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> | |
33 | #include "blk.h" | |
34 | #include "blk-mq.h" | |
9c1051aa | 35 | #include "blk-mq-debugfs.h" |
320ae51f | 36 | #include "blk-mq-tag.h" |
986d413b | 37 | #include "blk-pm.h" |
cf43e6be | 38 | #include "blk-stat.h" |
bd166ef1 | 39 | #include "blk-mq-sched.h" |
c1c80384 | 40 | #include "blk-rq-qos.h" |
320ae51f | 41 | |
34dbad5d OS |
42 | static void blk_mq_poll_stats_start(struct request_queue *q); |
43 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb); | |
44 | ||
720b8ccc SB |
45 | static int blk_mq_poll_stats_bkt(const struct request *rq) |
46 | { | |
47 | int ddir, bytes, bucket; | |
48 | ||
99c749a4 | 49 | ddir = rq_data_dir(rq); |
720b8ccc SB |
50 | bytes = blk_rq_bytes(rq); |
51 | ||
52 | bucket = ddir + 2*(ilog2(bytes) - 9); | |
53 | ||
54 | if (bucket < 0) | |
55 | return -1; | |
56 | else if (bucket >= BLK_MQ_POLL_STATS_BKTS) | |
57 | return ddir + BLK_MQ_POLL_STATS_BKTS - 2; | |
58 | ||
59 | return bucket; | |
60 | } | |
61 | ||
320ae51f | 62 | /* |
85fae294 YY |
63 | * Check if any of the ctx, dispatch list or elevator |
64 | * have pending work in this hardware queue. | |
320ae51f | 65 | */ |
79f720a7 | 66 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 67 | { |
79f720a7 JA |
68 | return !list_empty_careful(&hctx->dispatch) || |
69 | sbitmap_any_bit_set(&hctx->ctx_map) || | |
bd166ef1 | 70 | blk_mq_sched_has_work(hctx); |
1429d7c9 JA |
71 | } |
72 | ||
320ae51f JA |
73 | /* |
74 | * Mark this ctx as having pending work in this hardware queue | |
75 | */ | |
76 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
77 | struct blk_mq_ctx *ctx) | |
78 | { | |
f31967f0 JA |
79 | const int bit = ctx->index_hw[hctx->type]; |
80 | ||
81 | if (!sbitmap_test_bit(&hctx->ctx_map, bit)) | |
82 | sbitmap_set_bit(&hctx->ctx_map, bit); | |
1429d7c9 JA |
83 | } |
84 | ||
85 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
86 | struct blk_mq_ctx *ctx) | |
87 | { | |
f31967f0 JA |
88 | const int bit = ctx->index_hw[hctx->type]; |
89 | ||
90 | sbitmap_clear_bit(&hctx->ctx_map, bit); | |
320ae51f JA |
91 | } |
92 | ||
f299b7c7 JA |
93 | struct mq_inflight { |
94 | struct hd_struct *part; | |
95 | unsigned int *inflight; | |
96 | }; | |
97 | ||
7baa8572 | 98 | static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx, |
f299b7c7 JA |
99 | struct request *rq, void *priv, |
100 | bool reserved) | |
101 | { | |
102 | struct mq_inflight *mi = priv; | |
103 | ||
6131837b | 104 | /* |
e016b782 | 105 | * index[0] counts the specific partition that was asked for. |
6131837b OS |
106 | */ |
107 | if (rq->part == mi->part) | |
108 | mi->inflight[0]++; | |
7baa8572 JA |
109 | |
110 | return true; | |
f299b7c7 JA |
111 | } |
112 | ||
e016b782 | 113 | unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part) |
f299b7c7 | 114 | { |
e016b782 | 115 | unsigned inflight[2]; |
f299b7c7 JA |
116 | struct mq_inflight mi = { .part = part, .inflight = inflight, }; |
117 | ||
b8d62b3a | 118 | inflight[0] = inflight[1] = 0; |
f299b7c7 | 119 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
e016b782 MP |
120 | |
121 | return inflight[0]; | |
f299b7c7 JA |
122 | } |
123 | ||
7baa8572 | 124 | static bool blk_mq_check_inflight_rw(struct blk_mq_hw_ctx *hctx, |
bf0ddaba OS |
125 | struct request *rq, void *priv, |
126 | bool reserved) | |
127 | { | |
128 | struct mq_inflight *mi = priv; | |
129 | ||
130 | if (rq->part == mi->part) | |
131 | mi->inflight[rq_data_dir(rq)]++; | |
7baa8572 JA |
132 | |
133 | return true; | |
bf0ddaba OS |
134 | } |
135 | ||
136 | void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part, | |
137 | unsigned int inflight[2]) | |
138 | { | |
139 | struct mq_inflight mi = { .part = part, .inflight = inflight, }; | |
140 | ||
141 | inflight[0] = inflight[1] = 0; | |
142 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight_rw, &mi); | |
143 | } | |
144 | ||
1671d522 | 145 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 146 | { |
7996a8b5 BL |
147 | mutex_lock(&q->mq_freeze_lock); |
148 | if (++q->mq_freeze_depth == 1) { | |
3ef28e83 | 149 | percpu_ref_kill(&q->q_usage_counter); |
7996a8b5 | 150 | mutex_unlock(&q->mq_freeze_lock); |
344e9ffc | 151 | if (queue_is_mq(q)) |
055f6e18 | 152 | blk_mq_run_hw_queues(q, false); |
7996a8b5 BL |
153 | } else { |
154 | mutex_unlock(&q->mq_freeze_lock); | |
cddd5d17 | 155 | } |
f3af020b | 156 | } |
1671d522 | 157 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
f3af020b | 158 | |
6bae363e | 159 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 160 | { |
3ef28e83 | 161 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 162 | } |
6bae363e | 163 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 164 | |
f91328c4 KB |
165 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
166 | unsigned long timeout) | |
167 | { | |
168 | return wait_event_timeout(q->mq_freeze_wq, | |
169 | percpu_ref_is_zero(&q->q_usage_counter), | |
170 | timeout); | |
171 | } | |
172 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 173 | |
f3af020b TH |
174 | /* |
175 | * Guarantee no request is in use, so we can change any data structure of | |
176 | * the queue afterward. | |
177 | */ | |
3ef28e83 | 178 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 179 | { |
3ef28e83 DW |
180 | /* |
181 | * In the !blk_mq case we are only calling this to kill the | |
182 | * q_usage_counter, otherwise this increases the freeze depth | |
183 | * and waits for it to return to zero. For this reason there is | |
184 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
185 | * exported to drivers as the only user for unfreeze is blk_mq. | |
186 | */ | |
1671d522 | 187 | blk_freeze_queue_start(q); |
f3af020b TH |
188 | blk_mq_freeze_queue_wait(q); |
189 | } | |
3ef28e83 DW |
190 | |
191 | void blk_mq_freeze_queue(struct request_queue *q) | |
192 | { | |
193 | /* | |
194 | * ...just an alias to keep freeze and unfreeze actions balanced | |
195 | * in the blk_mq_* namespace | |
196 | */ | |
197 | blk_freeze_queue(q); | |
198 | } | |
c761d96b | 199 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 200 | |
b4c6a028 | 201 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 202 | { |
7996a8b5 BL |
203 | mutex_lock(&q->mq_freeze_lock); |
204 | q->mq_freeze_depth--; | |
205 | WARN_ON_ONCE(q->mq_freeze_depth < 0); | |
206 | if (!q->mq_freeze_depth) { | |
bdd63160 | 207 | percpu_ref_resurrect(&q->q_usage_counter); |
320ae51f | 208 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 209 | } |
7996a8b5 | 210 | mutex_unlock(&q->mq_freeze_lock); |
320ae51f | 211 | } |
b4c6a028 | 212 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 213 | |
852ec809 BVA |
214 | /* |
215 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
216 | * mpt3sas driver such that this function can be removed. | |
217 | */ | |
218 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
219 | { | |
8814ce8a | 220 | blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q); |
852ec809 BVA |
221 | } |
222 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
223 | ||
6a83e74d | 224 | /** |
69e07c4a | 225 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished |
6a83e74d BVA |
226 | * @q: request queue. |
227 | * | |
228 | * Note: this function does not prevent that the struct request end_io() | |
69e07c4a ML |
229 | * callback function is invoked. Once this function is returned, we make |
230 | * sure no dispatch can happen until the queue is unquiesced via | |
231 | * blk_mq_unquiesce_queue(). | |
6a83e74d BVA |
232 | */ |
233 | void blk_mq_quiesce_queue(struct request_queue *q) | |
234 | { | |
235 | struct blk_mq_hw_ctx *hctx; | |
236 | unsigned int i; | |
237 | bool rcu = false; | |
238 | ||
1d9e9bc6 | 239 | blk_mq_quiesce_queue_nowait(q); |
f4560ffe | 240 | |
6a83e74d BVA |
241 | queue_for_each_hw_ctx(q, hctx, i) { |
242 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
05707b64 | 243 | synchronize_srcu(hctx->srcu); |
6a83e74d BVA |
244 | else |
245 | rcu = true; | |
246 | } | |
247 | if (rcu) | |
248 | synchronize_rcu(); | |
249 | } | |
250 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
251 | ||
e4e73913 ML |
252 | /* |
253 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
254 | * @q: request queue. | |
255 | * | |
256 | * This function recovers queue into the state before quiescing | |
257 | * which is done by blk_mq_quiesce_queue. | |
258 | */ | |
259 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
260 | { | |
8814ce8a | 261 | blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q); |
f4560ffe | 262 | |
1d9e9bc6 ML |
263 | /* dispatch requests which are inserted during quiescing */ |
264 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
265 | } |
266 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
267 | ||
aed3ea94 JA |
268 | void blk_mq_wake_waiters(struct request_queue *q) |
269 | { | |
270 | struct blk_mq_hw_ctx *hctx; | |
271 | unsigned int i; | |
272 | ||
273 | queue_for_each_hw_ctx(q, hctx, i) | |
274 | if (blk_mq_hw_queue_mapped(hctx)) | |
275 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
276 | } | |
277 | ||
320ae51f JA |
278 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
279 | { | |
280 | return blk_mq_has_free_tags(hctx->tags); | |
281 | } | |
282 | EXPORT_SYMBOL(blk_mq_can_queue); | |
283 | ||
fe1f4526 JA |
284 | /* |
285 | * Only need start/end time stamping if we have stats enabled, or using | |
286 | * an IO scheduler. | |
287 | */ | |
288 | static inline bool blk_mq_need_time_stamp(struct request *rq) | |
289 | { | |
290 | return (rq->rq_flags & RQF_IO_STAT) || rq->q->elevator; | |
291 | } | |
292 | ||
e4cdf1a1 CH |
293 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
294 | unsigned int tag, unsigned int op) | |
320ae51f | 295 | { |
e4cdf1a1 CH |
296 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
297 | struct request *rq = tags->static_rqs[tag]; | |
bf9ae8c5 | 298 | req_flags_t rq_flags = 0; |
c3a148d2 | 299 | |
e4cdf1a1 CH |
300 | if (data->flags & BLK_MQ_REQ_INTERNAL) { |
301 | rq->tag = -1; | |
302 | rq->internal_tag = tag; | |
303 | } else { | |
d263ed99 | 304 | if (data->hctx->flags & BLK_MQ_F_TAG_SHARED) { |
bf9ae8c5 | 305 | rq_flags = RQF_MQ_INFLIGHT; |
e4cdf1a1 CH |
306 | atomic_inc(&data->hctx->nr_active); |
307 | } | |
308 | rq->tag = tag; | |
309 | rq->internal_tag = -1; | |
310 | data->hctx->tags->rqs[rq->tag] = rq; | |
311 | } | |
312 | ||
af76e555 | 313 | /* csd/requeue_work/fifo_time is initialized before use */ |
e4cdf1a1 CH |
314 | rq->q = data->q; |
315 | rq->mq_ctx = data->ctx; | |
ea4f995e | 316 | rq->mq_hctx = data->hctx; |
bf9ae8c5 | 317 | rq->rq_flags = rq_flags; |
ef295ecf | 318 | rq->cmd_flags = op; |
1b6d65a0 BVA |
319 | if (data->flags & BLK_MQ_REQ_PREEMPT) |
320 | rq->rq_flags |= RQF_PREEMPT; | |
e4cdf1a1 | 321 | if (blk_queue_io_stat(data->q)) |
e8064021 | 322 | rq->rq_flags |= RQF_IO_STAT; |
7c3fb70f | 323 | INIT_LIST_HEAD(&rq->queuelist); |
af76e555 CH |
324 | INIT_HLIST_NODE(&rq->hash); |
325 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
326 | rq->rq_disk = NULL; |
327 | rq->part = NULL; | |
fe1f4526 JA |
328 | if (blk_mq_need_time_stamp(rq)) |
329 | rq->start_time_ns = ktime_get_ns(); | |
330 | else | |
331 | rq->start_time_ns = 0; | |
544ccc8d | 332 | rq->io_start_time_ns = 0; |
af76e555 CH |
333 | rq->nr_phys_segments = 0; |
334 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
335 | rq->nr_integrity_segments = 0; | |
336 | #endif | |
af76e555 | 337 | /* tag was already set */ |
af76e555 | 338 | rq->extra_len = 0; |
079076b3 | 339 | WRITE_ONCE(rq->deadline, 0); |
af76e555 | 340 | |
f6be4fb4 JA |
341 | rq->timeout = 0; |
342 | ||
af76e555 CH |
343 | rq->end_io = NULL; |
344 | rq->end_io_data = NULL; | |
af76e555 | 345 | |
e4cdf1a1 | 346 | data->ctx->rq_dispatched[op_is_sync(op)]++; |
12f5b931 | 347 | refcount_set(&rq->ref, 1); |
e4cdf1a1 | 348 | return rq; |
5dee8577 CH |
349 | } |
350 | ||
d2c0d383 | 351 | static struct request *blk_mq_get_request(struct request_queue *q, |
f9afca4d JA |
352 | struct bio *bio, |
353 | struct blk_mq_alloc_data *data) | |
d2c0d383 CH |
354 | { |
355 | struct elevator_queue *e = q->elevator; | |
356 | struct request *rq; | |
e4cdf1a1 | 357 | unsigned int tag; |
c05f4220 | 358 | bool clear_ctx_on_error = false; |
d2c0d383 CH |
359 | |
360 | blk_queue_enter_live(q); | |
361 | data->q = q; | |
21e768b4 BVA |
362 | if (likely(!data->ctx)) { |
363 | data->ctx = blk_mq_get_ctx(q); | |
c05f4220 | 364 | clear_ctx_on_error = true; |
21e768b4 | 365 | } |
d2c0d383 | 366 | if (likely(!data->hctx)) |
f9afca4d | 367 | data->hctx = blk_mq_map_queue(q, data->cmd_flags, |
8ccdf4a3 | 368 | data->ctx); |
f9afca4d | 369 | if (data->cmd_flags & REQ_NOWAIT) |
03a07c92 | 370 | data->flags |= BLK_MQ_REQ_NOWAIT; |
d2c0d383 CH |
371 | |
372 | if (e) { | |
373 | data->flags |= BLK_MQ_REQ_INTERNAL; | |
374 | ||
375 | /* | |
376 | * Flush requests are special and go directly to the | |
17a51199 JA |
377 | * dispatch list. Don't include reserved tags in the |
378 | * limiting, as it isn't useful. | |
d2c0d383 | 379 | */ |
f9afca4d JA |
380 | if (!op_is_flush(data->cmd_flags) && |
381 | e->type->ops.limit_depth && | |
17a51199 | 382 | !(data->flags & BLK_MQ_REQ_RESERVED)) |
f9afca4d | 383 | e->type->ops.limit_depth(data->cmd_flags, data); |
d263ed99 JW |
384 | } else { |
385 | blk_mq_tag_busy(data->hctx); | |
d2c0d383 CH |
386 | } |
387 | ||
e4cdf1a1 CH |
388 | tag = blk_mq_get_tag(data); |
389 | if (tag == BLK_MQ_TAG_FAIL) { | |
c05f4220 | 390 | if (clear_ctx_on_error) |
1ad43c00 | 391 | data->ctx = NULL; |
037cebb8 CH |
392 | blk_queue_exit(q); |
393 | return NULL; | |
d2c0d383 CH |
394 | } |
395 | ||
f9afca4d JA |
396 | rq = blk_mq_rq_ctx_init(data, tag, data->cmd_flags); |
397 | if (!op_is_flush(data->cmd_flags)) { | |
037cebb8 | 398 | rq->elv.icq = NULL; |
f9cd4bfe | 399 | if (e && e->type->ops.prepare_request) { |
e2b3fa5a DLM |
400 | if (e->type->icq_cache) |
401 | blk_mq_sched_assign_ioc(rq); | |
44e8c2bf | 402 | |
f9cd4bfe | 403 | e->type->ops.prepare_request(rq, bio); |
5bbf4e5a | 404 | rq->rq_flags |= RQF_ELVPRIV; |
44e8c2bf | 405 | } |
037cebb8 CH |
406 | } |
407 | data->hctx->queued++; | |
408 | return rq; | |
d2c0d383 CH |
409 | } |
410 | ||
cd6ce148 | 411 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, |
9a95e4ef | 412 | blk_mq_req_flags_t flags) |
320ae51f | 413 | { |
f9afca4d | 414 | struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op }; |
bd166ef1 | 415 | struct request *rq; |
a492f075 | 416 | int ret; |
320ae51f | 417 | |
3a0a5299 | 418 | ret = blk_queue_enter(q, flags); |
a492f075 JL |
419 | if (ret) |
420 | return ERR_PTR(ret); | |
320ae51f | 421 | |
f9afca4d | 422 | rq = blk_mq_get_request(q, NULL, &alloc_data); |
3280d66a | 423 | blk_queue_exit(q); |
841bac2c | 424 | |
bd166ef1 | 425 | if (!rq) |
a492f075 | 426 | return ERR_PTR(-EWOULDBLOCK); |
0c4de0f3 CH |
427 | |
428 | rq->__data_len = 0; | |
429 | rq->__sector = (sector_t) -1; | |
430 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
431 | return rq; |
432 | } | |
4bb659b1 | 433 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 434 | |
cd6ce148 | 435 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
9a95e4ef | 436 | unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 437 | { |
f9afca4d | 438 | struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op }; |
1f5bd336 | 439 | struct request *rq; |
6d2809d5 | 440 | unsigned int cpu; |
1f5bd336 ML |
441 | int ret; |
442 | ||
443 | /* | |
444 | * If the tag allocator sleeps we could get an allocation for a | |
445 | * different hardware context. No need to complicate the low level | |
446 | * allocator for this for the rare use case of a command tied to | |
447 | * a specific queue. | |
448 | */ | |
449 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
450 | return ERR_PTR(-EINVAL); | |
451 | ||
452 | if (hctx_idx >= q->nr_hw_queues) | |
453 | return ERR_PTR(-EIO); | |
454 | ||
3a0a5299 | 455 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
456 | if (ret) |
457 | return ERR_PTR(ret); | |
458 | ||
c8712c6a CH |
459 | /* |
460 | * Check if the hardware context is actually mapped to anything. | |
461 | * If not tell the caller that it should skip this queue. | |
462 | */ | |
6d2809d5 OS |
463 | alloc_data.hctx = q->queue_hw_ctx[hctx_idx]; |
464 | if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) { | |
465 | blk_queue_exit(q); | |
466 | return ERR_PTR(-EXDEV); | |
c8712c6a | 467 | } |
20e4d813 | 468 | cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask); |
6d2809d5 | 469 | alloc_data.ctx = __blk_mq_get_ctx(q, cpu); |
1f5bd336 | 470 | |
f9afca4d | 471 | rq = blk_mq_get_request(q, NULL, &alloc_data); |
3280d66a | 472 | blk_queue_exit(q); |
c8712c6a | 473 | |
6d2809d5 OS |
474 | if (!rq) |
475 | return ERR_PTR(-EWOULDBLOCK); | |
476 | ||
477 | return rq; | |
1f5bd336 ML |
478 | } |
479 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
480 | ||
12f5b931 KB |
481 | static void __blk_mq_free_request(struct request *rq) |
482 | { | |
483 | struct request_queue *q = rq->q; | |
484 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 485 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
12f5b931 KB |
486 | const int sched_tag = rq->internal_tag; |
487 | ||
986d413b | 488 | blk_pm_mark_last_busy(rq); |
ea4f995e | 489 | rq->mq_hctx = NULL; |
12f5b931 KB |
490 | if (rq->tag != -1) |
491 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
492 | if (sched_tag != -1) | |
493 | blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag); | |
494 | blk_mq_sched_restart(hctx); | |
495 | blk_queue_exit(q); | |
496 | } | |
497 | ||
6af54051 | 498 | void blk_mq_free_request(struct request *rq) |
320ae51f | 499 | { |
320ae51f | 500 | struct request_queue *q = rq->q; |
6af54051 CH |
501 | struct elevator_queue *e = q->elevator; |
502 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 503 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
6af54051 | 504 | |
5bbf4e5a | 505 | if (rq->rq_flags & RQF_ELVPRIV) { |
f9cd4bfe JA |
506 | if (e && e->type->ops.finish_request) |
507 | e->type->ops.finish_request(rq); | |
6af54051 CH |
508 | if (rq->elv.icq) { |
509 | put_io_context(rq->elv.icq->ioc); | |
510 | rq->elv.icq = NULL; | |
511 | } | |
512 | } | |
320ae51f | 513 | |
6af54051 | 514 | ctx->rq_completed[rq_is_sync(rq)]++; |
e8064021 | 515 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 516 | atomic_dec(&hctx->nr_active); |
87760e5e | 517 | |
7beb2f84 JA |
518 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
519 | laptop_io_completion(q->backing_dev_info); | |
520 | ||
a7905043 | 521 | rq_qos_done(q, rq); |
0d2602ca | 522 | |
12f5b931 KB |
523 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
524 | if (refcount_dec_and_test(&rq->ref)) | |
525 | __blk_mq_free_request(rq); | |
320ae51f | 526 | } |
1a3b595a | 527 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 528 | |
2a842aca | 529 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
320ae51f | 530 | { |
fe1f4526 JA |
531 | u64 now = 0; |
532 | ||
533 | if (blk_mq_need_time_stamp(rq)) | |
534 | now = ktime_get_ns(); | |
522a7775 | 535 | |
4bc6339a OS |
536 | if (rq->rq_flags & RQF_STATS) { |
537 | blk_mq_poll_stats_start(rq->q); | |
522a7775 | 538 | blk_stat_add(rq, now); |
4bc6339a OS |
539 | } |
540 | ||
ed88660a OS |
541 | if (rq->internal_tag != -1) |
542 | blk_mq_sched_completed_request(rq, now); | |
543 | ||
522a7775 | 544 | blk_account_io_done(rq, now); |
0d11e6ac | 545 | |
91b63639 | 546 | if (rq->end_io) { |
a7905043 | 547 | rq_qos_done(rq->q, rq); |
320ae51f | 548 | rq->end_io(rq, error); |
91b63639 | 549 | } else { |
320ae51f | 550 | blk_mq_free_request(rq); |
91b63639 | 551 | } |
320ae51f | 552 | } |
c8a446ad | 553 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 554 | |
2a842aca | 555 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
556 | { |
557 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
558 | BUG(); | |
c8a446ad | 559 | __blk_mq_end_request(rq, error); |
63151a44 | 560 | } |
c8a446ad | 561 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 562 | |
30a91cb4 | 563 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 564 | { |
3d6efbf6 | 565 | struct request *rq = data; |
c7bb9ad1 | 566 | struct request_queue *q = rq->q; |
320ae51f | 567 | |
c7bb9ad1 | 568 | q->mq_ops->complete(rq); |
320ae51f | 569 | } |
320ae51f | 570 | |
453f8341 | 571 | static void __blk_mq_complete_request(struct request *rq) |
320ae51f JA |
572 | { |
573 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
c7bb9ad1 | 574 | struct request_queue *q = rq->q; |
38535201 | 575 | bool shared = false; |
320ae51f JA |
576 | int cpu; |
577 | ||
af78ff7c | 578 | WRITE_ONCE(rq->state, MQ_RQ_COMPLETE); |
36e76539 ML |
579 | /* |
580 | * Most of single queue controllers, there is only one irq vector | |
581 | * for handling IO completion, and the only irq's affinity is set | |
582 | * as all possible CPUs. On most of ARCHs, this affinity means the | |
583 | * irq is handled on one specific CPU. | |
584 | * | |
585 | * So complete IO reqeust in softirq context in case of single queue | |
586 | * for not degrading IO performance by irqsoff latency. | |
587 | */ | |
c7bb9ad1 | 588 | if (q->nr_hw_queues == 1) { |
36e76539 ML |
589 | __blk_complete_request(rq); |
590 | return; | |
591 | } | |
592 | ||
4ab32bf3 JA |
593 | /* |
594 | * For a polled request, always complete locallly, it's pointless | |
595 | * to redirect the completion. | |
596 | */ | |
597 | if ((rq->cmd_flags & REQ_HIPRI) || | |
598 | !test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags)) { | |
c7bb9ad1 | 599 | q->mq_ops->complete(rq); |
30a91cb4 CH |
600 | return; |
601 | } | |
320ae51f JA |
602 | |
603 | cpu = get_cpu(); | |
c7bb9ad1 | 604 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &q->queue_flags)) |
38535201 CH |
605 | shared = cpus_share_cache(cpu, ctx->cpu); |
606 | ||
607 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 608 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
609 | rq->csd.info = rq; |
610 | rq->csd.flags = 0; | |
c46fff2a | 611 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 612 | } else { |
c7bb9ad1 | 613 | q->mq_ops->complete(rq); |
3d6efbf6 | 614 | } |
320ae51f JA |
615 | put_cpu(); |
616 | } | |
30a91cb4 | 617 | |
04ced159 | 618 | static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx) |
b7435db8 | 619 | __releases(hctx->srcu) |
04ced159 JA |
620 | { |
621 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
622 | rcu_read_unlock(); | |
623 | else | |
05707b64 | 624 | srcu_read_unlock(hctx->srcu, srcu_idx); |
04ced159 JA |
625 | } |
626 | ||
627 | static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx) | |
b7435db8 | 628 | __acquires(hctx->srcu) |
04ced159 | 629 | { |
08b5a6e2 JA |
630 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
631 | /* shut up gcc false positive */ | |
632 | *srcu_idx = 0; | |
04ced159 | 633 | rcu_read_lock(); |
08b5a6e2 | 634 | } else |
05707b64 | 635 | *srcu_idx = srcu_read_lock(hctx->srcu); |
04ced159 JA |
636 | } |
637 | ||
30a91cb4 CH |
638 | /** |
639 | * blk_mq_complete_request - end I/O on a request | |
640 | * @rq: the request being processed | |
641 | * | |
642 | * Description: | |
643 | * Ends all I/O on a request. It does not handle partial completions. | |
644 | * The actual completion happens out-of-order, through a IPI handler. | |
645 | **/ | |
16c15eb1 | 646 | bool blk_mq_complete_request(struct request *rq) |
30a91cb4 | 647 | { |
12f5b931 | 648 | if (unlikely(blk_should_fake_timeout(rq->q))) |
16c15eb1 | 649 | return false; |
12f5b931 | 650 | __blk_mq_complete_request(rq); |
16c15eb1 | 651 | return true; |
30a91cb4 CH |
652 | } |
653 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 654 | |
1b8f21b7 ML |
655 | void blk_mq_complete_request_sync(struct request *rq) |
656 | { | |
657 | WRITE_ONCE(rq->state, MQ_RQ_COMPLETE); | |
658 | rq->q->mq_ops->complete(rq); | |
659 | } | |
660 | EXPORT_SYMBOL_GPL(blk_mq_complete_request_sync); | |
661 | ||
973c0191 KB |
662 | int blk_mq_request_started(struct request *rq) |
663 | { | |
5a61c363 | 664 | return blk_mq_rq_state(rq) != MQ_RQ_IDLE; |
973c0191 KB |
665 | } |
666 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
667 | ||
e2490073 | 668 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
669 | { |
670 | struct request_queue *q = rq->q; | |
671 | ||
672 | trace_block_rq_issue(q, rq); | |
673 | ||
cf43e6be | 674 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
544ccc8d OS |
675 | rq->io_start_time_ns = ktime_get_ns(); |
676 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW | |
677 | rq->throtl_size = blk_rq_sectors(rq); | |
678 | #endif | |
cf43e6be | 679 | rq->rq_flags |= RQF_STATS; |
a7905043 | 680 | rq_qos_issue(q, rq); |
cf43e6be JA |
681 | } |
682 | ||
1d9bd516 | 683 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 684 | |
1d9bd516 | 685 | blk_add_timer(rq); |
12f5b931 | 686 | WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); |
49f5baa5 CH |
687 | |
688 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
689 | /* | |
690 | * Make sure space for the drain appears. We know we can do | |
691 | * this because max_hw_segments has been adjusted to be one | |
692 | * fewer than the device can handle. | |
693 | */ | |
694 | rq->nr_phys_segments++; | |
695 | } | |
320ae51f | 696 | } |
e2490073 | 697 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 698 | |
ed0791b2 | 699 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
700 | { |
701 | struct request_queue *q = rq->q; | |
702 | ||
923218f6 ML |
703 | blk_mq_put_driver_tag(rq); |
704 | ||
320ae51f | 705 | trace_block_rq_requeue(q, rq); |
a7905043 | 706 | rq_qos_requeue(q, rq); |
49f5baa5 | 707 | |
12f5b931 KB |
708 | if (blk_mq_request_started(rq)) { |
709 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); | |
da661267 | 710 | rq->rq_flags &= ~RQF_TIMED_OUT; |
e2490073 CH |
711 | if (q->dma_drain_size && blk_rq_bytes(rq)) |
712 | rq->nr_phys_segments--; | |
713 | } | |
320ae51f JA |
714 | } |
715 | ||
2b053aca | 716 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 717 | { |
ed0791b2 | 718 | __blk_mq_requeue_request(rq); |
ed0791b2 | 719 | |
105976f5 ML |
720 | /* this request will be re-inserted to io scheduler queue */ |
721 | blk_mq_sched_requeue_request(rq); | |
722 | ||
7d692330 | 723 | BUG_ON(!list_empty(&rq->queuelist)); |
2b053aca | 724 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
725 | } |
726 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
727 | ||
6fca6a61 CH |
728 | static void blk_mq_requeue_work(struct work_struct *work) |
729 | { | |
730 | struct request_queue *q = | |
2849450a | 731 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
732 | LIST_HEAD(rq_list); |
733 | struct request *rq, *next; | |
6fca6a61 | 734 | |
18e9781d | 735 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 736 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 737 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
738 | |
739 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
aef1897c | 740 | if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP))) |
6fca6a61 CH |
741 | continue; |
742 | ||
e8064021 | 743 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 744 | list_del_init(&rq->queuelist); |
aef1897c JW |
745 | /* |
746 | * If RQF_DONTPREP, rq has contained some driver specific | |
747 | * data, so insert it to hctx dispatch list to avoid any | |
748 | * merge. | |
749 | */ | |
750 | if (rq->rq_flags & RQF_DONTPREP) | |
751 | blk_mq_request_bypass_insert(rq, false); | |
752 | else | |
753 | blk_mq_sched_insert_request(rq, true, false, false); | |
6fca6a61 CH |
754 | } |
755 | ||
756 | while (!list_empty(&rq_list)) { | |
757 | rq = list_entry(rq_list.next, struct request, queuelist); | |
758 | list_del_init(&rq->queuelist); | |
9e97d295 | 759 | blk_mq_sched_insert_request(rq, false, false, false); |
6fca6a61 CH |
760 | } |
761 | ||
52d7f1b5 | 762 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
763 | } |
764 | ||
2b053aca BVA |
765 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
766 | bool kick_requeue_list) | |
6fca6a61 CH |
767 | { |
768 | struct request_queue *q = rq->q; | |
769 | unsigned long flags; | |
770 | ||
771 | /* | |
772 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
ff821d27 | 773 | * request head insertion from the workqueue. |
6fca6a61 | 774 | */ |
e8064021 | 775 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
776 | |
777 | spin_lock_irqsave(&q->requeue_lock, flags); | |
778 | if (at_head) { | |
e8064021 | 779 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
780 | list_add(&rq->queuelist, &q->requeue_list); |
781 | } else { | |
782 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
783 | } | |
784 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
785 | |
786 | if (kick_requeue_list) | |
787 | blk_mq_kick_requeue_list(q); | |
6fca6a61 | 788 | } |
6fca6a61 CH |
789 | |
790 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
791 | { | |
ae943d20 | 792 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
793 | } |
794 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
795 | ||
2849450a MS |
796 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
797 | unsigned long msecs) | |
798 | { | |
d4acf365 BVA |
799 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
800 | msecs_to_jiffies(msecs)); | |
2849450a MS |
801 | } |
802 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
803 | ||
0e62f51f JA |
804 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
805 | { | |
88c7b2b7 JA |
806 | if (tag < tags->nr_tags) { |
807 | prefetch(tags->rqs[tag]); | |
4ee86bab | 808 | return tags->rqs[tag]; |
88c7b2b7 | 809 | } |
4ee86bab HR |
810 | |
811 | return NULL; | |
24d2f903 CH |
812 | } |
813 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
814 | ||
3c94d83c JA |
815 | static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq, |
816 | void *priv, bool reserved) | |
ae879912 JA |
817 | { |
818 | /* | |
3c94d83c JA |
819 | * If we find a request that is inflight and the queue matches, |
820 | * we know the queue is busy. Return false to stop the iteration. | |
ae879912 | 821 | */ |
3c94d83c | 822 | if (rq->state == MQ_RQ_IN_FLIGHT && rq->q == hctx->queue) { |
ae879912 JA |
823 | bool *busy = priv; |
824 | ||
825 | *busy = true; | |
826 | return false; | |
827 | } | |
828 | ||
829 | return true; | |
830 | } | |
831 | ||
3c94d83c | 832 | bool blk_mq_queue_inflight(struct request_queue *q) |
ae879912 JA |
833 | { |
834 | bool busy = false; | |
835 | ||
3c94d83c | 836 | blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy); |
ae879912 JA |
837 | return busy; |
838 | } | |
3c94d83c | 839 | EXPORT_SYMBOL_GPL(blk_mq_queue_inflight); |
ae879912 | 840 | |
358f70da | 841 | static void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 842 | { |
da661267 | 843 | req->rq_flags |= RQF_TIMED_OUT; |
d1210d5a CH |
844 | if (req->q->mq_ops->timeout) { |
845 | enum blk_eh_timer_return ret; | |
846 | ||
847 | ret = req->q->mq_ops->timeout(req, reserved); | |
848 | if (ret == BLK_EH_DONE) | |
849 | return; | |
850 | WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); | |
46f92d42 | 851 | } |
d1210d5a CH |
852 | |
853 | blk_add_timer(req); | |
87ee7b11 | 854 | } |
5b3f25fc | 855 | |
12f5b931 | 856 | static bool blk_mq_req_expired(struct request *rq, unsigned long *next) |
81481eb4 | 857 | { |
12f5b931 | 858 | unsigned long deadline; |
87ee7b11 | 859 | |
12f5b931 KB |
860 | if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) |
861 | return false; | |
da661267 CH |
862 | if (rq->rq_flags & RQF_TIMED_OUT) |
863 | return false; | |
a7af0af3 | 864 | |
079076b3 | 865 | deadline = READ_ONCE(rq->deadline); |
12f5b931 KB |
866 | if (time_after_eq(jiffies, deadline)) |
867 | return true; | |
a7af0af3 | 868 | |
12f5b931 KB |
869 | if (*next == 0) |
870 | *next = deadline; | |
871 | else if (time_after(*next, deadline)) | |
872 | *next = deadline; | |
873 | return false; | |
87ee7b11 JA |
874 | } |
875 | ||
7baa8572 | 876 | static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
1d9bd516 TH |
877 | struct request *rq, void *priv, bool reserved) |
878 | { | |
12f5b931 KB |
879 | unsigned long *next = priv; |
880 | ||
881 | /* | |
882 | * Just do a quick check if it is expired before locking the request in | |
883 | * so we're not unnecessarilly synchronizing across CPUs. | |
884 | */ | |
885 | if (!blk_mq_req_expired(rq, next)) | |
7baa8572 | 886 | return true; |
12f5b931 KB |
887 | |
888 | /* | |
889 | * We have reason to believe the request may be expired. Take a | |
890 | * reference on the request to lock this request lifetime into its | |
891 | * currently allocated context to prevent it from being reallocated in | |
892 | * the event the completion by-passes this timeout handler. | |
893 | * | |
894 | * If the reference was already released, then the driver beat the | |
895 | * timeout handler to posting a natural completion. | |
896 | */ | |
897 | if (!refcount_inc_not_zero(&rq->ref)) | |
7baa8572 | 898 | return true; |
12f5b931 | 899 | |
1d9bd516 | 900 | /* |
12f5b931 KB |
901 | * The request is now locked and cannot be reallocated underneath the |
902 | * timeout handler's processing. Re-verify this exact request is truly | |
903 | * expired; if it is not expired, then the request was completed and | |
904 | * reallocated as a new request. | |
1d9bd516 | 905 | */ |
12f5b931 | 906 | if (blk_mq_req_expired(rq, next)) |
1d9bd516 | 907 | blk_mq_rq_timed_out(rq, reserved); |
12f5b931 KB |
908 | if (refcount_dec_and_test(&rq->ref)) |
909 | __blk_mq_free_request(rq); | |
7baa8572 JA |
910 | |
911 | return true; | |
1d9bd516 TH |
912 | } |
913 | ||
287922eb | 914 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 915 | { |
287922eb CH |
916 | struct request_queue *q = |
917 | container_of(work, struct request_queue, timeout_work); | |
12f5b931 | 918 | unsigned long next = 0; |
1d9bd516 | 919 | struct blk_mq_hw_ctx *hctx; |
81481eb4 | 920 | int i; |
320ae51f | 921 | |
71f79fb3 GKB |
922 | /* A deadlock might occur if a request is stuck requiring a |
923 | * timeout at the same time a queue freeze is waiting | |
924 | * completion, since the timeout code would not be able to | |
925 | * acquire the queue reference here. | |
926 | * | |
927 | * That's why we don't use blk_queue_enter here; instead, we use | |
928 | * percpu_ref_tryget directly, because we need to be able to | |
929 | * obtain a reference even in the short window between the queue | |
930 | * starting to freeze, by dropping the first reference in | |
1671d522 | 931 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
932 | * consumed, marked by the instant q_usage_counter reaches |
933 | * zero. | |
934 | */ | |
935 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
936 | return; |
937 | ||
12f5b931 | 938 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next); |
320ae51f | 939 | |
12f5b931 KB |
940 | if (next != 0) { |
941 | mod_timer(&q->timeout, next); | |
0d2602ca | 942 | } else { |
fcd36c36 BVA |
943 | /* |
944 | * Request timeouts are handled as a forward rolling timer. If | |
945 | * we end up here it means that no requests are pending and | |
946 | * also that no request has been pending for a while. Mark | |
947 | * each hctx as idle. | |
948 | */ | |
f054b56c ML |
949 | queue_for_each_hw_ctx(q, hctx, i) { |
950 | /* the hctx may be unmapped, so check it here */ | |
951 | if (blk_mq_hw_queue_mapped(hctx)) | |
952 | blk_mq_tag_idle(hctx); | |
953 | } | |
0d2602ca | 954 | } |
287922eb | 955 | blk_queue_exit(q); |
320ae51f JA |
956 | } |
957 | ||
88459642 OS |
958 | struct flush_busy_ctx_data { |
959 | struct blk_mq_hw_ctx *hctx; | |
960 | struct list_head *list; | |
961 | }; | |
962 | ||
963 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
964 | { | |
965 | struct flush_busy_ctx_data *flush_data = data; | |
966 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
967 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 968 | enum hctx_type type = hctx->type; |
88459642 | 969 | |
88459642 | 970 | spin_lock(&ctx->lock); |
c16d6b5a | 971 | list_splice_tail_init(&ctx->rq_lists[type], flush_data->list); |
e9a99a63 | 972 | sbitmap_clear_bit(sb, bitnr); |
88459642 OS |
973 | spin_unlock(&ctx->lock); |
974 | return true; | |
975 | } | |
976 | ||
1429d7c9 JA |
977 | /* |
978 | * Process software queues that have been marked busy, splicing them | |
979 | * to the for-dispatch | |
980 | */ | |
2c3ad667 | 981 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 982 | { |
88459642 OS |
983 | struct flush_busy_ctx_data data = { |
984 | .hctx = hctx, | |
985 | .list = list, | |
986 | }; | |
1429d7c9 | 987 | |
88459642 | 988 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 989 | } |
2c3ad667 | 990 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 991 | |
b347689f ML |
992 | struct dispatch_rq_data { |
993 | struct blk_mq_hw_ctx *hctx; | |
994 | struct request *rq; | |
995 | }; | |
996 | ||
997 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
998 | void *data) | |
999 | { | |
1000 | struct dispatch_rq_data *dispatch_data = data; | |
1001 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
1002 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1003 | enum hctx_type type = hctx->type; |
b347689f ML |
1004 | |
1005 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
1006 | if (!list_empty(&ctx->rq_lists[type])) { |
1007 | dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next); | |
b347689f | 1008 | list_del_init(&dispatch_data->rq->queuelist); |
c16d6b5a | 1009 | if (list_empty(&ctx->rq_lists[type])) |
b347689f ML |
1010 | sbitmap_clear_bit(sb, bitnr); |
1011 | } | |
1012 | spin_unlock(&ctx->lock); | |
1013 | ||
1014 | return !dispatch_data->rq; | |
1015 | } | |
1016 | ||
1017 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
1018 | struct blk_mq_ctx *start) | |
1019 | { | |
f31967f0 | 1020 | unsigned off = start ? start->index_hw[hctx->type] : 0; |
b347689f ML |
1021 | struct dispatch_rq_data data = { |
1022 | .hctx = hctx, | |
1023 | .rq = NULL, | |
1024 | }; | |
1025 | ||
1026 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1027 | dispatch_rq_from_ctx, &data); | |
1028 | ||
1029 | return data.rq; | |
1030 | } | |
1031 | ||
703fd1c0 JA |
1032 | static inline unsigned int queued_to_index(unsigned int queued) |
1033 | { | |
1034 | if (!queued) | |
1035 | return 0; | |
1429d7c9 | 1036 | |
703fd1c0 | 1037 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
1038 | } |
1039 | ||
8ab6bb9e | 1040 | bool blk_mq_get_driver_tag(struct request *rq) |
bd166ef1 JA |
1041 | { |
1042 | struct blk_mq_alloc_data data = { | |
1043 | .q = rq->q, | |
ea4f995e | 1044 | .hctx = rq->mq_hctx, |
8ab6bb9e | 1045 | .flags = BLK_MQ_REQ_NOWAIT, |
f9afca4d | 1046 | .cmd_flags = rq->cmd_flags, |
bd166ef1 | 1047 | }; |
d263ed99 | 1048 | bool shared; |
5feeacdd | 1049 | |
81380ca1 OS |
1050 | if (rq->tag != -1) |
1051 | goto done; | |
bd166ef1 | 1052 | |
415b806d SG |
1053 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) |
1054 | data.flags |= BLK_MQ_REQ_RESERVED; | |
1055 | ||
d263ed99 | 1056 | shared = blk_mq_tag_busy(data.hctx); |
bd166ef1 JA |
1057 | rq->tag = blk_mq_get_tag(&data); |
1058 | if (rq->tag >= 0) { | |
d263ed99 | 1059 | if (shared) { |
200e86b3 JA |
1060 | rq->rq_flags |= RQF_MQ_INFLIGHT; |
1061 | atomic_inc(&data.hctx->nr_active); | |
1062 | } | |
bd166ef1 | 1063 | data.hctx->tags->rqs[rq->tag] = rq; |
bd166ef1 JA |
1064 | } |
1065 | ||
81380ca1 | 1066 | done: |
81380ca1 | 1067 | return rq->tag != -1; |
bd166ef1 JA |
1068 | } |
1069 | ||
eb619fdb JA |
1070 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1071 | int flags, void *key) | |
da55f2cc OS |
1072 | { |
1073 | struct blk_mq_hw_ctx *hctx; | |
1074 | ||
1075 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1076 | ||
5815839b | 1077 | spin_lock(&hctx->dispatch_wait_lock); |
e8618575 JA |
1078 | if (!list_empty(&wait->entry)) { |
1079 | struct sbitmap_queue *sbq; | |
1080 | ||
1081 | list_del_init(&wait->entry); | |
1082 | sbq = &hctx->tags->bitmap_tags; | |
1083 | atomic_dec(&sbq->ws_active); | |
1084 | } | |
5815839b ML |
1085 | spin_unlock(&hctx->dispatch_wait_lock); |
1086 | ||
da55f2cc OS |
1087 | blk_mq_run_hw_queue(hctx, true); |
1088 | return 1; | |
1089 | } | |
1090 | ||
f906a6a0 JA |
1091 | /* |
1092 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1093 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1094 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1095 | * marking us as waiting. |
1096 | */ | |
2278d69f | 1097 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, |
f906a6a0 | 1098 | struct request *rq) |
da55f2cc | 1099 | { |
e8618575 | 1100 | struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags; |
5815839b | 1101 | struct wait_queue_head *wq; |
f906a6a0 JA |
1102 | wait_queue_entry_t *wait; |
1103 | bool ret; | |
da55f2cc | 1104 | |
2278d69f | 1105 | if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) { |
684b7324 | 1106 | blk_mq_sched_mark_restart_hctx(hctx); |
f906a6a0 | 1107 | |
c27d53fb BVA |
1108 | /* |
1109 | * It's possible that a tag was freed in the window between the | |
1110 | * allocation failure and adding the hardware queue to the wait | |
1111 | * queue. | |
1112 | * | |
1113 | * Don't clear RESTART here, someone else could have set it. | |
1114 | * At most this will cost an extra queue run. | |
1115 | */ | |
8ab6bb9e | 1116 | return blk_mq_get_driver_tag(rq); |
eb619fdb | 1117 | } |
eb619fdb | 1118 | |
2278d69f | 1119 | wait = &hctx->dispatch_wait; |
c27d53fb BVA |
1120 | if (!list_empty_careful(&wait->entry)) |
1121 | return false; | |
1122 | ||
e8618575 | 1123 | wq = &bt_wait_ptr(sbq, hctx)->wait; |
5815839b ML |
1124 | |
1125 | spin_lock_irq(&wq->lock); | |
1126 | spin_lock(&hctx->dispatch_wait_lock); | |
c27d53fb | 1127 | if (!list_empty(&wait->entry)) { |
5815839b ML |
1128 | spin_unlock(&hctx->dispatch_wait_lock); |
1129 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1130 | return false; |
eb619fdb JA |
1131 | } |
1132 | ||
e8618575 | 1133 | atomic_inc(&sbq->ws_active); |
5815839b ML |
1134 | wait->flags &= ~WQ_FLAG_EXCLUSIVE; |
1135 | __add_wait_queue(wq, wait); | |
c27d53fb | 1136 | |
da55f2cc | 1137 | /* |
eb619fdb JA |
1138 | * It's possible that a tag was freed in the window between the |
1139 | * allocation failure and adding the hardware queue to the wait | |
1140 | * queue. | |
da55f2cc | 1141 | */ |
8ab6bb9e | 1142 | ret = blk_mq_get_driver_tag(rq); |
c27d53fb | 1143 | if (!ret) { |
5815839b ML |
1144 | spin_unlock(&hctx->dispatch_wait_lock); |
1145 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1146 | return false; |
eb619fdb | 1147 | } |
c27d53fb BVA |
1148 | |
1149 | /* | |
1150 | * We got a tag, remove ourselves from the wait queue to ensure | |
1151 | * someone else gets the wakeup. | |
1152 | */ | |
c27d53fb | 1153 | list_del_init(&wait->entry); |
e8618575 | 1154 | atomic_dec(&sbq->ws_active); |
5815839b ML |
1155 | spin_unlock(&hctx->dispatch_wait_lock); |
1156 | spin_unlock_irq(&wq->lock); | |
c27d53fb BVA |
1157 | |
1158 | return true; | |
da55f2cc OS |
1159 | } |
1160 | ||
6e768717 ML |
1161 | #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 |
1162 | #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 | |
1163 | /* | |
1164 | * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): | |
1165 | * - EWMA is one simple way to compute running average value | |
1166 | * - weight(7/8 and 1/8) is applied so that it can decrease exponentially | |
1167 | * - take 4 as factor for avoiding to get too small(0) result, and this | |
1168 | * factor doesn't matter because EWMA decreases exponentially | |
1169 | */ | |
1170 | static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) | |
1171 | { | |
1172 | unsigned int ewma; | |
1173 | ||
1174 | if (hctx->queue->elevator) | |
1175 | return; | |
1176 | ||
1177 | ewma = hctx->dispatch_busy; | |
1178 | ||
1179 | if (!ewma && !busy) | |
1180 | return; | |
1181 | ||
1182 | ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; | |
1183 | if (busy) | |
1184 | ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; | |
1185 | ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; | |
1186 | ||
1187 | hctx->dispatch_busy = ewma; | |
1188 | } | |
1189 | ||
86ff7c2a ML |
1190 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
1191 | ||
1f57f8d4 JA |
1192 | /* |
1193 | * Returns true if we did some work AND can potentially do more. | |
1194 | */ | |
de148297 | 1195 | bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list, |
eb619fdb | 1196 | bool got_budget) |
320ae51f | 1197 | { |
81380ca1 | 1198 | struct blk_mq_hw_ctx *hctx; |
6d6f167c | 1199 | struct request *rq, *nxt; |
eb619fdb | 1200 | bool no_tag = false; |
fc17b653 | 1201 | int errors, queued; |
86ff7c2a | 1202 | blk_status_t ret = BLK_STS_OK; |
320ae51f | 1203 | |
81380ca1 OS |
1204 | if (list_empty(list)) |
1205 | return false; | |
1206 | ||
de148297 ML |
1207 | WARN_ON(!list_is_singular(list) && got_budget); |
1208 | ||
320ae51f JA |
1209 | /* |
1210 | * Now process all the entries, sending them to the driver. | |
1211 | */ | |
93efe981 | 1212 | errors = queued = 0; |
81380ca1 | 1213 | do { |
74c45052 | 1214 | struct blk_mq_queue_data bd; |
320ae51f | 1215 | |
f04c3df3 | 1216 | rq = list_first_entry(list, struct request, queuelist); |
0bca799b | 1217 | |
ea4f995e | 1218 | hctx = rq->mq_hctx; |
0bca799b ML |
1219 | if (!got_budget && !blk_mq_get_dispatch_budget(hctx)) |
1220 | break; | |
1221 | ||
8ab6bb9e | 1222 | if (!blk_mq_get_driver_tag(rq)) { |
3c782d67 | 1223 | /* |
da55f2cc | 1224 | * The initial allocation attempt failed, so we need to |
eb619fdb JA |
1225 | * rerun the hardware queue when a tag is freed. The |
1226 | * waitqueue takes care of that. If the queue is run | |
1227 | * before we add this entry back on the dispatch list, | |
1228 | * we'll re-run it below. | |
3c782d67 | 1229 | */ |
2278d69f | 1230 | if (!blk_mq_mark_tag_wait(hctx, rq)) { |
0bca799b | 1231 | blk_mq_put_dispatch_budget(hctx); |
f906a6a0 JA |
1232 | /* |
1233 | * For non-shared tags, the RESTART check | |
1234 | * will suffice. | |
1235 | */ | |
1236 | if (hctx->flags & BLK_MQ_F_TAG_SHARED) | |
1237 | no_tag = true; | |
de148297 ML |
1238 | break; |
1239 | } | |
1240 | } | |
1241 | ||
320ae51f | 1242 | list_del_init(&rq->queuelist); |
320ae51f | 1243 | |
74c45052 | 1244 | bd.rq = rq; |
113285b4 JA |
1245 | |
1246 | /* | |
1247 | * Flag last if we have no more requests, or if we have more | |
1248 | * but can't assign a driver tag to it. | |
1249 | */ | |
1250 | if (list_empty(list)) | |
1251 | bd.last = true; | |
1252 | else { | |
113285b4 | 1253 | nxt = list_first_entry(list, struct request, queuelist); |
8ab6bb9e | 1254 | bd.last = !blk_mq_get_driver_tag(nxt); |
113285b4 | 1255 | } |
74c45052 JA |
1256 | |
1257 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
86ff7c2a | 1258 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) { |
6d6f167c JW |
1259 | /* |
1260 | * If an I/O scheduler has been configured and we got a | |
ff821d27 JA |
1261 | * driver tag for the next request already, free it |
1262 | * again. | |
6d6f167c JW |
1263 | */ |
1264 | if (!list_empty(list)) { | |
1265 | nxt = list_first_entry(list, struct request, queuelist); | |
1266 | blk_mq_put_driver_tag(nxt); | |
1267 | } | |
f04c3df3 | 1268 | list_add(&rq->queuelist, list); |
ed0791b2 | 1269 | __blk_mq_requeue_request(rq); |
320ae51f | 1270 | break; |
fc17b653 CH |
1271 | } |
1272 | ||
1273 | if (unlikely(ret != BLK_STS_OK)) { | |
93efe981 | 1274 | errors++; |
2a842aca | 1275 | blk_mq_end_request(rq, BLK_STS_IOERR); |
fc17b653 | 1276 | continue; |
320ae51f JA |
1277 | } |
1278 | ||
fc17b653 | 1279 | queued++; |
81380ca1 | 1280 | } while (!list_empty(list)); |
320ae51f | 1281 | |
703fd1c0 | 1282 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
1283 | |
1284 | /* | |
1285 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1286 | * that is where we will continue on next queue run. | |
1287 | */ | |
f04c3df3 | 1288 | if (!list_empty(list)) { |
86ff7c2a ML |
1289 | bool needs_restart; |
1290 | ||
d666ba98 JA |
1291 | /* |
1292 | * If we didn't flush the entire list, we could have told | |
1293 | * the driver there was more coming, but that turned out to | |
1294 | * be a lie. | |
1295 | */ | |
1296 | if (q->mq_ops->commit_rqs) | |
1297 | q->mq_ops->commit_rqs(hctx); | |
1298 | ||
320ae51f | 1299 | spin_lock(&hctx->lock); |
c13660a0 | 1300 | list_splice_init(list, &hctx->dispatch); |
320ae51f | 1301 | spin_unlock(&hctx->lock); |
f04c3df3 | 1302 | |
9ba52e58 | 1303 | /* |
710c785f BVA |
1304 | * If SCHED_RESTART was set by the caller of this function and |
1305 | * it is no longer set that means that it was cleared by another | |
1306 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1307 | * |
eb619fdb JA |
1308 | * If 'no_tag' is set, that means that we failed getting |
1309 | * a driver tag with an I/O scheduler attached. If our dispatch | |
1310 | * waitqueue is no longer active, ensure that we run the queue | |
1311 | * AFTER adding our entries back to the list. | |
bd166ef1 | 1312 | * |
710c785f BVA |
1313 | * If no I/O scheduler has been configured it is possible that |
1314 | * the hardware queue got stopped and restarted before requests | |
1315 | * were pushed back onto the dispatch list. Rerun the queue to | |
1316 | * avoid starvation. Notes: | |
1317 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1318 | * been stopped before rerunning a queue. | |
1319 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1320 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1321 | * and dm-rq. |
86ff7c2a ML |
1322 | * |
1323 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
1324 | * bit is set, run queue after a delay to avoid IO stalls | |
1325 | * that could otherwise occur if the queue is idle. | |
bd166ef1 | 1326 | */ |
86ff7c2a ML |
1327 | needs_restart = blk_mq_sched_needs_restart(hctx); |
1328 | if (!needs_restart || | |
eb619fdb | 1329 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
bd166ef1 | 1330 | blk_mq_run_hw_queue(hctx, true); |
86ff7c2a ML |
1331 | else if (needs_restart && (ret == BLK_STS_RESOURCE)) |
1332 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); | |
1f57f8d4 | 1333 | |
6e768717 | 1334 | blk_mq_update_dispatch_busy(hctx, true); |
1f57f8d4 | 1335 | return false; |
6e768717 ML |
1336 | } else |
1337 | blk_mq_update_dispatch_busy(hctx, false); | |
f04c3df3 | 1338 | |
1f57f8d4 JA |
1339 | /* |
1340 | * If the host/device is unable to accept more work, inform the | |
1341 | * caller of that. | |
1342 | */ | |
1343 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) | |
1344 | return false; | |
1345 | ||
93efe981 | 1346 | return (queued + errors) != 0; |
f04c3df3 JA |
1347 | } |
1348 | ||
6a83e74d BVA |
1349 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1350 | { | |
1351 | int srcu_idx; | |
1352 | ||
b7a71e66 JA |
1353 | /* |
1354 | * We should be running this queue from one of the CPUs that | |
1355 | * are mapped to it. | |
7df938fb ML |
1356 | * |
1357 | * There are at least two related races now between setting | |
1358 | * hctx->next_cpu from blk_mq_hctx_next_cpu() and running | |
1359 | * __blk_mq_run_hw_queue(): | |
1360 | * | |
1361 | * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(), | |
1362 | * but later it becomes online, then this warning is harmless | |
1363 | * at all | |
1364 | * | |
1365 | * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(), | |
1366 | * but later it becomes offline, then the warning can't be | |
1367 | * triggered, and we depend on blk-mq timeout handler to | |
1368 | * handle dispatched requests to this hctx | |
b7a71e66 | 1369 | */ |
7df938fb ML |
1370 | if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && |
1371 | cpu_online(hctx->next_cpu)) { | |
1372 | printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n", | |
1373 | raw_smp_processor_id(), | |
1374 | cpumask_empty(hctx->cpumask) ? "inactive": "active"); | |
1375 | dump_stack(); | |
1376 | } | |
6a83e74d | 1377 | |
b7a71e66 JA |
1378 | /* |
1379 | * We can't run the queue inline with ints disabled. Ensure that | |
1380 | * we catch bad users of this early. | |
1381 | */ | |
1382 | WARN_ON_ONCE(in_interrupt()); | |
1383 | ||
04ced159 | 1384 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1385 | |
04ced159 JA |
1386 | hctx_lock(hctx, &srcu_idx); |
1387 | blk_mq_sched_dispatch_requests(hctx); | |
1388 | hctx_unlock(hctx, srcu_idx); | |
6a83e74d BVA |
1389 | } |
1390 | ||
f82ddf19 ML |
1391 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
1392 | { | |
1393 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
1394 | ||
1395 | if (cpu >= nr_cpu_ids) | |
1396 | cpu = cpumask_first(hctx->cpumask); | |
1397 | return cpu; | |
1398 | } | |
1399 | ||
506e931f JA |
1400 | /* |
1401 | * It'd be great if the workqueue API had a way to pass | |
1402 | * in a mask and had some smarts for more clever placement. | |
1403 | * For now we just round-robin here, switching for every | |
1404 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1405 | */ | |
1406 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1407 | { | |
7bed4595 | 1408 | bool tried = false; |
476f8c98 | 1409 | int next_cpu = hctx->next_cpu; |
7bed4595 | 1410 | |
b657d7e6 CH |
1411 | if (hctx->queue->nr_hw_queues == 1) |
1412 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1413 | |
1414 | if (--hctx->next_cpu_batch <= 0) { | |
7bed4595 | 1415 | select_cpu: |
476f8c98 | 1416 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
20e4d813 | 1417 | cpu_online_mask); |
506e931f | 1418 | if (next_cpu >= nr_cpu_ids) |
f82ddf19 | 1419 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
1420 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
1421 | } | |
1422 | ||
7bed4595 ML |
1423 | /* |
1424 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
1425 | * and it should only happen in the path of handling CPU DEAD. | |
1426 | */ | |
476f8c98 | 1427 | if (!cpu_online(next_cpu)) { |
7bed4595 ML |
1428 | if (!tried) { |
1429 | tried = true; | |
1430 | goto select_cpu; | |
1431 | } | |
1432 | ||
1433 | /* | |
1434 | * Make sure to re-select CPU next time once after CPUs | |
1435 | * in hctx->cpumask become online again. | |
1436 | */ | |
476f8c98 | 1437 | hctx->next_cpu = next_cpu; |
7bed4595 ML |
1438 | hctx->next_cpu_batch = 1; |
1439 | return WORK_CPU_UNBOUND; | |
1440 | } | |
476f8c98 ML |
1441 | |
1442 | hctx->next_cpu = next_cpu; | |
1443 | return next_cpu; | |
506e931f JA |
1444 | } |
1445 | ||
7587a5ae BVA |
1446 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1447 | unsigned long msecs) | |
320ae51f | 1448 | { |
5435c023 | 1449 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f JA |
1450 | return; |
1451 | ||
1b792f2f | 1452 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1453 | int cpu = get_cpu(); |
1454 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1455 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1456 | put_cpu(); |
398205b8 PB |
1457 | return; |
1458 | } | |
e4043dcf | 1459 | |
2a90d4aa | 1460 | put_cpu(); |
e4043dcf | 1461 | } |
398205b8 | 1462 | |
ae943d20 BVA |
1463 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
1464 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
1465 | } |
1466 | ||
1467 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1468 | { | |
1469 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1470 | } | |
1471 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1472 | ||
79f720a7 | 1473 | bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 1474 | { |
24f5a90f ML |
1475 | int srcu_idx; |
1476 | bool need_run; | |
1477 | ||
1478 | /* | |
1479 | * When queue is quiesced, we may be switching io scheduler, or | |
1480 | * updating nr_hw_queues, or other things, and we can't run queue | |
1481 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
1482 | * | |
1483 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
1484 | * quiesced. | |
1485 | */ | |
04ced159 JA |
1486 | hctx_lock(hctx, &srcu_idx); |
1487 | need_run = !blk_queue_quiesced(hctx->queue) && | |
1488 | blk_mq_hctx_has_pending(hctx); | |
1489 | hctx_unlock(hctx, srcu_idx); | |
24f5a90f ML |
1490 | |
1491 | if (need_run) { | |
79f720a7 JA |
1492 | __blk_mq_delay_run_hw_queue(hctx, async, 0); |
1493 | return true; | |
1494 | } | |
1495 | ||
1496 | return false; | |
320ae51f | 1497 | } |
5b727272 | 1498 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1499 | |
b94ec296 | 1500 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1501 | { |
1502 | struct blk_mq_hw_ctx *hctx; | |
1503 | int i; | |
1504 | ||
1505 | queue_for_each_hw_ctx(q, hctx, i) { | |
79f720a7 | 1506 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1507 | continue; |
1508 | ||
b94ec296 | 1509 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1510 | } |
1511 | } | |
b94ec296 | 1512 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1513 | |
fd001443 BVA |
1514 | /** |
1515 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1516 | * @q: request queue. | |
1517 | * | |
1518 | * The caller is responsible for serializing this function against | |
1519 | * blk_mq_{start,stop}_hw_queue(). | |
1520 | */ | |
1521 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1522 | { | |
1523 | struct blk_mq_hw_ctx *hctx; | |
1524 | int i; | |
1525 | ||
1526 | queue_for_each_hw_ctx(q, hctx, i) | |
1527 | if (blk_mq_hctx_stopped(hctx)) | |
1528 | return true; | |
1529 | ||
1530 | return false; | |
1531 | } | |
1532 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1533 | ||
39a70c76 ML |
1534 | /* |
1535 | * This function is often used for pausing .queue_rq() by driver when | |
1536 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1537 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1538 | * |
1539 | * We do not guarantee that dispatch can be drained or blocked | |
1540 | * after blk_mq_stop_hw_queue() returns. Please use | |
1541 | * blk_mq_quiesce_queue() for that requirement. | |
1542 | */ | |
2719aa21 JA |
1543 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1544 | { | |
641a9ed6 | 1545 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 1546 | |
641a9ed6 | 1547 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 1548 | } |
641a9ed6 | 1549 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 1550 | |
39a70c76 ML |
1551 | /* |
1552 | * This function is often used for pausing .queue_rq() by driver when | |
1553 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1554 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1555 | * |
1556 | * We do not guarantee that dispatch can be drained or blocked | |
1557 | * after blk_mq_stop_hw_queues() returns. Please use | |
1558 | * blk_mq_quiesce_queue() for that requirement. | |
1559 | */ | |
2719aa21 JA |
1560 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1561 | { | |
641a9ed6 ML |
1562 | struct blk_mq_hw_ctx *hctx; |
1563 | int i; | |
1564 | ||
1565 | queue_for_each_hw_ctx(q, hctx, i) | |
1566 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
1567 | } |
1568 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1569 | ||
320ae51f JA |
1570 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1571 | { | |
1572 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1573 | |
0ffbce80 | 1574 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1575 | } |
1576 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1577 | ||
2f268556 CH |
1578 | void blk_mq_start_hw_queues(struct request_queue *q) |
1579 | { | |
1580 | struct blk_mq_hw_ctx *hctx; | |
1581 | int i; | |
1582 | ||
1583 | queue_for_each_hw_ctx(q, hctx, i) | |
1584 | blk_mq_start_hw_queue(hctx); | |
1585 | } | |
1586 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1587 | ||
ae911c5e JA |
1588 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1589 | { | |
1590 | if (!blk_mq_hctx_stopped(hctx)) | |
1591 | return; | |
1592 | ||
1593 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1594 | blk_mq_run_hw_queue(hctx, async); | |
1595 | } | |
1596 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1597 | ||
1b4a3258 | 1598 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1599 | { |
1600 | struct blk_mq_hw_ctx *hctx; | |
1601 | int i; | |
1602 | ||
ae911c5e JA |
1603 | queue_for_each_hw_ctx(q, hctx, i) |
1604 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1605 | } |
1606 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1607 | ||
70f4db63 | 1608 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1609 | { |
1610 | struct blk_mq_hw_ctx *hctx; | |
1611 | ||
9f993737 | 1612 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 1613 | |
21c6e939 | 1614 | /* |
15fe8a90 | 1615 | * If we are stopped, don't run the queue. |
21c6e939 | 1616 | */ |
15fe8a90 | 1617 | if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
0196d6b4 | 1618 | return; |
7587a5ae BVA |
1619 | |
1620 | __blk_mq_run_hw_queue(hctx); | |
1621 | } | |
1622 | ||
cfd0c552 | 1623 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1624 | struct request *rq, |
1625 | bool at_head) | |
320ae51f | 1626 | { |
e57690fe | 1627 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
c16d6b5a | 1628 | enum hctx_type type = hctx->type; |
e57690fe | 1629 | |
7b607814 BVA |
1630 | lockdep_assert_held(&ctx->lock); |
1631 | ||
01b983c9 JA |
1632 | trace_block_rq_insert(hctx->queue, rq); |
1633 | ||
72a0a36e | 1634 | if (at_head) |
c16d6b5a | 1635 | list_add(&rq->queuelist, &ctx->rq_lists[type]); |
72a0a36e | 1636 | else |
c16d6b5a | 1637 | list_add_tail(&rq->queuelist, &ctx->rq_lists[type]); |
cfd0c552 | 1638 | } |
4bb659b1 | 1639 | |
2c3ad667 JA |
1640 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1641 | bool at_head) | |
cfd0c552 ML |
1642 | { |
1643 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1644 | ||
7b607814 BVA |
1645 | lockdep_assert_held(&ctx->lock); |
1646 | ||
e57690fe | 1647 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1648 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1649 | } |
1650 | ||
157f377b JA |
1651 | /* |
1652 | * Should only be used carefully, when the caller knows we want to | |
1653 | * bypass a potential IO scheduler on the target device. | |
1654 | */ | |
b0850297 | 1655 | void blk_mq_request_bypass_insert(struct request *rq, bool run_queue) |
157f377b | 1656 | { |
ea4f995e | 1657 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
157f377b JA |
1658 | |
1659 | spin_lock(&hctx->lock); | |
1660 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
1661 | spin_unlock(&hctx->lock); | |
1662 | ||
b0850297 ML |
1663 | if (run_queue) |
1664 | blk_mq_run_hw_queue(hctx, false); | |
157f377b JA |
1665 | } |
1666 | ||
bd166ef1 JA |
1667 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1668 | struct list_head *list) | |
320ae51f JA |
1669 | |
1670 | { | |
3f0cedc7 | 1671 | struct request *rq; |
c16d6b5a | 1672 | enum hctx_type type = hctx->type; |
3f0cedc7 | 1673 | |
320ae51f JA |
1674 | /* |
1675 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1676 | * offline now | |
1677 | */ | |
3f0cedc7 | 1678 | list_for_each_entry(rq, list, queuelist) { |
e57690fe | 1679 | BUG_ON(rq->mq_ctx != ctx); |
3f0cedc7 | 1680 | trace_block_rq_insert(hctx->queue, rq); |
320ae51f | 1681 | } |
3f0cedc7 ML |
1682 | |
1683 | spin_lock(&ctx->lock); | |
c16d6b5a | 1684 | list_splice_tail_init(list, &ctx->rq_lists[type]); |
cfd0c552 | 1685 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1686 | spin_unlock(&ctx->lock); |
320ae51f JA |
1687 | } |
1688 | ||
3110fc79 | 1689 | static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b) |
320ae51f JA |
1690 | { |
1691 | struct request *rqa = container_of(a, struct request, queuelist); | |
1692 | struct request *rqb = container_of(b, struct request, queuelist); | |
1693 | ||
3110fc79 JA |
1694 | if (rqa->mq_ctx < rqb->mq_ctx) |
1695 | return -1; | |
1696 | else if (rqa->mq_ctx > rqb->mq_ctx) | |
1697 | return 1; | |
1698 | else if (rqa->mq_hctx < rqb->mq_hctx) | |
1699 | return -1; | |
1700 | else if (rqa->mq_hctx > rqb->mq_hctx) | |
1701 | return 1; | |
1702 | ||
1703 | return blk_rq_pos(rqa) > blk_rq_pos(rqb); | |
320ae51f JA |
1704 | } |
1705 | ||
1706 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1707 | { | |
67cae4c9 | 1708 | struct blk_mq_hw_ctx *this_hctx; |
320ae51f JA |
1709 | struct blk_mq_ctx *this_ctx; |
1710 | struct request_queue *this_q; | |
1711 | struct request *rq; | |
1712 | LIST_HEAD(list); | |
67cae4c9 | 1713 | LIST_HEAD(rq_list); |
320ae51f JA |
1714 | unsigned int depth; |
1715 | ||
1716 | list_splice_init(&plug->mq_list, &list); | |
1717 | ||
ce5b009c JA |
1718 | if (plug->rq_count > 2 && plug->multiple_queues) |
1719 | list_sort(NULL, &list, plug_rq_cmp); | |
320ae51f | 1720 | |
bcc816df DZ |
1721 | plug->rq_count = 0; |
1722 | ||
320ae51f | 1723 | this_q = NULL; |
67cae4c9 | 1724 | this_hctx = NULL; |
320ae51f JA |
1725 | this_ctx = NULL; |
1726 | depth = 0; | |
1727 | ||
1728 | while (!list_empty(&list)) { | |
1729 | rq = list_entry_rq(list.next); | |
1730 | list_del_init(&rq->queuelist); | |
1731 | BUG_ON(!rq->q); | |
67cae4c9 JA |
1732 | if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) { |
1733 | if (this_hctx) { | |
587562d0 | 1734 | trace_block_unplug(this_q, depth, !from_schedule); |
67cae4c9 JA |
1735 | blk_mq_sched_insert_requests(this_hctx, this_ctx, |
1736 | &rq_list, | |
bd166ef1 | 1737 | from_schedule); |
320ae51f JA |
1738 | } |
1739 | ||
320ae51f | 1740 | this_q = rq->q; |
67cae4c9 JA |
1741 | this_ctx = rq->mq_ctx; |
1742 | this_hctx = rq->mq_hctx; | |
320ae51f JA |
1743 | depth = 0; |
1744 | } | |
1745 | ||
1746 | depth++; | |
67cae4c9 | 1747 | list_add_tail(&rq->queuelist, &rq_list); |
320ae51f JA |
1748 | } |
1749 | ||
1750 | /* | |
67cae4c9 JA |
1751 | * If 'this_hctx' is set, we know we have entries to complete |
1752 | * on 'rq_list'. Do those. | |
320ae51f | 1753 | */ |
67cae4c9 | 1754 | if (this_hctx) { |
587562d0 | 1755 | trace_block_unplug(this_q, depth, !from_schedule); |
67cae4c9 | 1756 | blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list, |
bd166ef1 | 1757 | from_schedule); |
320ae51f JA |
1758 | } |
1759 | } | |
1760 | ||
14ccb66b CH |
1761 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio, |
1762 | unsigned int nr_segs) | |
320ae51f | 1763 | { |
f924cdde CH |
1764 | if (bio->bi_opf & REQ_RAHEAD) |
1765 | rq->cmd_flags |= REQ_FAILFAST_MASK; | |
1766 | ||
1767 | rq->__sector = bio->bi_iter.bi_sector; | |
1768 | rq->write_hint = bio->bi_write_hint; | |
14ccb66b | 1769 | blk_rq_bio_prep(rq, bio, nr_segs); |
4b570521 | 1770 | |
6e85eaf3 | 1771 | blk_account_io_start(rq, true); |
320ae51f JA |
1772 | } |
1773 | ||
0f95549c MS |
1774 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
1775 | struct request *rq, | |
be94f058 | 1776 | blk_qc_t *cookie, bool last) |
f984df1f | 1777 | { |
f984df1f | 1778 | struct request_queue *q = rq->q; |
f984df1f SL |
1779 | struct blk_mq_queue_data bd = { |
1780 | .rq = rq, | |
be94f058 | 1781 | .last = last, |
f984df1f | 1782 | }; |
bd166ef1 | 1783 | blk_qc_t new_cookie; |
f06345ad | 1784 | blk_status_t ret; |
0f95549c MS |
1785 | |
1786 | new_cookie = request_to_qc_t(hctx, rq); | |
1787 | ||
1788 | /* | |
1789 | * For OK queue, we are done. For error, caller may kill it. | |
1790 | * Any other error (busy), just add it to our list as we | |
1791 | * previously would have done. | |
1792 | */ | |
1793 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
1794 | switch (ret) { | |
1795 | case BLK_STS_OK: | |
6ce3dd6e | 1796 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
1797 | *cookie = new_cookie; |
1798 | break; | |
1799 | case BLK_STS_RESOURCE: | |
86ff7c2a | 1800 | case BLK_STS_DEV_RESOURCE: |
6ce3dd6e | 1801 | blk_mq_update_dispatch_busy(hctx, true); |
0f95549c MS |
1802 | __blk_mq_requeue_request(rq); |
1803 | break; | |
1804 | default: | |
6ce3dd6e | 1805 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
1806 | *cookie = BLK_QC_T_NONE; |
1807 | break; | |
1808 | } | |
1809 | ||
1810 | return ret; | |
1811 | } | |
1812 | ||
fd9c40f6 | 1813 | static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
0f95549c | 1814 | struct request *rq, |
396eaf21 | 1815 | blk_qc_t *cookie, |
fd9c40f6 | 1816 | bool bypass_insert, bool last) |
0f95549c MS |
1817 | { |
1818 | struct request_queue *q = rq->q; | |
d964f04a ML |
1819 | bool run_queue = true; |
1820 | ||
23d4ee19 | 1821 | /* |
fd9c40f6 | 1822 | * RCU or SRCU read lock is needed before checking quiesced flag. |
23d4ee19 | 1823 | * |
fd9c40f6 BVA |
1824 | * When queue is stopped or quiesced, ignore 'bypass_insert' from |
1825 | * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller, | |
1826 | * and avoid driver to try to dispatch again. | |
23d4ee19 | 1827 | */ |
fd9c40f6 | 1828 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { |
d964f04a | 1829 | run_queue = false; |
fd9c40f6 BVA |
1830 | bypass_insert = false; |
1831 | goto insert; | |
d964f04a | 1832 | } |
f984df1f | 1833 | |
fd9c40f6 BVA |
1834 | if (q->elevator && !bypass_insert) |
1835 | goto insert; | |
2253efc8 | 1836 | |
0bca799b | 1837 | if (!blk_mq_get_dispatch_budget(hctx)) |
fd9c40f6 | 1838 | goto insert; |
bd166ef1 | 1839 | |
8ab6bb9e | 1840 | if (!blk_mq_get_driver_tag(rq)) { |
0bca799b | 1841 | blk_mq_put_dispatch_budget(hctx); |
fd9c40f6 | 1842 | goto insert; |
88022d72 | 1843 | } |
de148297 | 1844 | |
fd9c40f6 BVA |
1845 | return __blk_mq_issue_directly(hctx, rq, cookie, last); |
1846 | insert: | |
1847 | if (bypass_insert) | |
1848 | return BLK_STS_RESOURCE; | |
1849 | ||
1850 | blk_mq_request_bypass_insert(rq, run_queue); | |
1851 | return BLK_STS_OK; | |
1852 | } | |
1853 | ||
1854 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, | |
1855 | struct request *rq, blk_qc_t *cookie) | |
1856 | { | |
1857 | blk_status_t ret; | |
1858 | int srcu_idx; | |
1859 | ||
1860 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); | |
1861 | ||
1862 | hctx_lock(hctx, &srcu_idx); | |
1863 | ||
1864 | ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true); | |
1865 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) | |
1866 | blk_mq_request_bypass_insert(rq, true); | |
1867 | else if (ret != BLK_STS_OK) | |
1868 | blk_mq_end_request(rq, ret); | |
1869 | ||
1870 | hctx_unlock(hctx, srcu_idx); | |
1871 | } | |
1872 | ||
1873 | blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last) | |
1874 | { | |
1875 | blk_status_t ret; | |
1876 | int srcu_idx; | |
1877 | blk_qc_t unused_cookie; | |
1878 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
1879 | ||
1880 | hctx_lock(hctx, &srcu_idx); | |
1881 | ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last); | |
04ced159 | 1882 | hctx_unlock(hctx, srcu_idx); |
7f556a44 JW |
1883 | |
1884 | return ret; | |
5eb6126e CH |
1885 | } |
1886 | ||
6ce3dd6e ML |
1887 | void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
1888 | struct list_head *list) | |
1889 | { | |
1890 | while (!list_empty(list)) { | |
fd9c40f6 | 1891 | blk_status_t ret; |
6ce3dd6e ML |
1892 | struct request *rq = list_first_entry(list, struct request, |
1893 | queuelist); | |
1894 | ||
1895 | list_del_init(&rq->queuelist); | |
fd9c40f6 BVA |
1896 | ret = blk_mq_request_issue_directly(rq, list_empty(list)); |
1897 | if (ret != BLK_STS_OK) { | |
1898 | if (ret == BLK_STS_RESOURCE || | |
1899 | ret == BLK_STS_DEV_RESOURCE) { | |
1900 | blk_mq_request_bypass_insert(rq, | |
c616cbee | 1901 | list_empty(list)); |
fd9c40f6 BVA |
1902 | break; |
1903 | } | |
1904 | blk_mq_end_request(rq, ret); | |
1905 | } | |
6ce3dd6e | 1906 | } |
d666ba98 JA |
1907 | |
1908 | /* | |
1909 | * If we didn't flush the entire list, we could have told | |
1910 | * the driver there was more coming, but that turned out to | |
1911 | * be a lie. | |
1912 | */ | |
fd9c40f6 | 1913 | if (!list_empty(list) && hctx->queue->mq_ops->commit_rqs) |
d666ba98 | 1914 | hctx->queue->mq_ops->commit_rqs(hctx); |
6ce3dd6e ML |
1915 | } |
1916 | ||
ce5b009c JA |
1917 | static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq) |
1918 | { | |
1919 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1920 | plug->rq_count++; | |
1921 | if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) { | |
1922 | struct request *tmp; | |
1923 | ||
1924 | tmp = list_first_entry(&plug->mq_list, struct request, | |
1925 | queuelist); | |
1926 | if (tmp->q != rq->q) | |
1927 | plug->multiple_queues = true; | |
1928 | } | |
1929 | } | |
1930 | ||
dece1635 | 1931 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1932 | { |
ef295ecf | 1933 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1934 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
7809167d | 1935 | struct blk_mq_alloc_data data = { .flags = 0}; |
07068d5b | 1936 | struct request *rq; |
f984df1f | 1937 | struct blk_plug *plug; |
5b3f341f | 1938 | struct request *same_queue_rq = NULL; |
14ccb66b | 1939 | unsigned int nr_segs; |
7b371636 | 1940 | blk_qc_t cookie; |
07068d5b JA |
1941 | |
1942 | blk_queue_bounce(q, &bio); | |
14ccb66b | 1943 | __blk_queue_split(q, &bio, &nr_segs); |
f36ea50c | 1944 | |
e23947bd | 1945 | if (!bio_integrity_prep(bio)) |
dece1635 | 1946 | return BLK_QC_T_NONE; |
07068d5b | 1947 | |
87c279e6 | 1948 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
14ccb66b | 1949 | blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq)) |
87c279e6 | 1950 | return BLK_QC_T_NONE; |
f984df1f | 1951 | |
14ccb66b | 1952 | if (blk_mq_sched_bio_merge(q, bio, nr_segs)) |
bd166ef1 JA |
1953 | return BLK_QC_T_NONE; |
1954 | ||
d5337560 | 1955 | rq_qos_throttle(q, bio); |
87760e5e | 1956 | |
7809167d | 1957 | data.cmd_flags = bio->bi_opf; |
f9afca4d | 1958 | rq = blk_mq_get_request(q, bio, &data); |
87760e5e | 1959 | if (unlikely(!rq)) { |
c1c80384 | 1960 | rq_qos_cleanup(q, bio); |
893a1c97 JA |
1961 | |
1962 | cookie = BLK_QC_T_NONE; | |
1963 | if (bio->bi_opf & REQ_NOWAIT_INLINE) | |
1964 | cookie = BLK_QC_T_EAGAIN; | |
1965 | else if (bio->bi_opf & REQ_NOWAIT) | |
03a07c92 | 1966 | bio_wouldblock_error(bio); |
893a1c97 | 1967 | return cookie; |
87760e5e JA |
1968 | } |
1969 | ||
d6f1dda2 XW |
1970 | trace_block_getrq(q, bio, bio->bi_opf); |
1971 | ||
c1c80384 | 1972 | rq_qos_track(q, rq, bio); |
07068d5b | 1973 | |
fd2d3326 | 1974 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b | 1975 | |
970d168d BVA |
1976 | blk_mq_bio_to_request(rq, bio, nr_segs); |
1977 | ||
b49773e7 | 1978 | plug = blk_mq_plug(q, bio); |
07068d5b | 1979 | if (unlikely(is_flush_fua)) { |
923218f6 ML |
1980 | /* bypass scheduler for flush rq */ |
1981 | blk_insert_flush(rq); | |
1982 | blk_mq_run_hw_queue(data.hctx, true); | |
b2c5d16b JA |
1983 | } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs)) { |
1984 | /* | |
1985 | * Use plugging if we have a ->commit_rqs() hook as well, as | |
1986 | * we know the driver uses bd->last in a smart fashion. | |
1987 | */ | |
5f0ed774 | 1988 | unsigned int request_count = plug->rq_count; |
600271d9 SL |
1989 | struct request *last = NULL; |
1990 | ||
676d0607 | 1991 | if (!request_count) |
e6c4438b | 1992 | trace_block_plug(q); |
600271d9 SL |
1993 | else |
1994 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c | 1995 | |
600271d9 SL |
1996 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1997 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1998 | blk_flush_plug_list(plug, false); |
1999 | trace_block_plug(q); | |
320ae51f | 2000 | } |
b094f89c | 2001 | |
ce5b009c | 2002 | blk_add_rq_to_plug(plug, rq); |
2299722c | 2003 | } else if (plug && !blk_queue_nomerges(q)) { |
07068d5b | 2004 | /* |
6a83e74d | 2005 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
2006 | * Otherwise the existing request in the plug list will be |
2007 | * issued. So the plug list will have one request at most | |
2299722c CH |
2008 | * The plug list might get flushed before this. If that happens, |
2009 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 2010 | */ |
2299722c CH |
2011 | if (list_empty(&plug->mq_list)) |
2012 | same_queue_rq = NULL; | |
4711b573 | 2013 | if (same_queue_rq) { |
2299722c | 2014 | list_del_init(&same_queue_rq->queuelist); |
4711b573 JA |
2015 | plug->rq_count--; |
2016 | } | |
ce5b009c | 2017 | blk_add_rq_to_plug(plug, rq); |
ff3b74b8 | 2018 | trace_block_plug(q); |
2299722c | 2019 | |
dad7a3be | 2020 | if (same_queue_rq) { |
ea4f995e | 2021 | data.hctx = same_queue_rq->mq_hctx; |
ff3b74b8 | 2022 | trace_block_unplug(q, 1, true); |
2299722c | 2023 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, |
fd9c40f6 | 2024 | &cookie); |
dad7a3be | 2025 | } |
6ce3dd6e ML |
2026 | } else if ((q->nr_hw_queues > 1 && is_sync) || (!q->elevator && |
2027 | !data.hctx->dispatch_busy)) { | |
fd9c40f6 | 2028 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); |
ab42f35d | 2029 | } else { |
8fa9f556 | 2030 | blk_mq_sched_insert_request(rq, false, true, true); |
ab42f35d | 2031 | } |
320ae51f | 2032 | |
7b371636 | 2033 | return cookie; |
320ae51f JA |
2034 | } |
2035 | ||
cc71a6f4 JA |
2036 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2037 | unsigned int hctx_idx) | |
95363efd | 2038 | { |
e9b267d9 | 2039 | struct page *page; |
320ae51f | 2040 | |
24d2f903 | 2041 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 2042 | int i; |
320ae51f | 2043 | |
24d2f903 | 2044 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
2045 | struct request *rq = tags->static_rqs[i]; |
2046 | ||
2047 | if (!rq) | |
e9b267d9 | 2048 | continue; |
d6296d39 | 2049 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 2050 | tags->static_rqs[i] = NULL; |
e9b267d9 | 2051 | } |
320ae51f | 2052 | } |
320ae51f | 2053 | |
24d2f903 CH |
2054 | while (!list_empty(&tags->page_list)) { |
2055 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 2056 | list_del_init(&page->lru); |
f75782e4 CM |
2057 | /* |
2058 | * Remove kmemleak object previously allocated in | |
273938bf | 2059 | * blk_mq_alloc_rqs(). |
f75782e4 CM |
2060 | */ |
2061 | kmemleak_free(page_address(page)); | |
320ae51f JA |
2062 | __free_pages(page, page->private); |
2063 | } | |
cc71a6f4 | 2064 | } |
320ae51f | 2065 | |
cc71a6f4 JA |
2066 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
2067 | { | |
24d2f903 | 2068 | kfree(tags->rqs); |
cc71a6f4 | 2069 | tags->rqs = NULL; |
2af8cbe3 JA |
2070 | kfree(tags->static_rqs); |
2071 | tags->static_rqs = NULL; | |
320ae51f | 2072 | |
24d2f903 | 2073 | blk_mq_free_tags(tags); |
320ae51f JA |
2074 | } |
2075 | ||
cc71a6f4 JA |
2076 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
2077 | unsigned int hctx_idx, | |
2078 | unsigned int nr_tags, | |
2079 | unsigned int reserved_tags) | |
320ae51f | 2080 | { |
24d2f903 | 2081 | struct blk_mq_tags *tags; |
59f082e4 | 2082 | int node; |
320ae51f | 2083 | |
7d76f856 | 2084 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx); |
59f082e4 SL |
2085 | if (node == NUMA_NO_NODE) |
2086 | node = set->numa_node; | |
2087 | ||
2088 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, | |
24391c0d | 2089 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
24d2f903 CH |
2090 | if (!tags) |
2091 | return NULL; | |
320ae51f | 2092 | |
590b5b7d | 2093 | tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
36e1f3d1 | 2094 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 2095 | node); |
24d2f903 CH |
2096 | if (!tags->rqs) { |
2097 | blk_mq_free_tags(tags); | |
2098 | return NULL; | |
2099 | } | |
320ae51f | 2100 | |
590b5b7d KC |
2101 | tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
2102 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
2103 | node); | |
2af8cbe3 JA |
2104 | if (!tags->static_rqs) { |
2105 | kfree(tags->rqs); | |
2106 | blk_mq_free_tags(tags); | |
2107 | return NULL; | |
2108 | } | |
2109 | ||
cc71a6f4 JA |
2110 | return tags; |
2111 | } | |
2112 | ||
2113 | static size_t order_to_size(unsigned int order) | |
2114 | { | |
2115 | return (size_t)PAGE_SIZE << order; | |
2116 | } | |
2117 | ||
1d9bd516 TH |
2118 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
2119 | unsigned int hctx_idx, int node) | |
2120 | { | |
2121 | int ret; | |
2122 | ||
2123 | if (set->ops->init_request) { | |
2124 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
2125 | if (ret) | |
2126 | return ret; | |
2127 | } | |
2128 | ||
12f5b931 | 2129 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
1d9bd516 TH |
2130 | return 0; |
2131 | } | |
2132 | ||
cc71a6f4 JA |
2133 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2134 | unsigned int hctx_idx, unsigned int depth) | |
2135 | { | |
2136 | unsigned int i, j, entries_per_page, max_order = 4; | |
2137 | size_t rq_size, left; | |
59f082e4 SL |
2138 | int node; |
2139 | ||
7d76f856 | 2140 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx); |
59f082e4 SL |
2141 | if (node == NUMA_NO_NODE) |
2142 | node = set->numa_node; | |
cc71a6f4 JA |
2143 | |
2144 | INIT_LIST_HEAD(&tags->page_list); | |
2145 | ||
320ae51f JA |
2146 | /* |
2147 | * rq_size is the size of the request plus driver payload, rounded | |
2148 | * to the cacheline size | |
2149 | */ | |
24d2f903 | 2150 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 2151 | cache_line_size()); |
cc71a6f4 | 2152 | left = rq_size * depth; |
320ae51f | 2153 | |
cc71a6f4 | 2154 | for (i = 0; i < depth; ) { |
320ae51f JA |
2155 | int this_order = max_order; |
2156 | struct page *page; | |
2157 | int to_do; | |
2158 | void *p; | |
2159 | ||
b3a834b1 | 2160 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
2161 | this_order--; |
2162 | ||
2163 | do { | |
59f082e4 | 2164 | page = alloc_pages_node(node, |
36e1f3d1 | 2165 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 2166 | this_order); |
320ae51f JA |
2167 | if (page) |
2168 | break; | |
2169 | if (!this_order--) | |
2170 | break; | |
2171 | if (order_to_size(this_order) < rq_size) | |
2172 | break; | |
2173 | } while (1); | |
2174 | ||
2175 | if (!page) | |
24d2f903 | 2176 | goto fail; |
320ae51f JA |
2177 | |
2178 | page->private = this_order; | |
24d2f903 | 2179 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
2180 | |
2181 | p = page_address(page); | |
f75782e4 CM |
2182 | /* |
2183 | * Allow kmemleak to scan these pages as they contain pointers | |
2184 | * to additional allocations like via ops->init_request(). | |
2185 | */ | |
36e1f3d1 | 2186 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 2187 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 2188 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
2189 | left -= to_do * rq_size; |
2190 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
2191 | struct request *rq = p; |
2192 | ||
2193 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
2194 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
2195 | tags->static_rqs[i] = NULL; | |
2196 | goto fail; | |
e9b267d9 CH |
2197 | } |
2198 | ||
320ae51f JA |
2199 | p += rq_size; |
2200 | i++; | |
2201 | } | |
2202 | } | |
cc71a6f4 | 2203 | return 0; |
320ae51f | 2204 | |
24d2f903 | 2205 | fail: |
cc71a6f4 JA |
2206 | blk_mq_free_rqs(set, tags, hctx_idx); |
2207 | return -ENOMEM; | |
320ae51f JA |
2208 | } |
2209 | ||
e57690fe JA |
2210 | /* |
2211 | * 'cpu' is going away. splice any existing rq_list entries from this | |
2212 | * software queue to the hw queue dispatch list, and ensure that it | |
2213 | * gets run. | |
2214 | */ | |
9467f859 | 2215 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 2216 | { |
9467f859 | 2217 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
2218 | struct blk_mq_ctx *ctx; |
2219 | LIST_HEAD(tmp); | |
c16d6b5a | 2220 | enum hctx_type type; |
484b4061 | 2221 | |
9467f859 | 2222 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 2223 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
c16d6b5a | 2224 | type = hctx->type; |
484b4061 JA |
2225 | |
2226 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
2227 | if (!list_empty(&ctx->rq_lists[type])) { |
2228 | list_splice_init(&ctx->rq_lists[type], &tmp); | |
484b4061 JA |
2229 | blk_mq_hctx_clear_pending(hctx, ctx); |
2230 | } | |
2231 | spin_unlock(&ctx->lock); | |
2232 | ||
2233 | if (list_empty(&tmp)) | |
9467f859 | 2234 | return 0; |
484b4061 | 2235 | |
e57690fe JA |
2236 | spin_lock(&hctx->lock); |
2237 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
2238 | spin_unlock(&hctx->lock); | |
484b4061 JA |
2239 | |
2240 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 2241 | return 0; |
484b4061 JA |
2242 | } |
2243 | ||
9467f859 | 2244 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 2245 | { |
9467f859 TG |
2246 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
2247 | &hctx->cpuhp_dead); | |
484b4061 JA |
2248 | } |
2249 | ||
c3b4afca | 2250 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
2251 | static void blk_mq_exit_hctx(struct request_queue *q, |
2252 | struct blk_mq_tag_set *set, | |
2253 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
2254 | { | |
8ab0b7dc ML |
2255 | if (blk_mq_hw_queue_mapped(hctx)) |
2256 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 2257 | |
f70ced09 | 2258 | if (set->ops->exit_request) |
d6296d39 | 2259 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); |
f70ced09 | 2260 | |
08e98fc6 ML |
2261 | if (set->ops->exit_hctx) |
2262 | set->ops->exit_hctx(hctx, hctx_idx); | |
2263 | ||
9467f859 | 2264 | blk_mq_remove_cpuhp(hctx); |
2f8f1336 ML |
2265 | |
2266 | spin_lock(&q->unused_hctx_lock); | |
2267 | list_add(&hctx->hctx_list, &q->unused_hctx_list); | |
2268 | spin_unlock(&q->unused_hctx_lock); | |
08e98fc6 ML |
2269 | } |
2270 | ||
624dbe47 ML |
2271 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
2272 | struct blk_mq_tag_set *set, int nr_queue) | |
2273 | { | |
2274 | struct blk_mq_hw_ctx *hctx; | |
2275 | unsigned int i; | |
2276 | ||
2277 | queue_for_each_hw_ctx(q, hctx, i) { | |
2278 | if (i == nr_queue) | |
2279 | break; | |
477e19de | 2280 | blk_mq_debugfs_unregister_hctx(hctx); |
08e98fc6 | 2281 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 2282 | } |
624dbe47 ML |
2283 | } |
2284 | ||
7c6c5b7c ML |
2285 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) |
2286 | { | |
2287 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2288 | ||
2289 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu), | |
2290 | __alignof__(struct blk_mq_hw_ctx)) != | |
2291 | sizeof(struct blk_mq_hw_ctx)); | |
2292 | ||
2293 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2294 | hw_ctx_size += sizeof(struct srcu_struct); | |
2295 | ||
2296 | return hw_ctx_size; | |
2297 | } | |
2298 | ||
08e98fc6 ML |
2299 | static int blk_mq_init_hctx(struct request_queue *q, |
2300 | struct blk_mq_tag_set *set, | |
2301 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 2302 | { |
7c6c5b7c ML |
2303 | hctx->queue_num = hctx_idx; |
2304 | ||
2305 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); | |
2306 | ||
2307 | hctx->tags = set->tags[hctx_idx]; | |
2308 | ||
2309 | if (set->ops->init_hctx && | |
2310 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
2311 | goto unregister_cpu_notifier; | |
08e98fc6 | 2312 | |
7c6c5b7c ML |
2313 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, |
2314 | hctx->numa_node)) | |
2315 | goto exit_hctx; | |
2316 | return 0; | |
2317 | ||
2318 | exit_hctx: | |
2319 | if (set->ops->exit_hctx) | |
2320 | set->ops->exit_hctx(hctx, hctx_idx); | |
2321 | unregister_cpu_notifier: | |
2322 | blk_mq_remove_cpuhp(hctx); | |
2323 | return -1; | |
2324 | } | |
2325 | ||
2326 | static struct blk_mq_hw_ctx * | |
2327 | blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set, | |
2328 | int node) | |
2329 | { | |
2330 | struct blk_mq_hw_ctx *hctx; | |
2331 | gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY; | |
2332 | ||
2333 | hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node); | |
2334 | if (!hctx) | |
2335 | goto fail_alloc_hctx; | |
2336 | ||
2337 | if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node)) | |
2338 | goto free_hctx; | |
2339 | ||
2340 | atomic_set(&hctx->nr_active, 0); | |
08e98fc6 | 2341 | if (node == NUMA_NO_NODE) |
7c6c5b7c ML |
2342 | node = set->numa_node; |
2343 | hctx->numa_node = node; | |
08e98fc6 | 2344 | |
9f993737 | 2345 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
2346 | spin_lock_init(&hctx->lock); |
2347 | INIT_LIST_HEAD(&hctx->dispatch); | |
2348 | hctx->queue = q; | |
2404e607 | 2349 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 2350 | |
2f8f1336 ML |
2351 | INIT_LIST_HEAD(&hctx->hctx_list); |
2352 | ||
320ae51f | 2353 | /* |
08e98fc6 ML |
2354 | * Allocate space for all possible cpus to avoid allocation at |
2355 | * runtime | |
320ae51f | 2356 | */ |
d904bfa7 | 2357 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
7c6c5b7c | 2358 | gfp, node); |
08e98fc6 | 2359 | if (!hctx->ctxs) |
7c6c5b7c | 2360 | goto free_cpumask; |
320ae51f | 2361 | |
5b202853 | 2362 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), |
7c6c5b7c | 2363 | gfp, node)) |
08e98fc6 | 2364 | goto free_ctxs; |
08e98fc6 | 2365 | hctx->nr_ctx = 0; |
320ae51f | 2366 | |
5815839b | 2367 | spin_lock_init(&hctx->dispatch_wait_lock); |
eb619fdb JA |
2368 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
2369 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
2370 | ||
5b202853 | 2371 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size, |
7c6c5b7c | 2372 | gfp); |
f70ced09 | 2373 | if (!hctx->fq) |
7c6c5b7c | 2374 | goto free_bitmap; |
320ae51f | 2375 | |
6a83e74d | 2376 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2377 | init_srcu_struct(hctx->srcu); |
7c6c5b7c | 2378 | blk_mq_hctx_kobj_init(hctx); |
6a83e74d | 2379 | |
7c6c5b7c | 2380 | return hctx; |
320ae51f | 2381 | |
08e98fc6 | 2382 | free_bitmap: |
88459642 | 2383 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2384 | free_ctxs: |
2385 | kfree(hctx->ctxs); | |
7c6c5b7c ML |
2386 | free_cpumask: |
2387 | free_cpumask_var(hctx->cpumask); | |
2388 | free_hctx: | |
2389 | kfree(hctx); | |
2390 | fail_alloc_hctx: | |
2391 | return NULL; | |
08e98fc6 | 2392 | } |
320ae51f | 2393 | |
320ae51f JA |
2394 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
2395 | unsigned int nr_hw_queues) | |
2396 | { | |
b3c661b1 JA |
2397 | struct blk_mq_tag_set *set = q->tag_set; |
2398 | unsigned int i, j; | |
320ae51f JA |
2399 | |
2400 | for_each_possible_cpu(i) { | |
2401 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2402 | struct blk_mq_hw_ctx *hctx; | |
c16d6b5a | 2403 | int k; |
320ae51f | 2404 | |
320ae51f JA |
2405 | __ctx->cpu = i; |
2406 | spin_lock_init(&__ctx->lock); | |
c16d6b5a ML |
2407 | for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++) |
2408 | INIT_LIST_HEAD(&__ctx->rq_lists[k]); | |
2409 | ||
320ae51f JA |
2410 | __ctx->queue = q; |
2411 | ||
320ae51f JA |
2412 | /* |
2413 | * Set local node, IFF we have more than one hw queue. If | |
2414 | * not, we remain on the home node of the device | |
2415 | */ | |
b3c661b1 JA |
2416 | for (j = 0; j < set->nr_maps; j++) { |
2417 | hctx = blk_mq_map_queue_type(q, j, i); | |
2418 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
2419 | hctx->numa_node = local_memory_node(cpu_to_node(i)); | |
2420 | } | |
320ae51f JA |
2421 | } |
2422 | } | |
2423 | ||
cc71a6f4 JA |
2424 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
2425 | { | |
2426 | int ret = 0; | |
2427 | ||
2428 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
2429 | set->queue_depth, set->reserved_tags); | |
2430 | if (!set->tags[hctx_idx]) | |
2431 | return false; | |
2432 | ||
2433 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2434 | set->queue_depth); | |
2435 | if (!ret) | |
2436 | return true; | |
2437 | ||
2438 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2439 | set->tags[hctx_idx] = NULL; | |
2440 | return false; | |
2441 | } | |
2442 | ||
2443 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2444 | unsigned int hctx_idx) | |
2445 | { | |
4e6db0f2 | 2446 | if (set->tags && set->tags[hctx_idx]) { |
bd166ef1 JA |
2447 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); |
2448 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2449 | set->tags[hctx_idx] = NULL; | |
2450 | } | |
cc71a6f4 JA |
2451 | } |
2452 | ||
4b855ad3 | 2453 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 2454 | { |
b3c661b1 | 2455 | unsigned int i, j, hctx_idx; |
320ae51f JA |
2456 | struct blk_mq_hw_ctx *hctx; |
2457 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2458 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2459 | |
60de074b AM |
2460 | /* |
2461 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
2462 | */ | |
2463 | mutex_lock(&q->sysfs_lock); | |
2464 | ||
320ae51f | 2465 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 2466 | cpumask_clear(hctx->cpumask); |
320ae51f | 2467 | hctx->nr_ctx = 0; |
d416c92c | 2468 | hctx->dispatch_from = NULL; |
320ae51f JA |
2469 | } |
2470 | ||
2471 | /* | |
4b855ad3 | 2472 | * Map software to hardware queues. |
4412efec ML |
2473 | * |
2474 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 2475 | */ |
20e4d813 | 2476 | for_each_possible_cpu(i) { |
7d76f856 | 2477 | hctx_idx = set->map[HCTX_TYPE_DEFAULT].mq_map[i]; |
4412efec ML |
2478 | /* unmapped hw queue can be remapped after CPU topo changed */ |
2479 | if (!set->tags[hctx_idx] && | |
2480 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
2481 | /* | |
2482 | * If tags initialization fail for some hctx, | |
2483 | * that hctx won't be brought online. In this | |
2484 | * case, remap the current ctx to hctx[0] which | |
2485 | * is guaranteed to always have tags allocated | |
2486 | */ | |
7d76f856 | 2487 | set->map[HCTX_TYPE_DEFAULT].mq_map[i] = 0; |
4412efec ML |
2488 | } |
2489 | ||
897bb0c7 | 2490 | ctx = per_cpu_ptr(q->queue_ctx, i); |
b3c661b1 | 2491 | for (j = 0; j < set->nr_maps; j++) { |
bb94aea1 JW |
2492 | if (!set->map[j].nr_queues) { |
2493 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
2494 | HCTX_TYPE_DEFAULT, i); | |
e5edd5f2 | 2495 | continue; |
bb94aea1 | 2496 | } |
e5edd5f2 | 2497 | |
b3c661b1 | 2498 | hctx = blk_mq_map_queue_type(q, j, i); |
8ccdf4a3 | 2499 | ctx->hctxs[j] = hctx; |
b3c661b1 JA |
2500 | /* |
2501 | * If the CPU is already set in the mask, then we've | |
2502 | * mapped this one already. This can happen if | |
2503 | * devices share queues across queue maps. | |
2504 | */ | |
2505 | if (cpumask_test_cpu(i, hctx->cpumask)) | |
2506 | continue; | |
2507 | ||
2508 | cpumask_set_cpu(i, hctx->cpumask); | |
2509 | hctx->type = j; | |
2510 | ctx->index_hw[hctx->type] = hctx->nr_ctx; | |
2511 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2512 | ||
2513 | /* | |
2514 | * If the nr_ctx type overflows, we have exceeded the | |
2515 | * amount of sw queues we can support. | |
2516 | */ | |
2517 | BUG_ON(!hctx->nr_ctx); | |
2518 | } | |
bb94aea1 JW |
2519 | |
2520 | for (; j < HCTX_MAX_TYPES; j++) | |
2521 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
2522 | HCTX_TYPE_DEFAULT, i); | |
320ae51f | 2523 | } |
506e931f | 2524 | |
60de074b AM |
2525 | mutex_unlock(&q->sysfs_lock); |
2526 | ||
506e931f | 2527 | queue_for_each_hw_ctx(q, hctx, i) { |
4412efec ML |
2528 | /* |
2529 | * If no software queues are mapped to this hardware queue, | |
2530 | * disable it and free the request entries. | |
2531 | */ | |
2532 | if (!hctx->nr_ctx) { | |
2533 | /* Never unmap queue 0. We need it as a | |
2534 | * fallback in case of a new remap fails | |
2535 | * allocation | |
2536 | */ | |
2537 | if (i && set->tags[i]) | |
2538 | blk_mq_free_map_and_requests(set, i); | |
2539 | ||
2540 | hctx->tags = NULL; | |
2541 | continue; | |
2542 | } | |
484b4061 | 2543 | |
2a34c087 ML |
2544 | hctx->tags = set->tags[i]; |
2545 | WARN_ON(!hctx->tags); | |
2546 | ||
889fa31f CY |
2547 | /* |
2548 | * Set the map size to the number of mapped software queues. | |
2549 | * This is more accurate and more efficient than looping | |
2550 | * over all possibly mapped software queues. | |
2551 | */ | |
88459642 | 2552 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2553 | |
484b4061 JA |
2554 | /* |
2555 | * Initialize batch roundrobin counts | |
2556 | */ | |
f82ddf19 | 2557 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
2558 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2559 | } | |
320ae51f JA |
2560 | } |
2561 | ||
8e8320c9 JA |
2562 | /* |
2563 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
2564 | * the queue isn't live yet. | |
2565 | */ | |
2404e607 | 2566 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2567 | { |
2568 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2569 | int i; |
2570 | ||
2404e607 | 2571 | queue_for_each_hw_ctx(q, hctx, i) { |
97889f9a | 2572 | if (shared) |
2404e607 | 2573 | hctx->flags |= BLK_MQ_F_TAG_SHARED; |
97889f9a | 2574 | else |
2404e607 JM |
2575 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; |
2576 | } | |
2577 | } | |
2578 | ||
8e8320c9 JA |
2579 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, |
2580 | bool shared) | |
2404e607 JM |
2581 | { |
2582 | struct request_queue *q; | |
0d2602ca | 2583 | |
705cda97 BVA |
2584 | lockdep_assert_held(&set->tag_list_lock); |
2585 | ||
0d2602ca JA |
2586 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2587 | blk_mq_freeze_queue(q); | |
2404e607 | 2588 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2589 | blk_mq_unfreeze_queue(q); |
2590 | } | |
2591 | } | |
2592 | ||
2593 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2594 | { | |
2595 | struct blk_mq_tag_set *set = q->tag_set; | |
2596 | ||
0d2602ca | 2597 | mutex_lock(&set->tag_list_lock); |
705cda97 | 2598 | list_del_rcu(&q->tag_set_list); |
2404e607 JM |
2599 | if (list_is_singular(&set->tag_list)) { |
2600 | /* just transitioned to unshared */ | |
2601 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2602 | /* update existing queue */ | |
2603 | blk_mq_update_tag_set_depth(set, false); | |
2604 | } | |
0d2602ca | 2605 | mutex_unlock(&set->tag_list_lock); |
a347c7ad | 2606 | INIT_LIST_HEAD(&q->tag_set_list); |
0d2602ca JA |
2607 | } |
2608 | ||
2609 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2610 | struct request_queue *q) | |
2611 | { | |
0d2602ca | 2612 | mutex_lock(&set->tag_list_lock); |
2404e607 | 2613 | |
ff821d27 JA |
2614 | /* |
2615 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
2616 | */ | |
2617 | if (!list_empty(&set->tag_list) && | |
2618 | !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2404e607 JM |
2619 | set->flags |= BLK_MQ_F_TAG_SHARED; |
2620 | /* update existing queue */ | |
2621 | blk_mq_update_tag_set_depth(set, true); | |
2622 | } | |
2623 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2624 | queue_set_hctx_shared(q, true); | |
705cda97 | 2625 | list_add_tail_rcu(&q->tag_set_list, &set->tag_list); |
2404e607 | 2626 | |
0d2602ca JA |
2627 | mutex_unlock(&set->tag_list_lock); |
2628 | } | |
2629 | ||
1db4909e ML |
2630 | /* All allocations will be freed in release handler of q->mq_kobj */ |
2631 | static int blk_mq_alloc_ctxs(struct request_queue *q) | |
2632 | { | |
2633 | struct blk_mq_ctxs *ctxs; | |
2634 | int cpu; | |
2635 | ||
2636 | ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL); | |
2637 | if (!ctxs) | |
2638 | return -ENOMEM; | |
2639 | ||
2640 | ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx); | |
2641 | if (!ctxs->queue_ctx) | |
2642 | goto fail; | |
2643 | ||
2644 | for_each_possible_cpu(cpu) { | |
2645 | struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu); | |
2646 | ctx->ctxs = ctxs; | |
2647 | } | |
2648 | ||
2649 | q->mq_kobj = &ctxs->kobj; | |
2650 | q->queue_ctx = ctxs->queue_ctx; | |
2651 | ||
2652 | return 0; | |
2653 | fail: | |
2654 | kfree(ctxs); | |
2655 | return -ENOMEM; | |
2656 | } | |
2657 | ||
e09aae7e ML |
2658 | /* |
2659 | * It is the actual release handler for mq, but we do it from | |
2660 | * request queue's release handler for avoiding use-after-free | |
2661 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2662 | * but we can't group ctx/kctx kobj without it. | |
2663 | */ | |
2664 | void blk_mq_release(struct request_queue *q) | |
2665 | { | |
2f8f1336 ML |
2666 | struct blk_mq_hw_ctx *hctx, *next; |
2667 | int i; | |
e09aae7e | 2668 | |
fbc2a15e ML |
2669 | cancel_delayed_work_sync(&q->requeue_work); |
2670 | ||
2f8f1336 ML |
2671 | queue_for_each_hw_ctx(q, hctx, i) |
2672 | WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list)); | |
2673 | ||
2674 | /* all hctx are in .unused_hctx_list now */ | |
2675 | list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) { | |
2676 | list_del_init(&hctx->hctx_list); | |
6c8b232e | 2677 | kobject_put(&hctx->kobj); |
c3b4afca | 2678 | } |
e09aae7e ML |
2679 | |
2680 | kfree(q->queue_hw_ctx); | |
2681 | ||
7ea5fe31 ML |
2682 | /* |
2683 | * release .mq_kobj and sw queue's kobject now because | |
2684 | * both share lifetime with request queue. | |
2685 | */ | |
2686 | blk_mq_sysfs_deinit(q); | |
e09aae7e ML |
2687 | } |
2688 | ||
24d2f903 | 2689 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2690 | { |
2691 | struct request_queue *uninit_q, *q; | |
2692 | ||
6d469642 | 2693 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); |
b62c21b7 MS |
2694 | if (!uninit_q) |
2695 | return ERR_PTR(-ENOMEM); | |
2696 | ||
2697 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2698 | if (IS_ERR(q)) | |
2699 | blk_cleanup_queue(uninit_q); | |
2700 | ||
2701 | return q; | |
2702 | } | |
2703 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2704 | ||
9316a9ed JA |
2705 | /* |
2706 | * Helper for setting up a queue with mq ops, given queue depth, and | |
2707 | * the passed in mq ops flags. | |
2708 | */ | |
2709 | struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set, | |
2710 | const struct blk_mq_ops *ops, | |
2711 | unsigned int queue_depth, | |
2712 | unsigned int set_flags) | |
2713 | { | |
2714 | struct request_queue *q; | |
2715 | int ret; | |
2716 | ||
2717 | memset(set, 0, sizeof(*set)); | |
2718 | set->ops = ops; | |
2719 | set->nr_hw_queues = 1; | |
b3c661b1 | 2720 | set->nr_maps = 1; |
9316a9ed JA |
2721 | set->queue_depth = queue_depth; |
2722 | set->numa_node = NUMA_NO_NODE; | |
2723 | set->flags = set_flags; | |
2724 | ||
2725 | ret = blk_mq_alloc_tag_set(set); | |
2726 | if (ret) | |
2727 | return ERR_PTR(ret); | |
2728 | ||
2729 | q = blk_mq_init_queue(set); | |
2730 | if (IS_ERR(q)) { | |
2731 | blk_mq_free_tag_set(set); | |
2732 | return q; | |
2733 | } | |
2734 | ||
2735 | return q; | |
2736 | } | |
2737 | EXPORT_SYMBOL(blk_mq_init_sq_queue); | |
2738 | ||
34d11ffa JW |
2739 | static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx( |
2740 | struct blk_mq_tag_set *set, struct request_queue *q, | |
2741 | int hctx_idx, int node) | |
2742 | { | |
2f8f1336 | 2743 | struct blk_mq_hw_ctx *hctx = NULL, *tmp; |
34d11ffa | 2744 | |
2f8f1336 ML |
2745 | /* reuse dead hctx first */ |
2746 | spin_lock(&q->unused_hctx_lock); | |
2747 | list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) { | |
2748 | if (tmp->numa_node == node) { | |
2749 | hctx = tmp; | |
2750 | break; | |
2751 | } | |
2752 | } | |
2753 | if (hctx) | |
2754 | list_del_init(&hctx->hctx_list); | |
2755 | spin_unlock(&q->unused_hctx_lock); | |
2756 | ||
2757 | if (!hctx) | |
2758 | hctx = blk_mq_alloc_hctx(q, set, node); | |
34d11ffa | 2759 | if (!hctx) |
7c6c5b7c | 2760 | goto fail; |
34d11ffa | 2761 | |
7c6c5b7c ML |
2762 | if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) |
2763 | goto free_hctx; | |
34d11ffa JW |
2764 | |
2765 | return hctx; | |
7c6c5b7c ML |
2766 | |
2767 | free_hctx: | |
2768 | kobject_put(&hctx->kobj); | |
2769 | fail: | |
2770 | return NULL; | |
34d11ffa JW |
2771 | } |
2772 | ||
868f2f0b KB |
2773 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2774 | struct request_queue *q) | |
320ae51f | 2775 | { |
e01ad46d | 2776 | int i, j, end; |
868f2f0b | 2777 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; |
f14bbe77 | 2778 | |
fb350e0a ML |
2779 | /* protect against switching io scheduler */ |
2780 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 2781 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2782 | int node; |
34d11ffa | 2783 | struct blk_mq_hw_ctx *hctx; |
868f2f0b | 2784 | |
7d76f856 | 2785 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i); |
34d11ffa JW |
2786 | /* |
2787 | * If the hw queue has been mapped to another numa node, | |
2788 | * we need to realloc the hctx. If allocation fails, fallback | |
2789 | * to use the previous one. | |
2790 | */ | |
2791 | if (hctxs[i] && (hctxs[i]->numa_node == node)) | |
2792 | continue; | |
868f2f0b | 2793 | |
34d11ffa JW |
2794 | hctx = blk_mq_alloc_and_init_hctx(set, q, i, node); |
2795 | if (hctx) { | |
2f8f1336 | 2796 | if (hctxs[i]) |
34d11ffa | 2797 | blk_mq_exit_hctx(q, set, hctxs[i], i); |
34d11ffa JW |
2798 | hctxs[i] = hctx; |
2799 | } else { | |
2800 | if (hctxs[i]) | |
2801 | pr_warn("Allocate new hctx on node %d fails,\ | |
2802 | fallback to previous one on node %d\n", | |
2803 | node, hctxs[i]->numa_node); | |
2804 | else | |
2805 | break; | |
868f2f0b | 2806 | } |
320ae51f | 2807 | } |
e01ad46d JW |
2808 | /* |
2809 | * Increasing nr_hw_queues fails. Free the newly allocated | |
2810 | * hctxs and keep the previous q->nr_hw_queues. | |
2811 | */ | |
2812 | if (i != set->nr_hw_queues) { | |
2813 | j = q->nr_hw_queues; | |
2814 | end = i; | |
2815 | } else { | |
2816 | j = i; | |
2817 | end = q->nr_hw_queues; | |
2818 | q->nr_hw_queues = set->nr_hw_queues; | |
2819 | } | |
34d11ffa | 2820 | |
e01ad46d | 2821 | for (; j < end; j++) { |
868f2f0b KB |
2822 | struct blk_mq_hw_ctx *hctx = hctxs[j]; |
2823 | ||
2824 | if (hctx) { | |
cc71a6f4 JA |
2825 | if (hctx->tags) |
2826 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 2827 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 2828 | hctxs[j] = NULL; |
868f2f0b KB |
2829 | } |
2830 | } | |
fb350e0a | 2831 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
2832 | } |
2833 | ||
392546ae JA |
2834 | /* |
2835 | * Maximum number of hardware queues we support. For single sets, we'll never | |
2836 | * have more than the CPUs (software queues). For multiple sets, the tag_set | |
2837 | * user may have set ->nr_hw_queues larger. | |
2838 | */ | |
2839 | static unsigned int nr_hw_queues(struct blk_mq_tag_set *set) | |
2840 | { | |
2841 | if (set->nr_maps == 1) | |
2842 | return nr_cpu_ids; | |
2843 | ||
2844 | return max(set->nr_hw_queues, nr_cpu_ids); | |
2845 | } | |
2846 | ||
868f2f0b KB |
2847 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, |
2848 | struct request_queue *q) | |
2849 | { | |
66841672 ML |
2850 | /* mark the queue as mq asap */ |
2851 | q->mq_ops = set->ops; | |
2852 | ||
34dbad5d | 2853 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
2854 | blk_mq_poll_stats_bkt, |
2855 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
2856 | if (!q->poll_cb) |
2857 | goto err_exit; | |
2858 | ||
1db4909e | 2859 | if (blk_mq_alloc_ctxs(q)) |
41de54c6 | 2860 | goto err_poll; |
868f2f0b | 2861 | |
737f98cf ML |
2862 | /* init q->mq_kobj and sw queues' kobjects */ |
2863 | blk_mq_sysfs_init(q); | |
2864 | ||
392546ae JA |
2865 | q->nr_queues = nr_hw_queues(set); |
2866 | q->queue_hw_ctx = kcalloc_node(q->nr_queues, sizeof(*(q->queue_hw_ctx)), | |
868f2f0b KB |
2867 | GFP_KERNEL, set->numa_node); |
2868 | if (!q->queue_hw_ctx) | |
1db4909e | 2869 | goto err_sys_init; |
868f2f0b | 2870 | |
2f8f1336 ML |
2871 | INIT_LIST_HEAD(&q->unused_hctx_list); |
2872 | spin_lock_init(&q->unused_hctx_lock); | |
2873 | ||
868f2f0b KB |
2874 | blk_mq_realloc_hw_ctxs(set, q); |
2875 | if (!q->nr_hw_queues) | |
2876 | goto err_hctxs; | |
320ae51f | 2877 | |
287922eb | 2878 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2879 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f | 2880 | |
a8908939 | 2881 | q->tag_set = set; |
320ae51f | 2882 | |
94eddfbe | 2883 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
cd19181b ML |
2884 | if (set->nr_maps > HCTX_TYPE_POLL && |
2885 | set->map[HCTX_TYPE_POLL].nr_queues) | |
6544d229 | 2886 | blk_queue_flag_set(QUEUE_FLAG_POLL, q); |
320ae51f | 2887 | |
1be036e9 CH |
2888 | q->sg_reserved_size = INT_MAX; |
2889 | ||
2849450a | 2890 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2891 | INIT_LIST_HEAD(&q->requeue_list); |
2892 | spin_lock_init(&q->requeue_lock); | |
2893 | ||
254d259d | 2894 | blk_queue_make_request(q, blk_mq_make_request); |
07068d5b | 2895 | |
eba71768 JA |
2896 | /* |
2897 | * Do this after blk_queue_make_request() overrides it... | |
2898 | */ | |
2899 | q->nr_requests = set->queue_depth; | |
2900 | ||
64f1c21e JA |
2901 | /* |
2902 | * Default to classic polling | |
2903 | */ | |
29ece8b4 | 2904 | q->poll_nsec = BLK_MQ_POLL_CLASSIC; |
64f1c21e | 2905 | |
24d2f903 | 2906 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 2907 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 2908 | blk_mq_map_swqueue(q); |
4593fdbe | 2909 | |
d3484991 JA |
2910 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { |
2911 | int ret; | |
2912 | ||
131d08e1 | 2913 | ret = elevator_init_mq(q); |
d3484991 JA |
2914 | if (ret) |
2915 | return ERR_PTR(ret); | |
2916 | } | |
2917 | ||
320ae51f | 2918 | return q; |
18741986 | 2919 | |
320ae51f | 2920 | err_hctxs: |
868f2f0b | 2921 | kfree(q->queue_hw_ctx); |
1db4909e ML |
2922 | err_sys_init: |
2923 | blk_mq_sysfs_deinit(q); | |
41de54c6 JS |
2924 | err_poll: |
2925 | blk_stat_free_callback(q->poll_cb); | |
2926 | q->poll_cb = NULL; | |
c7de5726 ML |
2927 | err_exit: |
2928 | q->mq_ops = NULL; | |
320ae51f JA |
2929 | return ERR_PTR(-ENOMEM); |
2930 | } | |
b62c21b7 | 2931 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f | 2932 | |
c7e2d94b ML |
2933 | /* tags can _not_ be used after returning from blk_mq_exit_queue */ |
2934 | void blk_mq_exit_queue(struct request_queue *q) | |
320ae51f | 2935 | { |
624dbe47 | 2936 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2937 | |
0d2602ca | 2938 | blk_mq_del_queue_tag_set(q); |
624dbe47 | 2939 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
320ae51f | 2940 | } |
320ae51f | 2941 | |
a5164405 JA |
2942 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2943 | { | |
2944 | int i; | |
2945 | ||
cc71a6f4 JA |
2946 | for (i = 0; i < set->nr_hw_queues; i++) |
2947 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 2948 | goto out_unwind; |
a5164405 JA |
2949 | |
2950 | return 0; | |
2951 | ||
2952 | out_unwind: | |
2953 | while (--i >= 0) | |
cc71a6f4 | 2954 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 2955 | |
a5164405 JA |
2956 | return -ENOMEM; |
2957 | } | |
2958 | ||
2959 | /* | |
2960 | * Allocate the request maps associated with this tag_set. Note that this | |
2961 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2962 | * will be updated to reflect the allocated depth. | |
2963 | */ | |
2964 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2965 | { | |
2966 | unsigned int depth; | |
2967 | int err; | |
2968 | ||
2969 | depth = set->queue_depth; | |
2970 | do { | |
2971 | err = __blk_mq_alloc_rq_maps(set); | |
2972 | if (!err) | |
2973 | break; | |
2974 | ||
2975 | set->queue_depth >>= 1; | |
2976 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2977 | err = -ENOMEM; | |
2978 | break; | |
2979 | } | |
2980 | } while (set->queue_depth); | |
2981 | ||
2982 | if (!set->queue_depth || err) { | |
2983 | pr_err("blk-mq: failed to allocate request map\n"); | |
2984 | return -ENOMEM; | |
2985 | } | |
2986 | ||
2987 | if (depth != set->queue_depth) | |
2988 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2989 | depth, set->queue_depth); | |
2990 | ||
2991 | return 0; | |
2992 | } | |
2993 | ||
ebe8bddb OS |
2994 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
2995 | { | |
59388702 | 2996 | if (set->ops->map_queues && !is_kdump_kernel()) { |
b3c661b1 JA |
2997 | int i; |
2998 | ||
7d4901a9 ML |
2999 | /* |
3000 | * transport .map_queues is usually done in the following | |
3001 | * way: | |
3002 | * | |
3003 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
3004 | * mask = get_cpu_mask(queue) | |
3005 | * for_each_cpu(cpu, mask) | |
b3c661b1 | 3006 | * set->map[x].mq_map[cpu] = queue; |
7d4901a9 ML |
3007 | * } |
3008 | * | |
3009 | * When we need to remap, the table has to be cleared for | |
3010 | * killing stale mapping since one CPU may not be mapped | |
3011 | * to any hw queue. | |
3012 | */ | |
b3c661b1 JA |
3013 | for (i = 0; i < set->nr_maps; i++) |
3014 | blk_mq_clear_mq_map(&set->map[i]); | |
7d4901a9 | 3015 | |
ebe8bddb | 3016 | return set->ops->map_queues(set); |
b3c661b1 JA |
3017 | } else { |
3018 | BUG_ON(set->nr_maps > 1); | |
7d76f856 | 3019 | return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
b3c661b1 | 3020 | } |
ebe8bddb OS |
3021 | } |
3022 | ||
a4391c64 JA |
3023 | /* |
3024 | * Alloc a tag set to be associated with one or more request queues. | |
3025 | * May fail with EINVAL for various error conditions. May adjust the | |
c018c84f | 3026 | * requested depth down, if it's too large. In that case, the set |
a4391c64 JA |
3027 | * value will be stored in set->queue_depth. |
3028 | */ | |
24d2f903 CH |
3029 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
3030 | { | |
b3c661b1 | 3031 | int i, ret; |
da695ba2 | 3032 | |
205fb5f5 BVA |
3033 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
3034 | ||
24d2f903 CH |
3035 | if (!set->nr_hw_queues) |
3036 | return -EINVAL; | |
a4391c64 | 3037 | if (!set->queue_depth) |
24d2f903 CH |
3038 | return -EINVAL; |
3039 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
3040 | return -EINVAL; | |
3041 | ||
7d7e0f90 | 3042 | if (!set->ops->queue_rq) |
24d2f903 CH |
3043 | return -EINVAL; |
3044 | ||
de148297 ML |
3045 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
3046 | return -EINVAL; | |
3047 | ||
a4391c64 JA |
3048 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
3049 | pr_info("blk-mq: reduced tag depth to %u\n", | |
3050 | BLK_MQ_MAX_DEPTH); | |
3051 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
3052 | } | |
24d2f903 | 3053 | |
b3c661b1 JA |
3054 | if (!set->nr_maps) |
3055 | set->nr_maps = 1; | |
3056 | else if (set->nr_maps > HCTX_MAX_TYPES) | |
3057 | return -EINVAL; | |
3058 | ||
6637fadf SL |
3059 | /* |
3060 | * If a crashdump is active, then we are potentially in a very | |
3061 | * memory constrained environment. Limit us to 1 queue and | |
3062 | * 64 tags to prevent using too much memory. | |
3063 | */ | |
3064 | if (is_kdump_kernel()) { | |
3065 | set->nr_hw_queues = 1; | |
59388702 | 3066 | set->nr_maps = 1; |
6637fadf SL |
3067 | set->queue_depth = min(64U, set->queue_depth); |
3068 | } | |
868f2f0b | 3069 | /* |
392546ae JA |
3070 | * There is no use for more h/w queues than cpus if we just have |
3071 | * a single map | |
868f2f0b | 3072 | */ |
392546ae | 3073 | if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids) |
868f2f0b | 3074 | set->nr_hw_queues = nr_cpu_ids; |
6637fadf | 3075 | |
392546ae | 3076 | set->tags = kcalloc_node(nr_hw_queues(set), sizeof(struct blk_mq_tags *), |
24d2f903 CH |
3077 | GFP_KERNEL, set->numa_node); |
3078 | if (!set->tags) | |
a5164405 | 3079 | return -ENOMEM; |
24d2f903 | 3080 | |
da695ba2 | 3081 | ret = -ENOMEM; |
b3c661b1 JA |
3082 | for (i = 0; i < set->nr_maps; i++) { |
3083 | set->map[i].mq_map = kcalloc_node(nr_cpu_ids, | |
07b35eb5 | 3084 | sizeof(set->map[i].mq_map[0]), |
b3c661b1 JA |
3085 | GFP_KERNEL, set->numa_node); |
3086 | if (!set->map[i].mq_map) | |
3087 | goto out_free_mq_map; | |
59388702 | 3088 | set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues; |
b3c661b1 | 3089 | } |
bdd17e75 | 3090 | |
ebe8bddb | 3091 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
3092 | if (ret) |
3093 | goto out_free_mq_map; | |
3094 | ||
3095 | ret = blk_mq_alloc_rq_maps(set); | |
3096 | if (ret) | |
bdd17e75 | 3097 | goto out_free_mq_map; |
24d2f903 | 3098 | |
0d2602ca JA |
3099 | mutex_init(&set->tag_list_lock); |
3100 | INIT_LIST_HEAD(&set->tag_list); | |
3101 | ||
24d2f903 | 3102 | return 0; |
bdd17e75 CH |
3103 | |
3104 | out_free_mq_map: | |
b3c661b1 JA |
3105 | for (i = 0; i < set->nr_maps; i++) { |
3106 | kfree(set->map[i].mq_map); | |
3107 | set->map[i].mq_map = NULL; | |
3108 | } | |
5676e7b6 RE |
3109 | kfree(set->tags); |
3110 | set->tags = NULL; | |
da695ba2 | 3111 | return ret; |
24d2f903 CH |
3112 | } |
3113 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
3114 | ||
3115 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
3116 | { | |
b3c661b1 | 3117 | int i, j; |
24d2f903 | 3118 | |
392546ae | 3119 | for (i = 0; i < nr_hw_queues(set); i++) |
cc71a6f4 | 3120 | blk_mq_free_map_and_requests(set, i); |
484b4061 | 3121 | |
b3c661b1 JA |
3122 | for (j = 0; j < set->nr_maps; j++) { |
3123 | kfree(set->map[j].mq_map); | |
3124 | set->map[j].mq_map = NULL; | |
3125 | } | |
bdd17e75 | 3126 | |
981bd189 | 3127 | kfree(set->tags); |
5676e7b6 | 3128 | set->tags = NULL; |
24d2f903 CH |
3129 | } |
3130 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
3131 | ||
e3a2b3f9 JA |
3132 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
3133 | { | |
3134 | struct blk_mq_tag_set *set = q->tag_set; | |
3135 | struct blk_mq_hw_ctx *hctx; | |
3136 | int i, ret; | |
3137 | ||
bd166ef1 | 3138 | if (!set) |
e3a2b3f9 JA |
3139 | return -EINVAL; |
3140 | ||
e5fa8140 AZ |
3141 | if (q->nr_requests == nr) |
3142 | return 0; | |
3143 | ||
70f36b60 | 3144 | blk_mq_freeze_queue(q); |
24f5a90f | 3145 | blk_mq_quiesce_queue(q); |
70f36b60 | 3146 | |
e3a2b3f9 JA |
3147 | ret = 0; |
3148 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
3149 | if (!hctx->tags) |
3150 | continue; | |
bd166ef1 JA |
3151 | /* |
3152 | * If we're using an MQ scheduler, just update the scheduler | |
3153 | * queue depth. This is similar to what the old code would do. | |
3154 | */ | |
70f36b60 | 3155 | if (!hctx->sched_tags) { |
c2e82a23 | 3156 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, |
70f36b60 JA |
3157 | false); |
3158 | } else { | |
3159 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
3160 | nr, true); | |
3161 | } | |
e3a2b3f9 JA |
3162 | if (ret) |
3163 | break; | |
77f1e0a5 JA |
3164 | if (q->elevator && q->elevator->type->ops.depth_updated) |
3165 | q->elevator->type->ops.depth_updated(hctx); | |
e3a2b3f9 JA |
3166 | } |
3167 | ||
3168 | if (!ret) | |
3169 | q->nr_requests = nr; | |
3170 | ||
24f5a90f | 3171 | blk_mq_unquiesce_queue(q); |
70f36b60 | 3172 | blk_mq_unfreeze_queue(q); |
70f36b60 | 3173 | |
e3a2b3f9 JA |
3174 | return ret; |
3175 | } | |
3176 | ||
d48ece20 JW |
3177 | /* |
3178 | * request_queue and elevator_type pair. | |
3179 | * It is just used by __blk_mq_update_nr_hw_queues to cache | |
3180 | * the elevator_type associated with a request_queue. | |
3181 | */ | |
3182 | struct blk_mq_qe_pair { | |
3183 | struct list_head node; | |
3184 | struct request_queue *q; | |
3185 | struct elevator_type *type; | |
3186 | }; | |
3187 | ||
3188 | /* | |
3189 | * Cache the elevator_type in qe pair list and switch the | |
3190 | * io scheduler to 'none' | |
3191 | */ | |
3192 | static bool blk_mq_elv_switch_none(struct list_head *head, | |
3193 | struct request_queue *q) | |
3194 | { | |
3195 | struct blk_mq_qe_pair *qe; | |
3196 | ||
3197 | if (!q->elevator) | |
3198 | return true; | |
3199 | ||
3200 | qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); | |
3201 | if (!qe) | |
3202 | return false; | |
3203 | ||
3204 | INIT_LIST_HEAD(&qe->node); | |
3205 | qe->q = q; | |
3206 | qe->type = q->elevator->type; | |
3207 | list_add(&qe->node, head); | |
3208 | ||
3209 | mutex_lock(&q->sysfs_lock); | |
3210 | /* | |
3211 | * After elevator_switch_mq, the previous elevator_queue will be | |
3212 | * released by elevator_release. The reference of the io scheduler | |
3213 | * module get by elevator_get will also be put. So we need to get | |
3214 | * a reference of the io scheduler module here to prevent it to be | |
3215 | * removed. | |
3216 | */ | |
3217 | __module_get(qe->type->elevator_owner); | |
3218 | elevator_switch_mq(q, NULL); | |
3219 | mutex_unlock(&q->sysfs_lock); | |
3220 | ||
3221 | return true; | |
3222 | } | |
3223 | ||
3224 | static void blk_mq_elv_switch_back(struct list_head *head, | |
3225 | struct request_queue *q) | |
3226 | { | |
3227 | struct blk_mq_qe_pair *qe; | |
3228 | struct elevator_type *t = NULL; | |
3229 | ||
3230 | list_for_each_entry(qe, head, node) | |
3231 | if (qe->q == q) { | |
3232 | t = qe->type; | |
3233 | break; | |
3234 | } | |
3235 | ||
3236 | if (!t) | |
3237 | return; | |
3238 | ||
3239 | list_del(&qe->node); | |
3240 | kfree(qe); | |
3241 | ||
3242 | mutex_lock(&q->sysfs_lock); | |
3243 | elevator_switch_mq(q, t); | |
3244 | mutex_unlock(&q->sysfs_lock); | |
3245 | } | |
3246 | ||
e4dc2b32 KB |
3247 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
3248 | int nr_hw_queues) | |
868f2f0b KB |
3249 | { |
3250 | struct request_queue *q; | |
d48ece20 | 3251 | LIST_HEAD(head); |
e01ad46d | 3252 | int prev_nr_hw_queues; |
868f2f0b | 3253 | |
705cda97 BVA |
3254 | lockdep_assert_held(&set->tag_list_lock); |
3255 | ||
392546ae | 3256 | if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids) |
868f2f0b KB |
3257 | nr_hw_queues = nr_cpu_ids; |
3258 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
3259 | return; | |
3260 | ||
3261 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3262 | blk_mq_freeze_queue(q); | |
f5bbbbe4 JW |
3263 | /* |
3264 | * Sync with blk_mq_queue_tag_busy_iter. | |
3265 | */ | |
3266 | synchronize_rcu(); | |
d48ece20 JW |
3267 | /* |
3268 | * Switch IO scheduler to 'none', cleaning up the data associated | |
3269 | * with the previous scheduler. We will switch back once we are done | |
3270 | * updating the new sw to hw queue mappings. | |
3271 | */ | |
3272 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3273 | if (!blk_mq_elv_switch_none(&head, q)) | |
3274 | goto switch_back; | |
868f2f0b | 3275 | |
477e19de JW |
3276 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3277 | blk_mq_debugfs_unregister_hctxs(q); | |
3278 | blk_mq_sysfs_unregister(q); | |
3279 | } | |
3280 | ||
e01ad46d | 3281 | prev_nr_hw_queues = set->nr_hw_queues; |
868f2f0b | 3282 | set->nr_hw_queues = nr_hw_queues; |
ebe8bddb | 3283 | blk_mq_update_queue_map(set); |
e01ad46d | 3284 | fallback: |
868f2f0b KB |
3285 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3286 | blk_mq_realloc_hw_ctxs(set, q); | |
e01ad46d JW |
3287 | if (q->nr_hw_queues != set->nr_hw_queues) { |
3288 | pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n", | |
3289 | nr_hw_queues, prev_nr_hw_queues); | |
3290 | set->nr_hw_queues = prev_nr_hw_queues; | |
7d76f856 | 3291 | blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
e01ad46d JW |
3292 | goto fallback; |
3293 | } | |
477e19de JW |
3294 | blk_mq_map_swqueue(q); |
3295 | } | |
3296 | ||
3297 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
3298 | blk_mq_sysfs_register(q); | |
3299 | blk_mq_debugfs_register_hctxs(q); | |
868f2f0b KB |
3300 | } |
3301 | ||
d48ece20 JW |
3302 | switch_back: |
3303 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3304 | blk_mq_elv_switch_back(&head, q); | |
3305 | ||
868f2f0b KB |
3306 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
3307 | blk_mq_unfreeze_queue(q); | |
3308 | } | |
e4dc2b32 KB |
3309 | |
3310 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
3311 | { | |
3312 | mutex_lock(&set->tag_list_lock); | |
3313 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
3314 | mutex_unlock(&set->tag_list_lock); | |
3315 | } | |
868f2f0b KB |
3316 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
3317 | ||
34dbad5d OS |
3318 | /* Enable polling stats and return whether they were already enabled. */ |
3319 | static bool blk_poll_stats_enable(struct request_queue *q) | |
3320 | { | |
3321 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
7dfdbc73 | 3322 | blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q)) |
34dbad5d OS |
3323 | return true; |
3324 | blk_stat_add_callback(q, q->poll_cb); | |
3325 | return false; | |
3326 | } | |
3327 | ||
3328 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
3329 | { | |
3330 | /* | |
3331 | * We don't arm the callback if polling stats are not enabled or the | |
3332 | * callback is already active. | |
3333 | */ | |
3334 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
3335 | blk_stat_is_active(q->poll_cb)) | |
3336 | return; | |
3337 | ||
3338 | blk_stat_activate_msecs(q->poll_cb, 100); | |
3339 | } | |
3340 | ||
3341 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
3342 | { | |
3343 | struct request_queue *q = cb->data; | |
720b8ccc | 3344 | int bucket; |
34dbad5d | 3345 | |
720b8ccc SB |
3346 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
3347 | if (cb->stat[bucket].nr_samples) | |
3348 | q->poll_stat[bucket] = cb->stat[bucket]; | |
3349 | } | |
34dbad5d OS |
3350 | } |
3351 | ||
64f1c21e JA |
3352 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
3353 | struct blk_mq_hw_ctx *hctx, | |
3354 | struct request *rq) | |
3355 | { | |
64f1c21e | 3356 | unsigned long ret = 0; |
720b8ccc | 3357 | int bucket; |
64f1c21e JA |
3358 | |
3359 | /* | |
3360 | * If stats collection isn't on, don't sleep but turn it on for | |
3361 | * future users | |
3362 | */ | |
34dbad5d | 3363 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
3364 | return 0; |
3365 | ||
64f1c21e JA |
3366 | /* |
3367 | * As an optimistic guess, use half of the mean service time | |
3368 | * for this type of request. We can (and should) make this smarter. | |
3369 | * For instance, if the completion latencies are tight, we can | |
3370 | * get closer than just half the mean. This is especially | |
3371 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
3372 | * than ~10 usec. We do use the stats for the relevant IO size |
3373 | * if available which does lead to better estimates. | |
64f1c21e | 3374 | */ |
720b8ccc SB |
3375 | bucket = blk_mq_poll_stats_bkt(rq); |
3376 | if (bucket < 0) | |
3377 | return ret; | |
3378 | ||
3379 | if (q->poll_stat[bucket].nr_samples) | |
3380 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
3381 | |
3382 | return ret; | |
3383 | } | |
3384 | ||
06426adf | 3385 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 3386 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
3387 | struct request *rq) |
3388 | { | |
3389 | struct hrtimer_sleeper hs; | |
3390 | enum hrtimer_mode mode; | |
64f1c21e | 3391 | unsigned int nsecs; |
06426adf JA |
3392 | ktime_t kt; |
3393 | ||
76a86f9d | 3394 | if (rq->rq_flags & RQF_MQ_POLL_SLEPT) |
64f1c21e JA |
3395 | return false; |
3396 | ||
3397 | /* | |
1052b8ac | 3398 | * If we get here, hybrid polling is enabled. Hence poll_nsec can be: |
64f1c21e | 3399 | * |
64f1c21e JA |
3400 | * 0: use half of prev avg |
3401 | * >0: use this specific value | |
3402 | */ | |
1052b8ac | 3403 | if (q->poll_nsec > 0) |
64f1c21e JA |
3404 | nsecs = q->poll_nsec; |
3405 | else | |
3406 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
3407 | ||
3408 | if (!nsecs) | |
06426adf JA |
3409 | return false; |
3410 | ||
76a86f9d | 3411 | rq->rq_flags |= RQF_MQ_POLL_SLEPT; |
06426adf JA |
3412 | |
3413 | /* | |
3414 | * This will be replaced with the stats tracking code, using | |
3415 | * 'avg_completion_time / 2' as the pre-sleep target. | |
3416 | */ | |
8b0e1953 | 3417 | kt = nsecs; |
06426adf JA |
3418 | |
3419 | mode = HRTIMER_MODE_REL; | |
3420 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
3421 | hrtimer_set_expires(&hs.timer, kt); | |
3422 | ||
3423 | hrtimer_init_sleeper(&hs, current); | |
3424 | do { | |
5a61c363 | 3425 | if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE) |
06426adf JA |
3426 | break; |
3427 | set_current_state(TASK_UNINTERRUPTIBLE); | |
3428 | hrtimer_start_expires(&hs.timer, mode); | |
3429 | if (hs.task) | |
3430 | io_schedule(); | |
3431 | hrtimer_cancel(&hs.timer); | |
3432 | mode = HRTIMER_MODE_ABS; | |
3433 | } while (hs.task && !signal_pending(current)); | |
3434 | ||
3435 | __set_current_state(TASK_RUNNING); | |
3436 | destroy_hrtimer_on_stack(&hs.timer); | |
3437 | return true; | |
3438 | } | |
3439 | ||
1052b8ac JA |
3440 | static bool blk_mq_poll_hybrid(struct request_queue *q, |
3441 | struct blk_mq_hw_ctx *hctx, blk_qc_t cookie) | |
bbd7bb70 | 3442 | { |
1052b8ac JA |
3443 | struct request *rq; |
3444 | ||
29ece8b4 | 3445 | if (q->poll_nsec == BLK_MQ_POLL_CLASSIC) |
1052b8ac JA |
3446 | return false; |
3447 | ||
3448 | if (!blk_qc_t_is_internal(cookie)) | |
3449 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
3450 | else { | |
3451 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); | |
3452 | /* | |
3453 | * With scheduling, if the request has completed, we'll | |
3454 | * get a NULL return here, as we clear the sched tag when | |
3455 | * that happens. The request still remains valid, like always, | |
3456 | * so we should be safe with just the NULL check. | |
3457 | */ | |
3458 | if (!rq) | |
3459 | return false; | |
3460 | } | |
3461 | ||
3462 | return blk_mq_poll_hybrid_sleep(q, hctx, rq); | |
3463 | } | |
3464 | ||
529262d5 CH |
3465 | /** |
3466 | * blk_poll - poll for IO completions | |
3467 | * @q: the queue | |
3468 | * @cookie: cookie passed back at IO submission time | |
3469 | * @spin: whether to spin for completions | |
3470 | * | |
3471 | * Description: | |
3472 | * Poll for completions on the passed in queue. Returns number of | |
3473 | * completed entries found. If @spin is true, then blk_poll will continue | |
3474 | * looping until at least one completion is found, unless the task is | |
3475 | * otherwise marked running (or we need to reschedule). | |
3476 | */ | |
3477 | int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin) | |
1052b8ac JA |
3478 | { |
3479 | struct blk_mq_hw_ctx *hctx; | |
bbd7bb70 JA |
3480 | long state; |
3481 | ||
529262d5 CH |
3482 | if (!blk_qc_t_valid(cookie) || |
3483 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
1052b8ac JA |
3484 | return 0; |
3485 | ||
529262d5 CH |
3486 | if (current->plug) |
3487 | blk_flush_plug_list(current->plug, false); | |
3488 | ||
1052b8ac JA |
3489 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; |
3490 | ||
06426adf JA |
3491 | /* |
3492 | * If we sleep, have the caller restart the poll loop to reset | |
3493 | * the state. Like for the other success return cases, the | |
3494 | * caller is responsible for checking if the IO completed. If | |
3495 | * the IO isn't complete, we'll get called again and will go | |
3496 | * straight to the busy poll loop. | |
3497 | */ | |
1052b8ac | 3498 | if (blk_mq_poll_hybrid(q, hctx, cookie)) |
85f4d4b6 | 3499 | return 1; |
06426adf | 3500 | |
bbd7bb70 JA |
3501 | hctx->poll_considered++; |
3502 | ||
3503 | state = current->state; | |
aa61bec3 | 3504 | do { |
bbd7bb70 JA |
3505 | int ret; |
3506 | ||
3507 | hctx->poll_invoked++; | |
3508 | ||
9743139c | 3509 | ret = q->mq_ops->poll(hctx); |
bbd7bb70 JA |
3510 | if (ret > 0) { |
3511 | hctx->poll_success++; | |
849a3700 | 3512 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 3513 | return ret; |
bbd7bb70 JA |
3514 | } |
3515 | ||
3516 | if (signal_pending_state(state, current)) | |
849a3700 | 3517 | __set_current_state(TASK_RUNNING); |
bbd7bb70 JA |
3518 | |
3519 | if (current->state == TASK_RUNNING) | |
85f4d4b6 | 3520 | return 1; |
0a1b8b87 | 3521 | if (ret < 0 || !spin) |
bbd7bb70 JA |
3522 | break; |
3523 | cpu_relax(); | |
aa61bec3 | 3524 | } while (!need_resched()); |
bbd7bb70 | 3525 | |
67b4110f | 3526 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 3527 | return 0; |
bbd7bb70 | 3528 | } |
529262d5 | 3529 | EXPORT_SYMBOL_GPL(blk_poll); |
bbd7bb70 | 3530 | |
9cf2bab6 JA |
3531 | unsigned int blk_mq_rq_cpu(struct request *rq) |
3532 | { | |
3533 | return rq->mq_ctx->cpu; | |
3534 | } | |
3535 | EXPORT_SYMBOL(blk_mq_rq_cpu); | |
3536 | ||
320ae51f JA |
3537 | static int __init blk_mq_init(void) |
3538 | { | |
9467f859 TG |
3539 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
3540 | blk_mq_hctx_notify_dead); | |
320ae51f JA |
3541 | return 0; |
3542 | } | |
3543 | subsys_initcall(blk_mq_init); |