f4 JA |
3001 | for (i = 0; i < set->nr_hw_queues; i++) |
3002 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 3003 | goto out_unwind; |
a5164405 JA |
3004 | |
3005 | return 0; | |
3006 | ||
3007 | out_unwind: | |
3008 | while (--i >= 0) | |
cc71a6f4 | 3009 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 3010 | |
a5164405 JA |
3011 | return -ENOMEM; |
3012 | } | |
3013 | ||
3014 | /* | |
3015 | * Allocate the request maps associated with this tag_set. Note that this | |
3016 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
3017 | * will be updated to reflect the allocated depth. | |
3018 | */ | |
3019 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
3020 | { | |
3021 | unsigned int depth; | |
3022 | int err; | |
3023 | ||
3024 | depth = set->queue_depth; | |
3025 | do { | |
3026 | err = __blk_mq_alloc_rq_maps(set); | |
3027 | if (!err) | |
3028 | break; | |
3029 | ||
3030 | set->queue_depth >>= 1; | |
3031 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
3032 | err = -ENOMEM; | |
3033 | break; | |
3034 | } | |
3035 | } while (set->queue_depth); | |
3036 | ||
3037 | if (!set->queue_depth || err) { | |
3038 | pr_err("blk-mq: failed to allocate request map\n"); | |
3039 | return -ENOMEM; | |
3040 | } | |
3041 | ||
3042 | if (depth != set->queue_depth) | |
3043 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
3044 | depth, set->queue_depth); | |
3045 | ||
3046 | return 0; | |
3047 | } | |
3048 | ||
ebe8bddb OS |
3049 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
3050 | { | |
6e66b493 BVA |
3051 | /* |
3052 | * blk_mq_map_queues() and multiple .map_queues() implementations | |
3053 | * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the | |
3054 | * number of hardware queues. | |
3055 | */ | |
3056 | if (set->nr_maps == 1) | |
3057 | set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues; | |
3058 | ||
59388702 | 3059 | if (set->ops->map_queues && !is_kdump_kernel()) { |
b3c661b1 JA |
3060 | int i; |
3061 | ||
7d4901a9 ML |
3062 | /* |
3063 | * transport .map_queues is usually done in the following | |
3064 | * way: | |
3065 | * | |
3066 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
3067 | * mask = get_cpu_mask(queue) | |
3068 | * for_each_cpu(cpu, mask) | |
b3c661b1 | 3069 | * set->map[x].mq_map[cpu] = queue; |
7d4901a9 ML |
3070 | * } |
3071 | * | |
3072 | * When we need to remap, the table has to be cleared for | |
3073 | * killing stale mapping since one CPU may not be mapped | |
3074 | * to any hw queue. | |
3075 | */ | |
b3c661b1 JA |
3076 | for (i = 0; i < set->nr_maps; i++) |
3077 | blk_mq_clear_mq_map(&set->map[i]); | |
7d4901a9 | 3078 | |
ebe8bddb | 3079 | return set->ops->map_queues(set); |
b3c661b1 JA |
3080 | } else { |
3081 | BUG_ON(set->nr_maps > 1); | |
7d76f856 | 3082 | return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
b3c661b1 | 3083 | } |
ebe8bddb OS |
3084 | } |
3085 | ||
f7e76dbc BVA |
3086 | static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set, |
3087 | int cur_nr_hw_queues, int new_nr_hw_queues) | |
3088 | { | |
3089 | struct blk_mq_tags **new_tags; | |
3090 | ||
3091 | if (cur_nr_hw_queues >= new_nr_hw_queues) | |
3092 | return 0; | |
3093 | ||
3094 | new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *), | |
3095 | GFP_KERNEL, set->numa_node); | |
3096 | if (!new_tags) | |
3097 | return -ENOMEM; | |
3098 | ||
3099 | if (set->tags) | |
3100 | memcpy(new_tags, set->tags, cur_nr_hw_queues * | |
3101 | sizeof(*set->tags)); | |
3102 | kfree(set->tags); | |
3103 | set->tags = new_tags; | |
3104 | set->nr_hw_queues = new_nr_hw_queues; | |
3105 | ||
3106 | return 0; | |
3107 | } | |
3108 | ||
a4391c64 JA |
3109 | /* |
3110 | * Alloc a tag set to be associated with one or more request queues. | |
3111 | * May fail with EINVAL for various error conditions. May adjust the | |
c018c84f | 3112 | * requested depth down, if it's too large. In that case, the set |
a4391c64 JA |
3113 | * value will be stored in set->queue_depth. |
3114 | */ | |
24d2f903 CH |
3115 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
3116 | { | |
b3c661b1 | 3117 | int i, ret; |
da695ba2 | 3118 | |
205fb5f5 BVA |
3119 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
3120 | ||
24d2f903 CH |
3121 | if (!set->nr_hw_queues) |
3122 | return -EINVAL; | |
a4391c64 | 3123 | if (!set->queue_depth) |
24d2f903 CH |
3124 | return -EINVAL; |
3125 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
3126 | return -EINVAL; | |
3127 | ||
7d7e0f90 | 3128 | if (!set->ops->queue_rq) |
24d2f903 CH |
3129 | return -EINVAL; |
3130 | ||
de148297 ML |
3131 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
3132 | return -EINVAL; | |
3133 | ||
a4391c64 JA |
3134 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
3135 | pr_info("blk-mq: reduced tag depth to %u\n", | |
3136 | BLK_MQ_MAX_DEPTH); | |
3137 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
3138 | } | |
24d2f903 | 3139 | |
b3c661b1 JA |
3140 | if (!set->nr_maps) |
3141 | set->nr_maps = 1; | |
3142 | else if (set->nr_maps > HCTX_MAX_TYPES) | |
3143 | return -EINVAL; | |
3144 | ||
6637fadf SL |
3145 | /* |
3146 | * If a crashdump is active, then we are potentially in a very | |
3147 | * memory constrained environment. Limit us to 1 queue and | |
3148 | * 64 tags to prevent using too much memory. | |
3149 | */ | |
3150 | if (is_kdump_kernel()) { | |
3151 | set->nr_hw_queues = 1; | |
59388702 | 3152 | set->nr_maps = 1; |
6637fadf SL |
3153 | set->queue_depth = min(64U, set->queue_depth); |
3154 | } | |
868f2f0b | 3155 | /* |
392546ae JA |
3156 | * There is no use for more h/w queues than cpus if we just have |
3157 | * a single map | |
868f2f0b | 3158 | */ |
392546ae | 3159 | if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids) |
868f2f0b | 3160 | set->nr_hw_queues = nr_cpu_ids; |
6637fadf | 3161 | |
f7e76dbc | 3162 | if (blk_mq_realloc_tag_set_tags(set, 0, set->nr_hw_queues) < 0) |
a5164405 | 3163 | return -ENOMEM; |
24d2f903 | 3164 | |
da695ba2 | 3165 | ret = -ENOMEM; |
b3c661b1 JA |
3166 | for (i = 0; i < set->nr_maps; i++) { |
3167 | set->map[i].mq_map = kcalloc_node(nr_cpu_ids, | |
07b35eb5 | 3168 | sizeof(set->map[i].mq_map[0]), |
b3c661b1 JA |
3169 | GFP_KERNEL, set->numa_node); |
3170 | if (!set->map[i].mq_map) | |
3171 | goto out_free_mq_map; | |
59388702 | 3172 | set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues; |
b3c661b1 | 3173 | } |
bdd17e75 | 3174 | |
ebe8bddb | 3175 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
3176 | if (ret) |
3177 | goto out_free_mq_map; | |
3178 | ||
3179 | ret = blk_mq_alloc_rq_maps(set); | |
3180 | if (ret) | |
bdd17e75 | 3181 | goto out_free_mq_map; |
24d2f903 | 3182 | |
0d2602ca JA |
3183 | mutex_init(&set->tag_list_lock); |
3184 | INIT_LIST_HEAD(&set->tag_list); | |
3185 | ||
24d2f903 | 3186 | return 0; |
bdd17e75 CH |
3187 | |
3188 | out_free_mq_map: | |
b3c661b1 JA |
3189 | for (i = 0; i < set->nr_maps; i++) { |
3190 | kfree(set->map[i].mq_map); | |
3191 | set->map[i].mq_map = NULL; | |
3192 | } | |
5676e7b6 RE |
3193 | kfree(set->tags); |
3194 | set->tags = NULL; | |
da695ba2 | 3195 | return ret; |
24d2f903 CH |
3196 | } |
3197 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
3198 | ||
3199 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
3200 | { | |
b3c661b1 | 3201 | int i, j; |
24d2f903 | 3202 | |
f7e76dbc | 3203 | for (i = 0; i < set->nr_hw_queues; i++) |
cc71a6f4 | 3204 | blk_mq_free_map_and_requests(set, i); |
484b4061 | 3205 | |
b3c661b1 JA |
3206 | for (j = 0; j < set->nr_maps; j++) { |
3207 | kfree(set->map[j].mq_map); | |
3208 | set->map[j].mq_map = NULL; | |
3209 | } | |
bdd17e75 | 3210 | |
981bd189 | 3211 | kfree(set->tags); |
5676e7b6 | 3212 | set->tags = NULL; |
24d2f903 CH |
3213 | } |
3214 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
3215 | ||
e3a2b3f9 JA |
3216 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
3217 | { | |
3218 | struct blk_mq_tag_set *set = q->tag_set; | |
3219 | struct blk_mq_hw_ctx *hctx; | |
3220 | int i, ret; | |
3221 | ||
bd166ef1 | 3222 | if (!set) |
e3a2b3f9 JA |
3223 | return -EINVAL; |
3224 | ||
e5fa8140 AZ |
3225 | if (q->nr_requests == nr) |
3226 | return 0; | |
3227 | ||
70f36b60 | 3228 | blk_mq_freeze_queue(q); |
24f5a90f | 3229 | blk_mq_quiesce_queue(q); |
70f36b60 | 3230 | |
e3a2b3f9 JA |
3231 | ret = 0; |
3232 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
3233 | if (!hctx->tags) |
3234 | continue; | |
bd166ef1 JA |
3235 | /* |
3236 | * If we're using an MQ scheduler, just update the scheduler | |
3237 | * queue depth. This is similar to what the old code would do. | |
3238 | */ | |
70f36b60 | 3239 | if (!hctx->sched_tags) { |
c2e82a23 | 3240 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, |
70f36b60 JA |
3241 | false); |
3242 | } else { | |
3243 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
3244 | nr, true); | |
3245 | } | |
e3a2b3f9 JA |
3246 | if (ret) |
3247 | break; | |
77f1e0a5 JA |
3248 | if (q->elevator && q->elevator->type->ops.depth_updated) |
3249 | q->elevator->type->ops.depth_updated(hctx); | |
e3a2b3f9 JA |
3250 | } |
3251 | ||
3252 | if (!ret) | |
3253 | q->nr_requests = nr; | |
3254 | ||
24f5a90f | 3255 | blk_mq_unquiesce_queue(q); |
70f36b60 | 3256 | blk_mq_unfreeze_queue(q); |
70f36b60 | 3257 | |
e3a2b3f9 JA |
3258 | return ret; |
3259 | } | |
3260 | ||
d48ece20 JW |
3261 | /* |
3262 | * request_queue and elevator_type pair. | |
3263 | * It is just used by __blk_mq_update_nr_hw_queues to cache | |
3264 | * the elevator_type associated with a request_queue. | |
3265 | */ | |
3266 | struct blk_mq_qe_pair { | |
3267 | struct list_head node; | |
3268 | struct request_queue *q; | |
3269 | struct elevator_type *type; | |
3270 | }; | |
3271 | ||
3272 | /* | |
3273 | * Cache the elevator_type in qe pair list and switch the | |
3274 | * io scheduler to 'none' | |
3275 | */ | |
3276 | static bool blk_mq_elv_switch_none(struct list_head *head, | |
3277 | struct request_queue *q) | |
3278 | { | |
3279 | struct blk_mq_qe_pair *qe; | |
3280 | ||
3281 | if (!q->elevator) | |
3282 | return true; | |
3283 | ||
3284 | qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); | |
3285 | if (!qe) | |
3286 | return false; | |
3287 | ||
3288 | INIT_LIST_HEAD(&qe->node); | |
3289 | qe->q = q; | |
3290 | qe->type = q->elevator->type; | |
3291 | list_add(&qe->node, head); | |
3292 | ||
3293 | mutex_lock(&q->sysfs_lock); | |
3294 | /* | |
3295 | * After elevator_switch_mq, the previous elevator_queue will be | |
3296 | * released by elevator_release. The reference of the io scheduler | |
3297 | * module get by elevator_get will also be put. So we need to get | |
3298 | * a reference of the io scheduler module here to prevent it to be | |
3299 | * removed. | |
3300 | */ | |
3301 | __module_get(qe->type->elevator_owner); | |
3302 | elevator_switch_mq(q, NULL); | |
3303 | mutex_unlock(&q->sysfs_lock); | |
3304 | ||
3305 | return true; | |
3306 | } | |
3307 | ||
3308 | static void blk_mq_elv_switch_back(struct list_head *head, | |
3309 | struct request_queue *q) | |
3310 | { | |
3311 | struct blk_mq_qe_pair *qe; | |
3312 | struct elevator_type *t = NULL; | |
3313 | ||
3314 | list_for_each_entry(qe, head, node) | |
3315 | if (qe->q == q) { | |
3316 | t = qe->type; | |
3317 | break; | |
3318 | } | |
3319 | ||
3320 | if (!t) | |
3321 | return; | |
3322 | ||
3323 | list_del(&qe->node); | |
3324 | kfree(qe); | |
3325 | ||
3326 | mutex_lock(&q->sysfs_lock); | |
3327 | elevator_switch_mq(q, t); | |
3328 | mutex_unlock(&q->sysfs_lock); | |
3329 | } | |
3330 | ||
e4dc2b32 KB |
3331 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
3332 | int nr_hw_queues) | |
868f2f0b KB |
3333 | { |
3334 | struct request_queue *q; | |
d48ece20 | 3335 | LIST_HEAD(head); |
e01ad46d | 3336 | int prev_nr_hw_queues; |
868f2f0b | 3337 | |
705cda97 BVA |
3338 | lockdep_assert_held(&set->tag_list_lock); |
3339 | ||
392546ae | 3340 | if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids) |
868f2f0b KB |
3341 | nr_hw_queues = nr_cpu_ids; |
3342 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
3343 | return; | |
3344 | ||
3345 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3346 | blk_mq_freeze_queue(q); | |
d48ece20 JW |
3347 | /* |
3348 | * Switch IO scheduler to 'none', cleaning up the data associated | |
3349 | * with the previous scheduler. We will switch back once we are done | |
3350 | * updating the new sw to hw queue mappings. | |
3351 | */ | |
3352 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3353 | if (!blk_mq_elv_switch_none(&head, q)) | |
3354 | goto switch_back; | |
868f2f0b | 3355 | |
477e19de JW |
3356 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3357 | blk_mq_debugfs_unregister_hctxs(q); | |
3358 | blk_mq_sysfs_unregister(q); | |
3359 | } | |
3360 | ||
f7e76dbc BVA |
3361 | if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) < |
3362 | 0) | |
3363 | goto reregister; | |
3364 | ||
e01ad46d | 3365 | prev_nr_hw_queues = set->nr_hw_queues; |
868f2f0b | 3366 | set->nr_hw_queues = nr_hw_queues; |
ebe8bddb | 3367 | blk_mq_update_queue_map(set); |
e01ad46d | 3368 | fallback: |
868f2f0b KB |
3369 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3370 | blk_mq_realloc_hw_ctxs(set, q); | |
e01ad46d JW |
3371 | if (q->nr_hw_queues != set->nr_hw_queues) { |
3372 | pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n", | |
3373 | nr_hw_queues, prev_nr_hw_queues); | |
3374 | set->nr_hw_queues = prev_nr_hw_queues; | |
7d76f856 | 3375 | blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
e01ad46d JW |
3376 | goto fallback; |
3377 | } | |
477e19de JW |
3378 | blk_mq_map_swqueue(q); |
3379 | } | |
3380 | ||
f7e76dbc | 3381 | reregister: |
477e19de JW |
3382 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3383 | blk_mq_sysfs_register(q); | |
3384 | blk_mq_debugfs_register_hctxs(q); | |
868f2f0b KB |
3385 | } |
3386 | ||
d48ece20 JW |
3387 | switch_back: |
3388 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3389 | blk_mq_elv_switch_back(&head, q); | |
3390 | ||
868f2f0b KB |
3391 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
3392 | blk_mq_unfreeze_queue(q); | |
3393 | } | |
e4dc2b32 KB |
3394 | |
3395 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
3396 | { | |
3397 | mutex_lock(&set->tag_list_lock); | |
3398 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
3399 | mutex_unlock(&set->tag_list_lock); | |
3400 | } | |
868f2f0b KB |
3401 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
3402 | ||
34dbad5d OS |
3403 | /* Enable polling stats and return whether they were already enabled. */ |
3404 | static bool blk_poll_stats_enable(struct request_queue *q) | |
3405 | { | |
3406 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
7dfdbc73 | 3407 | blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q)) |
34dbad5d OS |
3408 | return true; |
3409 | blk_stat_add_callback(q, q->poll_cb); | |
3410 | return false; | |
3411 | } | |
3412 | ||
3413 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
3414 | { | |
3415 | /* | |
3416 | * We don't arm the callback if polling stats are not enabled or the | |
3417 | * callback is already active. | |
3418 | */ | |
3419 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
3420 | blk_stat_is_active(q->poll_cb)) | |
3421 | return; | |
3422 | ||
3423 | blk_stat_activate_msecs(q->poll_cb, 100); | |
3424 | } | |
3425 | ||
3426 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
3427 | { | |
3428 | struct request_queue *q = cb->data; | |
720b8ccc | 3429 | int bucket; |
34dbad5d | 3430 | |
720b8ccc SB |
3431 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
3432 | if (cb->stat[bucket].nr_samples) | |
3433 | q->poll_stat[bucket] = cb->stat[bucket]; | |
3434 | } | |
34dbad5d OS |
3435 | } |
3436 | ||
64f1c21e | 3437 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
64f1c21e JA |
3438 | struct request *rq) |
3439 | { | |
64f1c21e | 3440 | unsigned long ret = 0; |
720b8ccc | 3441 | int bucket; |
64f1c21e JA |
3442 | |
3443 | /* | |
3444 | * If stats collection isn't on, don't sleep but turn it on for | |
3445 | * future users | |
3446 | */ | |
34dbad5d | 3447 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
3448 | return 0; |
3449 | ||
64f1c21e JA |
3450 | /* |
3451 | * As an optimistic guess, use half of the mean service time | |
3452 | * for this type of request. We can (and should) make this smarter. | |
3453 | * For instance, if the completion latencies are tight, we can | |
3454 | * get closer than just half the mean. This is especially | |
3455 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
3456 | * than ~10 usec. We do use the stats for the relevant IO size |
3457 | * if available which does lead to better estimates. | |
64f1c21e | 3458 | */ |
720b8ccc SB |
3459 | bucket = blk_mq_poll_stats_bkt(rq); |
3460 | if (bucket < 0) | |
3461 | return ret; | |
3462 | ||
3463 | if (q->poll_stat[bucket].nr_samples) | |
3464 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
3465 | |
3466 | return ret; | |
3467 | } | |
3468 | ||
06426adf JA |
3469 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
3470 | struct request *rq) | |
3471 | { | |
3472 | struct hrtimer_sleeper hs; | |
3473 | enum hrtimer_mode mode; | |
64f1c21e | 3474 | unsigned int nsecs; |
06426adf JA |
3475 | ktime_t kt; |
3476 | ||
76a86f9d | 3477 | if (rq->rq_flags & RQF_MQ_POLL_SLEPT) |
64f1c21e JA |
3478 | return false; |
3479 | ||
3480 | /* | |
1052b8ac | 3481 | * If we get here, hybrid polling is enabled. Hence poll_nsec can be: |
64f1c21e | 3482 | * |
64f1c21e JA |
3483 | * 0: use half of prev avg |
3484 | * >0: use this specific value | |
3485 | */ | |
1052b8ac | 3486 | if (q->poll_nsec > 0) |
64f1c21e JA |
3487 | nsecs = q->poll_nsec; |
3488 | else | |
cae740a0 | 3489 | nsecs = blk_mq_poll_nsecs(q, rq); |
64f1c21e JA |
3490 | |
3491 | if (!nsecs) | |
06426adf JA |
3492 | return false; |
3493 | ||
76a86f9d | 3494 | rq->rq_flags |= RQF_MQ_POLL_SLEPT; |
06426adf JA |
3495 | |
3496 | /* | |
3497 | * This will be replaced with the stats tracking code, using | |
3498 | * 'avg_completion_time / 2' as the pre-sleep target. | |
3499 | */ | |
8b0e1953 | 3500 | kt = nsecs; |
06426adf JA |
3501 | |
3502 | mode = HRTIMER_MODE_REL; | |
dbc1625f | 3503 | hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode); |
06426adf JA |
3504 | hrtimer_set_expires(&hs.timer, kt); |
3505 | ||
06426adf | 3506 | do { |
5a61c363 | 3507 | if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE) |
06426adf JA |
3508 | break; |
3509 | set_current_state(TASK_UNINTERRUPTIBLE); | |
9dd8813e | 3510 | hrtimer_sleeper_start_expires(&hs, mode); |
06426adf JA |
3511 | if (hs.task) |
3512 | io_schedule(); | |
3513 | hrtimer_cancel(&hs.timer); | |
3514 | mode = HRTIMER_MODE_ABS; | |
3515 | } while (hs.task && !signal_pending(current)); | |
3516 | ||
3517 | __set_current_state(TASK_RUNNING); | |
3518 | destroy_hrtimer_on_stack(&hs.timer); | |
3519 | return true; | |
3520 | } | |
3521 | ||
1052b8ac JA |
3522 | static bool blk_mq_poll_hybrid(struct request_queue *q, |
3523 | struct blk_mq_hw_ctx *hctx, blk_qc_t cookie) | |
bbd7bb70 | 3524 | { |
1052b8ac JA |
3525 | struct request *rq; |
3526 | ||
29ece8b4 | 3527 | if (q->poll_nsec == BLK_MQ_POLL_CLASSIC) |
1052b8ac JA |
3528 | return false; |
3529 | ||
3530 | if (!blk_qc_t_is_internal(cookie)) | |
3531 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
3532 | else { | |
3533 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); | |
3534 | /* | |
3535 | * With scheduling, if the request has completed, we'll | |
3536 | * get a NULL return here, as we clear the sched tag when | |
3537 | * that happens. The request still remains valid, like always, | |
3538 | * so we should be safe with just the NULL check. | |
3539 | */ | |
3540 | if (!rq) | |
3541 | return false; | |
3542 | } | |
3543 | ||
cae740a0 | 3544 | return blk_mq_poll_hybrid_sleep(q, rq); |
1052b8ac JA |
3545 | } |
3546 | ||
529262d5 CH |
3547 | /** |
3548 | * blk_poll - poll for IO completions | |
3549 | * @q: the queue | |
3550 | * @cookie: cookie passed back at IO submission time | |
3551 | * @spin: whether to spin for completions | |
3552 | * | |
3553 | * Description: | |
3554 | * Poll for completions on the passed in queue. Returns number of | |
3555 | * completed entries found. If @spin is true, then blk_poll will continue | |
3556 | * looping until at least one completion is found, unless the task is | |
3557 | * otherwise marked running (or we need to reschedule). | |
3558 | */ | |
3559 | int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin) | |
1052b8ac JA |
3560 | { |
3561 | struct blk_mq_hw_ctx *hctx; | |
bbd7bb70 JA |
3562 | long state; |
3563 | ||
529262d5 CH |
3564 | if (!blk_qc_t_valid(cookie) || |
3565 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
1052b8ac JA |
3566 | return 0; |
3567 | ||
529262d5 CH |
3568 | if (current->plug) |
3569 | blk_flush_plug_list(current->plug, false); | |
3570 | ||
1052b8ac JA |
3571 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; |
3572 | ||
06426adf JA |
3573 | /* |
3574 | * If we sleep, have the caller restart the poll loop to reset | |
3575 | * the state. Like for the other success return cases, the | |
3576 | * caller is responsible for checking if the IO completed. If | |
3577 | * the IO isn't complete, we'll get called again and will go | |
3578 | * straight to the busy poll loop. | |
3579 | */ | |
1052b8ac | 3580 | if (blk_mq_poll_hybrid(q, hctx, cookie)) |
85f4d4b6 | 3581 | return 1; |
06426adf | 3582 | |
bbd7bb70 JA |
3583 | hctx->poll_considered++; |
3584 | ||
3585 | state = current->state; | |
aa61bec3 | 3586 | do { |
bbd7bb70 JA |
3587 | int ret; |
3588 | ||
3589 | hctx->poll_invoked++; | |
3590 | ||
9743139c | 3591 | ret = q->mq_ops->poll(hctx); |
bbd7bb70 JA |
3592 | if (ret > 0) { |
3593 | hctx->poll_success++; | |
849a3700 | 3594 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 3595 | return ret; |
bbd7bb70 JA |
3596 | } |
3597 | ||
3598 | if (signal_pending_state(state, current)) | |
849a3700 | 3599 | __set_current_state(TASK_RUNNING); |
bbd7bb70 JA |
3600 | |
3601 | if (current->state == TASK_RUNNING) | |
85f4d4b6 | 3602 | return 1; |
0a1b8b87 | 3603 | if (ret < 0 || !spin) |
bbd7bb70 JA |
3604 | break; |
3605 | cpu_relax(); | |
aa61bec3 | 3606 | } while (!need_resched()); |
bbd7bb70 | 3607 | |
67b4110f | 3608 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 3609 | return 0; |
bbd7bb70 | 3610 | } |
529262d5 | 3611 | EXPORT_SYMBOL_GPL(blk_poll); |
bbd7bb70 | 3612 | |
9cf2bab6 JA |
3613 | unsigned int blk_mq_rq_cpu(struct request *rq) |
3614 | { | |
3615 | return rq->mq_ctx->cpu; | |
3616 | } | |
3617 | EXPORT_SYMBOL(blk_mq_rq_cpu); | |
3618 | ||
320ae51f JA |
3619 | static int __init blk_mq_init(void) |
3620 | { | |
9467f859 TG |
3621 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
3622 | blk_mq_hctx_notify_dead); | |
320ae51f JA |
3623 | return 0; |
3624 | } | |
3625 | subsys_initcall(blk_mq_init); |