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 */ |
af76e555 | 321 | rq->extra_len = 0; |
079076b3 | 322 | WRITE_ONCE(rq->deadline, 0); |
af76e555 | 323 | |
f6be4fb4 JA |
324 | rq->timeout = 0; |
325 | ||
af76e555 CH |
326 | rq->end_io = NULL; |
327 | rq->end_io_data = NULL; | |
af76e555 | 328 | |
e4cdf1a1 | 329 | data->ctx->rq_dispatched[op_is_sync(op)]++; |
12f5b931 | 330 | refcount_set(&rq->ref, 1); |
e4cdf1a1 | 331 | return rq; |
5dee8577 CH |
332 | } |
333 | ||
d2c0d383 | 334 | static struct request *blk_mq_get_request(struct request_queue *q, |
f9afca4d JA |
335 | struct bio *bio, |
336 | struct blk_mq_alloc_data *data) | |
d2c0d383 CH |
337 | { |
338 | struct elevator_queue *e = q->elevator; | |
339 | struct request *rq; | |
e4cdf1a1 | 340 | unsigned int tag; |
c05f4220 | 341 | bool clear_ctx_on_error = false; |
6f816b4b | 342 | u64 alloc_time_ns = 0; |
d2c0d383 CH |
343 | |
344 | blk_queue_enter_live(q); | |
6f816b4b TH |
345 | |
346 | /* alloc_time includes depth and tag waits */ | |
347 | if (blk_queue_rq_alloc_time(q)) | |
348 | alloc_time_ns = ktime_get_ns(); | |
349 | ||
d2c0d383 | 350 | data->q = q; |
21e768b4 BVA |
351 | if (likely(!data->ctx)) { |
352 | data->ctx = blk_mq_get_ctx(q); | |
c05f4220 | 353 | clear_ctx_on_error = true; |
21e768b4 | 354 | } |
d2c0d383 | 355 | if (likely(!data->hctx)) |
f9afca4d | 356 | data->hctx = blk_mq_map_queue(q, data->cmd_flags, |
8ccdf4a3 | 357 | data->ctx); |
f9afca4d | 358 | if (data->cmd_flags & REQ_NOWAIT) |
03a07c92 | 359 | data->flags |= BLK_MQ_REQ_NOWAIT; |
d2c0d383 CH |
360 | |
361 | if (e) { | |
362 | data->flags |= BLK_MQ_REQ_INTERNAL; | |
363 | ||
364 | /* | |
365 | * Flush requests are special and go directly to the | |
17a51199 JA |
366 | * dispatch list. Don't include reserved tags in the |
367 | * limiting, as it isn't useful. | |
d2c0d383 | 368 | */ |
f9afca4d JA |
369 | if (!op_is_flush(data->cmd_flags) && |
370 | e->type->ops.limit_depth && | |
17a51199 | 371 | !(data->flags & BLK_MQ_REQ_RESERVED)) |
f9afca4d | 372 | e->type->ops.limit_depth(data->cmd_flags, data); |
d263ed99 JW |
373 | } else { |
374 | blk_mq_tag_busy(data->hctx); | |
d2c0d383 CH |
375 | } |
376 | ||
e4cdf1a1 CH |
377 | tag = blk_mq_get_tag(data); |
378 | if (tag == BLK_MQ_TAG_FAIL) { | |
c05f4220 | 379 | if (clear_ctx_on_error) |
1ad43c00 | 380 | data->ctx = NULL; |
037cebb8 CH |
381 | blk_queue_exit(q); |
382 | return NULL; | |
d2c0d383 CH |
383 | } |
384 | ||
6f816b4b | 385 | rq = blk_mq_rq_ctx_init(data, tag, data->cmd_flags, alloc_time_ns); |
f9afca4d | 386 | if (!op_is_flush(data->cmd_flags)) { |
037cebb8 | 387 | rq->elv.icq = NULL; |
f9cd4bfe | 388 | if (e && e->type->ops.prepare_request) { |
e2b3fa5a DLM |
389 | if (e->type->icq_cache) |
390 | blk_mq_sched_assign_ioc(rq); | |
44e8c2bf | 391 | |
f9cd4bfe | 392 | e->type->ops.prepare_request(rq, bio); |
5bbf4e5a | 393 | rq->rq_flags |= RQF_ELVPRIV; |
44e8c2bf | 394 | } |
037cebb8 CH |
395 | } |
396 | data->hctx->queued++; | |
397 | return rq; | |
d2c0d383 CH |
398 | } |
399 | ||
cd6ce148 | 400 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, |
9a95e4ef | 401 | blk_mq_req_flags_t flags) |
320ae51f | 402 | { |
f9afca4d | 403 | struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op }; |
bd166ef1 | 404 | struct request *rq; |
a492f075 | 405 | int ret; |
320ae51f | 406 | |
3a0a5299 | 407 | ret = blk_queue_enter(q, flags); |
a492f075 JL |
408 | if (ret) |
409 | return ERR_PTR(ret); | |
320ae51f | 410 | |
f9afca4d | 411 | rq = blk_mq_get_request(q, NULL, &alloc_data); |
3280d66a | 412 | blk_queue_exit(q); |
841bac2c | 413 | |
bd166ef1 | 414 | if (!rq) |
a492f075 | 415 | return ERR_PTR(-EWOULDBLOCK); |
0c4de0f3 CH |
416 | |
417 | rq->__data_len = 0; | |
418 | rq->__sector = (sector_t) -1; | |
419 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
420 | return rq; |
421 | } | |
4bb659b1 | 422 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 423 | |
cd6ce148 | 424 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
9a95e4ef | 425 | unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 426 | { |
f9afca4d | 427 | struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op }; |
1f5bd336 | 428 | struct request *rq; |
6d2809d5 | 429 | unsigned int cpu; |
1f5bd336 ML |
430 | int ret; |
431 | ||
432 | /* | |
433 | * If the tag allocator sleeps we could get an allocation for a | |
434 | * different hardware context. No need to complicate the low level | |
435 | * allocator for this for the rare use case of a command tied to | |
436 | * a specific queue. | |
437 | */ | |
438 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
439 | return ERR_PTR(-EINVAL); | |
440 | ||
441 | if (hctx_idx >= q->nr_hw_queues) | |
442 | return ERR_PTR(-EIO); | |
443 | ||
3a0a5299 | 444 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
445 | if (ret) |
446 | return ERR_PTR(ret); | |
447 | ||
c8712c6a CH |
448 | /* |
449 | * Check if the hardware context is actually mapped to anything. | |
450 | * If not tell the caller that it should skip this queue. | |
451 | */ | |
6d2809d5 OS |
452 | alloc_data.hctx = q->queue_hw_ctx[hctx_idx]; |
453 | if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) { | |
454 | blk_queue_exit(q); | |
455 | return ERR_PTR(-EXDEV); | |
c8712c6a | 456 | } |
20e4d813 | 457 | cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask); |
6d2809d5 | 458 | alloc_data.ctx = __blk_mq_get_ctx(q, cpu); |
1f5bd336 | 459 | |
f9afca4d | 460 | rq = blk_mq_get_request(q, NULL, &alloc_data); |
3280d66a | 461 | blk_queue_exit(q); |
c8712c6a | 462 | |
6d2809d5 OS |
463 | if (!rq) |
464 | return ERR_PTR(-EWOULDBLOCK); | |
465 | ||
466 | return rq; | |
1f5bd336 ML |
467 | } |
468 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
469 | ||
12f5b931 KB |
470 | static void __blk_mq_free_request(struct request *rq) |
471 | { | |
472 | struct request_queue *q = rq->q; | |
473 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 474 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
12f5b931 KB |
475 | const int sched_tag = rq->internal_tag; |
476 | ||
986d413b | 477 | blk_pm_mark_last_busy(rq); |
ea4f995e | 478 | rq->mq_hctx = NULL; |
12f5b931 KB |
479 | if (rq->tag != -1) |
480 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
481 | if (sched_tag != -1) | |
482 | blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag); | |
483 | blk_mq_sched_restart(hctx); | |
484 | blk_queue_exit(q); | |
485 | } | |
486 | ||
6af54051 | 487 | void blk_mq_free_request(struct request *rq) |
320ae51f | 488 | { |
320ae51f | 489 | struct request_queue *q = rq->q; |
6af54051 CH |
490 | struct elevator_queue *e = q->elevator; |
491 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 492 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
6af54051 | 493 | |
5bbf4e5a | 494 | if (rq->rq_flags & RQF_ELVPRIV) { |
f9cd4bfe JA |
495 | if (e && e->type->ops.finish_request) |
496 | e->type->ops.finish_request(rq); | |
6af54051 CH |
497 | if (rq->elv.icq) { |
498 | put_io_context(rq->elv.icq->ioc); | |
499 | rq->elv.icq = NULL; | |
500 | } | |
501 | } | |
320ae51f | 502 | |
6af54051 | 503 | ctx->rq_completed[rq_is_sync(rq)]++; |
e8064021 | 504 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 505 | atomic_dec(&hctx->nr_active); |
87760e5e | 506 | |
7beb2f84 JA |
507 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
508 | laptop_io_completion(q->backing_dev_info); | |
509 | ||
a7905043 | 510 | rq_qos_done(q, rq); |
0d2602ca | 511 | |
12f5b931 KB |
512 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
513 | if (refcount_dec_and_test(&rq->ref)) | |
514 | __blk_mq_free_request(rq); | |
320ae51f | 515 | } |
1a3b595a | 516 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 517 | |
2a842aca | 518 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
320ae51f | 519 | { |
fe1f4526 JA |
520 | u64 now = 0; |
521 | ||
522 | if (blk_mq_need_time_stamp(rq)) | |
523 | now = ktime_get_ns(); | |
522a7775 | 524 | |
4bc6339a OS |
525 | if (rq->rq_flags & RQF_STATS) { |
526 | blk_mq_poll_stats_start(rq->q); | |
522a7775 | 527 | blk_stat_add(rq, now); |
4bc6339a OS |
528 | } |
529 | ||
ed88660a OS |
530 | if (rq->internal_tag != -1) |
531 | blk_mq_sched_completed_request(rq, now); | |
532 | ||
522a7775 | 533 | blk_account_io_done(rq, now); |
0d11e6ac | 534 | |
91b63639 | 535 | if (rq->end_io) { |
a7905043 | 536 | rq_qos_done(rq->q, rq); |
320ae51f | 537 | rq->end_io(rq, error); |
91b63639 | 538 | } else { |
320ae51f | 539 | blk_mq_free_request(rq); |
91b63639 | 540 | } |
320ae51f | 541 | } |
c8a446ad | 542 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 543 | |
2a842aca | 544 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
545 | { |
546 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
547 | BUG(); | |
c8a446ad | 548 | __blk_mq_end_request(rq, error); |
63151a44 | 549 | } |
c8a446ad | 550 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 551 | |
30a91cb4 | 552 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 553 | { |
3d6efbf6 | 554 | struct request *rq = data; |
c7bb9ad1 | 555 | struct request_queue *q = rq->q; |
320ae51f | 556 | |
c7bb9ad1 | 557 | q->mq_ops->complete(rq); |
320ae51f | 558 | } |
320ae51f | 559 | |
453f8341 | 560 | static void __blk_mq_complete_request(struct request *rq) |
320ae51f JA |
561 | { |
562 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
c7bb9ad1 | 563 | struct request_queue *q = rq->q; |
38535201 | 564 | bool shared = false; |
320ae51f JA |
565 | int cpu; |
566 | ||
af78ff7c | 567 | WRITE_ONCE(rq->state, MQ_RQ_COMPLETE); |
36e76539 ML |
568 | /* |
569 | * Most of single queue controllers, there is only one irq vector | |
570 | * for handling IO completion, and the only irq's affinity is set | |
571 | * as all possible CPUs. On most of ARCHs, this affinity means the | |
572 | * irq is handled on one specific CPU. | |
573 | * | |
574 | * So complete IO reqeust in softirq context in case of single queue | |
575 | * for not degrading IO performance by irqsoff latency. | |
576 | */ | |
c7bb9ad1 | 577 | if (q->nr_hw_queues == 1) { |
36e76539 ML |
578 | __blk_complete_request(rq); |
579 | return; | |
580 | } | |
581 | ||
4ab32bf3 JA |
582 | /* |
583 | * For a polled request, always complete locallly, it's pointless | |
584 | * to redirect the completion. | |
585 | */ | |
586 | if ((rq->cmd_flags & REQ_HIPRI) || | |
587 | !test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags)) { | |
c7bb9ad1 | 588 | q->mq_ops->complete(rq); |
30a91cb4 CH |
589 | return; |
590 | } | |
320ae51f JA |
591 | |
592 | cpu = get_cpu(); | |
c7bb9ad1 | 593 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &q->queue_flags)) |
38535201 CH |
594 | shared = cpus_share_cache(cpu, ctx->cpu); |
595 | ||
596 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 597 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
598 | rq->csd.info = rq; |
599 | rq->csd.flags = 0; | |
c46fff2a | 600 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 601 | } else { |
c7bb9ad1 | 602 | q->mq_ops->complete(rq); |
3d6efbf6 | 603 | } |
320ae51f JA |
604 | put_cpu(); |
605 | } | |
30a91cb4 | 606 | |
04ced159 | 607 | static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx) |
b7435db8 | 608 | __releases(hctx->srcu) |
04ced159 JA |
609 | { |
610 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
611 | rcu_read_unlock(); | |
612 | else | |
05707b64 | 613 | srcu_read_unlock(hctx->srcu, srcu_idx); |
04ced159 JA |
614 | } |
615 | ||
616 | static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx) | |
b7435db8 | 617 | __acquires(hctx->srcu) |
04ced159 | 618 | { |
08b5a6e2 JA |
619 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
620 | /* shut up gcc false positive */ | |
621 | *srcu_idx = 0; | |
04ced159 | 622 | rcu_read_lock(); |
08b5a6e2 | 623 | } else |
05707b64 | 624 | *srcu_idx = srcu_read_lock(hctx->srcu); |
04ced159 JA |
625 | } |
626 | ||
30a91cb4 CH |
627 | /** |
628 | * blk_mq_complete_request - end I/O on a request | |
629 | * @rq: the request being processed | |
630 | * | |
631 | * Description: | |
632 | * Ends all I/O on a request. It does not handle partial completions. | |
633 | * The actual completion happens out-of-order, through a IPI handler. | |
634 | **/ | |
16c15eb1 | 635 | bool blk_mq_complete_request(struct request *rq) |
30a91cb4 | 636 | { |
12f5b931 | 637 | if (unlikely(blk_should_fake_timeout(rq->q))) |
16c15eb1 | 638 | return false; |
12f5b931 | 639 | __blk_mq_complete_request(rq); |
16c15eb1 | 640 | return true; |
30a91cb4 CH |
641 | } |
642 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 643 | |
e2490073 | 644 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
645 | { |
646 | struct request_queue *q = rq->q; | |
647 | ||
648 | trace_block_rq_issue(q, rq); | |
649 | ||
cf43e6be | 650 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
544ccc8d | 651 | rq->io_start_time_ns = ktime_get_ns(); |
3d244306 | 652 | rq->stats_sectors = blk_rq_sectors(rq); |
cf43e6be | 653 | rq->rq_flags |= RQF_STATS; |
a7905043 | 654 | rq_qos_issue(q, rq); |
cf43e6be JA |
655 | } |
656 | ||
1d9bd516 | 657 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 658 | |
1d9bd516 | 659 | blk_add_timer(rq); |
12f5b931 | 660 | WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); |
49f5baa5 CH |
661 | |
662 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
663 | /* | |
664 | * Make sure space for the drain appears. We know we can do | |
665 | * this because max_hw_segments has been adjusted to be one | |
666 | * fewer than the device can handle. | |
667 | */ | |
668 | rq->nr_phys_segments++; | |
669 | } | |
54d4e6ab MG |
670 | |
671 | #ifdef CONFIG_BLK_DEV_INTEGRITY | |
672 | if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE) | |
673 | q->integrity.profile->prepare_fn(rq); | |
674 | #endif | |
320ae51f | 675 | } |
e2490073 | 676 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 677 | |
ed0791b2 | 678 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
679 | { |
680 | struct request_queue *q = rq->q; | |
681 | ||
923218f6 ML |
682 | blk_mq_put_driver_tag(rq); |
683 | ||
320ae51f | 684 | trace_block_rq_requeue(q, rq); |
a7905043 | 685 | rq_qos_requeue(q, rq); |
49f5baa5 | 686 | |
12f5b931 KB |
687 | if (blk_mq_request_started(rq)) { |
688 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); | |
da661267 | 689 | rq->rq_flags &= ~RQF_TIMED_OUT; |
e2490073 CH |
690 | if (q->dma_drain_size && blk_rq_bytes(rq)) |
691 | rq->nr_phys_segments--; | |
692 | } | |
320ae51f JA |
693 | } |
694 | ||
2b053aca | 695 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 696 | { |
ed0791b2 | 697 | __blk_mq_requeue_request(rq); |
ed0791b2 | 698 | |
105976f5 ML |
699 | /* this request will be re-inserted to io scheduler queue */ |
700 | blk_mq_sched_requeue_request(rq); | |
701 | ||
7d692330 | 702 | BUG_ON(!list_empty(&rq->queuelist)); |
2b053aca | 703 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
704 | } |
705 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
706 | ||
6fca6a61 CH |
707 | static void blk_mq_requeue_work(struct work_struct *work) |
708 | { | |
709 | struct request_queue *q = | |
2849450a | 710 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
711 | LIST_HEAD(rq_list); |
712 | struct request *rq, *next; | |
6fca6a61 | 713 | |
18e9781d | 714 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 715 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 716 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
717 | |
718 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
aef1897c | 719 | if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP))) |
6fca6a61 CH |
720 | continue; |
721 | ||
e8064021 | 722 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 723 | list_del_init(&rq->queuelist); |
aef1897c JW |
724 | /* |
725 | * If RQF_DONTPREP, rq has contained some driver specific | |
726 | * data, so insert it to hctx dispatch list to avoid any | |
727 | * merge. | |
728 | */ | |
729 | if (rq->rq_flags & RQF_DONTPREP) | |
730 | blk_mq_request_bypass_insert(rq, false); | |
731 | else | |
732 | blk_mq_sched_insert_request(rq, true, false, false); | |
6fca6a61 CH |
733 | } |
734 | ||
735 | while (!list_empty(&rq_list)) { | |
736 | rq = list_entry(rq_list.next, struct request, queuelist); | |
737 | list_del_init(&rq->queuelist); | |
9e97d295 | 738 | blk_mq_sched_insert_request(rq, false, false, false); |
6fca6a61 CH |
739 | } |
740 | ||
52d7f1b5 | 741 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
742 | } |
743 | ||
2b053aca BVA |
744 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
745 | bool kick_requeue_list) | |
6fca6a61 CH |
746 | { |
747 | struct request_queue *q = rq->q; | |
748 | unsigned long flags; | |
749 | ||
750 | /* | |
751 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
ff821d27 | 752 | * request head insertion from the workqueue. |
6fca6a61 | 753 | */ |
e8064021 | 754 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
755 | |
756 | spin_lock_irqsave(&q->requeue_lock, flags); | |
757 | if (at_head) { | |
e8064021 | 758 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
759 | list_add(&rq->queuelist, &q->requeue_list); |
760 | } else { | |
761 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
762 | } | |
763 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
764 | |
765 | if (kick_requeue_list) | |
766 | blk_mq_kick_requeue_list(q); | |
6fca6a61 | 767 | } |
6fca6a61 CH |
768 | |
769 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
770 | { | |
ae943d20 | 771 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
772 | } |
773 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
774 | ||
2849450a MS |
775 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
776 | unsigned long msecs) | |
777 | { | |
d4acf365 BVA |
778 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
779 | msecs_to_jiffies(msecs)); | |
2849450a MS |
780 | } |
781 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
782 | ||
0e62f51f JA |
783 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
784 | { | |
88c7b2b7 JA |
785 | if (tag < tags->nr_tags) { |
786 | prefetch(tags->rqs[tag]); | |
4ee86bab | 787 | return tags->rqs[tag]; |
88c7b2b7 | 788 | } |
4ee86bab HR |
789 | |
790 | return NULL; | |
24d2f903 CH |
791 | } |
792 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
793 | ||
3c94d83c JA |
794 | static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq, |
795 | void *priv, bool reserved) | |
ae879912 JA |
796 | { |
797 | /* | |
3c94d83c JA |
798 | * If we find a request that is inflight and the queue matches, |
799 | * we know the queue is busy. Return false to stop the iteration. | |
ae879912 | 800 | */ |
3c94d83c | 801 | if (rq->state == MQ_RQ_IN_FLIGHT && rq->q == hctx->queue) { |
ae879912 JA |
802 | bool *busy = priv; |
803 | ||
804 | *busy = true; | |
805 | return false; | |
806 | } | |
807 | ||
808 | return true; | |
809 | } | |
810 | ||
3c94d83c | 811 | bool blk_mq_queue_inflight(struct request_queue *q) |
ae879912 JA |
812 | { |
813 | bool busy = false; | |
814 | ||
3c94d83c | 815 | blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy); |
ae879912 JA |
816 | return busy; |
817 | } | |
3c94d83c | 818 | EXPORT_SYMBOL_GPL(blk_mq_queue_inflight); |
ae879912 | 819 | |
358f70da | 820 | static void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 821 | { |
da661267 | 822 | req->rq_flags |= RQF_TIMED_OUT; |
d1210d5a CH |
823 | if (req->q->mq_ops->timeout) { |
824 | enum blk_eh_timer_return ret; | |
825 | ||
826 | ret = req->q->mq_ops->timeout(req, reserved); | |
827 | if (ret == BLK_EH_DONE) | |
828 | return; | |
829 | WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); | |
46f92d42 | 830 | } |
d1210d5a CH |
831 | |
832 | blk_add_timer(req); | |
87ee7b11 | 833 | } |
5b3f25fc | 834 | |
12f5b931 | 835 | static bool blk_mq_req_expired(struct request *rq, unsigned long *next) |
81481eb4 | 836 | { |
12f5b931 | 837 | unsigned long deadline; |
87ee7b11 | 838 | |
12f5b931 KB |
839 | if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) |
840 | return false; | |
da661267 CH |
841 | if (rq->rq_flags & RQF_TIMED_OUT) |
842 | return false; | |
a7af0af3 | 843 | |
079076b3 | 844 | deadline = READ_ONCE(rq->deadline); |
12f5b931 KB |
845 | if (time_after_eq(jiffies, deadline)) |
846 | return true; | |
a7af0af3 | 847 | |
12f5b931 KB |
848 | if (*next == 0) |
849 | *next = deadline; | |
850 | else if (time_after(*next, deadline)) | |
851 | *next = deadline; | |
852 | return false; | |
87ee7b11 JA |
853 | } |
854 | ||
7baa8572 | 855 | static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
1d9bd516 TH |
856 | struct request *rq, void *priv, bool reserved) |
857 | { | |
12f5b931 KB |
858 | unsigned long *next = priv; |
859 | ||
860 | /* | |
861 | * Just do a quick check if it is expired before locking the request in | |
862 | * so we're not unnecessarilly synchronizing across CPUs. | |
863 | */ | |
864 | if (!blk_mq_req_expired(rq, next)) | |
7baa8572 | 865 | return true; |
12f5b931 KB |
866 | |
867 | /* | |
868 | * We have reason to believe the request may be expired. Take a | |
869 | * reference on the request to lock this request lifetime into its | |
870 | * currently allocated context to prevent it from being reallocated in | |
871 | * the event the completion by-passes this timeout handler. | |
872 | * | |
873 | * If the reference was already released, then the driver beat the | |
874 | * timeout handler to posting a natural completion. | |
875 | */ | |
876 | if (!refcount_inc_not_zero(&rq->ref)) | |
7baa8572 | 877 | return true; |
12f5b931 | 878 | |
1d9bd516 | 879 | /* |
12f5b931 KB |
880 | * The request is now locked and cannot be reallocated underneath the |
881 | * timeout handler's processing. Re-verify this exact request is truly | |
882 | * expired; if it is not expired, then the request was completed and | |
883 | * reallocated as a new request. | |
1d9bd516 | 884 | */ |
12f5b931 | 885 | if (blk_mq_req_expired(rq, next)) |
1d9bd516 | 886 | blk_mq_rq_timed_out(rq, reserved); |
8d699663 YY |
887 | |
888 | if (is_flush_rq(rq, hctx)) | |
889 | rq->end_io(rq, 0); | |
890 | else if (refcount_dec_and_test(&rq->ref)) | |
12f5b931 | 891 | __blk_mq_free_request(rq); |
7baa8572 JA |
892 | |
893 | return true; | |
1d9bd516 TH |
894 | } |
895 | ||
287922eb | 896 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 897 | { |
287922eb CH |
898 | struct request_queue *q = |
899 | container_of(work, struct request_queue, timeout_work); | |
12f5b931 | 900 | unsigned long next = 0; |
1d9bd516 | 901 | struct blk_mq_hw_ctx *hctx; |
81481eb4 | 902 | int i; |
320ae51f | 903 | |
71f79fb3 GKB |
904 | /* A deadlock might occur if a request is stuck requiring a |
905 | * timeout at the same time a queue freeze is waiting | |
906 | * completion, since the timeout code would not be able to | |
907 | * acquire the queue reference here. | |
908 | * | |
909 | * That's why we don't use blk_queue_enter here; instead, we use | |
910 | * percpu_ref_tryget directly, because we need to be able to | |
911 | * obtain a reference even in the short window between the queue | |
912 | * starting to freeze, by dropping the first reference in | |
1671d522 | 913 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
914 | * consumed, marked by the instant q_usage_counter reaches |
915 | * zero. | |
916 | */ | |
917 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
918 | return; |
919 | ||
12f5b931 | 920 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next); |
320ae51f | 921 | |
12f5b931 KB |
922 | if (next != 0) { |
923 | mod_timer(&q->timeout, next); | |
0d2602ca | 924 | } else { |
fcd36c36 BVA |
925 | /* |
926 | * Request timeouts are handled as a forward rolling timer. If | |
927 | * we end up here it means that no requests are pending and | |
928 | * also that no request has been pending for a while. Mark | |
929 | * each hctx as idle. | |
930 | */ | |
f054b56c ML |
931 | queue_for_each_hw_ctx(q, hctx, i) { |
932 | /* the hctx may be unmapped, so check it here */ | |
933 | if (blk_mq_hw_queue_mapped(hctx)) | |
934 | blk_mq_tag_idle(hctx); | |
935 | } | |
0d2602ca | 936 | } |
287922eb | 937 | blk_queue_exit(q); |
320ae51f JA |
938 | } |
939 | ||
88459642 OS |
940 | struct flush_busy_ctx_data { |
941 | struct blk_mq_hw_ctx *hctx; | |
942 | struct list_head *list; | |
943 | }; | |
944 | ||
945 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
946 | { | |
947 | struct flush_busy_ctx_data *flush_data = data; | |
948 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
949 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 950 | enum hctx_type type = hctx->type; |
88459642 | 951 | |
88459642 | 952 | spin_lock(&ctx->lock); |
c16d6b5a | 953 | list_splice_tail_init(&ctx->rq_lists[type], flush_data->list); |
e9a99a63 | 954 | sbitmap_clear_bit(sb, bitnr); |
88459642 OS |
955 | spin_unlock(&ctx->lock); |
956 | return true; | |
957 | } | |
958 | ||
1429d7c9 JA |
959 | /* |
960 | * Process software queues that have been marked busy, splicing them | |
961 | * to the for-dispatch | |
962 | */ | |
2c3ad667 | 963 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 964 | { |
88459642 OS |
965 | struct flush_busy_ctx_data data = { |
966 | .hctx = hctx, | |
967 | .list = list, | |
968 | }; | |
1429d7c9 | 969 | |
88459642 | 970 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 971 | } |
2c3ad667 | 972 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 973 | |
b347689f ML |
974 | struct dispatch_rq_data { |
975 | struct blk_mq_hw_ctx *hctx; | |
976 | struct request *rq; | |
977 | }; | |
978 | ||
979 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
980 | void *data) | |
981 | { | |
982 | struct dispatch_rq_data *dispatch_data = data; | |
983 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
984 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 985 | enum hctx_type type = hctx->type; |
b347689f ML |
986 | |
987 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
988 | if (!list_empty(&ctx->rq_lists[type])) { |
989 | dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next); | |
b347689f | 990 | list_del_init(&dispatch_data->rq->queuelist); |
c16d6b5a | 991 | if (list_empty(&ctx->rq_lists[type])) |
b347689f ML |
992 | sbitmap_clear_bit(sb, bitnr); |
993 | } | |
994 | spin_unlock(&ctx->lock); | |
995 | ||
996 | return !dispatch_data->rq; | |
997 | } | |
998 | ||
999 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
1000 | struct blk_mq_ctx *start) | |
1001 | { | |
f31967f0 | 1002 | unsigned off = start ? start->index_hw[hctx->type] : 0; |
b347689f ML |
1003 | struct dispatch_rq_data data = { |
1004 | .hctx = hctx, | |
1005 | .rq = NULL, | |
1006 | }; | |
1007 | ||
1008 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1009 | dispatch_rq_from_ctx, &data); | |
1010 | ||
1011 | return data.rq; | |
1012 | } | |
1013 | ||
703fd1c0 JA |
1014 | static inline unsigned int queued_to_index(unsigned int queued) |
1015 | { | |
1016 | if (!queued) | |
1017 | return 0; | |
1429d7c9 | 1018 | |
703fd1c0 | 1019 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
1020 | } |
1021 | ||
8ab6bb9e | 1022 | bool blk_mq_get_driver_tag(struct request *rq) |
bd166ef1 JA |
1023 | { |
1024 | struct blk_mq_alloc_data data = { | |
1025 | .q = rq->q, | |
ea4f995e | 1026 | .hctx = rq->mq_hctx, |
8ab6bb9e | 1027 | .flags = BLK_MQ_REQ_NOWAIT, |
f9afca4d | 1028 | .cmd_flags = rq->cmd_flags, |
bd166ef1 | 1029 | }; |
d263ed99 | 1030 | bool shared; |
5feeacdd | 1031 | |
81380ca1 | 1032 | if (rq->tag != -1) |
1fead718 | 1033 | return true; |
bd166ef1 | 1034 | |
415b806d SG |
1035 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) |
1036 | data.flags |= BLK_MQ_REQ_RESERVED; | |
1037 | ||
d263ed99 | 1038 | shared = blk_mq_tag_busy(data.hctx); |
bd166ef1 JA |
1039 | rq->tag = blk_mq_get_tag(&data); |
1040 | if (rq->tag >= 0) { | |
d263ed99 | 1041 | if (shared) { |
200e86b3 JA |
1042 | rq->rq_flags |= RQF_MQ_INFLIGHT; |
1043 | atomic_inc(&data.hctx->nr_active); | |
1044 | } | |
bd166ef1 | 1045 | data.hctx->tags->rqs[rq->tag] = rq; |
bd166ef1 JA |
1046 | } |
1047 | ||
81380ca1 | 1048 | return rq->tag != -1; |
bd166ef1 JA |
1049 | } |
1050 | ||
eb619fdb JA |
1051 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1052 | int flags, void *key) | |
da55f2cc OS |
1053 | { |
1054 | struct blk_mq_hw_ctx *hctx; | |
1055 | ||
1056 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1057 | ||
5815839b | 1058 | spin_lock(&hctx->dispatch_wait_lock); |
e8618575 JA |
1059 | if (!list_empty(&wait->entry)) { |
1060 | struct sbitmap_queue *sbq; | |
1061 | ||
1062 | list_del_init(&wait->entry); | |
1063 | sbq = &hctx->tags->bitmap_tags; | |
1064 | atomic_dec(&sbq->ws_active); | |
1065 | } | |
5815839b ML |
1066 | spin_unlock(&hctx->dispatch_wait_lock); |
1067 | ||
da55f2cc OS |
1068 | blk_mq_run_hw_queue(hctx, true); |
1069 | return 1; | |
1070 | } | |
1071 | ||
f906a6a0 JA |
1072 | /* |
1073 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1074 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1075 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1076 | * marking us as waiting. |
1077 | */ | |
2278d69f | 1078 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, |
f906a6a0 | 1079 | struct request *rq) |
da55f2cc | 1080 | { |
e8618575 | 1081 | struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags; |
5815839b | 1082 | struct wait_queue_head *wq; |
f906a6a0 JA |
1083 | wait_queue_entry_t *wait; |
1084 | bool ret; | |
da55f2cc | 1085 | |
2278d69f | 1086 | if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) { |
684b7324 | 1087 | blk_mq_sched_mark_restart_hctx(hctx); |
f906a6a0 | 1088 | |
c27d53fb BVA |
1089 | /* |
1090 | * It's possible that a tag was freed in the window between the | |
1091 | * allocation failure and adding the hardware queue to the wait | |
1092 | * queue. | |
1093 | * | |
1094 | * Don't clear RESTART here, someone else could have set it. | |
1095 | * At most this will cost an extra queue run. | |
1096 | */ | |
8ab6bb9e | 1097 | return blk_mq_get_driver_tag(rq); |
eb619fdb | 1098 | } |
eb619fdb | 1099 | |
2278d69f | 1100 | wait = &hctx->dispatch_wait; |
c27d53fb BVA |
1101 | if (!list_empty_careful(&wait->entry)) |
1102 | return false; | |
1103 | ||
e8618575 | 1104 | wq = &bt_wait_ptr(sbq, hctx)->wait; |
5815839b ML |
1105 | |
1106 | spin_lock_irq(&wq->lock); | |
1107 | spin_lock(&hctx->dispatch_wait_lock); | |
c27d53fb | 1108 | if (!list_empty(&wait->entry)) { |
5815839b ML |
1109 | spin_unlock(&hctx->dispatch_wait_lock); |
1110 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1111 | return false; |
eb619fdb JA |
1112 | } |
1113 | ||
e8618575 | 1114 | atomic_inc(&sbq->ws_active); |
5815839b ML |
1115 | wait->flags &= ~WQ_FLAG_EXCLUSIVE; |
1116 | __add_wait_queue(wq, wait); | |
c27d53fb | 1117 | |
da55f2cc | 1118 | /* |
eb619fdb JA |
1119 | * It's possible that a tag was freed in the window between the |
1120 | * allocation failure and adding the hardware queue to the wait | |
1121 | * queue. | |
da55f2cc | 1122 | */ |
8ab6bb9e | 1123 | ret = blk_mq_get_driver_tag(rq); |
c27d53fb | 1124 | if (!ret) { |
5815839b ML |
1125 | spin_unlock(&hctx->dispatch_wait_lock); |
1126 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1127 | return false; |
eb619fdb | 1128 | } |
c27d53fb BVA |
1129 | |
1130 | /* | |
1131 | * We got a tag, remove ourselves from the wait queue to ensure | |
1132 | * someone else gets the wakeup. | |
1133 | */ | |
c27d53fb | 1134 | list_del_init(&wait->entry); |
e8618575 | 1135 | atomic_dec(&sbq->ws_active); |
5815839b ML |
1136 | spin_unlock(&hctx->dispatch_wait_lock); |
1137 | spin_unlock_irq(&wq->lock); | |
c27d53fb BVA |
1138 | |
1139 | return true; | |
da55f2cc OS |
1140 | } |
1141 | ||
6e768717 ML |
1142 | #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 |
1143 | #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 | |
1144 | /* | |
1145 | * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): | |
1146 | * - EWMA is one simple way to compute running average value | |
1147 | * - weight(7/8 and 1/8) is applied so that it can decrease exponentially | |
1148 | * - take 4 as factor for avoiding to get too small(0) result, and this | |
1149 | * factor doesn't matter because EWMA decreases exponentially | |
1150 | */ | |
1151 | static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) | |
1152 | { | |
1153 | unsigned int ewma; | |
1154 | ||
1155 | if (hctx->queue->elevator) | |
1156 | return; | |
1157 | ||
1158 | ewma = hctx->dispatch_busy; | |
1159 | ||
1160 | if (!ewma && !busy) | |
1161 | return; | |
1162 | ||
1163 | ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; | |
1164 | if (busy) | |
1165 | ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; | |
1166 | ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; | |
1167 | ||
1168 | hctx->dispatch_busy = ewma; | |
1169 | } | |
1170 | ||
86ff7c2a ML |
1171 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
1172 | ||
1f57f8d4 JA |
1173 | /* |
1174 | * Returns true if we did some work AND can potentially do more. | |
1175 | */ | |
de148297 | 1176 | bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list, |
eb619fdb | 1177 | bool got_budget) |
320ae51f | 1178 | { |
81380ca1 | 1179 | struct blk_mq_hw_ctx *hctx; |
6d6f167c | 1180 | struct request *rq, *nxt; |
eb619fdb | 1181 | bool no_tag = false; |
fc17b653 | 1182 | int errors, queued; |
86ff7c2a | 1183 | blk_status_t ret = BLK_STS_OK; |
320ae51f | 1184 | |
81380ca1 OS |
1185 | if (list_empty(list)) |
1186 | return false; | |
1187 | ||
de148297 ML |
1188 | WARN_ON(!list_is_singular(list) && got_budget); |
1189 | ||
320ae51f JA |
1190 | /* |
1191 | * Now process all the entries, sending them to the driver. | |
1192 | */ | |
93efe981 | 1193 | errors = queued = 0; |
81380ca1 | 1194 | do { |
74c45052 | 1195 | struct blk_mq_queue_data bd; |
320ae51f | 1196 | |
f04c3df3 | 1197 | rq = list_first_entry(list, struct request, queuelist); |
0bca799b | 1198 | |
ea4f995e | 1199 | hctx = rq->mq_hctx; |
0bca799b ML |
1200 | if (!got_budget && !blk_mq_get_dispatch_budget(hctx)) |
1201 | break; | |
1202 | ||
8ab6bb9e | 1203 | if (!blk_mq_get_driver_tag(rq)) { |
3c782d67 | 1204 | /* |
da55f2cc | 1205 | * The initial allocation attempt failed, so we need to |
eb619fdb JA |
1206 | * rerun the hardware queue when a tag is freed. The |
1207 | * waitqueue takes care of that. If the queue is run | |
1208 | * before we add this entry back on the dispatch list, | |
1209 | * we'll re-run it below. | |
3c782d67 | 1210 | */ |
2278d69f | 1211 | if (!blk_mq_mark_tag_wait(hctx, rq)) { |
0bca799b | 1212 | blk_mq_put_dispatch_budget(hctx); |
f906a6a0 JA |
1213 | /* |
1214 | * For non-shared tags, the RESTART check | |
1215 | * will suffice. | |
1216 | */ | |
1217 | if (hctx->flags & BLK_MQ_F_TAG_SHARED) | |
1218 | no_tag = true; | |
de148297 ML |
1219 | break; |
1220 | } | |
1221 | } | |
1222 | ||
320ae51f | 1223 | list_del_init(&rq->queuelist); |
320ae51f | 1224 | |
74c45052 | 1225 | bd.rq = rq; |
113285b4 JA |
1226 | |
1227 | /* | |
1228 | * Flag last if we have no more requests, or if we have more | |
1229 | * but can't assign a driver tag to it. | |
1230 | */ | |
1231 | if (list_empty(list)) | |
1232 | bd.last = true; | |
1233 | else { | |
113285b4 | 1234 | nxt = list_first_entry(list, struct request, queuelist); |
8ab6bb9e | 1235 | bd.last = !blk_mq_get_driver_tag(nxt); |
113285b4 | 1236 | } |
74c45052 JA |
1237 | |
1238 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
86ff7c2a | 1239 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) { |
6d6f167c JW |
1240 | /* |
1241 | * If an I/O scheduler has been configured and we got a | |
ff821d27 JA |
1242 | * driver tag for the next request already, free it |
1243 | * again. | |
6d6f167c JW |
1244 | */ |
1245 | if (!list_empty(list)) { | |
1246 | nxt = list_first_entry(list, struct request, queuelist); | |
1247 | blk_mq_put_driver_tag(nxt); | |
1248 | } | |
f04c3df3 | 1249 | list_add(&rq->queuelist, list); |
ed0791b2 | 1250 | __blk_mq_requeue_request(rq); |
320ae51f | 1251 | break; |
fc17b653 CH |
1252 | } |
1253 | ||
1254 | if (unlikely(ret != BLK_STS_OK)) { | |
93efe981 | 1255 | errors++; |
2a842aca | 1256 | blk_mq_end_request(rq, BLK_STS_IOERR); |
fc17b653 | 1257 | continue; |
320ae51f JA |
1258 | } |
1259 | ||
fc17b653 | 1260 | queued++; |
81380ca1 | 1261 | } while (!list_empty(list)); |
320ae51f | 1262 | |
703fd1c0 | 1263 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
1264 | |
1265 | /* | |
1266 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1267 | * that is where we will continue on next queue run. | |
1268 | */ | |
f04c3df3 | 1269 | if (!list_empty(list)) { |
86ff7c2a ML |
1270 | bool needs_restart; |
1271 | ||
d666ba98 JA |
1272 | /* |
1273 | * If we didn't flush the entire list, we could have told | |
1274 | * the driver there was more coming, but that turned out to | |
1275 | * be a lie. | |
1276 | */ | |
1277 | if (q->mq_ops->commit_rqs) | |
1278 | q->mq_ops->commit_rqs(hctx); | |
1279 | ||
320ae51f | 1280 | spin_lock(&hctx->lock); |
c13660a0 | 1281 | list_splice_init(list, &hctx->dispatch); |
320ae51f | 1282 | spin_unlock(&hctx->lock); |
f04c3df3 | 1283 | |
9ba52e58 | 1284 | /* |
710c785f BVA |
1285 | * If SCHED_RESTART was set by the caller of this function and |
1286 | * it is no longer set that means that it was cleared by another | |
1287 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1288 | * |
eb619fdb JA |
1289 | * If 'no_tag' is set, that means that we failed getting |
1290 | * a driver tag with an I/O scheduler attached. If our dispatch | |
1291 | * waitqueue is no longer active, ensure that we run the queue | |
1292 | * AFTER adding our entries back to the list. | |
bd166ef1 | 1293 | * |
710c785f BVA |
1294 | * If no I/O scheduler has been configured it is possible that |
1295 | * the hardware queue got stopped and restarted before requests | |
1296 | * were pushed back onto the dispatch list. Rerun the queue to | |
1297 | * avoid starvation. Notes: | |
1298 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1299 | * been stopped before rerunning a queue. | |
1300 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1301 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1302 | * and dm-rq. |
86ff7c2a ML |
1303 | * |
1304 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
1305 | * bit is set, run queue after a delay to avoid IO stalls | |
1306 | * that could otherwise occur if the queue is idle. | |
bd166ef1 | 1307 | */ |
86ff7c2a ML |
1308 | needs_restart = blk_mq_sched_needs_restart(hctx); |
1309 | if (!needs_restart || | |
eb619fdb | 1310 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
bd166ef1 | 1311 | blk_mq_run_hw_queue(hctx, true); |
86ff7c2a ML |
1312 | else if (needs_restart && (ret == BLK_STS_RESOURCE)) |
1313 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); | |
1f57f8d4 | 1314 | |
6e768717 | 1315 | blk_mq_update_dispatch_busy(hctx, true); |
1f57f8d4 | 1316 | return false; |
6e768717 ML |
1317 | } else |
1318 | blk_mq_update_dispatch_busy(hctx, false); | |
f04c3df3 | 1319 | |
1f57f8d4 JA |
1320 | /* |
1321 | * If the host/device is unable to accept more work, inform the | |
1322 | * caller of that. | |
1323 | */ | |
1324 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) | |
1325 | return false; | |
1326 | ||
93efe981 | 1327 | return (queued + errors) != 0; |
f04c3df3 JA |
1328 | } |
1329 | ||
6a83e74d BVA |
1330 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1331 | { | |
1332 | int srcu_idx; | |
1333 | ||
b7a71e66 JA |
1334 | /* |
1335 | * We should be running this queue from one of the CPUs that | |
1336 | * are mapped to it. | |
7df938fb ML |
1337 | * |
1338 | * There are at least two related races now between setting | |
1339 | * hctx->next_cpu from blk_mq_hctx_next_cpu() and running | |
1340 | * __blk_mq_run_hw_queue(): | |
1341 | * | |
1342 | * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(), | |
1343 | * but later it becomes online, then this warning is harmless | |
1344 | * at all | |
1345 | * | |
1346 | * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(), | |
1347 | * but later it becomes offline, then the warning can't be | |
1348 | * triggered, and we depend on blk-mq timeout handler to | |
1349 | * handle dispatched requests to this hctx | |
b7a71e66 | 1350 | */ |
7df938fb ML |
1351 | if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && |
1352 | cpu_online(hctx->next_cpu)) { | |
1353 | printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n", | |
1354 | raw_smp_processor_id(), | |
1355 | cpumask_empty(hctx->cpumask) ? "inactive": "active"); | |
1356 | dump_stack(); | |
1357 | } | |
6a83e74d | 1358 | |
b7a71e66 JA |
1359 | /* |
1360 | * We can't run the queue inline with ints disabled. Ensure that | |
1361 | * we catch bad users of this early. | |
1362 | */ | |
1363 | WARN_ON_ONCE(in_interrupt()); | |
1364 | ||
04ced159 | 1365 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1366 | |
04ced159 JA |
1367 | hctx_lock(hctx, &srcu_idx); |
1368 | blk_mq_sched_dispatch_requests(hctx); | |
1369 | hctx_unlock(hctx, srcu_idx); | |
6a83e74d BVA |
1370 | } |
1371 | ||
f82ddf19 ML |
1372 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
1373 | { | |
1374 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
1375 | ||
1376 | if (cpu >= nr_cpu_ids) | |
1377 | cpu = cpumask_first(hctx->cpumask); | |
1378 | return cpu; | |
1379 | } | |
1380 | ||
506e931f JA |
1381 | /* |
1382 | * It'd be great if the workqueue API had a way to pass | |
1383 | * in a mask and had some smarts for more clever placement. | |
1384 | * For now we just round-robin here, switching for every | |
1385 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1386 | */ | |
1387 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1388 | { | |
7bed4595 | 1389 | bool tried = false; |
476f8c98 | 1390 | int next_cpu = hctx->next_cpu; |
7bed4595 | 1391 | |
b657d7e6 CH |
1392 | if (hctx->queue->nr_hw_queues == 1) |
1393 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1394 | |
1395 | if (--hctx->next_cpu_batch <= 0) { | |
7bed4595 | 1396 | select_cpu: |
476f8c98 | 1397 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
20e4d813 | 1398 | cpu_online_mask); |
506e931f | 1399 | if (next_cpu >= nr_cpu_ids) |
f82ddf19 | 1400 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
1401 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
1402 | } | |
1403 | ||
7bed4595 ML |
1404 | /* |
1405 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
1406 | * and it should only happen in the path of handling CPU DEAD. | |
1407 | */ | |
476f8c98 | 1408 | if (!cpu_online(next_cpu)) { |
7bed4595 ML |
1409 | if (!tried) { |
1410 | tried = true; | |
1411 | goto select_cpu; | |
1412 | } | |
1413 | ||
1414 | /* | |
1415 | * Make sure to re-select CPU next time once after CPUs | |
1416 | * in hctx->cpumask become online again. | |
1417 | */ | |
476f8c98 | 1418 | hctx->next_cpu = next_cpu; |
7bed4595 ML |
1419 | hctx->next_cpu_batch = 1; |
1420 | return WORK_CPU_UNBOUND; | |
1421 | } | |
476f8c98 ML |
1422 | |
1423 | hctx->next_cpu = next_cpu; | |
1424 | return next_cpu; | |
506e931f JA |
1425 | } |
1426 | ||
7587a5ae BVA |
1427 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1428 | unsigned long msecs) | |
320ae51f | 1429 | { |
5435c023 | 1430 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f JA |
1431 | return; |
1432 | ||
1b792f2f | 1433 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1434 | int cpu = get_cpu(); |
1435 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1436 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1437 | put_cpu(); |
398205b8 PB |
1438 | return; |
1439 | } | |
e4043dcf | 1440 | |
2a90d4aa | 1441 | put_cpu(); |
e4043dcf | 1442 | } |
398205b8 | 1443 | |
ae943d20 BVA |
1444 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
1445 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
1446 | } |
1447 | ||
1448 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1449 | { | |
1450 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1451 | } | |
1452 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1453 | ||
626fb735 | 1454 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 1455 | { |
24f5a90f ML |
1456 | int srcu_idx; |
1457 | bool need_run; | |
1458 | ||
1459 | /* | |
1460 | * When queue is quiesced, we may be switching io scheduler, or | |
1461 | * updating nr_hw_queues, or other things, and we can't run queue | |
1462 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
1463 | * | |
1464 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
1465 | * quiesced. | |
1466 | */ | |
04ced159 JA |
1467 | hctx_lock(hctx, &srcu_idx); |
1468 | need_run = !blk_queue_quiesced(hctx->queue) && | |
1469 | blk_mq_hctx_has_pending(hctx); | |
1470 | hctx_unlock(hctx, srcu_idx); | |
24f5a90f | 1471 | |
626fb735 | 1472 | if (need_run) |
79f720a7 | 1473 | __blk_mq_delay_run_hw_queue(hctx, async, 0); |
320ae51f | 1474 | } |
5b727272 | 1475 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1476 | |
b94ec296 | 1477 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1478 | { |
1479 | struct blk_mq_hw_ctx *hctx; | |
1480 | int i; | |
1481 | ||
1482 | queue_for_each_hw_ctx(q, hctx, i) { | |
79f720a7 | 1483 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1484 | continue; |
1485 | ||
b94ec296 | 1486 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1487 | } |
1488 | } | |
b94ec296 | 1489 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1490 | |
fd001443 BVA |
1491 | /** |
1492 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1493 | * @q: request queue. | |
1494 | * | |
1495 | * The caller is responsible for serializing this function against | |
1496 | * blk_mq_{start,stop}_hw_queue(). | |
1497 | */ | |
1498 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1499 | { | |
1500 | struct blk_mq_hw_ctx *hctx; | |
1501 | int i; | |
1502 | ||
1503 | queue_for_each_hw_ctx(q, hctx, i) | |
1504 | if (blk_mq_hctx_stopped(hctx)) | |
1505 | return true; | |
1506 | ||
1507 | return false; | |
1508 | } | |
1509 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1510 | ||
39a70c76 ML |
1511 | /* |
1512 | * This function is often used for pausing .queue_rq() by driver when | |
1513 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1514 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1515 | * |
1516 | * We do not guarantee that dispatch can be drained or blocked | |
1517 | * after blk_mq_stop_hw_queue() returns. Please use | |
1518 | * blk_mq_quiesce_queue() for that requirement. | |
1519 | */ | |
2719aa21 JA |
1520 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1521 | { | |
641a9ed6 | 1522 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 1523 | |
641a9ed6 | 1524 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 1525 | } |
641a9ed6 | 1526 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 1527 | |
39a70c76 ML |
1528 | /* |
1529 | * This function is often used for pausing .queue_rq() by driver when | |
1530 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1531 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1532 | * |
1533 | * We do not guarantee that dispatch can be drained or blocked | |
1534 | * after blk_mq_stop_hw_queues() returns. Please use | |
1535 | * blk_mq_quiesce_queue() for that requirement. | |
1536 | */ | |
2719aa21 JA |
1537 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1538 | { | |
641a9ed6 ML |
1539 | struct blk_mq_hw_ctx *hctx; |
1540 | int i; | |
1541 | ||
1542 | queue_for_each_hw_ctx(q, hctx, i) | |
1543 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
1544 | } |
1545 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1546 | ||
320ae51f JA |
1547 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1548 | { | |
1549 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1550 | |
0ffbce80 | 1551 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1552 | } |
1553 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1554 | ||
2f268556 CH |
1555 | void blk_mq_start_hw_queues(struct request_queue *q) |
1556 | { | |
1557 | struct blk_mq_hw_ctx *hctx; | |
1558 | int i; | |
1559 | ||
1560 | queue_for_each_hw_ctx(q, hctx, i) | |
1561 | blk_mq_start_hw_queue(hctx); | |
1562 | } | |
1563 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1564 | ||
ae911c5e JA |
1565 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1566 | { | |
1567 | if (!blk_mq_hctx_stopped(hctx)) | |
1568 | return; | |
1569 | ||
1570 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1571 | blk_mq_run_hw_queue(hctx, async); | |
1572 | } | |
1573 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1574 | ||
1b4a3258 | 1575 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1576 | { |
1577 | struct blk_mq_hw_ctx *hctx; | |
1578 | int i; | |
1579 | ||
ae911c5e JA |
1580 | queue_for_each_hw_ctx(q, hctx, i) |
1581 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1582 | } |
1583 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1584 | ||
70f4db63 | 1585 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1586 | { |
1587 | struct blk_mq_hw_ctx *hctx; | |
1588 | ||
9f993737 | 1589 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 1590 | |
21c6e939 | 1591 | /* |
15fe8a90 | 1592 | * If we are stopped, don't run the queue. |
21c6e939 | 1593 | */ |
15fe8a90 | 1594 | if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
0196d6b4 | 1595 | return; |
7587a5ae BVA |
1596 | |
1597 | __blk_mq_run_hw_queue(hctx); | |
1598 | } | |
1599 | ||
cfd0c552 | 1600 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1601 | struct request *rq, |
1602 | bool at_head) | |
320ae51f | 1603 | { |
e57690fe | 1604 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
c16d6b5a | 1605 | enum hctx_type type = hctx->type; |
e57690fe | 1606 | |
7b607814 BVA |
1607 | lockdep_assert_held(&ctx->lock); |
1608 | ||
01b983c9 JA |
1609 | trace_block_rq_insert(hctx->queue, rq); |
1610 | ||
72a0a36e | 1611 | if (at_head) |
c16d6b5a | 1612 | list_add(&rq->queuelist, &ctx->rq_lists[type]); |
72a0a36e | 1613 | else |
c16d6b5a | 1614 | list_add_tail(&rq->queuelist, &ctx->rq_lists[type]); |
cfd0c552 | 1615 | } |
4bb659b1 | 1616 | |
2c3ad667 JA |
1617 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1618 | bool at_head) | |
cfd0c552 ML |
1619 | { |
1620 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1621 | ||
7b607814 BVA |
1622 | lockdep_assert_held(&ctx->lock); |
1623 | ||
e57690fe | 1624 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1625 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1626 | } |
1627 | ||
157f377b JA |
1628 | /* |
1629 | * Should only be used carefully, when the caller knows we want to | |
1630 | * bypass a potential IO scheduler on the target device. | |
1631 | */ | |
b0850297 | 1632 | void blk_mq_request_bypass_insert(struct request *rq, bool run_queue) |
157f377b | 1633 | { |
ea4f995e | 1634 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
157f377b JA |
1635 | |
1636 | spin_lock(&hctx->lock); | |
1637 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
1638 | spin_unlock(&hctx->lock); | |
1639 | ||
b0850297 ML |
1640 | if (run_queue) |
1641 | blk_mq_run_hw_queue(hctx, false); | |
157f377b JA |
1642 | } |
1643 | ||
bd166ef1 JA |
1644 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1645 | struct list_head *list) | |
320ae51f JA |
1646 | |
1647 | { | |
3f0cedc7 | 1648 | struct request *rq; |
c16d6b5a | 1649 | enum hctx_type type = hctx->type; |
3f0cedc7 | 1650 | |
320ae51f JA |
1651 | /* |
1652 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1653 | * offline now | |
1654 | */ | |
3f0cedc7 | 1655 | list_for_each_entry(rq, list, queuelist) { |
e57690fe | 1656 | BUG_ON(rq->mq_ctx != ctx); |
3f0cedc7 | 1657 | trace_block_rq_insert(hctx->queue, rq); |
320ae51f | 1658 | } |
3f0cedc7 ML |
1659 | |
1660 | spin_lock(&ctx->lock); | |
c16d6b5a | 1661 | list_splice_tail_init(list, &ctx->rq_lists[type]); |
cfd0c552 | 1662 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1663 | spin_unlock(&ctx->lock); |
320ae51f JA |
1664 | } |
1665 | ||
3110fc79 | 1666 | static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b) |
320ae51f JA |
1667 | { |
1668 | struct request *rqa = container_of(a, struct request, queuelist); | |
1669 | struct request *rqb = container_of(b, struct request, queuelist); | |
1670 | ||
3110fc79 JA |
1671 | if (rqa->mq_ctx < rqb->mq_ctx) |
1672 | return -1; | |
1673 | else if (rqa->mq_ctx > rqb->mq_ctx) | |
1674 | return 1; | |
1675 | else if (rqa->mq_hctx < rqb->mq_hctx) | |
1676 | return -1; | |
1677 | else if (rqa->mq_hctx > rqb->mq_hctx) | |
1678 | return 1; | |
1679 | ||
1680 | return blk_rq_pos(rqa) > blk_rq_pos(rqb); | |
320ae51f JA |
1681 | } |
1682 | ||
1683 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1684 | { | |
67cae4c9 | 1685 | struct blk_mq_hw_ctx *this_hctx; |
320ae51f JA |
1686 | struct blk_mq_ctx *this_ctx; |
1687 | struct request_queue *this_q; | |
1688 | struct request *rq; | |
1689 | LIST_HEAD(list); | |
67cae4c9 | 1690 | LIST_HEAD(rq_list); |
320ae51f JA |
1691 | unsigned int depth; |
1692 | ||
1693 | list_splice_init(&plug->mq_list, &list); | |
1694 | ||
ce5b009c JA |
1695 | if (plug->rq_count > 2 && plug->multiple_queues) |
1696 | list_sort(NULL, &list, plug_rq_cmp); | |
320ae51f | 1697 | |
bcc816df DZ |
1698 | plug->rq_count = 0; |
1699 | ||
320ae51f | 1700 | this_q = NULL; |
67cae4c9 | 1701 | this_hctx = NULL; |
320ae51f JA |
1702 | this_ctx = NULL; |
1703 | depth = 0; | |
1704 | ||
1705 | while (!list_empty(&list)) { | |
1706 | rq = list_entry_rq(list.next); | |
1707 | list_del_init(&rq->queuelist); | |
1708 | BUG_ON(!rq->q); | |
67cae4c9 JA |
1709 | if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) { |
1710 | if (this_hctx) { | |
587562d0 | 1711 | trace_block_unplug(this_q, depth, !from_schedule); |
67cae4c9 JA |
1712 | blk_mq_sched_insert_requests(this_hctx, this_ctx, |
1713 | &rq_list, | |
bd166ef1 | 1714 | from_schedule); |
320ae51f JA |
1715 | } |
1716 | ||
320ae51f | 1717 | this_q = rq->q; |
67cae4c9 JA |
1718 | this_ctx = rq->mq_ctx; |
1719 | this_hctx = rq->mq_hctx; | |
320ae51f JA |
1720 | depth = 0; |
1721 | } | |
1722 | ||
1723 | depth++; | |
67cae4c9 | 1724 | list_add_tail(&rq->queuelist, &rq_list); |
320ae51f JA |
1725 | } |
1726 | ||
1727 | /* | |
67cae4c9 JA |
1728 | * If 'this_hctx' is set, we know we have entries to complete |
1729 | * on 'rq_list'. Do those. | |
320ae51f | 1730 | */ |
67cae4c9 | 1731 | if (this_hctx) { |
587562d0 | 1732 | trace_block_unplug(this_q, depth, !from_schedule); |
67cae4c9 | 1733 | blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list, |
bd166ef1 | 1734 | from_schedule); |
320ae51f JA |
1735 | } |
1736 | } | |
1737 | ||
14ccb66b CH |
1738 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio, |
1739 | unsigned int nr_segs) | |
320ae51f | 1740 | { |
f924cdde CH |
1741 | if (bio->bi_opf & REQ_RAHEAD) |
1742 | rq->cmd_flags |= REQ_FAILFAST_MASK; | |
1743 | ||
1744 | rq->__sector = bio->bi_iter.bi_sector; | |
1745 | rq->write_hint = bio->bi_write_hint; | |
14ccb66b | 1746 | blk_rq_bio_prep(rq, bio, nr_segs); |
4b570521 | 1747 | |
6e85eaf3 | 1748 | blk_account_io_start(rq, true); |
320ae51f JA |
1749 | } |
1750 | ||
0f95549c MS |
1751 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
1752 | struct request *rq, | |
be94f058 | 1753 | blk_qc_t *cookie, bool last) |
f984df1f | 1754 | { |
f984df1f | 1755 | struct request_queue *q = rq->q; |
f984df1f SL |
1756 | struct blk_mq_queue_data bd = { |
1757 | .rq = rq, | |
be94f058 | 1758 | .last = last, |
f984df1f | 1759 | }; |
bd166ef1 | 1760 | blk_qc_t new_cookie; |
f06345ad | 1761 | blk_status_t ret; |
0f95549c MS |
1762 | |
1763 | new_cookie = request_to_qc_t(hctx, rq); | |
1764 | ||
1765 | /* | |
1766 | * For OK queue, we are done. For error, caller may kill it. | |
1767 | * Any other error (busy), just add it to our list as we | |
1768 | * previously would have done. | |
1769 | */ | |
1770 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
1771 | switch (ret) { | |
1772 | case BLK_STS_OK: | |
6ce3dd6e | 1773 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
1774 | *cookie = new_cookie; |
1775 | break; | |
1776 | case BLK_STS_RESOURCE: | |
86ff7c2a | 1777 | case BLK_STS_DEV_RESOURCE: |
6ce3dd6e | 1778 | blk_mq_update_dispatch_busy(hctx, true); |
0f95549c MS |
1779 | __blk_mq_requeue_request(rq); |
1780 | break; | |
1781 | default: | |
6ce3dd6e | 1782 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
1783 | *cookie = BLK_QC_T_NONE; |
1784 | break; | |
1785 | } | |
1786 | ||
1787 | return ret; | |
1788 | } | |
1789 | ||
fd9c40f6 | 1790 | static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
0f95549c | 1791 | struct request *rq, |
396eaf21 | 1792 | blk_qc_t *cookie, |
fd9c40f6 | 1793 | bool bypass_insert, bool last) |
0f95549c MS |
1794 | { |
1795 | struct request_queue *q = rq->q; | |
d964f04a ML |
1796 | bool run_queue = true; |
1797 | ||
23d4ee19 | 1798 | /* |
fd9c40f6 | 1799 | * RCU or SRCU read lock is needed before checking quiesced flag. |
23d4ee19 | 1800 | * |
fd9c40f6 BVA |
1801 | * When queue is stopped or quiesced, ignore 'bypass_insert' from |
1802 | * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller, | |
1803 | * and avoid driver to try to dispatch again. | |
23d4ee19 | 1804 | */ |
fd9c40f6 | 1805 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { |
d964f04a | 1806 | run_queue = false; |
fd9c40f6 BVA |
1807 | bypass_insert = false; |
1808 | goto insert; | |
d964f04a | 1809 | } |
f984df1f | 1810 | |
fd9c40f6 BVA |
1811 | if (q->elevator && !bypass_insert) |
1812 | goto insert; | |
2253efc8 | 1813 | |
0bca799b | 1814 | if (!blk_mq_get_dispatch_budget(hctx)) |
fd9c40f6 | 1815 | goto insert; |
bd166ef1 | 1816 | |
8ab6bb9e | 1817 | if (!blk_mq_get_driver_tag(rq)) { |
0bca799b | 1818 | blk_mq_put_dispatch_budget(hctx); |
fd9c40f6 | 1819 | goto insert; |
88022d72 | 1820 | } |
de148297 | 1821 | |
fd9c40f6 BVA |
1822 | return __blk_mq_issue_directly(hctx, rq, cookie, last); |
1823 | insert: | |
1824 | if (bypass_insert) | |
1825 | return BLK_STS_RESOURCE; | |
1826 | ||
1827 | blk_mq_request_bypass_insert(rq, run_queue); | |
1828 | return BLK_STS_OK; | |
1829 | } | |
1830 | ||
1831 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, | |
1832 | struct request *rq, blk_qc_t *cookie) | |
1833 | { | |
1834 | blk_status_t ret; | |
1835 | int srcu_idx; | |
1836 | ||
1837 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); | |
1838 | ||
1839 | hctx_lock(hctx, &srcu_idx); | |
1840 | ||
1841 | ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true); | |
1842 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) | |
1843 | blk_mq_request_bypass_insert(rq, true); | |
1844 | else if (ret != BLK_STS_OK) | |
1845 | blk_mq_end_request(rq, ret); | |
1846 | ||
1847 | hctx_unlock(hctx, srcu_idx); | |
1848 | } | |
1849 | ||
1850 | blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last) | |
1851 | { | |
1852 | blk_status_t ret; | |
1853 | int srcu_idx; | |
1854 | blk_qc_t unused_cookie; | |
1855 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
1856 | ||
1857 | hctx_lock(hctx, &srcu_idx); | |
1858 | ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last); | |
04ced159 | 1859 | hctx_unlock(hctx, srcu_idx); |
7f556a44 JW |
1860 | |
1861 | return ret; | |
5eb6126e CH |
1862 | } |
1863 | ||
6ce3dd6e ML |
1864 | void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
1865 | struct list_head *list) | |
1866 | { | |
1867 | while (!list_empty(list)) { | |
fd9c40f6 | 1868 | blk_status_t ret; |
6ce3dd6e ML |
1869 | struct request *rq = list_first_entry(list, struct request, |
1870 | queuelist); | |
1871 | ||
1872 | list_del_init(&rq->queuelist); | |
fd9c40f6 BVA |
1873 | ret = blk_mq_request_issue_directly(rq, list_empty(list)); |
1874 | if (ret != BLK_STS_OK) { | |
1875 | if (ret == BLK_STS_RESOURCE || | |
1876 | ret == BLK_STS_DEV_RESOURCE) { | |
1877 | blk_mq_request_bypass_insert(rq, | |
c616cbee | 1878 | list_empty(list)); |
fd9c40f6 BVA |
1879 | break; |
1880 | } | |
1881 | blk_mq_end_request(rq, ret); | |
1882 | } | |
6ce3dd6e | 1883 | } |
d666ba98 JA |
1884 | |
1885 | /* | |
1886 | * If we didn't flush the entire list, we could have told | |
1887 | * the driver there was more coming, but that turned out to | |
1888 | * be a lie. | |
1889 | */ | |
fd9c40f6 | 1890 | if (!list_empty(list) && hctx->queue->mq_ops->commit_rqs) |
d666ba98 | 1891 | hctx->queue->mq_ops->commit_rqs(hctx); |
6ce3dd6e ML |
1892 | } |
1893 | ||
ce5b009c JA |
1894 | static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq) |
1895 | { | |
1896 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1897 | plug->rq_count++; | |
1898 | if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) { | |
1899 | struct request *tmp; | |
1900 | ||
1901 | tmp = list_first_entry(&plug->mq_list, struct request, | |
1902 | queuelist); | |
1903 | if (tmp->q != rq->q) | |
1904 | plug->multiple_queues = true; | |
1905 | } | |
1906 | } | |
1907 | ||
dece1635 | 1908 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1909 | { |
ef295ecf | 1910 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1911 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
7809167d | 1912 | struct blk_mq_alloc_data data = { .flags = 0}; |
07068d5b | 1913 | struct request *rq; |
f984df1f | 1914 | struct blk_plug *plug; |
5b3f341f | 1915 | struct request *same_queue_rq = NULL; |
14ccb66b | 1916 | unsigned int nr_segs; |
7b371636 | 1917 | blk_qc_t cookie; |
07068d5b JA |
1918 | |
1919 | blk_queue_bounce(q, &bio); | |
14ccb66b | 1920 | __blk_queue_split(q, &bio, &nr_segs); |
f36ea50c | 1921 | |
e23947bd | 1922 | if (!bio_integrity_prep(bio)) |
dece1635 | 1923 | return BLK_QC_T_NONE; |
07068d5b | 1924 | |
87c279e6 | 1925 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
14ccb66b | 1926 | blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq)) |
87c279e6 | 1927 | return BLK_QC_T_NONE; |
f984df1f | 1928 | |
14ccb66b | 1929 | if (blk_mq_sched_bio_merge(q, bio, nr_segs)) |
bd166ef1 JA |
1930 | return BLK_QC_T_NONE; |
1931 | ||
d5337560 | 1932 | rq_qos_throttle(q, bio); |
87760e5e | 1933 | |
7809167d | 1934 | data.cmd_flags = bio->bi_opf; |
f9afca4d | 1935 | rq = blk_mq_get_request(q, bio, &data); |
87760e5e | 1936 | if (unlikely(!rq)) { |
c1c80384 | 1937 | rq_qos_cleanup(q, bio); |
7b6620d7 | 1938 | if (bio->bi_opf & REQ_NOWAIT) |
03a07c92 | 1939 | bio_wouldblock_error(bio); |
7b6620d7 | 1940 | return BLK_QC_T_NONE; |
87760e5e JA |
1941 | } |
1942 | ||
d6f1dda2 XW |
1943 | trace_block_getrq(q, bio, bio->bi_opf); |
1944 | ||
c1c80384 | 1945 | rq_qos_track(q, rq, bio); |
07068d5b | 1946 | |
fd2d3326 | 1947 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b | 1948 | |
970d168d BVA |
1949 | blk_mq_bio_to_request(rq, bio, nr_segs); |
1950 | ||
b49773e7 | 1951 | plug = blk_mq_plug(q, bio); |
07068d5b | 1952 | if (unlikely(is_flush_fua)) { |
923218f6 ML |
1953 | /* bypass scheduler for flush rq */ |
1954 | blk_insert_flush(rq); | |
1955 | blk_mq_run_hw_queue(data.hctx, true); | |
3154df26 ML |
1956 | } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs || |
1957 | !blk_queue_nonrot(q))) { | |
b2c5d16b JA |
1958 | /* |
1959 | * Use plugging if we have a ->commit_rqs() hook as well, as | |
1960 | * we know the driver uses bd->last in a smart fashion. | |
3154df26 ML |
1961 | * |
1962 | * Use normal plugging if this disk is slow HDD, as sequential | |
1963 | * IO may benefit a lot from plug merging. | |
b2c5d16b | 1964 | */ |
5f0ed774 | 1965 | unsigned int request_count = plug->rq_count; |
600271d9 SL |
1966 | struct request *last = NULL; |
1967 | ||
676d0607 | 1968 | if (!request_count) |
e6c4438b | 1969 | trace_block_plug(q); |
600271d9 SL |
1970 | else |
1971 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c | 1972 | |
600271d9 SL |
1973 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1974 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1975 | blk_flush_plug_list(plug, false); |
1976 | trace_block_plug(q); | |
320ae51f | 1977 | } |
b094f89c | 1978 | |
ce5b009c | 1979 | blk_add_rq_to_plug(plug, rq); |
a12de1d4 ML |
1980 | } else if (q->elevator) { |
1981 | blk_mq_sched_insert_request(rq, false, true, true); | |
2299722c | 1982 | } else if (plug && !blk_queue_nomerges(q)) { |
07068d5b | 1983 | /* |
6a83e74d | 1984 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1985 | * Otherwise the existing request in the plug list will be |
1986 | * issued. So the plug list will have one request at most | |
2299722c CH |
1987 | * The plug list might get flushed before this. If that happens, |
1988 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 1989 | */ |
2299722c CH |
1990 | if (list_empty(&plug->mq_list)) |
1991 | same_queue_rq = NULL; | |
4711b573 | 1992 | if (same_queue_rq) { |
2299722c | 1993 | list_del_init(&same_queue_rq->queuelist); |
4711b573 JA |
1994 | plug->rq_count--; |
1995 | } | |
ce5b009c | 1996 | blk_add_rq_to_plug(plug, rq); |
ff3b74b8 | 1997 | trace_block_plug(q); |
2299722c | 1998 | |
dad7a3be | 1999 | if (same_queue_rq) { |
ea4f995e | 2000 | data.hctx = same_queue_rq->mq_hctx; |
ff3b74b8 | 2001 | trace_block_unplug(q, 1, true); |
2299722c | 2002 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, |
fd9c40f6 | 2003 | &cookie); |
dad7a3be | 2004 | } |
a12de1d4 ML |
2005 | } else if ((q->nr_hw_queues > 1 && is_sync) || |
2006 | !data.hctx->dispatch_busy) { | |
fd9c40f6 | 2007 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); |
ab42f35d | 2008 | } else { |
8fa9f556 | 2009 | blk_mq_sched_insert_request(rq, false, true, true); |
ab42f35d | 2010 | } |
320ae51f | 2011 | |
7b371636 | 2012 | return cookie; |
320ae51f JA |
2013 | } |
2014 | ||
cc71a6f4 JA |
2015 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2016 | unsigned int hctx_idx) | |
95363efd | 2017 | { |
e9b267d9 | 2018 | struct page *page; |
320ae51f | 2019 | |
24d2f903 | 2020 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 2021 | int i; |
320ae51f | 2022 | |
24d2f903 | 2023 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
2024 | struct request *rq = tags->static_rqs[i]; |
2025 | ||
2026 | if (!rq) | |
e9b267d9 | 2027 | continue; |
d6296d39 | 2028 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 2029 | tags->static_rqs[i] = NULL; |
e9b267d9 | 2030 | } |
320ae51f | 2031 | } |
320ae51f | 2032 | |
24d2f903 CH |
2033 | while (!list_empty(&tags->page_list)) { |
2034 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 2035 | list_del_init(&page->lru); |
f75782e4 CM |
2036 | /* |
2037 | * Remove kmemleak object previously allocated in | |
273938bf | 2038 | * blk_mq_alloc_rqs(). |
f75782e4 CM |
2039 | */ |
2040 | kmemleak_free(page_address(page)); | |
320ae51f JA |
2041 | __free_pages(page, page->private); |
2042 | } | |
cc71a6f4 | 2043 | } |
320ae51f | 2044 | |
cc71a6f4 JA |
2045 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
2046 | { | |
24d2f903 | 2047 | kfree(tags->rqs); |
cc71a6f4 | 2048 | tags->rqs = NULL; |
2af8cbe3 JA |
2049 | kfree(tags->static_rqs); |
2050 | tags->static_rqs = NULL; | |
320ae51f | 2051 | |
24d2f903 | 2052 | blk_mq_free_tags(tags); |
320ae51f JA |
2053 | } |
2054 | ||
cc71a6f4 JA |
2055 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
2056 | unsigned int hctx_idx, | |
2057 | unsigned int nr_tags, | |
2058 | unsigned int reserved_tags) | |
320ae51f | 2059 | { |
24d2f903 | 2060 | struct blk_mq_tags *tags; |
59f082e4 | 2061 | int node; |
320ae51f | 2062 | |
7d76f856 | 2063 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx); |
59f082e4 SL |
2064 | if (node == NUMA_NO_NODE) |
2065 | node = set->numa_node; | |
2066 | ||
2067 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, | |
24391c0d | 2068 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
24d2f903 CH |
2069 | if (!tags) |
2070 | return NULL; | |
320ae51f | 2071 | |
590b5b7d | 2072 | tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
36e1f3d1 | 2073 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 2074 | node); |
24d2f903 CH |
2075 | if (!tags->rqs) { |
2076 | blk_mq_free_tags(tags); | |
2077 | return NULL; | |
2078 | } | |
320ae51f | 2079 | |
590b5b7d KC |
2080 | tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
2081 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
2082 | node); | |
2af8cbe3 JA |
2083 | if (!tags->static_rqs) { |
2084 | kfree(tags->rqs); | |
2085 | blk_mq_free_tags(tags); | |
2086 | return NULL; | |
2087 | } | |
2088 | ||
cc71a6f4 JA |
2089 | return tags; |
2090 | } | |
2091 | ||
2092 | static size_t order_to_size(unsigned int order) | |
2093 | { | |
2094 | return (size_t)PAGE_SIZE << order; | |
2095 | } | |
2096 | ||
1d9bd516 TH |
2097 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
2098 | unsigned int hctx_idx, int node) | |
2099 | { | |
2100 | int ret; | |
2101 | ||
2102 | if (set->ops->init_request) { | |
2103 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
2104 | if (ret) | |
2105 | return ret; | |
2106 | } | |
2107 | ||
12f5b931 | 2108 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
1d9bd516 TH |
2109 | return 0; |
2110 | } | |
2111 | ||
cc71a6f4 JA |
2112 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2113 | unsigned int hctx_idx, unsigned int depth) | |
2114 | { | |
2115 | unsigned int i, j, entries_per_page, max_order = 4; | |
2116 | size_t rq_size, left; | |
59f082e4 SL |
2117 | int node; |
2118 | ||
7d76f856 | 2119 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx); |
59f082e4 SL |
2120 | if (node == NUMA_NO_NODE) |
2121 | node = set->numa_node; | |
cc71a6f4 JA |
2122 | |
2123 | INIT_LIST_HEAD(&tags->page_list); | |
2124 | ||
320ae51f JA |
2125 | /* |
2126 | * rq_size is the size of the request plus driver payload, rounded | |
2127 | * to the cacheline size | |
2128 | */ | |
24d2f903 | 2129 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 2130 | cache_line_size()); |
cc71a6f4 | 2131 | left = rq_size * depth; |
320ae51f | 2132 | |
cc71a6f4 | 2133 | for (i = 0; i < depth; ) { |
320ae51f JA |
2134 | int this_order = max_order; |
2135 | struct page *page; | |
2136 | int to_do; | |
2137 | void *p; | |
2138 | ||
b3a834b1 | 2139 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
2140 | this_order--; |
2141 | ||
2142 | do { | |
59f082e4 | 2143 | page = alloc_pages_node(node, |
36e1f3d1 | 2144 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 2145 | this_order); |
320ae51f JA |
2146 | if (page) |
2147 | break; | |
2148 | if (!this_order--) | |
2149 | break; | |
2150 | if (order_to_size(this_order) < rq_size) | |
2151 | break; | |
2152 | } while (1); | |
2153 | ||
2154 | if (!page) | |
24d2f903 | 2155 | goto fail; |
320ae51f JA |
2156 | |
2157 | page->private = this_order; | |
24d2f903 | 2158 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
2159 | |
2160 | p = page_address(page); | |
f75782e4 CM |
2161 | /* |
2162 | * Allow kmemleak to scan these pages as they contain pointers | |
2163 | * to additional allocations like via ops->init_request(). | |
2164 | */ | |
36e1f3d1 | 2165 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 2166 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 2167 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
2168 | left -= to_do * rq_size; |
2169 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
2170 | struct request *rq = p; |
2171 | ||
2172 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
2173 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
2174 | tags->static_rqs[i] = NULL; | |
2175 | goto fail; | |
e9b267d9 CH |
2176 | } |
2177 | ||
320ae51f JA |
2178 | p += rq_size; |
2179 | i++; | |
2180 | } | |
2181 | } | |
cc71a6f4 | 2182 | return 0; |
320ae51f | 2183 | |
24d2f903 | 2184 | fail: |
cc71a6f4 JA |
2185 | blk_mq_free_rqs(set, tags, hctx_idx); |
2186 | return -ENOMEM; | |
320ae51f JA |
2187 | } |
2188 | ||
e57690fe JA |
2189 | /* |
2190 | * 'cpu' is going away. splice any existing rq_list entries from this | |
2191 | * software queue to the hw queue dispatch list, and ensure that it | |
2192 | * gets run. | |
2193 | */ | |
9467f859 | 2194 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 2195 | { |
9467f859 | 2196 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
2197 | struct blk_mq_ctx *ctx; |
2198 | LIST_HEAD(tmp); | |
c16d6b5a | 2199 | enum hctx_type type; |
484b4061 | 2200 | |
9467f859 | 2201 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 2202 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
c16d6b5a | 2203 | type = hctx->type; |
484b4061 JA |
2204 | |
2205 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
2206 | if (!list_empty(&ctx->rq_lists[type])) { |
2207 | list_splice_init(&ctx->rq_lists[type], &tmp); | |
484b4061 JA |
2208 | blk_mq_hctx_clear_pending(hctx, ctx); |
2209 | } | |
2210 | spin_unlock(&ctx->lock); | |
2211 | ||
2212 | if (list_empty(&tmp)) | |
9467f859 | 2213 | return 0; |
484b4061 | 2214 | |
e57690fe JA |
2215 | spin_lock(&hctx->lock); |
2216 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
2217 | spin_unlock(&hctx->lock); | |
484b4061 JA |
2218 | |
2219 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 2220 | return 0; |
484b4061 JA |
2221 | } |
2222 | ||
9467f859 | 2223 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 2224 | { |
9467f859 TG |
2225 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
2226 | &hctx->cpuhp_dead); | |
484b4061 JA |
2227 | } |
2228 | ||
c3b4afca | 2229 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
2230 | static void blk_mq_exit_hctx(struct request_queue *q, |
2231 | struct blk_mq_tag_set *set, | |
2232 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
2233 | { | |
8ab0b7dc ML |
2234 | if (blk_mq_hw_queue_mapped(hctx)) |
2235 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 2236 | |
f70ced09 | 2237 | if (set->ops->exit_request) |
d6296d39 | 2238 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); |
f70ced09 | 2239 | |
08e98fc6 ML |
2240 | if (set->ops->exit_hctx) |
2241 | set->ops->exit_hctx(hctx, hctx_idx); | |
2242 | ||
9467f859 | 2243 | blk_mq_remove_cpuhp(hctx); |
2f8f1336 ML |
2244 | |
2245 | spin_lock(&q->unused_hctx_lock); | |
2246 | list_add(&hctx->hctx_list, &q->unused_hctx_list); | |
2247 | spin_unlock(&q->unused_hctx_lock); | |
08e98fc6 ML |
2248 | } |
2249 | ||
624dbe47 ML |
2250 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
2251 | struct blk_mq_tag_set *set, int nr_queue) | |
2252 | { | |
2253 | struct blk_mq_hw_ctx *hctx; | |
2254 | unsigned int i; | |
2255 | ||
2256 | queue_for_each_hw_ctx(q, hctx, i) { | |
2257 | if (i == nr_queue) | |
2258 | break; | |
477e19de | 2259 | blk_mq_debugfs_unregister_hctx(hctx); |
08e98fc6 | 2260 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 2261 | } |
624dbe47 ML |
2262 | } |
2263 | ||
7c6c5b7c ML |
2264 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) |
2265 | { | |
2266 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2267 | ||
2268 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu), | |
2269 | __alignof__(struct blk_mq_hw_ctx)) != | |
2270 | sizeof(struct blk_mq_hw_ctx)); | |
2271 | ||
2272 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2273 | hw_ctx_size += sizeof(struct srcu_struct); | |
2274 | ||
2275 | return hw_ctx_size; | |
2276 | } | |
2277 | ||
08e98fc6 ML |
2278 | static int blk_mq_init_hctx(struct request_queue *q, |
2279 | struct blk_mq_tag_set *set, | |
2280 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 2281 | { |
7c6c5b7c ML |
2282 | hctx->queue_num = hctx_idx; |
2283 | ||
2284 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); | |
2285 | ||
2286 | hctx->tags = set->tags[hctx_idx]; | |
2287 | ||
2288 | if (set->ops->init_hctx && | |
2289 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
2290 | goto unregister_cpu_notifier; | |
08e98fc6 | 2291 | |
7c6c5b7c ML |
2292 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, |
2293 | hctx->numa_node)) | |
2294 | goto exit_hctx; | |
2295 | return 0; | |
2296 | ||
2297 | exit_hctx: | |
2298 | if (set->ops->exit_hctx) | |
2299 | set->ops->exit_hctx(hctx, hctx_idx); | |
2300 | unregister_cpu_notifier: | |
2301 | blk_mq_remove_cpuhp(hctx); | |
2302 | return -1; | |
2303 | } | |
2304 | ||
2305 | static struct blk_mq_hw_ctx * | |
2306 | blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set, | |
2307 | int node) | |
2308 | { | |
2309 | struct blk_mq_hw_ctx *hctx; | |
2310 | gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY; | |
2311 | ||
2312 | hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node); | |
2313 | if (!hctx) | |
2314 | goto fail_alloc_hctx; | |
2315 | ||
2316 | if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node)) | |
2317 | goto free_hctx; | |
2318 | ||
2319 | atomic_set(&hctx->nr_active, 0); | |
08e98fc6 | 2320 | if (node == NUMA_NO_NODE) |
7c6c5b7c ML |
2321 | node = set->numa_node; |
2322 | hctx->numa_node = node; | |
08e98fc6 | 2323 | |
9f993737 | 2324 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
2325 | spin_lock_init(&hctx->lock); |
2326 | INIT_LIST_HEAD(&hctx->dispatch); | |
2327 | hctx->queue = q; | |
2404e607 | 2328 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 2329 | |
2f8f1336 ML |
2330 | INIT_LIST_HEAD(&hctx->hctx_list); |
2331 | ||
320ae51f | 2332 | /* |
08e98fc6 ML |
2333 | * Allocate space for all possible cpus to avoid allocation at |
2334 | * runtime | |
320ae51f | 2335 | */ |
d904bfa7 | 2336 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
7c6c5b7c | 2337 | gfp, node); |
08e98fc6 | 2338 | if (!hctx->ctxs) |
7c6c5b7c | 2339 | goto free_cpumask; |
320ae51f | 2340 | |
5b202853 | 2341 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), |
7c6c5b7c | 2342 | gfp, node)) |
08e98fc6 | 2343 | goto free_ctxs; |
08e98fc6 | 2344 | hctx->nr_ctx = 0; |
320ae51f | 2345 | |
5815839b | 2346 | spin_lock_init(&hctx->dispatch_wait_lock); |
eb619fdb JA |
2347 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
2348 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
2349 | ||
5b202853 | 2350 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size, |
7c6c5b7c | 2351 | gfp); |
f70ced09 | 2352 | if (!hctx->fq) |
7c6c5b7c | 2353 | goto free_bitmap; |
320ae51f | 2354 | |
6a83e74d | 2355 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2356 | init_srcu_struct(hctx->srcu); |
7c6c5b7c | 2357 | blk_mq_hctx_kobj_init(hctx); |
6a83e74d | 2358 | |
7c6c5b7c | 2359 | return hctx; |
320ae51f | 2360 | |
08e98fc6 | 2361 | free_bitmap: |
88459642 | 2362 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2363 | free_ctxs: |
2364 | kfree(hctx->ctxs); | |
7c6c5b7c ML |
2365 | free_cpumask: |
2366 | free_cpumask_var(hctx->cpumask); | |
2367 | free_hctx: | |
2368 | kfree(hctx); | |
2369 | fail_alloc_hctx: | |
2370 | return NULL; | |
08e98fc6 | 2371 | } |
320ae51f | 2372 | |
320ae51f JA |
2373 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
2374 | unsigned int nr_hw_queues) | |
2375 | { | |
b3c661b1 JA |
2376 | struct blk_mq_tag_set *set = q->tag_set; |
2377 | unsigned int i, j; | |
320ae51f JA |
2378 | |
2379 | for_each_possible_cpu(i) { | |
2380 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2381 | struct blk_mq_hw_ctx *hctx; | |
c16d6b5a | 2382 | int k; |
320ae51f | 2383 | |
320ae51f JA |
2384 | __ctx->cpu = i; |
2385 | spin_lock_init(&__ctx->lock); | |
c16d6b5a ML |
2386 | for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++) |
2387 | INIT_LIST_HEAD(&__ctx->rq_lists[k]); | |
2388 | ||
320ae51f JA |
2389 | __ctx->queue = q; |
2390 | ||
320ae51f JA |
2391 | /* |
2392 | * Set local node, IFF we have more than one hw queue. If | |
2393 | * not, we remain on the home node of the device | |
2394 | */ | |
b3c661b1 JA |
2395 | for (j = 0; j < set->nr_maps; j++) { |
2396 | hctx = blk_mq_map_queue_type(q, j, i); | |
2397 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
2398 | hctx->numa_node = local_memory_node(cpu_to_node(i)); | |
2399 | } | |
320ae51f JA |
2400 | } |
2401 | } | |
2402 | ||
cc71a6f4 JA |
2403 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
2404 | { | |
2405 | int ret = 0; | |
2406 | ||
2407 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
2408 | set->queue_depth, set->reserved_tags); | |
2409 | if (!set->tags[hctx_idx]) | |
2410 | return false; | |
2411 | ||
2412 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2413 | set->queue_depth); | |
2414 | if (!ret) | |
2415 | return true; | |
2416 | ||
2417 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2418 | set->tags[hctx_idx] = NULL; | |
2419 | return false; | |
2420 | } | |
2421 | ||
2422 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2423 | unsigned int hctx_idx) | |
2424 | { | |
4e6db0f2 | 2425 | if (set->tags && set->tags[hctx_idx]) { |
bd166ef1 JA |
2426 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); |
2427 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2428 | set->tags[hctx_idx] = NULL; | |
2429 | } | |
cc71a6f4 JA |
2430 | } |
2431 | ||
4b855ad3 | 2432 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 2433 | { |
b3c661b1 | 2434 | unsigned int i, j, hctx_idx; |
320ae51f JA |
2435 | struct blk_mq_hw_ctx *hctx; |
2436 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2437 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f JA |
2438 | |
2439 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 2440 | cpumask_clear(hctx->cpumask); |
320ae51f | 2441 | hctx->nr_ctx = 0; |
d416c92c | 2442 | hctx->dispatch_from = NULL; |
320ae51f JA |
2443 | } |
2444 | ||
2445 | /* | |
4b855ad3 | 2446 | * Map software to hardware queues. |
4412efec ML |
2447 | * |
2448 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 2449 | */ |
20e4d813 | 2450 | for_each_possible_cpu(i) { |
7d76f856 | 2451 | hctx_idx = set->map[HCTX_TYPE_DEFAULT].mq_map[i]; |
4412efec ML |
2452 | /* unmapped hw queue can be remapped after CPU topo changed */ |
2453 | if (!set->tags[hctx_idx] && | |
2454 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
2455 | /* | |
2456 | * If tags initialization fail for some hctx, | |
2457 | * that hctx won't be brought online. In this | |
2458 | * case, remap the current ctx to hctx[0] which | |
2459 | * is guaranteed to always have tags allocated | |
2460 | */ | |
7d76f856 | 2461 | set->map[HCTX_TYPE_DEFAULT].mq_map[i] = 0; |
4412efec ML |
2462 | } |
2463 | ||
897bb0c7 | 2464 | ctx = per_cpu_ptr(q->queue_ctx, i); |
b3c661b1 | 2465 | for (j = 0; j < set->nr_maps; j++) { |
bb94aea1 JW |
2466 | if (!set->map[j].nr_queues) { |
2467 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
2468 | HCTX_TYPE_DEFAULT, i); | |
e5edd5f2 | 2469 | continue; |
bb94aea1 | 2470 | } |
e5edd5f2 | 2471 | |
b3c661b1 | 2472 | hctx = blk_mq_map_queue_type(q, j, i); |
8ccdf4a3 | 2473 | ctx->hctxs[j] = hctx; |
b3c661b1 JA |
2474 | /* |
2475 | * If the CPU is already set in the mask, then we've | |
2476 | * mapped this one already. This can happen if | |
2477 | * devices share queues across queue maps. | |
2478 | */ | |
2479 | if (cpumask_test_cpu(i, hctx->cpumask)) | |
2480 | continue; | |
2481 | ||
2482 | cpumask_set_cpu(i, hctx->cpumask); | |
2483 | hctx->type = j; | |
2484 | ctx->index_hw[hctx->type] = hctx->nr_ctx; | |
2485 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2486 | ||
2487 | /* | |
2488 | * If the nr_ctx type overflows, we have exceeded the | |
2489 | * amount of sw queues we can support. | |
2490 | */ | |
2491 | BUG_ON(!hctx->nr_ctx); | |
2492 | } | |
bb94aea1 JW |
2493 | |
2494 | for (; j < HCTX_MAX_TYPES; j++) | |
2495 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
2496 | HCTX_TYPE_DEFAULT, i); | |
320ae51f | 2497 | } |
506e931f JA |
2498 | |
2499 | queue_for_each_hw_ctx(q, hctx, i) { | |
4412efec ML |
2500 | /* |
2501 | * If no software queues are mapped to this hardware queue, | |
2502 | * disable it and free the request entries. | |
2503 | */ | |
2504 | if (!hctx->nr_ctx) { | |
2505 | /* Never unmap queue 0. We need it as a | |
2506 | * fallback in case of a new remap fails | |
2507 | * allocation | |
2508 | */ | |
2509 | if (i && set->tags[i]) | |
2510 | blk_mq_free_map_and_requests(set, i); | |
2511 | ||
2512 | hctx->tags = NULL; | |
2513 | continue; | |
2514 | } | |
484b4061 | 2515 | |
2a34c087 ML |
2516 | hctx->tags = set->tags[i]; |
2517 | WARN_ON(!hctx->tags); | |
2518 | ||
889fa31f CY |
2519 | /* |
2520 | * Set the map size to the number of mapped software queues. | |
2521 | * This is more accurate and more efficient than looping | |
2522 | * over all possibly mapped software queues. | |
2523 | */ | |
88459642 | 2524 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2525 | |
484b4061 JA |
2526 | /* |
2527 | * Initialize batch roundrobin counts | |
2528 | */ | |
f82ddf19 | 2529 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
2530 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2531 | } | |
320ae51f JA |
2532 | } |
2533 | ||
8e8320c9 JA |
2534 | /* |
2535 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
2536 | * the queue isn't live yet. | |
2537 | */ | |
2404e607 | 2538 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2539 | { |
2540 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2541 | int i; |
2542 | ||
2404e607 | 2543 | queue_for_each_hw_ctx(q, hctx, i) { |
97889f9a | 2544 | if (shared) |
2404e607 | 2545 | hctx->flags |= BLK_MQ_F_TAG_SHARED; |
97889f9a | 2546 | else |
2404e607 JM |
2547 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; |
2548 | } | |
2549 | } | |
2550 | ||
8e8320c9 JA |
2551 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, |
2552 | bool shared) | |
2404e607 JM |
2553 | { |
2554 | struct request_queue *q; | |
0d2602ca | 2555 | |
705cda97 BVA |
2556 | lockdep_assert_held(&set->tag_list_lock); |
2557 | ||
0d2602ca JA |
2558 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2559 | blk_mq_freeze_queue(q); | |
2404e607 | 2560 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2561 | blk_mq_unfreeze_queue(q); |
2562 | } | |
2563 | } | |
2564 | ||
2565 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2566 | { | |
2567 | struct blk_mq_tag_set *set = q->tag_set; | |
2568 | ||
0d2602ca | 2569 | mutex_lock(&set->tag_list_lock); |
705cda97 | 2570 | list_del_rcu(&q->tag_set_list); |
2404e607 JM |
2571 | if (list_is_singular(&set->tag_list)) { |
2572 | /* just transitioned to unshared */ | |
2573 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2574 | /* update existing queue */ | |
2575 | blk_mq_update_tag_set_depth(set, false); | |
2576 | } | |
0d2602ca | 2577 | mutex_unlock(&set->tag_list_lock); |
a347c7ad | 2578 | INIT_LIST_HEAD(&q->tag_set_list); |
0d2602ca JA |
2579 | } |
2580 | ||
2581 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2582 | struct request_queue *q) | |
2583 | { | |
0d2602ca | 2584 | mutex_lock(&set->tag_list_lock); |
2404e607 | 2585 | |
ff821d27 JA |
2586 | /* |
2587 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
2588 | */ | |
2589 | if (!list_empty(&set->tag_list) && | |
2590 | !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2404e607 JM |
2591 | set->flags |= BLK_MQ_F_TAG_SHARED; |
2592 | /* update existing queue */ | |
2593 | blk_mq_update_tag_set_depth(set, true); | |
2594 | } | |
2595 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2596 | queue_set_hctx_shared(q, true); | |
705cda97 | 2597 | list_add_tail_rcu(&q->tag_set_list, &set->tag_list); |
2404e607 | 2598 | |
0d2602ca JA |
2599 | mutex_unlock(&set->tag_list_lock); |
2600 | } | |
2601 | ||
1db4909e ML |
2602 | /* All allocations will be freed in release handler of q->mq_kobj */ |
2603 | static int blk_mq_alloc_ctxs(struct request_queue *q) | |
2604 | { | |
2605 | struct blk_mq_ctxs *ctxs; | |
2606 | int cpu; | |
2607 | ||
2608 | ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL); | |
2609 | if (!ctxs) | |
2610 | return -ENOMEM; | |
2611 | ||
2612 | ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx); | |
2613 | if (!ctxs->queue_ctx) | |
2614 | goto fail; | |
2615 | ||
2616 | for_each_possible_cpu(cpu) { | |
2617 | struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu); | |
2618 | ctx->ctxs = ctxs; | |
2619 | } | |
2620 | ||
2621 | q->mq_kobj = &ctxs->kobj; | |
2622 | q->queue_ctx = ctxs->queue_ctx; | |
2623 | ||
2624 | return 0; | |
2625 | fail: | |
2626 | kfree(ctxs); | |
2627 | return -ENOMEM; | |
2628 | } | |
2629 | ||
e09aae7e ML |
2630 | /* |
2631 | * It is the actual release handler for mq, but we do it from | |
2632 | * request queue's release handler for avoiding use-after-free | |
2633 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2634 | * but we can't group ctx/kctx kobj without it. | |
2635 | */ | |
2636 | void blk_mq_release(struct request_queue *q) | |
2637 | { | |
2f8f1336 ML |
2638 | struct blk_mq_hw_ctx *hctx, *next; |
2639 | int i; | |
e09aae7e | 2640 | |
2f8f1336 ML |
2641 | queue_for_each_hw_ctx(q, hctx, i) |
2642 | WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list)); | |
2643 | ||
2644 | /* all hctx are in .unused_hctx_list now */ | |
2645 | list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) { | |
2646 | list_del_init(&hctx->hctx_list); | |
6c8b232e | 2647 | kobject_put(&hctx->kobj); |
c3b4afca | 2648 | } |
e09aae7e ML |
2649 | |
2650 | kfree(q->queue_hw_ctx); | |
2651 | ||
7ea5fe31 ML |
2652 | /* |
2653 | * release .mq_kobj and sw queue's kobject now because | |
2654 | * both share lifetime with request queue. | |
2655 | */ | |
2656 | blk_mq_sysfs_deinit(q); | |
e09aae7e ML |
2657 | } |
2658 | ||
24d2f903 | 2659 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2660 | { |
2661 | struct request_queue *uninit_q, *q; | |
2662 | ||
6d469642 | 2663 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); |
b62c21b7 MS |
2664 | if (!uninit_q) |
2665 | return ERR_PTR(-ENOMEM); | |
2666 | ||
737eb78e DLM |
2667 | /* |
2668 | * Initialize the queue without an elevator. device_add_disk() will do | |
2669 | * the initialization. | |
2670 | */ | |
2671 | q = blk_mq_init_allocated_queue(set, uninit_q, false); | |
b62c21b7 MS |
2672 | if (IS_ERR(q)) |
2673 | blk_cleanup_queue(uninit_q); | |
2674 | ||
2675 | return q; | |
2676 | } | |
2677 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2678 | ||
9316a9ed JA |
2679 | /* |
2680 | * Helper for setting up a queue with mq ops, given queue depth, and | |
2681 | * the passed in mq ops flags. | |
2682 | */ | |
2683 | struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set, | |
2684 | const struct blk_mq_ops *ops, | |
2685 | unsigned int queue_depth, | |
2686 | unsigned int set_flags) | |
2687 | { | |
2688 | struct request_queue *q; | |
2689 | int ret; | |
2690 | ||
2691 | memset(set, 0, sizeof(*set)); | |
2692 | set->ops = ops; | |
2693 | set->nr_hw_queues = 1; | |
b3c661b1 | 2694 | set->nr_maps = 1; |
9316a9ed JA |
2695 | set->queue_depth = queue_depth; |
2696 | set->numa_node = NUMA_NO_NODE; | |
2697 | set->flags = set_flags; | |
2698 | ||
2699 | ret = blk_mq_alloc_tag_set(set); | |
2700 | if (ret) | |
2701 | return ERR_PTR(ret); | |
2702 | ||
2703 | q = blk_mq_init_queue(set); | |
2704 | if (IS_ERR(q)) { | |
2705 | blk_mq_free_tag_set(set); | |
2706 | return q; | |
2707 | } | |
2708 | ||
2709 | return q; | |
2710 | } | |
2711 | EXPORT_SYMBOL(blk_mq_init_sq_queue); | |
2712 | ||
34d11ffa JW |
2713 | static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx( |
2714 | struct blk_mq_tag_set *set, struct request_queue *q, | |
2715 | int hctx_idx, int node) | |
2716 | { | |
2f8f1336 | 2717 | struct blk_mq_hw_ctx *hctx = NULL, *tmp; |
34d11ffa | 2718 | |
2f8f1336 ML |
2719 | /* reuse dead hctx first */ |
2720 | spin_lock(&q->unused_hctx_lock); | |
2721 | list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) { | |
2722 | if (tmp->numa_node == node) { | |
2723 | hctx = tmp; | |
2724 | break; | |
2725 | } | |
2726 | } | |
2727 | if (hctx) | |
2728 | list_del_init(&hctx->hctx_list); | |
2729 | spin_unlock(&q->unused_hctx_lock); | |
2730 | ||
2731 | if (!hctx) | |
2732 | hctx = blk_mq_alloc_hctx(q, set, node); | |
34d11ffa | 2733 | if (!hctx) |
7c6c5b7c | 2734 | goto fail; |
34d11ffa | 2735 | |
7c6c5b7c ML |
2736 | if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) |
2737 | goto free_hctx; | |
34d11ffa JW |
2738 | |
2739 | return hctx; | |
7c6c5b7c ML |
2740 | |
2741 | free_hctx: | |
2742 | kobject_put(&hctx->kobj); | |
2743 | fail: | |
2744 | return NULL; | |
34d11ffa JW |
2745 | } |
2746 | ||
868f2f0b KB |
2747 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2748 | struct request_queue *q) | |
320ae51f | 2749 | { |
e01ad46d | 2750 | int i, j, end; |
868f2f0b | 2751 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; |
f14bbe77 | 2752 | |
ac0d6b92 BVA |
2753 | if (q->nr_hw_queues < set->nr_hw_queues) { |
2754 | struct blk_mq_hw_ctx **new_hctxs; | |
2755 | ||
2756 | new_hctxs = kcalloc_node(set->nr_hw_queues, | |
2757 | sizeof(*new_hctxs), GFP_KERNEL, | |
2758 | set->numa_node); | |
2759 | if (!new_hctxs) | |
2760 | return; | |
2761 | if (hctxs) | |
2762 | memcpy(new_hctxs, hctxs, q->nr_hw_queues * | |
2763 | sizeof(*hctxs)); | |
2764 | q->queue_hw_ctx = new_hctxs; | |
2765 | q->nr_hw_queues = set->nr_hw_queues; | |
2766 | kfree(hctxs); | |
2767 | hctxs = new_hctxs; | |
2768 | } | |
2769 | ||
fb350e0a ML |
2770 | /* protect against switching io scheduler */ |
2771 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 2772 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2773 | int node; |
34d11ffa | 2774 | struct blk_mq_hw_ctx *hctx; |
868f2f0b | 2775 | |
7d76f856 | 2776 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i); |
34d11ffa JW |
2777 | /* |
2778 | * If the hw queue has been mapped to another numa node, | |
2779 | * we need to realloc the hctx. If allocation fails, fallback | |
2780 | * to use the previous one. | |
2781 | */ | |
2782 | if (hctxs[i] && (hctxs[i]->numa_node == node)) | |
2783 | continue; | |
868f2f0b | 2784 | |
34d11ffa JW |
2785 | hctx = blk_mq_alloc_and_init_hctx(set, q, i, node); |
2786 | if (hctx) { | |
2f8f1336 | 2787 | if (hctxs[i]) |
34d11ffa | 2788 | blk_mq_exit_hctx(q, set, hctxs[i], i); |
34d11ffa JW |
2789 | hctxs[i] = hctx; |
2790 | } else { | |
2791 | if (hctxs[i]) | |
2792 | pr_warn("Allocate new hctx on node %d fails,\ | |
2793 | fallback to previous one on node %d\n", | |
2794 | node, hctxs[i]->numa_node); | |
2795 | else | |
2796 | break; | |
868f2f0b | 2797 | } |
320ae51f | 2798 | } |
e01ad46d JW |
2799 | /* |
2800 | * Increasing nr_hw_queues fails. Free the newly allocated | |
2801 | * hctxs and keep the previous q->nr_hw_queues. | |
2802 | */ | |
2803 | if (i != set->nr_hw_queues) { | |
2804 | j = q->nr_hw_queues; | |
2805 | end = i; | |
2806 | } else { | |
2807 | j = i; | |
2808 | end = q->nr_hw_queues; | |
2809 | q->nr_hw_queues = set->nr_hw_queues; | |
2810 | } | |
34d11ffa | 2811 | |
e01ad46d | 2812 | for (; j < end; j++) { |
868f2f0b KB |
2813 | struct blk_mq_hw_ctx *hctx = hctxs[j]; |
2814 | ||
2815 | if (hctx) { | |
cc71a6f4 JA |
2816 | if (hctx->tags) |
2817 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 2818 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 2819 | hctxs[j] = NULL; |
868f2f0b KB |
2820 | } |
2821 | } | |
fb350e0a | 2822 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
2823 | } |
2824 | ||
2825 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
737eb78e DLM |
2826 | struct request_queue *q, |
2827 | bool elevator_init) | |
868f2f0b | 2828 | { |
66841672 ML |
2829 | /* mark the queue as mq asap */ |
2830 | q->mq_ops = set->ops; | |
2831 | ||
34dbad5d | 2832 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
2833 | blk_mq_poll_stats_bkt, |
2834 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
2835 | if (!q->poll_cb) |
2836 | goto err_exit; | |
2837 | ||
1db4909e | 2838 | if (blk_mq_alloc_ctxs(q)) |
41de54c6 | 2839 | goto err_poll; |
868f2f0b | 2840 | |
737f98cf ML |
2841 | /* init q->mq_kobj and sw queues' kobjects */ |
2842 | blk_mq_sysfs_init(q); | |
2843 | ||
2f8f1336 ML |
2844 | INIT_LIST_HEAD(&q->unused_hctx_list); |
2845 | spin_lock_init(&q->unused_hctx_lock); | |
2846 | ||
868f2f0b KB |
2847 | blk_mq_realloc_hw_ctxs(set, q); |
2848 | if (!q->nr_hw_queues) | |
2849 | goto err_hctxs; | |
320ae51f | 2850 | |
287922eb | 2851 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2852 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f | 2853 | |
a8908939 | 2854 | q->tag_set = set; |
320ae51f | 2855 | |
94eddfbe | 2856 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
cd19181b ML |
2857 | if (set->nr_maps > HCTX_TYPE_POLL && |
2858 | set->map[HCTX_TYPE_POLL].nr_queues) | |
6544d229 | 2859 | blk_queue_flag_set(QUEUE_FLAG_POLL, q); |
320ae51f | 2860 | |
1be036e9 CH |
2861 | q->sg_reserved_size = INT_MAX; |
2862 | ||
2849450a | 2863 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2864 | INIT_LIST_HEAD(&q->requeue_list); |
2865 | spin_lock_init(&q->requeue_lock); | |
2866 | ||
254d259d | 2867 | blk_queue_make_request(q, blk_mq_make_request); |
07068d5b | 2868 | |
eba71768 JA |
2869 | /* |
2870 | * Do this after blk_queue_make_request() overrides it... | |
2871 | */ | |
2872 | q->nr_requests = set->queue_depth; | |
2873 | ||
64f1c21e JA |
2874 | /* |
2875 | * Default to classic polling | |
2876 | */ | |
29ece8b4 | 2877 | q->poll_nsec = BLK_MQ_POLL_CLASSIC; |
64f1c21e | 2878 | |
24d2f903 | 2879 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 2880 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 2881 | blk_mq_map_swqueue(q); |
4593fdbe | 2882 | |
737eb78e DLM |
2883 | if (elevator_init) |
2884 | elevator_init_mq(q); | |
d3484991 | 2885 | |
320ae51f | 2886 | return q; |
18741986 | 2887 | |
320ae51f | 2888 | err_hctxs: |
868f2f0b | 2889 | kfree(q->queue_hw_ctx); |
73d9c8d4 | 2890 | q->nr_hw_queues = 0; |
1db4909e | 2891 | blk_mq_sysfs_deinit(q); |
41de54c6 JS |
2892 | err_poll: |
2893 | blk_stat_free_callback(q->poll_cb); | |
2894 | q->poll_cb = NULL; | |
c7de5726 ML |
2895 | err_exit: |
2896 | q->mq_ops = NULL; | |
320ae51f JA |
2897 | return ERR_PTR(-ENOMEM); |
2898 | } | |
b62c21b7 | 2899 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f | 2900 | |
c7e2d94b ML |
2901 | /* tags can _not_ be used after returning from blk_mq_exit_queue */ |
2902 | void blk_mq_exit_queue(struct request_queue *q) | |
320ae51f | 2903 | { |
624dbe47 | 2904 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2905 | |
0d2602ca | 2906 | blk_mq_del_queue_tag_set(q); |
624dbe47 | 2907 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
320ae51f | 2908 | } |
320ae51f | 2909 | |
a5164405 JA |
2910 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2911 | { | |
2912 | int i; | |
2913 | ||
cc71a6f4 JA |
2914 | for (i = 0; i < set->nr_hw_queues; i++) |
2915 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 2916 | goto out_unwind; |
a5164405 JA |
2917 | |
2918 | return 0; | |
2919 | ||
2920 | out_unwind: | |
2921 | while (--i >= 0) | |
cc71a6f4 | 2922 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 2923 | |
a5164405 JA |
2924 | return -ENOMEM; |
2925 | } | |
2926 | ||
2927 | /* | |
2928 | * Allocate the request maps associated with this tag_set. Note that this | |
2929 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2930 | * will be updated to reflect the allocated depth. | |
2931 | */ | |
2932 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2933 | { | |
2934 | unsigned int depth; | |
2935 | int err; | |
2936 | ||
2937 | depth = set->queue_depth; | |
2938 | do { | |
2939 | err = __blk_mq_alloc_rq_maps(set); | |
2940 | if (!err) | |
2941 | break; | |
2942 | ||
2943 | set->queue_depth >>= 1; | |
2944 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2945 | err = -ENOMEM; | |
2946 | break; | |
2947 | } | |
2948 | } while (set->queue_depth); | |
2949 | ||
2950 | if (!set->queue_depth || err) { | |
2951 | pr_err("blk-mq: failed to allocate request map\n"); | |
2952 | return -ENOMEM; | |
2953 | } | |
2954 | ||
2955 | if (depth != set->queue_depth) | |
2956 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2957 | depth, set->queue_depth); | |
2958 | ||
2959 | return 0; | |
2960 | } | |
2961 | ||
ebe8bddb OS |
2962 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
2963 | { | |
59388702 | 2964 | if (set->ops->map_queues && !is_kdump_kernel()) { |
b3c661b1 JA |
2965 | int i; |
2966 | ||
7d4901a9 ML |
2967 | /* |
2968 | * transport .map_queues is usually done in the following | |
2969 | * way: | |
2970 | * | |
2971 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
2972 | * mask = get_cpu_mask(queue) | |
2973 | * for_each_cpu(cpu, mask) | |
b3c661b1 | 2974 | * set->map[x].mq_map[cpu] = queue; |
7d4901a9 ML |
2975 | * } |
2976 | * | |
2977 | * When we need to remap, the table has to be cleared for | |
2978 | * killing stale mapping since one CPU may not be mapped | |
2979 | * to any hw queue. | |
2980 | */ | |
b3c661b1 JA |
2981 | for (i = 0; i < set->nr_maps; i++) |
2982 | blk_mq_clear_mq_map(&set->map[i]); | |
7d4901a9 | 2983 | |
ebe8bddb | 2984 | return set->ops->map_queues(set); |
b3c661b1 JA |
2985 | } else { |
2986 | BUG_ON(set->nr_maps > 1); | |
7d76f856 | 2987 | return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
b3c661b1 | 2988 | } |
ebe8bddb OS |
2989 | } |
2990 | ||
f7e76dbc BVA |
2991 | static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set, |
2992 | int cur_nr_hw_queues, int new_nr_hw_queues) | |
2993 | { | |
2994 | struct blk_mq_tags **new_tags; | |
2995 | ||
2996 | if (cur_nr_hw_queues >= new_nr_hw_queues) | |
2997 | return 0; | |
2998 | ||
2999 | new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *), | |
3000 | GFP_KERNEL, set->numa_node); | |
3001 | if (!new_tags) | |
3002 | return -ENOMEM; | |
3003 | ||
3004 | if (set->tags) | |
3005 | memcpy(new_tags, set->tags, cur_nr_hw_queues * | |
3006 | sizeof(*set->tags)); | |
3007 | kfree(set->tags); | |
3008 | set->tags = new_tags; | |
3009 | set->nr_hw_queues = new_nr_hw_queues; | |
3010 | ||
3011 | return 0; | |
3012 | } | |
3013 | ||
a4391c64 JA |
3014 | /* |
3015 | * Alloc a tag set to be associated with one or more request queues. | |
3016 | * May fail with EINVAL for various error conditions. May adjust the | |
c018c84f | 3017 | * requested depth down, if it's too large. In that case, the set |
a4391c64 JA |
3018 | * value will be stored in set->queue_depth. |
3019 | */ | |
24d2f903 CH |
3020 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
3021 | { | |
b3c661b1 | 3022 | int i, ret; |
da695ba2 | 3023 | |
205fb5f5 BVA |
3024 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
3025 | ||
24d2f903 CH |
3026 | if (!set->nr_hw_queues) |
3027 | return -EINVAL; | |
a4391c64 | 3028 | if (!set->queue_depth) |
24d2f903 CH |
3029 | return -EINVAL; |
3030 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
3031 | return -EINVAL; | |
3032 | ||
7d7e0f90 | 3033 | if (!set->ops->queue_rq) |
24d2f903 CH |
3034 | return -EINVAL; |
3035 | ||
de148297 ML |
3036 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
3037 | return -EINVAL; | |
3038 | ||
a4391c64 JA |
3039 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
3040 | pr_info("blk-mq: reduced tag depth to %u\n", | |
3041 | BLK_MQ_MAX_DEPTH); | |
3042 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
3043 | } | |
24d2f903 | 3044 | |
b3c661b1 JA |
3045 | if (!set->nr_maps) |
3046 | set->nr_maps = 1; | |
3047 | else if (set->nr_maps > HCTX_MAX_TYPES) | |
3048 | return -EINVAL; | |
3049 | ||
6637fadf SL |
3050 | /* |
3051 | * If a crashdump is active, then we are potentially in a very | |
3052 | * memory constrained environment. Limit us to 1 queue and | |
3053 | * 64 tags to prevent using too much memory. | |
3054 | */ | |
3055 | if (is_kdump_kernel()) { | |
3056 | set->nr_hw_queues = 1; | |
59388702 | 3057 | set->nr_maps = 1; |
6637fadf SL |
3058 | set->queue_depth = min(64U, set->queue_depth); |
3059 | } | |
868f2f0b | 3060 | /* |
392546ae JA |
3061 | * There is no use for more h/w queues than cpus if we just have |
3062 | * a single map | |
868f2f0b | 3063 | */ |
392546ae | 3064 | if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids) |
868f2f0b | 3065 | set->nr_hw_queues = nr_cpu_ids; |
6637fadf | 3066 | |
f7e76dbc | 3067 | if (blk_mq_realloc_tag_set_tags(set, 0, set->nr_hw_queues) < 0) |
a5164405 | 3068 | return -ENOMEM; |
24d2f903 | 3069 | |
da695ba2 | 3070 | ret = -ENOMEM; |
b3c661b1 JA |
3071 | for (i = 0; i < set->nr_maps; i++) { |
3072 | set->map[i].mq_map = kcalloc_node(nr_cpu_ids, | |
07b35eb5 | 3073 | sizeof(set->map[i].mq_map[0]), |
b3c661b1 JA |
3074 | GFP_KERNEL, set->numa_node); |
3075 | if (!set->map[i].mq_map) | |
3076 | goto out_free_mq_map; | |
59388702 | 3077 | set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues; |
b3c661b1 | 3078 | } |
bdd17e75 | 3079 | |
ebe8bddb | 3080 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
3081 | if (ret) |
3082 | goto out_free_mq_map; | |
3083 | ||
3084 | ret = blk_mq_alloc_rq_maps(set); | |
3085 | if (ret) | |
bdd17e75 | 3086 | goto out_free_mq_map; |
24d2f903 | 3087 | |
0d2602ca JA |
3088 | mutex_init(&set->tag_list_lock); |
3089 | INIT_LIST_HEAD(&set->tag_list); | |
3090 | ||
24d2f903 | 3091 | return 0; |
bdd17e75 CH |
3092 | |
3093 | out_free_mq_map: | |
b3c661b1 JA |
3094 | for (i = 0; i < set->nr_maps; i++) { |
3095 | kfree(set->map[i].mq_map); | |
3096 | set->map[i].mq_map = NULL; | |
3097 | } | |
5676e7b6 RE |
3098 | kfree(set->tags); |
3099 | set->tags = NULL; | |
da695ba2 | 3100 | return ret; |
24d2f903 CH |
3101 | } |
3102 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
3103 | ||
3104 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
3105 | { | |
b3c661b1 | 3106 | int i, j; |
24d2f903 | 3107 | |
f7e76dbc | 3108 | for (i = 0; i < set->nr_hw_queues; i++) |
cc71a6f4 | 3109 | blk_mq_free_map_and_requests(set, i); |
484b4061 | 3110 | |
b3c661b1 JA |
3111 | for (j = 0; j < set->nr_maps; j++) { |
3112 | kfree(set->map[j].mq_map); | |
3113 | set->map[j].mq_map = NULL; | |
3114 | } | |
bdd17e75 | 3115 | |
981bd189 | 3116 | kfree(set->tags); |
5676e7b6 | 3117 | set->tags = NULL; |
24d2f903 CH |
3118 | } |
3119 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
3120 | ||
e3a2b3f9 JA |
3121 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
3122 | { | |
3123 | struct blk_mq_tag_set *set = q->tag_set; | |
3124 | struct blk_mq_hw_ctx *hctx; | |
3125 | int i, ret; | |
3126 | ||
bd166ef1 | 3127 | if (!set) |
e3a2b3f9 JA |
3128 | return -EINVAL; |
3129 | ||
e5fa8140 AZ |
3130 | if (q->nr_requests == nr) |
3131 | return 0; | |
3132 | ||
70f36b60 | 3133 | blk_mq_freeze_queue(q); |
24f5a90f | 3134 | blk_mq_quiesce_queue(q); |
70f36b60 | 3135 | |
e3a2b3f9 JA |
3136 | ret = 0; |
3137 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
3138 | if (!hctx->tags) |
3139 | continue; | |
bd166ef1 JA |
3140 | /* |
3141 | * If we're using an MQ scheduler, just update the scheduler | |
3142 | * queue depth. This is similar to what the old code would do. | |
3143 | */ | |
70f36b60 | 3144 | if (!hctx->sched_tags) { |
c2e82a23 | 3145 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, |
70f36b60 JA |
3146 | false); |
3147 | } else { | |
3148 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
3149 | nr, true); | |
3150 | } | |
e3a2b3f9 JA |
3151 | if (ret) |
3152 | break; | |
77f1e0a5 JA |
3153 | if (q->elevator && q->elevator->type->ops.depth_updated) |
3154 | q->elevator->type->ops.depth_updated(hctx); | |
e3a2b3f9 JA |
3155 | } |
3156 | ||
3157 | if (!ret) | |
3158 | q->nr_requests = nr; | |
3159 | ||
24f5a90f | 3160 | blk_mq_unquiesce_queue(q); |
70f36b60 | 3161 | blk_mq_unfreeze_queue(q); |
70f36b60 | 3162 | |
e3a2b3f9 JA |
3163 | return ret; |
3164 | } | |
3165 | ||
d48ece20 JW |
3166 | /* |
3167 | * request_queue and elevator_type pair. | |
3168 | * It is just used by __blk_mq_update_nr_hw_queues to cache | |
3169 | * the elevator_type associated with a request_queue. | |
3170 | */ | |
3171 | struct blk_mq_qe_pair { | |
3172 | struct list_head node; | |
3173 | struct request_queue *q; | |
3174 | struct elevator_type *type; | |
3175 | }; | |
3176 | ||
3177 | /* | |
3178 | * Cache the elevator_type in qe pair list and switch the | |
3179 | * io scheduler to 'none' | |
3180 | */ | |
3181 | static bool blk_mq_elv_switch_none(struct list_head *head, | |
3182 | struct request_queue *q) | |
3183 | { | |
3184 | struct blk_mq_qe_pair *qe; | |
3185 | ||
3186 | if (!q->elevator) | |
3187 | return true; | |
3188 | ||
3189 | qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); | |
3190 | if (!qe) | |
3191 | return false; | |
3192 | ||
3193 | INIT_LIST_HEAD(&qe->node); | |
3194 | qe->q = q; | |
3195 | qe->type = q->elevator->type; | |
3196 | list_add(&qe->node, head); | |
3197 | ||
3198 | mutex_lock(&q->sysfs_lock); | |
3199 | /* | |
3200 | * After elevator_switch_mq, the previous elevator_queue will be | |
3201 | * released by elevator_release. The reference of the io scheduler | |
3202 | * module get by elevator_get will also be put. So we need to get | |
3203 | * a reference of the io scheduler module here to prevent it to be | |
3204 | * removed. | |
3205 | */ | |
3206 | __module_get(qe->type->elevator_owner); | |
3207 | elevator_switch_mq(q, NULL); | |
3208 | mutex_unlock(&q->sysfs_lock); | |
3209 | ||
3210 | return true; | |
3211 | } | |
3212 | ||
3213 | static void blk_mq_elv_switch_back(struct list_head *head, | |
3214 | struct request_queue *q) | |
3215 | { | |
3216 | struct blk_mq_qe_pair *qe; | |
3217 | struct elevator_type *t = NULL; | |
3218 | ||
3219 | list_for_each_entry(qe, head, node) | |
3220 | if (qe->q == q) { | |
3221 | t = qe->type; | |
3222 | break; | |
3223 | } | |
3224 | ||
3225 | if (!t) | |
3226 | return; | |
3227 | ||
3228 | list_del(&qe->node); | |
3229 | kfree(qe); | |
3230 | ||
3231 | mutex_lock(&q->sysfs_lock); | |
3232 | elevator_switch_mq(q, t); | |
3233 | mutex_unlock(&q->sysfs_lock); | |
3234 | } | |
3235 | ||
e4dc2b32 KB |
3236 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
3237 | int nr_hw_queues) | |
868f2f0b KB |
3238 | { |
3239 | struct request_queue *q; | |
d48ece20 | 3240 | LIST_HEAD(head); |
e01ad46d | 3241 | int prev_nr_hw_queues; |
868f2f0b | 3242 | |
705cda97 BVA |
3243 | lockdep_assert_held(&set->tag_list_lock); |
3244 | ||
392546ae | 3245 | if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids) |
868f2f0b KB |
3246 | nr_hw_queues = nr_cpu_ids; |
3247 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
3248 | return; | |
3249 | ||
3250 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3251 | blk_mq_freeze_queue(q); | |
d48ece20 JW |
3252 | /* |
3253 | * Switch IO scheduler to 'none', cleaning up the data associated | |
3254 | * with the previous scheduler. We will switch back once we are done | |
3255 | * updating the new sw to hw queue mappings. | |
3256 | */ | |
3257 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3258 | if (!blk_mq_elv_switch_none(&head, q)) | |
3259 | goto switch_back; | |
868f2f0b | 3260 | |
477e19de JW |
3261 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3262 | blk_mq_debugfs_unregister_hctxs(q); | |
3263 | blk_mq_sysfs_unregister(q); | |
3264 | } | |
3265 | ||
f7e76dbc BVA |
3266 | if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) < |
3267 | 0) | |
3268 | goto reregister; | |
3269 | ||
e01ad46d | 3270 | prev_nr_hw_queues = set->nr_hw_queues; |
868f2f0b | 3271 | set->nr_hw_queues = nr_hw_queues; |
ebe8bddb | 3272 | blk_mq_update_queue_map(set); |
e01ad46d | 3273 | fallback: |
868f2f0b KB |
3274 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3275 | blk_mq_realloc_hw_ctxs(set, q); | |
e01ad46d JW |
3276 | if (q->nr_hw_queues != set->nr_hw_queues) { |
3277 | pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n", | |
3278 | nr_hw_queues, prev_nr_hw_queues); | |
3279 | set->nr_hw_queues = prev_nr_hw_queues; | |
7d76f856 | 3280 | blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
e01ad46d JW |
3281 | goto fallback; |
3282 | } | |
477e19de JW |
3283 | blk_mq_map_swqueue(q); |
3284 | } | |
3285 | ||
f7e76dbc | 3286 | reregister: |
477e19de JW |
3287 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3288 | blk_mq_sysfs_register(q); | |
3289 | blk_mq_debugfs_register_hctxs(q); | |
868f2f0b KB |
3290 | } |
3291 | ||
d48ece20 JW |
3292 | switch_back: |
3293 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3294 | blk_mq_elv_switch_back(&head, q); | |
3295 | ||
868f2f0b KB |
3296 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
3297 | blk_mq_unfreeze_queue(q); | |
3298 | } | |
e4dc2b32 KB |
3299 | |
3300 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
3301 | { | |
3302 | mutex_lock(&set->tag_list_lock); | |
3303 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
3304 | mutex_unlock(&set->tag_list_lock); | |
3305 | } | |
868f2f0b KB |
3306 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
3307 | ||
34dbad5d OS |
3308 | /* Enable polling stats and return whether they were already enabled. */ |
3309 | static bool blk_poll_stats_enable(struct request_queue *q) | |
3310 | { | |
3311 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
7dfdbc73 | 3312 | blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q)) |
34dbad5d OS |
3313 | return true; |
3314 | blk_stat_add_callback(q, q->poll_cb); | |
3315 | return false; | |
3316 | } | |
3317 | ||
3318 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
3319 | { | |
3320 | /* | |
3321 | * We don't arm the callback if polling stats are not enabled or the | |
3322 | * callback is already active. | |
3323 | */ | |
3324 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
3325 | blk_stat_is_active(q->poll_cb)) | |
3326 | return; | |
3327 | ||
3328 | blk_stat_activate_msecs(q->poll_cb, 100); | |
3329 | } | |
3330 | ||
3331 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
3332 | { | |
3333 | struct request_queue *q = cb->data; | |
720b8ccc | 3334 | int bucket; |
34dbad5d | 3335 | |
720b8ccc SB |
3336 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
3337 | if (cb->stat[bucket].nr_samples) | |
3338 | q->poll_stat[bucket] = cb->stat[bucket]; | |
3339 | } | |
34dbad5d OS |
3340 | } |
3341 | ||
64f1c21e JA |
3342 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
3343 | struct blk_mq_hw_ctx *hctx, | |
3344 | struct request *rq) | |
3345 | { | |
64f1c21e | 3346 | unsigned long ret = 0; |
720b8ccc | 3347 | int bucket; |
64f1c21e JA |
3348 | |
3349 | /* | |
3350 | * If stats collection isn't on, don't sleep but turn it on for | |
3351 | * future users | |
3352 | */ | |
34dbad5d | 3353 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
3354 | return 0; |
3355 | ||
64f1c21e JA |
3356 | /* |
3357 | * As an optimistic guess, use half of the mean service time | |
3358 | * for this type of request. We can (and should) make this smarter. | |
3359 | * For instance, if the completion latencies are tight, we can | |
3360 | * get closer than just half the mean. This is especially | |
3361 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
3362 | * than ~10 usec. We do use the stats for the relevant IO size |
3363 | * if available which does lead to better estimates. | |
64f1c21e | 3364 | */ |
720b8ccc SB |
3365 | bucket = blk_mq_poll_stats_bkt(rq); |
3366 | if (bucket < 0) | |
3367 | return ret; | |
3368 | ||
3369 | if (q->poll_stat[bucket].nr_samples) | |
3370 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
3371 | |
3372 | return ret; | |
3373 | } | |
3374 | ||
06426adf | 3375 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 3376 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
3377 | struct request *rq) |
3378 | { | |
3379 | struct hrtimer_sleeper hs; | |
3380 | enum hrtimer_mode mode; | |
64f1c21e | 3381 | unsigned int nsecs; |
06426adf JA |
3382 | ktime_t kt; |
3383 | ||
76a86f9d | 3384 | if (rq->rq_flags & RQF_MQ_POLL_SLEPT) |
64f1c21e JA |
3385 | return false; |
3386 | ||
3387 | /* | |
1052b8ac | 3388 | * If we get here, hybrid polling is enabled. Hence poll_nsec can be: |
64f1c21e | 3389 | * |
64f1c21e JA |
3390 | * 0: use half of prev avg |
3391 | * >0: use this specific value | |
3392 | */ | |
1052b8ac | 3393 | if (q->poll_nsec > 0) |
64f1c21e JA |
3394 | nsecs = q->poll_nsec; |
3395 | else | |
3396 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
3397 | ||
3398 | if (!nsecs) | |
06426adf JA |
3399 | return false; |
3400 | ||
76a86f9d | 3401 | rq->rq_flags |= RQF_MQ_POLL_SLEPT; |
06426adf JA |
3402 | |
3403 | /* | |
3404 | * This will be replaced with the stats tracking code, using | |
3405 | * 'avg_completion_time / 2' as the pre-sleep target. | |
3406 | */ | |
8b0e1953 | 3407 | kt = nsecs; |
06426adf JA |
3408 | |
3409 | mode = HRTIMER_MODE_REL; | |
dbc1625f | 3410 | hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode); |
06426adf JA |
3411 | hrtimer_set_expires(&hs.timer, kt); |
3412 | ||
06426adf | 3413 | do { |
5a61c363 | 3414 | if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE) |
06426adf JA |
3415 | break; |
3416 | set_current_state(TASK_UNINTERRUPTIBLE); | |
9dd8813e | 3417 | hrtimer_sleeper_start_expires(&hs, mode); |
06426adf JA |
3418 | if (hs.task) |
3419 | io_schedule(); | |
3420 | hrtimer_cancel(&hs.timer); | |
3421 | mode = HRTIMER_MODE_ABS; | |
3422 | } while (hs.task && !signal_pending(current)); | |
3423 | ||
3424 | __set_current_state(TASK_RUNNING); | |
3425 | destroy_hrtimer_on_stack(&hs.timer); | |
3426 | return true; | |
3427 | } | |
3428 | ||
1052b8ac JA |
3429 | static bool blk_mq_poll_hybrid(struct request_queue *q, |
3430 | struct blk_mq_hw_ctx *hctx, blk_qc_t cookie) | |
bbd7bb70 | 3431 | { |
1052b8ac JA |
3432 | struct request *rq; |
3433 | ||
29ece8b4 | 3434 | if (q->poll_nsec == BLK_MQ_POLL_CLASSIC) |
1052b8ac JA |
3435 | return false; |
3436 | ||
3437 | if (!blk_qc_t_is_internal(cookie)) | |
3438 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
3439 | else { | |
3440 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); | |
3441 | /* | |
3442 | * With scheduling, if the request has completed, we'll | |
3443 | * get a NULL return here, as we clear the sched tag when | |
3444 | * that happens. The request still remains valid, like always, | |
3445 | * so we should be safe with just the NULL check. | |
3446 | */ | |
3447 | if (!rq) | |
3448 | return false; | |
3449 | } | |
3450 | ||
3451 | return blk_mq_poll_hybrid_sleep(q, hctx, rq); | |
3452 | } | |
3453 | ||
529262d5 CH |
3454 | /** |
3455 | * blk_poll - poll for IO completions | |
3456 | * @q: the queue | |
3457 | * @cookie: cookie passed back at IO submission time | |
3458 | * @spin: whether to spin for completions | |
3459 | * | |
3460 | * Description: | |
3461 | * Poll for completions on the passed in queue. Returns number of | |
3462 | * completed entries found. If @spin is true, then blk_poll will continue | |
3463 | * looping until at least one completion is found, unless the task is | |
3464 | * otherwise marked running (or we need to reschedule). | |
3465 | */ | |
3466 | int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin) | |
1052b8ac JA |
3467 | { |
3468 | struct blk_mq_hw_ctx *hctx; | |
bbd7bb70 JA |
3469 | long state; |
3470 | ||
529262d5 CH |
3471 | if (!blk_qc_t_valid(cookie) || |
3472 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
1052b8ac JA |
3473 | return 0; |
3474 | ||
529262d5 CH |
3475 | if (current->plug) |
3476 | blk_flush_plug_list(current->plug, false); | |
3477 | ||
1052b8ac JA |
3478 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; |
3479 | ||
06426adf JA |
3480 | /* |
3481 | * If we sleep, have the caller restart the poll loop to reset | |
3482 | * the state. Like for the other success return cases, the | |
3483 | * caller is responsible for checking if the IO completed. If | |
3484 | * the IO isn't complete, we'll get called again and will go | |
3485 | * straight to the busy poll loop. | |
3486 | */ | |
1052b8ac | 3487 | if (blk_mq_poll_hybrid(q, hctx, cookie)) |
85f4d4b6 | 3488 | return 1; |
06426adf | 3489 | |
bbd7bb70 JA |
3490 | hctx->poll_considered++; |
3491 | ||
3492 | state = current->state; | |
aa61bec3 | 3493 | do { |
bbd7bb70 JA |
3494 | int ret; |
3495 | ||
3496 | hctx->poll_invoked++; | |
3497 | ||
9743139c | 3498 | ret = q->mq_ops->poll(hctx); |
bbd7bb70 JA |
3499 | if (ret > 0) { |
3500 | hctx->poll_success++; | |
849a3700 | 3501 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 3502 | return ret; |
bbd7bb70 JA |
3503 | } |
3504 | ||
3505 | if (signal_pending_state(state, current)) | |
849a3700 | 3506 | __set_current_state(TASK_RUNNING); |
bbd7bb70 JA |
3507 | |
3508 | if (current->state == TASK_RUNNING) | |
85f4d4b6 | 3509 | return 1; |
0a1b8b87 | 3510 | if (ret < 0 || !spin) |
bbd7bb70 JA |
3511 | break; |
3512 | cpu_relax(); | |
aa61bec3 | 3513 | } while (!need_resched()); |
bbd7bb70 | 3514 | |
67b4110f | 3515 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 3516 | return 0; |
bbd7bb70 | 3517 | } |
529262d5 | 3518 | EXPORT_SYMBOL_GPL(blk_poll); |
bbd7bb70 | 3519 | |
9cf2bab6 JA |
3520 | unsigned int blk_mq_rq_cpu(struct request *rq) |
3521 | { | |
3522 | return rq->mq_ctx->cpu; | |
3523 | } | |
3524 | EXPORT_SYMBOL(blk_mq_rq_cpu); | |
3525 | ||
320ae51f JA |
3526 | static int __init blk_mq_init(void) |
3527 | { | |
9467f859 TG |
3528 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
3529 | blk_mq_hctx_notify_dead); | |
320ae51f JA |
3530 | return 0; |
3531 | } | |
3532 | subsys_initcall(blk_mq_init); |