Commit | Line | Data |
---|---|---|
75bb4625 JA |
1 | /* |
2 | * Block multiqueue core code | |
3 | * | |
4 | * Copyright (C) 2013-2014 Jens Axboe | |
5 | * Copyright (C) 2013-2014 Christoph Hellwig | |
6 | */ | |
320ae51f JA |
7 | #include <linux/kernel.h> |
8 | #include <linux/module.h> | |
9 | #include <linux/backing-dev.h> | |
10 | #include <linux/bio.h> | |
11 | #include <linux/blkdev.h> | |
f75782e4 | 12 | #include <linux/kmemleak.h> |
320ae51f JA |
13 | #include <linux/mm.h> |
14 | #include <linux/init.h> | |
15 | #include <linux/slab.h> | |
16 | #include <linux/workqueue.h> | |
17 | #include <linux/smp.h> | |
18 | #include <linux/llist.h> | |
19 | #include <linux/list_sort.h> | |
20 | #include <linux/cpu.h> | |
21 | #include <linux/cache.h> | |
22 | #include <linux/sched/sysctl.h> | |
105ab3d8 | 23 | #include <linux/sched/topology.h> |
174cd4b1 | 24 | #include <linux/sched/signal.h> |
320ae51f | 25 | #include <linux/delay.h> |
aedcd72f | 26 | #include <linux/crash_dump.h> |
88c7b2b7 | 27 | #include <linux/prefetch.h> |
320ae51f JA |
28 | |
29 | #include <trace/events/block.h> | |
30 | ||
31 | #include <linux/blk-mq.h> | |
32 | #include "blk.h" | |
33 | #include "blk-mq.h" | |
9c1051aa | 34 | #include "blk-mq-debugfs.h" |
320ae51f | 35 | #include "blk-mq-tag.h" |
986d413b | 36 | #include "blk-pm.h" |
cf43e6be | 37 | #include "blk-stat.h" |
bd166ef1 | 38 | #include "blk-mq-sched.h" |
c1c80384 | 39 | #include "blk-rq-qos.h" |
320ae51f | 40 | |
34dbad5d OS |
41 | static void blk_mq_poll_stats_start(struct request_queue *q); |
42 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb); | |
43 | ||
720b8ccc SB |
44 | static int blk_mq_poll_stats_bkt(const struct request *rq) |
45 | { | |
46 | int ddir, bytes, bucket; | |
47 | ||
99c749a4 | 48 | ddir = rq_data_dir(rq); |
720b8ccc SB |
49 | bytes = blk_rq_bytes(rq); |
50 | ||
51 | bucket = ddir + 2*(ilog2(bytes) - 9); | |
52 | ||
53 | if (bucket < 0) | |
54 | return -1; | |
55 | else if (bucket >= BLK_MQ_POLL_STATS_BKTS) | |
56 | return ddir + BLK_MQ_POLL_STATS_BKTS - 2; | |
57 | ||
58 | return bucket; | |
59 | } | |
60 | ||
320ae51f | 61 | /* |
85fae294 YY |
62 | * Check if any of the ctx, dispatch list or elevator |
63 | * have pending work in this hardware queue. | |
320ae51f | 64 | */ |
79f720a7 | 65 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 66 | { |
79f720a7 JA |
67 | return !list_empty_careful(&hctx->dispatch) || |
68 | sbitmap_any_bit_set(&hctx->ctx_map) || | |
bd166ef1 | 69 | blk_mq_sched_has_work(hctx); |
1429d7c9 JA |
70 | } |
71 | ||
320ae51f JA |
72 | /* |
73 | * Mark this ctx as having pending work in this hardware queue | |
74 | */ | |
75 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
76 | struct blk_mq_ctx *ctx) | |
77 | { | |
f31967f0 JA |
78 | const int bit = ctx->index_hw[hctx->type]; |
79 | ||
80 | if (!sbitmap_test_bit(&hctx->ctx_map, bit)) | |
81 | sbitmap_set_bit(&hctx->ctx_map, bit); | |
1429d7c9 JA |
82 | } |
83 | ||
84 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
85 | struct blk_mq_ctx *ctx) | |
86 | { | |
f31967f0 JA |
87 | const int bit = ctx->index_hw[hctx->type]; |
88 | ||
89 | sbitmap_clear_bit(&hctx->ctx_map, bit); | |
320ae51f JA |
90 | } |
91 | ||
f299b7c7 JA |
92 | struct mq_inflight { |
93 | struct hd_struct *part; | |
94 | unsigned int *inflight; | |
95 | }; | |
96 | ||
7baa8572 | 97 | static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx, |
f299b7c7 JA |
98 | struct request *rq, void *priv, |
99 | bool reserved) | |
100 | { | |
101 | struct mq_inflight *mi = priv; | |
102 | ||
6131837b | 103 | /* |
e016b782 | 104 | * index[0] counts the specific partition that was asked for. |
6131837b OS |
105 | */ |
106 | if (rq->part == mi->part) | |
107 | mi->inflight[0]++; | |
7baa8572 JA |
108 | |
109 | return true; | |
f299b7c7 JA |
110 | } |
111 | ||
e016b782 | 112 | unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part) |
f299b7c7 | 113 | { |
e016b782 | 114 | unsigned inflight[2]; |
f299b7c7 JA |
115 | struct mq_inflight mi = { .part = part, .inflight = inflight, }; |
116 | ||
b8d62b3a | 117 | inflight[0] = inflight[1] = 0; |
f299b7c7 | 118 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
e016b782 MP |
119 | |
120 | return inflight[0]; | |
f299b7c7 JA |
121 | } |
122 | ||
7baa8572 | 123 | static bool blk_mq_check_inflight_rw(struct blk_mq_hw_ctx *hctx, |
bf0ddaba OS |
124 | struct request *rq, void *priv, |
125 | bool reserved) | |
126 | { | |
127 | struct mq_inflight *mi = priv; | |
128 | ||
129 | if (rq->part == mi->part) | |
130 | mi->inflight[rq_data_dir(rq)]++; | |
7baa8572 JA |
131 | |
132 | return true; | |
bf0ddaba OS |
133 | } |
134 | ||
135 | void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part, | |
136 | unsigned int inflight[2]) | |
137 | { | |
138 | struct mq_inflight mi = { .part = part, .inflight = inflight, }; | |
139 | ||
140 | inflight[0] = inflight[1] = 0; | |
141 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight_rw, &mi); | |
142 | } | |
143 | ||
1671d522 | 144 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 145 | { |
4ecd4fef | 146 | int freeze_depth; |
cddd5d17 | 147 | |
4ecd4fef CH |
148 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
149 | if (freeze_depth == 1) { | |
3ef28e83 | 150 | percpu_ref_kill(&q->q_usage_counter); |
344e9ffc | 151 | if (queue_is_mq(q)) |
055f6e18 | 152 | blk_mq_run_hw_queues(q, false); |
cddd5d17 | 153 | } |
f3af020b | 154 | } |
1671d522 | 155 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
f3af020b | 156 | |
6bae363e | 157 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 158 | { |
3ef28e83 | 159 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 160 | } |
6bae363e | 161 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 162 | |
f91328c4 KB |
163 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
164 | unsigned long timeout) | |
165 | { | |
166 | return wait_event_timeout(q->mq_freeze_wq, | |
167 | percpu_ref_is_zero(&q->q_usage_counter), | |
168 | timeout); | |
169 | } | |
170 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 171 | |
f3af020b TH |
172 | /* |
173 | * Guarantee no request is in use, so we can change any data structure of | |
174 | * the queue afterward. | |
175 | */ | |
3ef28e83 | 176 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 177 | { |
3ef28e83 DW |
178 | /* |
179 | * In the !blk_mq case we are only calling this to kill the | |
180 | * q_usage_counter, otherwise this increases the freeze depth | |
181 | * and waits for it to return to zero. For this reason there is | |
182 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
183 | * exported to drivers as the only user for unfreeze is blk_mq. | |
184 | */ | |
1671d522 | 185 | blk_freeze_queue_start(q); |
f3af020b TH |
186 | blk_mq_freeze_queue_wait(q); |
187 | } | |
3ef28e83 DW |
188 | |
189 | void blk_mq_freeze_queue(struct request_queue *q) | |
190 | { | |
191 | /* | |
192 | * ...just an alias to keep freeze and unfreeze actions balanced | |
193 | * in the blk_mq_* namespace | |
194 | */ | |
195 | blk_freeze_queue(q); | |
196 | } | |
c761d96b | 197 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 198 | |
b4c6a028 | 199 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 200 | { |
4ecd4fef | 201 | int freeze_depth; |
320ae51f | 202 | |
4ecd4fef CH |
203 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
204 | WARN_ON_ONCE(freeze_depth < 0); | |
205 | if (!freeze_depth) { | |
bdd63160 | 206 | percpu_ref_resurrect(&q->q_usage_counter); |
320ae51f | 207 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 208 | } |
320ae51f | 209 | } |
b4c6a028 | 210 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 211 | |
852ec809 BVA |
212 | /* |
213 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
214 | * mpt3sas driver such that this function can be removed. | |
215 | */ | |
216 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
217 | { | |
8814ce8a | 218 | blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q); |
852ec809 BVA |
219 | } |
220 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
221 | ||
6a83e74d | 222 | /** |
69e07c4a | 223 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished |
6a83e74d BVA |
224 | * @q: request queue. |
225 | * | |
226 | * Note: this function does not prevent that the struct request end_io() | |
69e07c4a ML |
227 | * callback function is invoked. Once this function is returned, we make |
228 | * sure no dispatch can happen until the queue is unquiesced via | |
229 | * blk_mq_unquiesce_queue(). | |
6a83e74d BVA |
230 | */ |
231 | void blk_mq_quiesce_queue(struct request_queue *q) | |
232 | { | |
233 | struct blk_mq_hw_ctx *hctx; | |
234 | unsigned int i; | |
235 | bool rcu = false; | |
236 | ||
1d9e9bc6 | 237 | blk_mq_quiesce_queue_nowait(q); |
f4560ffe | 238 | |
6a83e74d BVA |
239 | queue_for_each_hw_ctx(q, hctx, i) { |
240 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
05707b64 | 241 | synchronize_srcu(hctx->srcu); |
6a83e74d BVA |
242 | else |
243 | rcu = true; | |
244 | } | |
245 | if (rcu) | |
246 | synchronize_rcu(); | |
247 | } | |
248 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
249 | ||
e4e73913 ML |
250 | /* |
251 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
252 | * @q: request queue. | |
253 | * | |
254 | * This function recovers queue into the state before quiescing | |
255 | * which is done by blk_mq_quiesce_queue. | |
256 | */ | |
257 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
258 | { | |
8814ce8a | 259 | blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q); |
f4560ffe | 260 | |
1d9e9bc6 ML |
261 | /* dispatch requests which are inserted during quiescing */ |
262 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
263 | } |
264 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
265 | ||
aed3ea94 JA |
266 | void blk_mq_wake_waiters(struct request_queue *q) |
267 | { | |
268 | struct blk_mq_hw_ctx *hctx; | |
269 | unsigned int i; | |
270 | ||
271 | queue_for_each_hw_ctx(q, hctx, i) | |
272 | if (blk_mq_hw_queue_mapped(hctx)) | |
273 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
274 | } | |
275 | ||
320ae51f JA |
276 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
277 | { | |
278 | return blk_mq_has_free_tags(hctx->tags); | |
279 | } | |
280 | EXPORT_SYMBOL(blk_mq_can_queue); | |
281 | ||
fe1f4526 JA |
282 | /* |
283 | * Only need start/end time stamping if we have stats enabled, or using | |
284 | * an IO scheduler. | |
285 | */ | |
286 | static inline bool blk_mq_need_time_stamp(struct request *rq) | |
287 | { | |
288 | return (rq->rq_flags & RQF_IO_STAT) || rq->q->elevator; | |
289 | } | |
290 | ||
e4cdf1a1 CH |
291 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
292 | unsigned int tag, unsigned int op) | |
320ae51f | 293 | { |
e4cdf1a1 CH |
294 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
295 | struct request *rq = tags->static_rqs[tag]; | |
bf9ae8c5 | 296 | req_flags_t rq_flags = 0; |
c3a148d2 | 297 | |
e4cdf1a1 CH |
298 | if (data->flags & BLK_MQ_REQ_INTERNAL) { |
299 | rq->tag = -1; | |
300 | rq->internal_tag = tag; | |
301 | } else { | |
d263ed99 | 302 | if (data->hctx->flags & BLK_MQ_F_TAG_SHARED) { |
bf9ae8c5 | 303 | rq_flags = RQF_MQ_INFLIGHT; |
e4cdf1a1 CH |
304 | atomic_inc(&data->hctx->nr_active); |
305 | } | |
306 | rq->tag = tag; | |
307 | rq->internal_tag = -1; | |
308 | data->hctx->tags->rqs[rq->tag] = rq; | |
309 | } | |
310 | ||
af76e555 | 311 | /* csd/requeue_work/fifo_time is initialized before use */ |
e4cdf1a1 CH |
312 | rq->q = data->q; |
313 | rq->mq_ctx = data->ctx; | |
ea4f995e | 314 | rq->mq_hctx = data->hctx; |
bf9ae8c5 | 315 | rq->rq_flags = rq_flags; |
ef295ecf | 316 | rq->cmd_flags = op; |
1b6d65a0 BVA |
317 | if (data->flags & BLK_MQ_REQ_PREEMPT) |
318 | rq->rq_flags |= RQF_PREEMPT; | |
e4cdf1a1 | 319 | if (blk_queue_io_stat(data->q)) |
e8064021 | 320 | rq->rq_flags |= RQF_IO_STAT; |
7c3fb70f | 321 | INIT_LIST_HEAD(&rq->queuelist); |
af76e555 CH |
322 | INIT_HLIST_NODE(&rq->hash); |
323 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
324 | rq->rq_disk = NULL; |
325 | rq->part = NULL; | |
fe1f4526 JA |
326 | if (blk_mq_need_time_stamp(rq)) |
327 | rq->start_time_ns = ktime_get_ns(); | |
328 | else | |
329 | rq->start_time_ns = 0; | |
544ccc8d | 330 | rq->io_start_time_ns = 0; |
af76e555 CH |
331 | rq->nr_phys_segments = 0; |
332 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
333 | rq->nr_integrity_segments = 0; | |
334 | #endif | |
af76e555 | 335 | /* tag was already set */ |
af76e555 | 336 | rq->extra_len = 0; |
079076b3 | 337 | WRITE_ONCE(rq->deadline, 0); |
af76e555 | 338 | |
f6be4fb4 JA |
339 | rq->timeout = 0; |
340 | ||
af76e555 CH |
341 | rq->end_io = NULL; |
342 | rq->end_io_data = NULL; | |
af76e555 | 343 | |
e4cdf1a1 | 344 | data->ctx->rq_dispatched[op_is_sync(op)]++; |
12f5b931 | 345 | refcount_set(&rq->ref, 1); |
e4cdf1a1 | 346 | return rq; |
5dee8577 CH |
347 | } |
348 | ||
d2c0d383 | 349 | static struct request *blk_mq_get_request(struct request_queue *q, |
f9afca4d JA |
350 | struct bio *bio, |
351 | struct blk_mq_alloc_data *data) | |
d2c0d383 CH |
352 | { |
353 | struct elevator_queue *e = q->elevator; | |
354 | struct request *rq; | |
e4cdf1a1 | 355 | unsigned int tag; |
21e768b4 | 356 | bool put_ctx_on_error = false; |
d2c0d383 CH |
357 | |
358 | blk_queue_enter_live(q); | |
359 | data->q = q; | |
21e768b4 BVA |
360 | if (likely(!data->ctx)) { |
361 | data->ctx = blk_mq_get_ctx(q); | |
362 | put_ctx_on_error = true; | |
363 | } | |
d2c0d383 | 364 | if (likely(!data->hctx)) |
f9afca4d | 365 | data->hctx = blk_mq_map_queue(q, data->cmd_flags, |
8ccdf4a3 | 366 | data->ctx); |
f9afca4d | 367 | if (data->cmd_flags & REQ_NOWAIT) |
03a07c92 | 368 | data->flags |= BLK_MQ_REQ_NOWAIT; |
d2c0d383 CH |
369 | |
370 | if (e) { | |
371 | data->flags |= BLK_MQ_REQ_INTERNAL; | |
372 | ||
373 | /* | |
374 | * Flush requests are special and go directly to the | |
17a51199 JA |
375 | * dispatch list. Don't include reserved tags in the |
376 | * limiting, as it isn't useful. | |
d2c0d383 | 377 | */ |
f9afca4d JA |
378 | if (!op_is_flush(data->cmd_flags) && |
379 | e->type->ops.limit_depth && | |
17a51199 | 380 | !(data->flags & BLK_MQ_REQ_RESERVED)) |
f9afca4d | 381 | e->type->ops.limit_depth(data->cmd_flags, data); |
d263ed99 JW |
382 | } else { |
383 | blk_mq_tag_busy(data->hctx); | |
d2c0d383 CH |
384 | } |
385 | ||
e4cdf1a1 CH |
386 | tag = blk_mq_get_tag(data); |
387 | if (tag == BLK_MQ_TAG_FAIL) { | |
21e768b4 BVA |
388 | if (put_ctx_on_error) { |
389 | blk_mq_put_ctx(data->ctx); | |
1ad43c00 ML |
390 | data->ctx = NULL; |
391 | } | |
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 | 427 | |
1ad43c00 | 428 | blk_mq_put_ctx(alloc_data.ctx); |
1ad43c00 | 429 | |
0c4de0f3 CH |
430 | rq->__data_len = 0; |
431 | rq->__sector = (sector_t) -1; | |
432 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
433 | return rq; |
434 | } | |
4bb659b1 | 435 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 436 | |
cd6ce148 | 437 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
9a95e4ef | 438 | unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 439 | { |
f9afca4d | 440 | struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op }; |
1f5bd336 | 441 | struct request *rq; |
6d2809d5 | 442 | unsigned int cpu; |
1f5bd336 ML |
443 | int ret; |
444 | ||
445 | /* | |
446 | * If the tag allocator sleeps we could get an allocation for a | |
447 | * different hardware context. No need to complicate the low level | |
448 | * allocator for this for the rare use case of a command tied to | |
449 | * a specific queue. | |
450 | */ | |
451 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
452 | return ERR_PTR(-EINVAL); | |
453 | ||
454 | if (hctx_idx >= q->nr_hw_queues) | |
455 | return ERR_PTR(-EIO); | |
456 | ||
3a0a5299 | 457 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
458 | if (ret) |
459 | return ERR_PTR(ret); | |
460 | ||
c8712c6a CH |
461 | /* |
462 | * Check if the hardware context is actually mapped to anything. | |
463 | * If not tell the caller that it should skip this queue. | |
464 | */ | |
6d2809d5 OS |
465 | alloc_data.hctx = q->queue_hw_ctx[hctx_idx]; |
466 | if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) { | |
467 | blk_queue_exit(q); | |
468 | return ERR_PTR(-EXDEV); | |
c8712c6a | 469 | } |
20e4d813 | 470 | cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask); |
6d2809d5 | 471 | alloc_data.ctx = __blk_mq_get_ctx(q, cpu); |
1f5bd336 | 472 | |
f9afca4d | 473 | rq = blk_mq_get_request(q, NULL, &alloc_data); |
3280d66a | 474 | blk_queue_exit(q); |
c8712c6a | 475 | |
6d2809d5 OS |
476 | if (!rq) |
477 | return ERR_PTR(-EWOULDBLOCK); | |
478 | ||
479 | return rq; | |
1f5bd336 ML |
480 | } |
481 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
482 | ||
12f5b931 KB |
483 | static void __blk_mq_free_request(struct request *rq) |
484 | { | |
485 | struct request_queue *q = rq->q; | |
486 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 487 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
12f5b931 KB |
488 | const int sched_tag = rq->internal_tag; |
489 | ||
986d413b | 490 | blk_pm_mark_last_busy(rq); |
ea4f995e | 491 | rq->mq_hctx = NULL; |
12f5b931 KB |
492 | if (rq->tag != -1) |
493 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
494 | if (sched_tag != -1) | |
495 | blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag); | |
496 | blk_mq_sched_restart(hctx); | |
497 | blk_queue_exit(q); | |
498 | } | |
499 | ||
6af54051 | 500 | void blk_mq_free_request(struct request *rq) |
320ae51f | 501 | { |
320ae51f | 502 | struct request_queue *q = rq->q; |
6af54051 CH |
503 | struct elevator_queue *e = q->elevator; |
504 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 505 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
6af54051 | 506 | |
5bbf4e5a | 507 | if (rq->rq_flags & RQF_ELVPRIV) { |
f9cd4bfe JA |
508 | if (e && e->type->ops.finish_request) |
509 | e->type->ops.finish_request(rq); | |
6af54051 CH |
510 | if (rq->elv.icq) { |
511 | put_io_context(rq->elv.icq->ioc); | |
512 | rq->elv.icq = NULL; | |
513 | } | |
514 | } | |
320ae51f | 515 | |
6af54051 | 516 | ctx->rq_completed[rq_is_sync(rq)]++; |
e8064021 | 517 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 518 | atomic_dec(&hctx->nr_active); |
87760e5e | 519 | |
7beb2f84 JA |
520 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
521 | laptop_io_completion(q->backing_dev_info); | |
522 | ||
a7905043 | 523 | rq_qos_done(q, rq); |
0d2602ca | 524 | |
12f5b931 KB |
525 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
526 | if (refcount_dec_and_test(&rq->ref)) | |
527 | __blk_mq_free_request(rq); | |
320ae51f | 528 | } |
1a3b595a | 529 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 530 | |
2a842aca | 531 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
320ae51f | 532 | { |
fe1f4526 JA |
533 | u64 now = 0; |
534 | ||
535 | if (blk_mq_need_time_stamp(rq)) | |
536 | now = ktime_get_ns(); | |
522a7775 | 537 | |
4bc6339a OS |
538 | if (rq->rq_flags & RQF_STATS) { |
539 | blk_mq_poll_stats_start(rq->q); | |
522a7775 | 540 | blk_stat_add(rq, now); |
4bc6339a OS |
541 | } |
542 | ||
ed88660a OS |
543 | if (rq->internal_tag != -1) |
544 | blk_mq_sched_completed_request(rq, now); | |
545 | ||
522a7775 | 546 | blk_account_io_done(rq, now); |
0d11e6ac | 547 | |
91b63639 | 548 | if (rq->end_io) { |
a7905043 | 549 | rq_qos_done(rq->q, rq); |
320ae51f | 550 | rq->end_io(rq, error); |
91b63639 | 551 | } else { |
320ae51f | 552 | blk_mq_free_request(rq); |
91b63639 | 553 | } |
320ae51f | 554 | } |
c8a446ad | 555 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 556 | |
2a842aca | 557 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
558 | { |
559 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
560 | BUG(); | |
c8a446ad | 561 | __blk_mq_end_request(rq, error); |
63151a44 | 562 | } |
c8a446ad | 563 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 564 | |
30a91cb4 | 565 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 566 | { |
3d6efbf6 | 567 | struct request *rq = data; |
c7bb9ad1 | 568 | struct request_queue *q = rq->q; |
320ae51f | 569 | |
c7bb9ad1 | 570 | q->mq_ops->complete(rq); |
320ae51f | 571 | } |
320ae51f | 572 | |
453f8341 | 573 | static void __blk_mq_complete_request(struct request *rq) |
320ae51f JA |
574 | { |
575 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
c7bb9ad1 | 576 | struct request_queue *q = rq->q; |
38535201 | 577 | bool shared = false; |
320ae51f JA |
578 | int cpu; |
579 | ||
af78ff7c | 580 | WRITE_ONCE(rq->state, MQ_RQ_COMPLETE); |
36e76539 ML |
581 | /* |
582 | * Most of single queue controllers, there is only one irq vector | |
583 | * for handling IO completion, and the only irq's affinity is set | |
584 | * as all possible CPUs. On most of ARCHs, this affinity means the | |
585 | * irq is handled on one specific CPU. | |
586 | * | |
587 | * So complete IO reqeust in softirq context in case of single queue | |
588 | * for not degrading IO performance by irqsoff latency. | |
589 | */ | |
c7bb9ad1 | 590 | if (q->nr_hw_queues == 1) { |
36e76539 ML |
591 | __blk_complete_request(rq); |
592 | return; | |
593 | } | |
594 | ||
4ab32bf3 JA |
595 | /* |
596 | * For a polled request, always complete locallly, it's pointless | |
597 | * to redirect the completion. | |
598 | */ | |
599 | if ((rq->cmd_flags & REQ_HIPRI) || | |
600 | !test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags)) { | |
c7bb9ad1 | 601 | q->mq_ops->complete(rq); |
30a91cb4 CH |
602 | return; |
603 | } | |
320ae51f JA |
604 | |
605 | cpu = get_cpu(); | |
c7bb9ad1 | 606 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &q->queue_flags)) |
38535201 CH |
607 | shared = cpus_share_cache(cpu, ctx->cpu); |
608 | ||
609 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 610 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
611 | rq->csd.info = rq; |
612 | rq->csd.flags = 0; | |
c46fff2a | 613 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 614 | } else { |
c7bb9ad1 | 615 | q->mq_ops->complete(rq); |
3d6efbf6 | 616 | } |
320ae51f JA |
617 | put_cpu(); |
618 | } | |
30a91cb4 | 619 | |
04ced159 | 620 | static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx) |
b7435db8 | 621 | __releases(hctx->srcu) |
04ced159 JA |
622 | { |
623 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
624 | rcu_read_unlock(); | |
625 | else | |
05707b64 | 626 | srcu_read_unlock(hctx->srcu, srcu_idx); |
04ced159 JA |
627 | } |
628 | ||
629 | static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx) | |
b7435db8 | 630 | __acquires(hctx->srcu) |
04ced159 | 631 | { |
08b5a6e2 JA |
632 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
633 | /* shut up gcc false positive */ | |
634 | *srcu_idx = 0; | |
04ced159 | 635 | rcu_read_lock(); |
08b5a6e2 | 636 | } else |
05707b64 | 637 | *srcu_idx = srcu_read_lock(hctx->srcu); |
04ced159 JA |
638 | } |
639 | ||
30a91cb4 CH |
640 | /** |
641 | * blk_mq_complete_request - end I/O on a request | |
642 | * @rq: the request being processed | |
643 | * | |
644 | * Description: | |
645 | * Ends all I/O on a request. It does not handle partial completions. | |
646 | * The actual completion happens out-of-order, through a IPI handler. | |
647 | **/ | |
16c15eb1 | 648 | bool blk_mq_complete_request(struct request *rq) |
30a91cb4 | 649 | { |
12f5b931 | 650 | if (unlikely(blk_should_fake_timeout(rq->q))) |
16c15eb1 | 651 | return false; |
12f5b931 | 652 | __blk_mq_complete_request(rq); |
16c15eb1 | 653 | return true; |
30a91cb4 CH |
654 | } |
655 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 656 | |
973c0191 KB |
657 | int blk_mq_request_started(struct request *rq) |
658 | { | |
5a61c363 | 659 | return blk_mq_rq_state(rq) != MQ_RQ_IDLE; |
973c0191 KB |
660 | } |
661 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
662 | ||
e2490073 | 663 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
664 | { |
665 | struct request_queue *q = rq->q; | |
666 | ||
bd166ef1 JA |
667 | blk_mq_sched_started_request(rq); |
668 | ||
320ae51f JA |
669 | trace_block_rq_issue(q, rq); |
670 | ||
cf43e6be | 671 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
544ccc8d OS |
672 | rq->io_start_time_ns = ktime_get_ns(); |
673 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW | |
674 | rq->throtl_size = blk_rq_sectors(rq); | |
675 | #endif | |
cf43e6be | 676 | rq->rq_flags |= RQF_STATS; |
a7905043 | 677 | rq_qos_issue(q, rq); |
cf43e6be JA |
678 | } |
679 | ||
1d9bd516 | 680 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 681 | |
1d9bd516 | 682 | blk_add_timer(rq); |
12f5b931 | 683 | WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); |
49f5baa5 CH |
684 | |
685 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
686 | /* | |
687 | * Make sure space for the drain appears. We know we can do | |
688 | * this because max_hw_segments has been adjusted to be one | |
689 | * fewer than the device can handle. | |
690 | */ | |
691 | rq->nr_phys_segments++; | |
692 | } | |
320ae51f | 693 | } |
e2490073 | 694 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 695 | |
ed0791b2 | 696 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
697 | { |
698 | struct request_queue *q = rq->q; | |
699 | ||
923218f6 ML |
700 | blk_mq_put_driver_tag(rq); |
701 | ||
320ae51f | 702 | trace_block_rq_requeue(q, rq); |
a7905043 | 703 | rq_qos_requeue(q, rq); |
49f5baa5 | 704 | |
12f5b931 KB |
705 | if (blk_mq_request_started(rq)) { |
706 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); | |
da661267 | 707 | rq->rq_flags &= ~RQF_TIMED_OUT; |
e2490073 CH |
708 | if (q->dma_drain_size && blk_rq_bytes(rq)) |
709 | rq->nr_phys_segments--; | |
710 | } | |
320ae51f JA |
711 | } |
712 | ||
2b053aca | 713 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 714 | { |
ed0791b2 | 715 | __blk_mq_requeue_request(rq); |
ed0791b2 | 716 | |
105976f5 ML |
717 | /* this request will be re-inserted to io scheduler queue */ |
718 | blk_mq_sched_requeue_request(rq); | |
719 | ||
7d692330 | 720 | BUG_ON(!list_empty(&rq->queuelist)); |
2b053aca | 721 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
722 | } |
723 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
724 | ||
6fca6a61 CH |
725 | static void blk_mq_requeue_work(struct work_struct *work) |
726 | { | |
727 | struct request_queue *q = | |
2849450a | 728 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
729 | LIST_HEAD(rq_list); |
730 | struct request *rq, *next; | |
6fca6a61 | 731 | |
18e9781d | 732 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 733 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 734 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
735 | |
736 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
aef1897c | 737 | if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP))) |
6fca6a61 CH |
738 | continue; |
739 | ||
e8064021 | 740 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 741 | list_del_init(&rq->queuelist); |
aef1897c JW |
742 | /* |
743 | * If RQF_DONTPREP, rq has contained some driver specific | |
744 | * data, so insert it to hctx dispatch list to avoid any | |
745 | * merge. | |
746 | */ | |
747 | if (rq->rq_flags & RQF_DONTPREP) | |
748 | blk_mq_request_bypass_insert(rq, false); | |
749 | else | |
750 | blk_mq_sched_insert_request(rq, true, false, false); | |
6fca6a61 CH |
751 | } |
752 | ||
753 | while (!list_empty(&rq_list)) { | |
754 | rq = list_entry(rq_list.next, struct request, queuelist); | |
755 | list_del_init(&rq->queuelist); | |
9e97d295 | 756 | blk_mq_sched_insert_request(rq, false, false, false); |
6fca6a61 CH |
757 | } |
758 | ||
52d7f1b5 | 759 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
760 | } |
761 | ||
2b053aca BVA |
762 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
763 | bool kick_requeue_list) | |
6fca6a61 CH |
764 | { |
765 | struct request_queue *q = rq->q; | |
766 | unsigned long flags; | |
767 | ||
768 | /* | |
769 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
ff821d27 | 770 | * request head insertion from the workqueue. |
6fca6a61 | 771 | */ |
e8064021 | 772 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
773 | |
774 | spin_lock_irqsave(&q->requeue_lock, flags); | |
775 | if (at_head) { | |
e8064021 | 776 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
777 | list_add(&rq->queuelist, &q->requeue_list); |
778 | } else { | |
779 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
780 | } | |
781 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
782 | |
783 | if (kick_requeue_list) | |
784 | blk_mq_kick_requeue_list(q); | |
6fca6a61 | 785 | } |
6fca6a61 CH |
786 | |
787 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
788 | { | |
ae943d20 | 789 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
790 | } |
791 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
792 | ||
2849450a MS |
793 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
794 | unsigned long msecs) | |
795 | { | |
d4acf365 BVA |
796 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
797 | msecs_to_jiffies(msecs)); | |
2849450a MS |
798 | } |
799 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
800 | ||
0e62f51f JA |
801 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
802 | { | |
88c7b2b7 JA |
803 | if (tag < tags->nr_tags) { |
804 | prefetch(tags->rqs[tag]); | |
4ee86bab | 805 | return tags->rqs[tag]; |
88c7b2b7 | 806 | } |
4ee86bab HR |
807 | |
808 | return NULL; | |
24d2f903 CH |
809 | } |
810 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
811 | ||
3c94d83c JA |
812 | static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq, |
813 | void *priv, bool reserved) | |
ae879912 JA |
814 | { |
815 | /* | |
3c94d83c JA |
816 | * If we find a request that is inflight and the queue matches, |
817 | * we know the queue is busy. Return false to stop the iteration. | |
ae879912 | 818 | */ |
3c94d83c | 819 | if (rq->state == MQ_RQ_IN_FLIGHT && rq->q == hctx->queue) { |
ae879912 JA |
820 | bool *busy = priv; |
821 | ||
822 | *busy = true; | |
823 | return false; | |
824 | } | |
825 | ||
826 | return true; | |
827 | } | |
828 | ||
3c94d83c | 829 | bool blk_mq_queue_inflight(struct request_queue *q) |
ae879912 JA |
830 | { |
831 | bool busy = false; | |
832 | ||
3c94d83c | 833 | blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy); |
ae879912 JA |
834 | return busy; |
835 | } | |
3c94d83c | 836 | EXPORT_SYMBOL_GPL(blk_mq_queue_inflight); |
ae879912 | 837 | |
358f70da | 838 | static void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 839 | { |
da661267 | 840 | req->rq_flags |= RQF_TIMED_OUT; |
d1210d5a CH |
841 | if (req->q->mq_ops->timeout) { |
842 | enum blk_eh_timer_return ret; | |
843 | ||
844 | ret = req->q->mq_ops->timeout(req, reserved); | |
845 | if (ret == BLK_EH_DONE) | |
846 | return; | |
847 | WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); | |
46f92d42 | 848 | } |
d1210d5a CH |
849 | |
850 | blk_add_timer(req); | |
87ee7b11 | 851 | } |
5b3f25fc | 852 | |
12f5b931 | 853 | static bool blk_mq_req_expired(struct request *rq, unsigned long *next) |
81481eb4 | 854 | { |
12f5b931 | 855 | unsigned long deadline; |
87ee7b11 | 856 | |
12f5b931 KB |
857 | if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) |
858 | return false; | |
da661267 CH |
859 | if (rq->rq_flags & RQF_TIMED_OUT) |
860 | return false; | |
a7af0af3 | 861 | |
079076b3 | 862 | deadline = READ_ONCE(rq->deadline); |
12f5b931 KB |
863 | if (time_after_eq(jiffies, deadline)) |
864 | return true; | |
a7af0af3 | 865 | |
12f5b931 KB |
866 | if (*next == 0) |
867 | *next = deadline; | |
868 | else if (time_after(*next, deadline)) | |
869 | *next = deadline; | |
870 | return false; | |
87ee7b11 JA |
871 | } |
872 | ||
7baa8572 | 873 | static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
1d9bd516 TH |
874 | struct request *rq, void *priv, bool reserved) |
875 | { | |
12f5b931 KB |
876 | unsigned long *next = priv; |
877 | ||
878 | /* | |
879 | * Just do a quick check if it is expired before locking the request in | |
880 | * so we're not unnecessarilly synchronizing across CPUs. | |
881 | */ | |
882 | if (!blk_mq_req_expired(rq, next)) | |
7baa8572 | 883 | return true; |
12f5b931 KB |
884 | |
885 | /* | |
886 | * We have reason to believe the request may be expired. Take a | |
887 | * reference on the request to lock this request lifetime into its | |
888 | * currently allocated context to prevent it from being reallocated in | |
889 | * the event the completion by-passes this timeout handler. | |
890 | * | |
891 | * If the reference was already released, then the driver beat the | |
892 | * timeout handler to posting a natural completion. | |
893 | */ | |
894 | if (!refcount_inc_not_zero(&rq->ref)) | |
7baa8572 | 895 | return true; |
12f5b931 | 896 | |
1d9bd516 | 897 | /* |
12f5b931 KB |
898 | * The request is now locked and cannot be reallocated underneath the |
899 | * timeout handler's processing. Re-verify this exact request is truly | |
900 | * expired; if it is not expired, then the request was completed and | |
901 | * reallocated as a new request. | |
1d9bd516 | 902 | */ |
12f5b931 | 903 | if (blk_mq_req_expired(rq, next)) |
1d9bd516 | 904 | blk_mq_rq_timed_out(rq, reserved); |
12f5b931 KB |
905 | if (refcount_dec_and_test(&rq->ref)) |
906 | __blk_mq_free_request(rq); | |
7baa8572 JA |
907 | |
908 | return true; | |
1d9bd516 TH |
909 | } |
910 | ||
287922eb | 911 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 912 | { |
287922eb CH |
913 | struct request_queue *q = |
914 | container_of(work, struct request_queue, timeout_work); | |
12f5b931 | 915 | unsigned long next = 0; |
1d9bd516 | 916 | struct blk_mq_hw_ctx *hctx; |
81481eb4 | 917 | int i; |
320ae51f | 918 | |
71f79fb3 GKB |
919 | /* A deadlock might occur if a request is stuck requiring a |
920 | * timeout at the same time a queue freeze is waiting | |
921 | * completion, since the timeout code would not be able to | |
922 | * acquire the queue reference here. | |
923 | * | |
924 | * That's why we don't use blk_queue_enter here; instead, we use | |
925 | * percpu_ref_tryget directly, because we need to be able to | |
926 | * obtain a reference even in the short window between the queue | |
927 | * starting to freeze, by dropping the first reference in | |
1671d522 | 928 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
929 | * consumed, marked by the instant q_usage_counter reaches |
930 | * zero. | |
931 | */ | |
932 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
933 | return; |
934 | ||
12f5b931 | 935 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next); |
320ae51f | 936 | |
12f5b931 KB |
937 | if (next != 0) { |
938 | mod_timer(&q->timeout, next); | |
0d2602ca | 939 | } else { |
fcd36c36 BVA |
940 | /* |
941 | * Request timeouts are handled as a forward rolling timer. If | |
942 | * we end up here it means that no requests are pending and | |
943 | * also that no request has been pending for a while. Mark | |
944 | * each hctx as idle. | |
945 | */ | |
f054b56c ML |
946 | queue_for_each_hw_ctx(q, hctx, i) { |
947 | /* the hctx may be unmapped, so check it here */ | |
948 | if (blk_mq_hw_queue_mapped(hctx)) | |
949 | blk_mq_tag_idle(hctx); | |
950 | } | |
0d2602ca | 951 | } |
287922eb | 952 | blk_queue_exit(q); |
320ae51f JA |
953 | } |
954 | ||
88459642 OS |
955 | struct flush_busy_ctx_data { |
956 | struct blk_mq_hw_ctx *hctx; | |
957 | struct list_head *list; | |
958 | }; | |
959 | ||
960 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
961 | { | |
962 | struct flush_busy_ctx_data *flush_data = data; | |
963 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
964 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 965 | enum hctx_type type = hctx->type; |
88459642 | 966 | |
88459642 | 967 | spin_lock(&ctx->lock); |
c16d6b5a | 968 | list_splice_tail_init(&ctx->rq_lists[type], flush_data->list); |
e9a99a63 | 969 | sbitmap_clear_bit(sb, bitnr); |
88459642 OS |
970 | spin_unlock(&ctx->lock); |
971 | return true; | |
972 | } | |
973 | ||
1429d7c9 JA |
974 | /* |
975 | * Process software queues that have been marked busy, splicing them | |
976 | * to the for-dispatch | |
977 | */ | |
2c3ad667 | 978 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 979 | { |
88459642 OS |
980 | struct flush_busy_ctx_data data = { |
981 | .hctx = hctx, | |
982 | .list = list, | |
983 | }; | |
1429d7c9 | 984 | |
88459642 | 985 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 986 | } |
2c3ad667 | 987 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 988 | |
b347689f ML |
989 | struct dispatch_rq_data { |
990 | struct blk_mq_hw_ctx *hctx; | |
991 | struct request *rq; | |
992 | }; | |
993 | ||
994 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
995 | void *data) | |
996 | { | |
997 | struct dispatch_rq_data *dispatch_data = data; | |
998 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
999 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1000 | enum hctx_type type = hctx->type; |
b347689f ML |
1001 | |
1002 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
1003 | if (!list_empty(&ctx->rq_lists[type])) { |
1004 | dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next); | |
b347689f | 1005 | list_del_init(&dispatch_data->rq->queuelist); |
c16d6b5a | 1006 | if (list_empty(&ctx->rq_lists[type])) |
b347689f ML |
1007 | sbitmap_clear_bit(sb, bitnr); |
1008 | } | |
1009 | spin_unlock(&ctx->lock); | |
1010 | ||
1011 | return !dispatch_data->rq; | |
1012 | } | |
1013 | ||
1014 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
1015 | struct blk_mq_ctx *start) | |
1016 | { | |
f31967f0 | 1017 | unsigned off = start ? start->index_hw[hctx->type] : 0; |
b347689f ML |
1018 | struct dispatch_rq_data data = { |
1019 | .hctx = hctx, | |
1020 | .rq = NULL, | |
1021 | }; | |
1022 | ||
1023 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1024 | dispatch_rq_from_ctx, &data); | |
1025 | ||
1026 | return data.rq; | |
1027 | } | |
1028 | ||
703fd1c0 JA |
1029 | static inline unsigned int queued_to_index(unsigned int queued) |
1030 | { | |
1031 | if (!queued) | |
1032 | return 0; | |
1429d7c9 | 1033 | |
703fd1c0 | 1034 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
1035 | } |
1036 | ||
8ab6bb9e | 1037 | bool blk_mq_get_driver_tag(struct request *rq) |
bd166ef1 JA |
1038 | { |
1039 | struct blk_mq_alloc_data data = { | |
1040 | .q = rq->q, | |
ea4f995e | 1041 | .hctx = rq->mq_hctx, |
8ab6bb9e | 1042 | .flags = BLK_MQ_REQ_NOWAIT, |
f9afca4d | 1043 | .cmd_flags = rq->cmd_flags, |
bd166ef1 | 1044 | }; |
d263ed99 | 1045 | bool shared; |
5feeacdd | 1046 | |
81380ca1 OS |
1047 | if (rq->tag != -1) |
1048 | goto done; | |
bd166ef1 | 1049 | |
415b806d SG |
1050 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) |
1051 | data.flags |= BLK_MQ_REQ_RESERVED; | |
1052 | ||
d263ed99 | 1053 | shared = blk_mq_tag_busy(data.hctx); |
bd166ef1 JA |
1054 | rq->tag = blk_mq_get_tag(&data); |
1055 | if (rq->tag >= 0) { | |
d263ed99 | 1056 | if (shared) { |
200e86b3 JA |
1057 | rq->rq_flags |= RQF_MQ_INFLIGHT; |
1058 | atomic_inc(&data.hctx->nr_active); | |
1059 | } | |
bd166ef1 | 1060 | data.hctx->tags->rqs[rq->tag] = rq; |
bd166ef1 JA |
1061 | } |
1062 | ||
81380ca1 | 1063 | done: |
81380ca1 | 1064 | return rq->tag != -1; |
bd166ef1 JA |
1065 | } |
1066 | ||
eb619fdb JA |
1067 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1068 | int flags, void *key) | |
da55f2cc OS |
1069 | { |
1070 | struct blk_mq_hw_ctx *hctx; | |
1071 | ||
1072 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1073 | ||
5815839b | 1074 | spin_lock(&hctx->dispatch_wait_lock); |
e8618575 JA |
1075 | if (!list_empty(&wait->entry)) { |
1076 | struct sbitmap_queue *sbq; | |
1077 | ||
1078 | list_del_init(&wait->entry); | |
1079 | sbq = &hctx->tags->bitmap_tags; | |
1080 | atomic_dec(&sbq->ws_active); | |
1081 | } | |
5815839b ML |
1082 | spin_unlock(&hctx->dispatch_wait_lock); |
1083 | ||
da55f2cc OS |
1084 | blk_mq_run_hw_queue(hctx, true); |
1085 | return 1; | |
1086 | } | |
1087 | ||
f906a6a0 JA |
1088 | /* |
1089 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1090 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1091 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1092 | * marking us as waiting. |
1093 | */ | |
2278d69f | 1094 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, |
f906a6a0 | 1095 | struct request *rq) |
da55f2cc | 1096 | { |
e8618575 | 1097 | struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags; |
5815839b | 1098 | struct wait_queue_head *wq; |
f906a6a0 JA |
1099 | wait_queue_entry_t *wait; |
1100 | bool ret; | |
da55f2cc | 1101 | |
2278d69f | 1102 | if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) { |
684b7324 | 1103 | blk_mq_sched_mark_restart_hctx(hctx); |
f906a6a0 | 1104 | |
c27d53fb BVA |
1105 | /* |
1106 | * It's possible that a tag was freed in the window between the | |
1107 | * allocation failure and adding the hardware queue to the wait | |
1108 | * queue. | |
1109 | * | |
1110 | * Don't clear RESTART here, someone else could have set it. | |
1111 | * At most this will cost an extra queue run. | |
1112 | */ | |
8ab6bb9e | 1113 | return blk_mq_get_driver_tag(rq); |
eb619fdb | 1114 | } |
eb619fdb | 1115 | |
2278d69f | 1116 | wait = &hctx->dispatch_wait; |
c27d53fb BVA |
1117 | if (!list_empty_careful(&wait->entry)) |
1118 | return false; | |
1119 | ||
e8618575 | 1120 | wq = &bt_wait_ptr(sbq, hctx)->wait; |
5815839b ML |
1121 | |
1122 | spin_lock_irq(&wq->lock); | |
1123 | spin_lock(&hctx->dispatch_wait_lock); | |
c27d53fb | 1124 | if (!list_empty(&wait->entry)) { |
5815839b ML |
1125 | spin_unlock(&hctx->dispatch_wait_lock); |
1126 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1127 | return false; |
eb619fdb JA |
1128 | } |
1129 | ||
e8618575 | 1130 | atomic_inc(&sbq->ws_active); |
5815839b ML |
1131 | wait->flags &= ~WQ_FLAG_EXCLUSIVE; |
1132 | __add_wait_queue(wq, wait); | |
c27d53fb | 1133 | |
da55f2cc | 1134 | /* |
eb619fdb JA |
1135 | * It's possible that a tag was freed in the window between the |
1136 | * allocation failure and adding the hardware queue to the wait | |
1137 | * queue. | |
da55f2cc | 1138 | */ |
8ab6bb9e | 1139 | ret = blk_mq_get_driver_tag(rq); |
c27d53fb | 1140 | if (!ret) { |
5815839b ML |
1141 | spin_unlock(&hctx->dispatch_wait_lock); |
1142 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1143 | return false; |
eb619fdb | 1144 | } |
c27d53fb BVA |
1145 | |
1146 | /* | |
1147 | * We got a tag, remove ourselves from the wait queue to ensure | |
1148 | * someone else gets the wakeup. | |
1149 | */ | |
c27d53fb | 1150 | list_del_init(&wait->entry); |
e8618575 | 1151 | atomic_dec(&sbq->ws_active); |
5815839b ML |
1152 | spin_unlock(&hctx->dispatch_wait_lock); |
1153 | spin_unlock_irq(&wq->lock); | |
c27d53fb BVA |
1154 | |
1155 | return true; | |
da55f2cc OS |
1156 | } |
1157 | ||
6e768717 ML |
1158 | #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 |
1159 | #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 | |
1160 | /* | |
1161 | * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): | |
1162 | * - EWMA is one simple way to compute running average value | |
1163 | * - weight(7/8 and 1/8) is applied so that it can decrease exponentially | |
1164 | * - take 4 as factor for avoiding to get too small(0) result, and this | |
1165 | * factor doesn't matter because EWMA decreases exponentially | |
1166 | */ | |
1167 | static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) | |
1168 | { | |
1169 | unsigned int ewma; | |
1170 | ||
1171 | if (hctx->queue->elevator) | |
1172 | return; | |
1173 | ||
1174 | ewma = hctx->dispatch_busy; | |
1175 | ||
1176 | if (!ewma && !busy) | |
1177 | return; | |
1178 | ||
1179 | ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; | |
1180 | if (busy) | |
1181 | ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; | |
1182 | ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; | |
1183 | ||
1184 | hctx->dispatch_busy = ewma; | |
1185 | } | |
1186 | ||
86ff7c2a ML |
1187 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
1188 | ||
1f57f8d4 JA |
1189 | /* |
1190 | * Returns true if we did some work AND can potentially do more. | |
1191 | */ | |
de148297 | 1192 | bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list, |
eb619fdb | 1193 | bool got_budget) |
320ae51f | 1194 | { |
81380ca1 | 1195 | struct blk_mq_hw_ctx *hctx; |
6d6f167c | 1196 | struct request *rq, *nxt; |
eb619fdb | 1197 | bool no_tag = false; |
fc17b653 | 1198 | int errors, queued; |
86ff7c2a | 1199 | blk_status_t ret = BLK_STS_OK; |
320ae51f | 1200 | |
81380ca1 OS |
1201 | if (list_empty(list)) |
1202 | return false; | |
1203 | ||
de148297 ML |
1204 | WARN_ON(!list_is_singular(list) && got_budget); |
1205 | ||
320ae51f JA |
1206 | /* |
1207 | * Now process all the entries, sending them to the driver. | |
1208 | */ | |
93efe981 | 1209 | errors = queued = 0; |
81380ca1 | 1210 | do { |
74c45052 | 1211 | struct blk_mq_queue_data bd; |
320ae51f | 1212 | |
f04c3df3 | 1213 | rq = list_first_entry(list, struct request, queuelist); |
0bca799b | 1214 | |
ea4f995e | 1215 | hctx = rq->mq_hctx; |
0bca799b ML |
1216 | if (!got_budget && !blk_mq_get_dispatch_budget(hctx)) |
1217 | break; | |
1218 | ||
8ab6bb9e | 1219 | if (!blk_mq_get_driver_tag(rq)) { |
3c782d67 | 1220 | /* |
da55f2cc | 1221 | * The initial allocation attempt failed, so we need to |
eb619fdb JA |
1222 | * rerun the hardware queue when a tag is freed. The |
1223 | * waitqueue takes care of that. If the queue is run | |
1224 | * before we add this entry back on the dispatch list, | |
1225 | * we'll re-run it below. | |
3c782d67 | 1226 | */ |
2278d69f | 1227 | if (!blk_mq_mark_tag_wait(hctx, rq)) { |
0bca799b | 1228 | blk_mq_put_dispatch_budget(hctx); |
f906a6a0 JA |
1229 | /* |
1230 | * For non-shared tags, the RESTART check | |
1231 | * will suffice. | |
1232 | */ | |
1233 | if (hctx->flags & BLK_MQ_F_TAG_SHARED) | |
1234 | no_tag = true; | |
de148297 ML |
1235 | break; |
1236 | } | |
1237 | } | |
1238 | ||
320ae51f | 1239 | list_del_init(&rq->queuelist); |
320ae51f | 1240 | |
74c45052 | 1241 | bd.rq = rq; |
113285b4 JA |
1242 | |
1243 | /* | |
1244 | * Flag last if we have no more requests, or if we have more | |
1245 | * but can't assign a driver tag to it. | |
1246 | */ | |
1247 | if (list_empty(list)) | |
1248 | bd.last = true; | |
1249 | else { | |
113285b4 | 1250 | nxt = list_first_entry(list, struct request, queuelist); |
8ab6bb9e | 1251 | bd.last = !blk_mq_get_driver_tag(nxt); |
113285b4 | 1252 | } |
74c45052 JA |
1253 | |
1254 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
86ff7c2a | 1255 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) { |
6d6f167c JW |
1256 | /* |
1257 | * If an I/O scheduler has been configured and we got a | |
ff821d27 JA |
1258 | * driver tag for the next request already, free it |
1259 | * again. | |
6d6f167c JW |
1260 | */ |
1261 | if (!list_empty(list)) { | |
1262 | nxt = list_first_entry(list, struct request, queuelist); | |
1263 | blk_mq_put_driver_tag(nxt); | |
1264 | } | |
f04c3df3 | 1265 | list_add(&rq->queuelist, list); |
ed0791b2 | 1266 | __blk_mq_requeue_request(rq); |
320ae51f | 1267 | break; |
fc17b653 CH |
1268 | } |
1269 | ||
1270 | if (unlikely(ret != BLK_STS_OK)) { | |
93efe981 | 1271 | errors++; |
2a842aca | 1272 | blk_mq_end_request(rq, BLK_STS_IOERR); |
fc17b653 | 1273 | continue; |
320ae51f JA |
1274 | } |
1275 | ||
fc17b653 | 1276 | queued++; |
81380ca1 | 1277 | } while (!list_empty(list)); |
320ae51f | 1278 | |
703fd1c0 | 1279 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
1280 | |
1281 | /* | |
1282 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1283 | * that is where we will continue on next queue run. | |
1284 | */ | |
f04c3df3 | 1285 | if (!list_empty(list)) { |
86ff7c2a ML |
1286 | bool needs_restart; |
1287 | ||
d666ba98 JA |
1288 | /* |
1289 | * If we didn't flush the entire list, we could have told | |
1290 | * the driver there was more coming, but that turned out to | |
1291 | * be a lie. | |
1292 | */ | |
1293 | if (q->mq_ops->commit_rqs) | |
1294 | q->mq_ops->commit_rqs(hctx); | |
1295 | ||
320ae51f | 1296 | spin_lock(&hctx->lock); |
c13660a0 | 1297 | list_splice_init(list, &hctx->dispatch); |
320ae51f | 1298 | spin_unlock(&hctx->lock); |
f04c3df3 | 1299 | |
9ba52e58 | 1300 | /* |
710c785f BVA |
1301 | * If SCHED_RESTART was set by the caller of this function and |
1302 | * it is no longer set that means that it was cleared by another | |
1303 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1304 | * |
eb619fdb JA |
1305 | * If 'no_tag' is set, that means that we failed getting |
1306 | * a driver tag with an I/O scheduler attached. If our dispatch | |
1307 | * waitqueue is no longer active, ensure that we run the queue | |
1308 | * AFTER adding our entries back to the list. | |
bd166ef1 | 1309 | * |
710c785f BVA |
1310 | * If no I/O scheduler has been configured it is possible that |
1311 | * the hardware queue got stopped and restarted before requests | |
1312 | * were pushed back onto the dispatch list. Rerun the queue to | |
1313 | * avoid starvation. Notes: | |
1314 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1315 | * been stopped before rerunning a queue. | |
1316 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1317 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1318 | * and dm-rq. |
86ff7c2a ML |
1319 | * |
1320 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
1321 | * bit is set, run queue after a delay to avoid IO stalls | |
1322 | * that could otherwise occur if the queue is idle. | |
bd166ef1 | 1323 | */ |
86ff7c2a ML |
1324 | needs_restart = blk_mq_sched_needs_restart(hctx); |
1325 | if (!needs_restart || | |
eb619fdb | 1326 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
bd166ef1 | 1327 | blk_mq_run_hw_queue(hctx, true); |
86ff7c2a ML |
1328 | else if (needs_restart && (ret == BLK_STS_RESOURCE)) |
1329 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); | |
1f57f8d4 | 1330 | |
6e768717 | 1331 | blk_mq_update_dispatch_busy(hctx, true); |
1f57f8d4 | 1332 | return false; |
6e768717 ML |
1333 | } else |
1334 | blk_mq_update_dispatch_busy(hctx, false); | |
f04c3df3 | 1335 | |
1f57f8d4 JA |
1336 | /* |
1337 | * If the host/device is unable to accept more work, inform the | |
1338 | * caller of that. | |
1339 | */ | |
1340 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) | |
1341 | return false; | |
1342 | ||
93efe981 | 1343 | return (queued + errors) != 0; |
f04c3df3 JA |
1344 | } |
1345 | ||
6a83e74d BVA |
1346 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1347 | { | |
1348 | int srcu_idx; | |
1349 | ||
b7a71e66 JA |
1350 | /* |
1351 | * We should be running this queue from one of the CPUs that | |
1352 | * are mapped to it. | |
7df938fb ML |
1353 | * |
1354 | * There are at least two related races now between setting | |
1355 | * hctx->next_cpu from blk_mq_hctx_next_cpu() and running | |
1356 | * __blk_mq_run_hw_queue(): | |
1357 | * | |
1358 | * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(), | |
1359 | * but later it becomes online, then this warning is harmless | |
1360 | * at all | |
1361 | * | |
1362 | * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(), | |
1363 | * but later it becomes offline, then the warning can't be | |
1364 | * triggered, and we depend on blk-mq timeout handler to | |
1365 | * handle dispatched requests to this hctx | |
b7a71e66 | 1366 | */ |
7df938fb ML |
1367 | if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && |
1368 | cpu_online(hctx->next_cpu)) { | |
1369 | printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n", | |
1370 | raw_smp_processor_id(), | |
1371 | cpumask_empty(hctx->cpumask) ? "inactive": "active"); | |
1372 | dump_stack(); | |
1373 | } | |
6a83e74d | 1374 | |
b7a71e66 JA |
1375 | /* |
1376 | * We can't run the queue inline with ints disabled. Ensure that | |
1377 | * we catch bad users of this early. | |
1378 | */ | |
1379 | WARN_ON_ONCE(in_interrupt()); | |
1380 | ||
04ced159 | 1381 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1382 | |
04ced159 JA |
1383 | hctx_lock(hctx, &srcu_idx); |
1384 | blk_mq_sched_dispatch_requests(hctx); | |
1385 | hctx_unlock(hctx, srcu_idx); | |
6a83e74d BVA |
1386 | } |
1387 | ||
f82ddf19 ML |
1388 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
1389 | { | |
1390 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
1391 | ||
1392 | if (cpu >= nr_cpu_ids) | |
1393 | cpu = cpumask_first(hctx->cpumask); | |
1394 | return cpu; | |
1395 | } | |
1396 | ||
506e931f JA |
1397 | /* |
1398 | * It'd be great if the workqueue API had a way to pass | |
1399 | * in a mask and had some smarts for more clever placement. | |
1400 | * For now we just round-robin here, switching for every | |
1401 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1402 | */ | |
1403 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1404 | { | |
7bed4595 | 1405 | bool tried = false; |
476f8c98 | 1406 | int next_cpu = hctx->next_cpu; |
7bed4595 | 1407 | |
b657d7e6 CH |
1408 | if (hctx->queue->nr_hw_queues == 1) |
1409 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1410 | |
1411 | if (--hctx->next_cpu_batch <= 0) { | |
7bed4595 | 1412 | select_cpu: |
476f8c98 | 1413 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
20e4d813 | 1414 | cpu_online_mask); |
506e931f | 1415 | if (next_cpu >= nr_cpu_ids) |
f82ddf19 | 1416 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
1417 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
1418 | } | |
1419 | ||
7bed4595 ML |
1420 | /* |
1421 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
1422 | * and it should only happen in the path of handling CPU DEAD. | |
1423 | */ | |
476f8c98 | 1424 | if (!cpu_online(next_cpu)) { |
7bed4595 ML |
1425 | if (!tried) { |
1426 | tried = true; | |
1427 | goto select_cpu; | |
1428 | } | |
1429 | ||
1430 | /* | |
1431 | * Make sure to re-select CPU next time once after CPUs | |
1432 | * in hctx->cpumask become online again. | |
1433 | */ | |
476f8c98 | 1434 | hctx->next_cpu = next_cpu; |
7bed4595 ML |
1435 | hctx->next_cpu_batch = 1; |
1436 | return WORK_CPU_UNBOUND; | |
1437 | } | |
476f8c98 ML |
1438 | |
1439 | hctx->next_cpu = next_cpu; | |
1440 | return next_cpu; | |
506e931f JA |
1441 | } |
1442 | ||
7587a5ae BVA |
1443 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1444 | unsigned long msecs) | |
320ae51f | 1445 | { |
5435c023 | 1446 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f JA |
1447 | return; |
1448 | ||
1b792f2f | 1449 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1450 | int cpu = get_cpu(); |
1451 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1452 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1453 | put_cpu(); |
398205b8 PB |
1454 | return; |
1455 | } | |
e4043dcf | 1456 | |
2a90d4aa | 1457 | put_cpu(); |
e4043dcf | 1458 | } |
398205b8 | 1459 | |
ae943d20 BVA |
1460 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
1461 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
1462 | } |
1463 | ||
1464 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1465 | { | |
1466 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1467 | } | |
1468 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1469 | ||
79f720a7 | 1470 | bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 1471 | { |
24f5a90f ML |
1472 | int srcu_idx; |
1473 | bool need_run; | |
1474 | ||
1475 | /* | |
1476 | * When queue is quiesced, we may be switching io scheduler, or | |
1477 | * updating nr_hw_queues, or other things, and we can't run queue | |
1478 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
1479 | * | |
1480 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
1481 | * quiesced. | |
1482 | */ | |
04ced159 JA |
1483 | hctx_lock(hctx, &srcu_idx); |
1484 | need_run = !blk_queue_quiesced(hctx->queue) && | |
1485 | blk_mq_hctx_has_pending(hctx); | |
1486 | hctx_unlock(hctx, srcu_idx); | |
24f5a90f ML |
1487 | |
1488 | if (need_run) { | |
79f720a7 JA |
1489 | __blk_mq_delay_run_hw_queue(hctx, async, 0); |
1490 | return true; | |
1491 | } | |
1492 | ||
1493 | return false; | |
320ae51f | 1494 | } |
5b727272 | 1495 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1496 | |
b94ec296 | 1497 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1498 | { |
1499 | struct blk_mq_hw_ctx *hctx; | |
1500 | int i; | |
1501 | ||
1502 | queue_for_each_hw_ctx(q, hctx, i) { | |
79f720a7 | 1503 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1504 | continue; |
1505 | ||
b94ec296 | 1506 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1507 | } |
1508 | } | |
b94ec296 | 1509 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1510 | |
fd001443 BVA |
1511 | /** |
1512 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1513 | * @q: request queue. | |
1514 | * | |
1515 | * The caller is responsible for serializing this function against | |
1516 | * blk_mq_{start,stop}_hw_queue(). | |
1517 | */ | |
1518 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1519 | { | |
1520 | struct blk_mq_hw_ctx *hctx; | |
1521 | int i; | |
1522 | ||
1523 | queue_for_each_hw_ctx(q, hctx, i) | |
1524 | if (blk_mq_hctx_stopped(hctx)) | |
1525 | return true; | |
1526 | ||
1527 | return false; | |
1528 | } | |
1529 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1530 | ||
39a70c76 ML |
1531 | /* |
1532 | * This function is often used for pausing .queue_rq() by driver when | |
1533 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1534 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1535 | * |
1536 | * We do not guarantee that dispatch can be drained or blocked | |
1537 | * after blk_mq_stop_hw_queue() returns. Please use | |
1538 | * blk_mq_quiesce_queue() for that requirement. | |
1539 | */ | |
2719aa21 JA |
1540 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1541 | { | |
641a9ed6 | 1542 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 1543 | |
641a9ed6 | 1544 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 1545 | } |
641a9ed6 | 1546 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 1547 | |
39a70c76 ML |
1548 | /* |
1549 | * This function is often used for pausing .queue_rq() by driver when | |
1550 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1551 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1552 | * |
1553 | * We do not guarantee that dispatch can be drained or blocked | |
1554 | * after blk_mq_stop_hw_queues() returns. Please use | |
1555 | * blk_mq_quiesce_queue() for that requirement. | |
1556 | */ | |
2719aa21 JA |
1557 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1558 | { | |
641a9ed6 ML |
1559 | struct blk_mq_hw_ctx *hctx; |
1560 | int i; | |
1561 | ||
1562 | queue_for_each_hw_ctx(q, hctx, i) | |
1563 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
1564 | } |
1565 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1566 | ||
320ae51f JA |
1567 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1568 | { | |
1569 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1570 | |
0ffbce80 | 1571 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1572 | } |
1573 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1574 | ||
2f268556 CH |
1575 | void blk_mq_start_hw_queues(struct request_queue *q) |
1576 | { | |
1577 | struct blk_mq_hw_ctx *hctx; | |
1578 | int i; | |
1579 | ||
1580 | queue_for_each_hw_ctx(q, hctx, i) | |
1581 | blk_mq_start_hw_queue(hctx); | |
1582 | } | |
1583 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1584 | ||
ae911c5e JA |
1585 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1586 | { | |
1587 | if (!blk_mq_hctx_stopped(hctx)) | |
1588 | return; | |
1589 | ||
1590 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1591 | blk_mq_run_hw_queue(hctx, async); | |
1592 | } | |
1593 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1594 | ||
1b4a3258 | 1595 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1596 | { |
1597 | struct blk_mq_hw_ctx *hctx; | |
1598 | int i; | |
1599 | ||
ae911c5e JA |
1600 | queue_for_each_hw_ctx(q, hctx, i) |
1601 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1602 | } |
1603 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1604 | ||
70f4db63 | 1605 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1606 | { |
1607 | struct blk_mq_hw_ctx *hctx; | |
1608 | ||
9f993737 | 1609 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 1610 | |
21c6e939 | 1611 | /* |
15fe8a90 | 1612 | * If we are stopped, don't run the queue. |
21c6e939 | 1613 | */ |
15fe8a90 | 1614 | if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
0196d6b4 | 1615 | return; |
7587a5ae BVA |
1616 | |
1617 | __blk_mq_run_hw_queue(hctx); | |
1618 | } | |
1619 | ||
cfd0c552 | 1620 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1621 | struct request *rq, |
1622 | bool at_head) | |
320ae51f | 1623 | { |
e57690fe | 1624 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
c16d6b5a | 1625 | enum hctx_type type = hctx->type; |
e57690fe | 1626 | |
7b607814 BVA |
1627 | lockdep_assert_held(&ctx->lock); |
1628 | ||
01b983c9 JA |
1629 | trace_block_rq_insert(hctx->queue, rq); |
1630 | ||
72a0a36e | 1631 | if (at_head) |
c16d6b5a | 1632 | list_add(&rq->queuelist, &ctx->rq_lists[type]); |
72a0a36e | 1633 | else |
c16d6b5a | 1634 | list_add_tail(&rq->queuelist, &ctx->rq_lists[type]); |
cfd0c552 | 1635 | } |
4bb659b1 | 1636 | |
2c3ad667 JA |
1637 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1638 | bool at_head) | |
cfd0c552 ML |
1639 | { |
1640 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1641 | ||
7b607814 BVA |
1642 | lockdep_assert_held(&ctx->lock); |
1643 | ||
e57690fe | 1644 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1645 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1646 | } |
1647 | ||
157f377b JA |
1648 | /* |
1649 | * Should only be used carefully, when the caller knows we want to | |
1650 | * bypass a potential IO scheduler on the target device. | |
1651 | */ | |
b0850297 | 1652 | void blk_mq_request_bypass_insert(struct request *rq, bool run_queue) |
157f377b | 1653 | { |
ea4f995e | 1654 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
157f377b JA |
1655 | |
1656 | spin_lock(&hctx->lock); | |
1657 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
1658 | spin_unlock(&hctx->lock); | |
1659 | ||
b0850297 ML |
1660 | if (run_queue) |
1661 | blk_mq_run_hw_queue(hctx, false); | |
157f377b JA |
1662 | } |
1663 | ||
bd166ef1 JA |
1664 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1665 | struct list_head *list) | |
320ae51f JA |
1666 | |
1667 | { | |
3f0cedc7 | 1668 | struct request *rq; |
c16d6b5a | 1669 | enum hctx_type type = hctx->type; |
3f0cedc7 | 1670 | |
320ae51f JA |
1671 | /* |
1672 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1673 | * offline now | |
1674 | */ | |
3f0cedc7 | 1675 | list_for_each_entry(rq, list, queuelist) { |
e57690fe | 1676 | BUG_ON(rq->mq_ctx != ctx); |
3f0cedc7 | 1677 | trace_block_rq_insert(hctx->queue, rq); |
320ae51f | 1678 | } |
3f0cedc7 ML |
1679 | |
1680 | spin_lock(&ctx->lock); | |
c16d6b5a | 1681 | list_splice_tail_init(list, &ctx->rq_lists[type]); |
cfd0c552 | 1682 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1683 | spin_unlock(&ctx->lock); |
320ae51f JA |
1684 | } |
1685 | ||
3110fc79 | 1686 | static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b) |
320ae51f JA |
1687 | { |
1688 | struct request *rqa = container_of(a, struct request, queuelist); | |
1689 | struct request *rqb = container_of(b, struct request, queuelist); | |
1690 | ||
3110fc79 JA |
1691 | if (rqa->mq_ctx < rqb->mq_ctx) |
1692 | return -1; | |
1693 | else if (rqa->mq_ctx > rqb->mq_ctx) | |
1694 | return 1; | |
1695 | else if (rqa->mq_hctx < rqb->mq_hctx) | |
1696 | return -1; | |
1697 | else if (rqa->mq_hctx > rqb->mq_hctx) | |
1698 | return 1; | |
1699 | ||
1700 | return blk_rq_pos(rqa) > blk_rq_pos(rqb); | |
320ae51f JA |
1701 | } |
1702 | ||
1703 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1704 | { | |
67cae4c9 | 1705 | struct blk_mq_hw_ctx *this_hctx; |
320ae51f JA |
1706 | struct blk_mq_ctx *this_ctx; |
1707 | struct request_queue *this_q; | |
1708 | struct request *rq; | |
1709 | LIST_HEAD(list); | |
67cae4c9 | 1710 | LIST_HEAD(rq_list); |
320ae51f JA |
1711 | unsigned int depth; |
1712 | ||
1713 | list_splice_init(&plug->mq_list, &list); | |
1714 | ||
ce5b009c JA |
1715 | if (plug->rq_count > 2 && plug->multiple_queues) |
1716 | list_sort(NULL, &list, plug_rq_cmp); | |
320ae51f | 1717 | |
bcc816df DZ |
1718 | plug->rq_count = 0; |
1719 | ||
320ae51f | 1720 | this_q = NULL; |
67cae4c9 | 1721 | this_hctx = NULL; |
320ae51f JA |
1722 | this_ctx = NULL; |
1723 | depth = 0; | |
1724 | ||
1725 | while (!list_empty(&list)) { | |
1726 | rq = list_entry_rq(list.next); | |
1727 | list_del_init(&rq->queuelist); | |
1728 | BUG_ON(!rq->q); | |
67cae4c9 JA |
1729 | if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) { |
1730 | if (this_hctx) { | |
587562d0 | 1731 | trace_block_unplug(this_q, depth, !from_schedule); |
67cae4c9 JA |
1732 | blk_mq_sched_insert_requests(this_hctx, this_ctx, |
1733 | &rq_list, | |
bd166ef1 | 1734 | from_schedule); |
320ae51f JA |
1735 | } |
1736 | ||
320ae51f | 1737 | this_q = rq->q; |
67cae4c9 JA |
1738 | this_ctx = rq->mq_ctx; |
1739 | this_hctx = rq->mq_hctx; | |
320ae51f JA |
1740 | depth = 0; |
1741 | } | |
1742 | ||
1743 | depth++; | |
67cae4c9 | 1744 | list_add_tail(&rq->queuelist, &rq_list); |
320ae51f JA |
1745 | } |
1746 | ||
1747 | /* | |
67cae4c9 JA |
1748 | * If 'this_hctx' is set, we know we have entries to complete |
1749 | * on 'rq_list'. Do those. | |
320ae51f | 1750 | */ |
67cae4c9 | 1751 | if (this_hctx) { |
587562d0 | 1752 | trace_block_unplug(this_q, depth, !from_schedule); |
67cae4c9 | 1753 | blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list, |
bd166ef1 | 1754 | from_schedule); |
320ae51f JA |
1755 | } |
1756 | } | |
1757 | ||
1758 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1759 | { | |
da8d7f07 | 1760 | blk_init_request_from_bio(rq, bio); |
4b570521 | 1761 | |
6e85eaf3 | 1762 | blk_account_io_start(rq, true); |
320ae51f JA |
1763 | } |
1764 | ||
0f95549c MS |
1765 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
1766 | struct request *rq, | |
be94f058 | 1767 | blk_qc_t *cookie, bool last) |
f984df1f | 1768 | { |
f984df1f | 1769 | struct request_queue *q = rq->q; |
f984df1f SL |
1770 | struct blk_mq_queue_data bd = { |
1771 | .rq = rq, | |
be94f058 | 1772 | .last = last, |
f984df1f | 1773 | }; |
bd166ef1 | 1774 | blk_qc_t new_cookie; |
f06345ad | 1775 | blk_status_t ret; |
0f95549c MS |
1776 | |
1777 | new_cookie = request_to_qc_t(hctx, rq); | |
1778 | ||
1779 | /* | |
1780 | * For OK queue, we are done. For error, caller may kill it. | |
1781 | * Any other error (busy), just add it to our list as we | |
1782 | * previously would have done. | |
1783 | */ | |
1784 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
1785 | switch (ret) { | |
1786 | case BLK_STS_OK: | |
6ce3dd6e | 1787 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
1788 | *cookie = new_cookie; |
1789 | break; | |
1790 | case BLK_STS_RESOURCE: | |
86ff7c2a | 1791 | case BLK_STS_DEV_RESOURCE: |
6ce3dd6e | 1792 | blk_mq_update_dispatch_busy(hctx, true); |
0f95549c MS |
1793 | __blk_mq_requeue_request(rq); |
1794 | break; | |
1795 | default: | |
6ce3dd6e | 1796 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
1797 | *cookie = BLK_QC_T_NONE; |
1798 | break; | |
1799 | } | |
1800 | ||
1801 | return ret; | |
1802 | } | |
1803 | ||
fd9c40f6 | 1804 | static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
0f95549c | 1805 | struct request *rq, |
396eaf21 | 1806 | blk_qc_t *cookie, |
fd9c40f6 | 1807 | bool bypass_insert, bool last) |
0f95549c MS |
1808 | { |
1809 | struct request_queue *q = rq->q; | |
d964f04a ML |
1810 | bool run_queue = true; |
1811 | ||
23d4ee19 | 1812 | /* |
fd9c40f6 | 1813 | * RCU or SRCU read lock is needed before checking quiesced flag. |
23d4ee19 | 1814 | * |
fd9c40f6 BVA |
1815 | * When queue is stopped or quiesced, ignore 'bypass_insert' from |
1816 | * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller, | |
1817 | * and avoid driver to try to dispatch again. | |
23d4ee19 | 1818 | */ |
fd9c40f6 | 1819 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { |
d964f04a | 1820 | run_queue = false; |
fd9c40f6 BVA |
1821 | bypass_insert = false; |
1822 | goto insert; | |
d964f04a | 1823 | } |
f984df1f | 1824 | |
fd9c40f6 BVA |
1825 | if (q->elevator && !bypass_insert) |
1826 | goto insert; | |
2253efc8 | 1827 | |
0bca799b | 1828 | if (!blk_mq_get_dispatch_budget(hctx)) |
fd9c40f6 | 1829 | goto insert; |
bd166ef1 | 1830 | |
8ab6bb9e | 1831 | if (!blk_mq_get_driver_tag(rq)) { |
0bca799b | 1832 | blk_mq_put_dispatch_budget(hctx); |
fd9c40f6 | 1833 | goto insert; |
88022d72 | 1834 | } |
de148297 | 1835 | |
fd9c40f6 BVA |
1836 | return __blk_mq_issue_directly(hctx, rq, cookie, last); |
1837 | insert: | |
1838 | if (bypass_insert) | |
1839 | return BLK_STS_RESOURCE; | |
1840 | ||
1841 | blk_mq_request_bypass_insert(rq, run_queue); | |
1842 | return BLK_STS_OK; | |
1843 | } | |
1844 | ||
1845 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, | |
1846 | struct request *rq, blk_qc_t *cookie) | |
1847 | { | |
1848 | blk_status_t ret; | |
1849 | int srcu_idx; | |
1850 | ||
1851 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); | |
1852 | ||
1853 | hctx_lock(hctx, &srcu_idx); | |
1854 | ||
1855 | ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true); | |
1856 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) | |
1857 | blk_mq_request_bypass_insert(rq, true); | |
1858 | else if (ret != BLK_STS_OK) | |
1859 | blk_mq_end_request(rq, ret); | |
1860 | ||
1861 | hctx_unlock(hctx, srcu_idx); | |
1862 | } | |
1863 | ||
1864 | blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last) | |
1865 | { | |
1866 | blk_status_t ret; | |
1867 | int srcu_idx; | |
1868 | blk_qc_t unused_cookie; | |
1869 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
1870 | ||
1871 | hctx_lock(hctx, &srcu_idx); | |
1872 | ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last); | |
04ced159 | 1873 | hctx_unlock(hctx, srcu_idx); |
7f556a44 JW |
1874 | |
1875 | return ret; | |
5eb6126e CH |
1876 | } |
1877 | ||
6ce3dd6e ML |
1878 | void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
1879 | struct list_head *list) | |
1880 | { | |
1881 | while (!list_empty(list)) { | |
fd9c40f6 | 1882 | blk_status_t ret; |
6ce3dd6e ML |
1883 | struct request *rq = list_first_entry(list, struct request, |
1884 | queuelist); | |
1885 | ||
1886 | list_del_init(&rq->queuelist); | |
fd9c40f6 BVA |
1887 | ret = blk_mq_request_issue_directly(rq, list_empty(list)); |
1888 | if (ret != BLK_STS_OK) { | |
1889 | if (ret == BLK_STS_RESOURCE || | |
1890 | ret == BLK_STS_DEV_RESOURCE) { | |
1891 | blk_mq_request_bypass_insert(rq, | |
c616cbee | 1892 | list_empty(list)); |
fd9c40f6 BVA |
1893 | break; |
1894 | } | |
1895 | blk_mq_end_request(rq, ret); | |
1896 | } | |
6ce3dd6e | 1897 | } |
d666ba98 JA |
1898 | |
1899 | /* | |
1900 | * If we didn't flush the entire list, we could have told | |
1901 | * the driver there was more coming, but that turned out to | |
1902 | * be a lie. | |
1903 | */ | |
fd9c40f6 | 1904 | if (!list_empty(list) && hctx->queue->mq_ops->commit_rqs) |
d666ba98 | 1905 | hctx->queue->mq_ops->commit_rqs(hctx); |
6ce3dd6e ML |
1906 | } |
1907 | ||
ce5b009c JA |
1908 | static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq) |
1909 | { | |
1910 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1911 | plug->rq_count++; | |
1912 | if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) { | |
1913 | struct request *tmp; | |
1914 | ||
1915 | tmp = list_first_entry(&plug->mq_list, struct request, | |
1916 | queuelist); | |
1917 | if (tmp->q != rq->q) | |
1918 | plug->multiple_queues = true; | |
1919 | } | |
1920 | } | |
1921 | ||
dece1635 | 1922 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1923 | { |
ef295ecf | 1924 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1925 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
7809167d | 1926 | struct blk_mq_alloc_data data = { .flags = 0}; |
07068d5b | 1927 | struct request *rq; |
f984df1f | 1928 | struct blk_plug *plug; |
5b3f341f | 1929 | struct request *same_queue_rq = NULL; |
7b371636 | 1930 | blk_qc_t cookie; |
07068d5b JA |
1931 | |
1932 | blk_queue_bounce(q, &bio); | |
1933 | ||
af67c31f | 1934 | blk_queue_split(q, &bio); |
f36ea50c | 1935 | |
e23947bd | 1936 | if (!bio_integrity_prep(bio)) |
dece1635 | 1937 | return BLK_QC_T_NONE; |
07068d5b | 1938 | |
87c279e6 | 1939 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
5f0ed774 | 1940 | blk_attempt_plug_merge(q, bio, &same_queue_rq)) |
87c279e6 | 1941 | return BLK_QC_T_NONE; |
f984df1f | 1942 | |
bd166ef1 JA |
1943 | if (blk_mq_sched_bio_merge(q, bio)) |
1944 | return BLK_QC_T_NONE; | |
1945 | ||
d5337560 | 1946 | rq_qos_throttle(q, bio); |
87760e5e | 1947 | |
7809167d | 1948 | data.cmd_flags = bio->bi_opf; |
f9afca4d | 1949 | rq = blk_mq_get_request(q, bio, &data); |
87760e5e | 1950 | if (unlikely(!rq)) { |
c1c80384 | 1951 | rq_qos_cleanup(q, bio); |
03a07c92 GR |
1952 | if (bio->bi_opf & REQ_NOWAIT) |
1953 | bio_wouldblock_error(bio); | |
dece1635 | 1954 | return BLK_QC_T_NONE; |
87760e5e JA |
1955 | } |
1956 | ||
d6f1dda2 XW |
1957 | trace_block_getrq(q, bio, bio->bi_opf); |
1958 | ||
c1c80384 | 1959 | rq_qos_track(q, rq, bio); |
07068d5b | 1960 | |
fd2d3326 | 1961 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b | 1962 | |
f984df1f | 1963 | plug = current->plug; |
07068d5b | 1964 | if (unlikely(is_flush_fua)) { |
f984df1f | 1965 | blk_mq_put_ctx(data.ctx); |
07068d5b | 1966 | blk_mq_bio_to_request(rq, bio); |
923218f6 ML |
1967 | |
1968 | /* bypass scheduler for flush rq */ | |
1969 | blk_insert_flush(rq); | |
1970 | blk_mq_run_hw_queue(data.hctx, true); | |
b2c5d16b JA |
1971 | } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs)) { |
1972 | /* | |
1973 | * Use plugging if we have a ->commit_rqs() hook as well, as | |
1974 | * we know the driver uses bd->last in a smart fashion. | |
1975 | */ | |
5f0ed774 | 1976 | unsigned int request_count = plug->rq_count; |
600271d9 SL |
1977 | struct request *last = NULL; |
1978 | ||
b00c53e8 | 1979 | blk_mq_put_ctx(data.ctx); |
e6c4438b | 1980 | blk_mq_bio_to_request(rq, bio); |
0a6219a9 | 1981 | |
676d0607 | 1982 | if (!request_count) |
e6c4438b | 1983 | trace_block_plug(q); |
600271d9 SL |
1984 | else |
1985 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c | 1986 | |
600271d9 SL |
1987 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1988 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1989 | blk_flush_plug_list(plug, false); |
1990 | trace_block_plug(q); | |
320ae51f | 1991 | } |
b094f89c | 1992 | |
ce5b009c | 1993 | blk_add_rq_to_plug(plug, rq); |
2299722c | 1994 | } else if (plug && !blk_queue_nomerges(q)) { |
bd166ef1 | 1995 | blk_mq_bio_to_request(rq, bio); |
07068d5b | 1996 | |
07068d5b | 1997 | /* |
6a83e74d | 1998 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1999 | * Otherwise the existing request in the plug list will be |
2000 | * issued. So the plug list will have one request at most | |
2299722c CH |
2001 | * The plug list might get flushed before this. If that happens, |
2002 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 2003 | */ |
2299722c CH |
2004 | if (list_empty(&plug->mq_list)) |
2005 | same_queue_rq = NULL; | |
4711b573 | 2006 | if (same_queue_rq) { |
2299722c | 2007 | list_del_init(&same_queue_rq->queuelist); |
4711b573 JA |
2008 | plug->rq_count--; |
2009 | } | |
ce5b009c | 2010 | blk_add_rq_to_plug(plug, rq); |
ff3b74b8 | 2011 | trace_block_plug(q); |
2299722c | 2012 | |
bf4907c0 JA |
2013 | blk_mq_put_ctx(data.ctx); |
2014 | ||
dad7a3be | 2015 | if (same_queue_rq) { |
ea4f995e | 2016 | data.hctx = same_queue_rq->mq_hctx; |
ff3b74b8 | 2017 | trace_block_unplug(q, 1, true); |
2299722c | 2018 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, |
fd9c40f6 | 2019 | &cookie); |
dad7a3be | 2020 | } |
6ce3dd6e ML |
2021 | } else if ((q->nr_hw_queues > 1 && is_sync) || (!q->elevator && |
2022 | !data.hctx->dispatch_busy)) { | |
bf4907c0 | 2023 | blk_mq_put_ctx(data.ctx); |
2299722c | 2024 | blk_mq_bio_to_request(rq, bio); |
fd9c40f6 | 2025 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); |
ab42f35d | 2026 | } else { |
b00c53e8 | 2027 | blk_mq_put_ctx(data.ctx); |
ab42f35d | 2028 | blk_mq_bio_to_request(rq, bio); |
8fa9f556 | 2029 | blk_mq_sched_insert_request(rq, false, true, true); |
ab42f35d | 2030 | } |
320ae51f | 2031 | |
7b371636 | 2032 | return cookie; |
320ae51f JA |
2033 | } |
2034 | ||
cc71a6f4 JA |
2035 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2036 | unsigned int hctx_idx) | |
95363efd | 2037 | { |
e9b267d9 | 2038 | struct page *page; |
320ae51f | 2039 | |
24d2f903 | 2040 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 2041 | int i; |
320ae51f | 2042 | |
24d2f903 | 2043 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
2044 | struct request *rq = tags->static_rqs[i]; |
2045 | ||
2046 | if (!rq) | |
e9b267d9 | 2047 | continue; |
d6296d39 | 2048 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 2049 | tags->static_rqs[i] = NULL; |
e9b267d9 | 2050 | } |
320ae51f | 2051 | } |
320ae51f | 2052 | |
24d2f903 CH |
2053 | while (!list_empty(&tags->page_list)) { |
2054 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 2055 | list_del_init(&page->lru); |
f75782e4 CM |
2056 | /* |
2057 | * Remove kmemleak object previously allocated in | |
2058 | * blk_mq_init_rq_map(). | |
2059 | */ | |
2060 | kmemleak_free(page_address(page)); | |
320ae51f JA |
2061 | __free_pages(page, page->private); |
2062 | } | |
cc71a6f4 | 2063 | } |
320ae51f | 2064 | |
cc71a6f4 JA |
2065 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
2066 | { | |
24d2f903 | 2067 | kfree(tags->rqs); |
cc71a6f4 | 2068 | tags->rqs = NULL; |
2af8cbe3 JA |
2069 | kfree(tags->static_rqs); |
2070 | tags->static_rqs = NULL; | |
320ae51f | 2071 | |
24d2f903 | 2072 | blk_mq_free_tags(tags); |
320ae51f JA |
2073 | } |
2074 | ||
cc71a6f4 JA |
2075 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
2076 | unsigned int hctx_idx, | |
2077 | unsigned int nr_tags, | |
2078 | unsigned int reserved_tags) | |
320ae51f | 2079 | { |
24d2f903 | 2080 | struct blk_mq_tags *tags; |
59f082e4 | 2081 | int node; |
320ae51f | 2082 | |
7d76f856 | 2083 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx); |
59f082e4 SL |
2084 | if (node == NUMA_NO_NODE) |
2085 | node = set->numa_node; | |
2086 | ||
2087 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, | |
24391c0d | 2088 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
24d2f903 CH |
2089 | if (!tags) |
2090 | return NULL; | |
320ae51f | 2091 | |
590b5b7d | 2092 | tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
36e1f3d1 | 2093 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 2094 | node); |
24d2f903 CH |
2095 | if (!tags->rqs) { |
2096 | blk_mq_free_tags(tags); | |
2097 | return NULL; | |
2098 | } | |
320ae51f | 2099 | |
590b5b7d KC |
2100 | tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
2101 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
2102 | node); | |
2af8cbe3 JA |
2103 | if (!tags->static_rqs) { |
2104 | kfree(tags->rqs); | |
2105 | blk_mq_free_tags(tags); | |
2106 | return NULL; | |
2107 | } | |
2108 | ||
cc71a6f4 JA |
2109 | return tags; |
2110 | } | |
2111 | ||
2112 | static size_t order_to_size(unsigned int order) | |
2113 | { | |
2114 | return (size_t)PAGE_SIZE << order; | |
2115 | } | |
2116 | ||
1d9bd516 TH |
2117 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
2118 | unsigned int hctx_idx, int node) | |
2119 | { | |
2120 | int ret; | |
2121 | ||
2122 | if (set->ops->init_request) { | |
2123 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
2124 | if (ret) | |
2125 | return ret; | |
2126 | } | |
2127 | ||
12f5b931 | 2128 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
1d9bd516 TH |
2129 | return 0; |
2130 | } | |
2131 | ||
cc71a6f4 JA |
2132 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2133 | unsigned int hctx_idx, unsigned int depth) | |
2134 | { | |
2135 | unsigned int i, j, entries_per_page, max_order = 4; | |
2136 | size_t rq_size, left; | |
59f082e4 SL |
2137 | int node; |
2138 | ||
7d76f856 | 2139 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx); |
59f082e4 SL |
2140 | if (node == NUMA_NO_NODE) |
2141 | node = set->numa_node; | |
cc71a6f4 JA |
2142 | |
2143 | INIT_LIST_HEAD(&tags->page_list); | |
2144 | ||
320ae51f JA |
2145 | /* |
2146 | * rq_size is the size of the request plus driver payload, rounded | |
2147 | * to the cacheline size | |
2148 | */ | |
24d2f903 | 2149 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 2150 | cache_line_size()); |
cc71a6f4 | 2151 | left = rq_size * depth; |
320ae51f | 2152 | |
cc71a6f4 | 2153 | for (i = 0; i < depth; ) { |
320ae51f JA |
2154 | int this_order = max_order; |
2155 | struct page *page; | |
2156 | int to_do; | |
2157 | void *p; | |
2158 | ||
b3a834b1 | 2159 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
2160 | this_order--; |
2161 | ||
2162 | do { | |
59f082e4 | 2163 | page = alloc_pages_node(node, |
36e1f3d1 | 2164 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 2165 | this_order); |
320ae51f JA |
2166 | if (page) |
2167 | break; | |
2168 | if (!this_order--) | |
2169 | break; | |
2170 | if (order_to_size(this_order) < rq_size) | |
2171 | break; | |
2172 | } while (1); | |
2173 | ||
2174 | if (!page) | |
24d2f903 | 2175 | goto fail; |
320ae51f JA |
2176 | |
2177 | page->private = this_order; | |
24d2f903 | 2178 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
2179 | |
2180 | p = page_address(page); | |
f75782e4 CM |
2181 | /* |
2182 | * Allow kmemleak to scan these pages as they contain pointers | |
2183 | * to additional allocations like via ops->init_request(). | |
2184 | */ | |
36e1f3d1 | 2185 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 2186 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 2187 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
2188 | left -= to_do * rq_size; |
2189 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
2190 | struct request *rq = p; |
2191 | ||
2192 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
2193 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
2194 | tags->static_rqs[i] = NULL; | |
2195 | goto fail; | |
e9b267d9 CH |
2196 | } |
2197 | ||
320ae51f JA |
2198 | p += rq_size; |
2199 | i++; | |
2200 | } | |
2201 | } | |
cc71a6f4 | 2202 | return 0; |
320ae51f | 2203 | |
24d2f903 | 2204 | fail: |
cc71a6f4 JA |
2205 | blk_mq_free_rqs(set, tags, hctx_idx); |
2206 | return -ENOMEM; | |
320ae51f JA |
2207 | } |
2208 | ||
e57690fe JA |
2209 | /* |
2210 | * 'cpu' is going away. splice any existing rq_list entries from this | |
2211 | * software queue to the hw queue dispatch list, and ensure that it | |
2212 | * gets run. | |
2213 | */ | |
9467f859 | 2214 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 2215 | { |
9467f859 | 2216 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
2217 | struct blk_mq_ctx *ctx; |
2218 | LIST_HEAD(tmp); | |
c16d6b5a | 2219 | enum hctx_type type; |
484b4061 | 2220 | |
9467f859 | 2221 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 2222 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
c16d6b5a | 2223 | type = hctx->type; |
484b4061 JA |
2224 | |
2225 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
2226 | if (!list_empty(&ctx->rq_lists[type])) { |
2227 | list_splice_init(&ctx->rq_lists[type], &tmp); | |
484b4061 JA |
2228 | blk_mq_hctx_clear_pending(hctx, ctx); |
2229 | } | |
2230 | spin_unlock(&ctx->lock); | |
2231 | ||
2232 | if (list_empty(&tmp)) | |
9467f859 | 2233 | return 0; |
484b4061 | 2234 | |
e57690fe JA |
2235 | spin_lock(&hctx->lock); |
2236 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
2237 | spin_unlock(&hctx->lock); | |
484b4061 JA |
2238 | |
2239 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 2240 | return 0; |
484b4061 JA |
2241 | } |
2242 | ||
9467f859 | 2243 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 2244 | { |
9467f859 TG |
2245 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
2246 | &hctx->cpuhp_dead); | |
484b4061 JA |
2247 | } |
2248 | ||
c3b4afca | 2249 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
2250 | static void blk_mq_exit_hctx(struct request_queue *q, |
2251 | struct blk_mq_tag_set *set, | |
2252 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
2253 | { | |
8ab0b7dc ML |
2254 | if (blk_mq_hw_queue_mapped(hctx)) |
2255 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 2256 | |
f70ced09 | 2257 | if (set->ops->exit_request) |
d6296d39 | 2258 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); |
f70ced09 | 2259 | |
08e98fc6 ML |
2260 | if (set->ops->exit_hctx) |
2261 | set->ops->exit_hctx(hctx, hctx_idx); | |
2262 | ||
6a83e74d | 2263 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2264 | cleanup_srcu_struct(hctx->srcu); |
6a83e74d | 2265 | |
9467f859 | 2266 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 2267 | blk_free_flush_queue(hctx->fq); |
88459642 | 2268 | sbitmap_free(&hctx->ctx_map); |
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 | ||
08e98fc6 ML |
2285 | static int blk_mq_init_hctx(struct request_queue *q, |
2286 | struct blk_mq_tag_set *set, | |
2287 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 2288 | { |
08e98fc6 ML |
2289 | int node; |
2290 | ||
2291 | node = hctx->numa_node; | |
2292 | if (node == NUMA_NO_NODE) | |
2293 | node = hctx->numa_node = set->numa_node; | |
2294 | ||
9f993737 | 2295 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
2296 | spin_lock_init(&hctx->lock); |
2297 | INIT_LIST_HEAD(&hctx->dispatch); | |
2298 | hctx->queue = q; | |
2404e607 | 2299 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 2300 | |
9467f859 | 2301 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
2302 | |
2303 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
2304 | |
2305 | /* | |
08e98fc6 ML |
2306 | * Allocate space for all possible cpus to avoid allocation at |
2307 | * runtime | |
320ae51f | 2308 | */ |
d904bfa7 | 2309 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
5b202853 | 2310 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, node); |
08e98fc6 ML |
2311 | if (!hctx->ctxs) |
2312 | goto unregister_cpu_notifier; | |
320ae51f | 2313 | |
5b202853 JW |
2314 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), |
2315 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, node)) | |
08e98fc6 | 2316 | goto free_ctxs; |
320ae51f | 2317 | |
08e98fc6 | 2318 | hctx->nr_ctx = 0; |
320ae51f | 2319 | |
5815839b | 2320 | spin_lock_init(&hctx->dispatch_wait_lock); |
eb619fdb JA |
2321 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
2322 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
2323 | ||
08e98fc6 ML |
2324 | if (set->ops->init_hctx && |
2325 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
2326 | goto free_bitmap; | |
320ae51f | 2327 | |
5b202853 JW |
2328 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size, |
2329 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); | |
f70ced09 | 2330 | if (!hctx->fq) |
d48ece20 | 2331 | goto exit_hctx; |
320ae51f | 2332 | |
1d9bd516 | 2333 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node)) |
f70ced09 | 2334 | goto free_fq; |
320ae51f | 2335 | |
6a83e74d | 2336 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2337 | init_srcu_struct(hctx->srcu); |
6a83e74d | 2338 | |
08e98fc6 | 2339 | return 0; |
320ae51f | 2340 | |
f70ced09 | 2341 | free_fq: |
b9a1ff50 | 2342 | blk_free_flush_queue(hctx->fq); |
f70ced09 ML |
2343 | exit_hctx: |
2344 | if (set->ops->exit_hctx) | |
2345 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 2346 | free_bitmap: |
88459642 | 2347 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2348 | free_ctxs: |
2349 | kfree(hctx->ctxs); | |
2350 | unregister_cpu_notifier: | |
9467f859 | 2351 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
2352 | return -1; |
2353 | } | |
320ae51f | 2354 | |
320ae51f JA |
2355 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
2356 | unsigned int nr_hw_queues) | |
2357 | { | |
b3c661b1 JA |
2358 | struct blk_mq_tag_set *set = q->tag_set; |
2359 | unsigned int i, j; | |
320ae51f JA |
2360 | |
2361 | for_each_possible_cpu(i) { | |
2362 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2363 | struct blk_mq_hw_ctx *hctx; | |
c16d6b5a | 2364 | int k; |
320ae51f | 2365 | |
320ae51f JA |
2366 | __ctx->cpu = i; |
2367 | spin_lock_init(&__ctx->lock); | |
c16d6b5a ML |
2368 | for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++) |
2369 | INIT_LIST_HEAD(&__ctx->rq_lists[k]); | |
2370 | ||
320ae51f JA |
2371 | __ctx->queue = q; |
2372 | ||
320ae51f JA |
2373 | /* |
2374 | * Set local node, IFF we have more than one hw queue. If | |
2375 | * not, we remain on the home node of the device | |
2376 | */ | |
b3c661b1 JA |
2377 | for (j = 0; j < set->nr_maps; j++) { |
2378 | hctx = blk_mq_map_queue_type(q, j, i); | |
2379 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
2380 | hctx->numa_node = local_memory_node(cpu_to_node(i)); | |
2381 | } | |
320ae51f JA |
2382 | } |
2383 | } | |
2384 | ||
cc71a6f4 JA |
2385 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
2386 | { | |
2387 | int ret = 0; | |
2388 | ||
2389 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
2390 | set->queue_depth, set->reserved_tags); | |
2391 | if (!set->tags[hctx_idx]) | |
2392 | return false; | |
2393 | ||
2394 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2395 | set->queue_depth); | |
2396 | if (!ret) | |
2397 | return true; | |
2398 | ||
2399 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2400 | set->tags[hctx_idx] = NULL; | |
2401 | return false; | |
2402 | } | |
2403 | ||
2404 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2405 | unsigned int hctx_idx) | |
2406 | { | |
4e6db0f2 | 2407 | if (set->tags && set->tags[hctx_idx]) { |
bd166ef1 JA |
2408 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); |
2409 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2410 | set->tags[hctx_idx] = NULL; | |
2411 | } | |
cc71a6f4 JA |
2412 | } |
2413 | ||
4b855ad3 | 2414 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 2415 | { |
b3c661b1 | 2416 | unsigned int i, j, hctx_idx; |
320ae51f JA |
2417 | struct blk_mq_hw_ctx *hctx; |
2418 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2419 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2420 | |
60de074b AM |
2421 | /* |
2422 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
2423 | */ | |
2424 | mutex_lock(&q->sysfs_lock); | |
2425 | ||
320ae51f | 2426 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 2427 | cpumask_clear(hctx->cpumask); |
320ae51f | 2428 | hctx->nr_ctx = 0; |
d416c92c | 2429 | hctx->dispatch_from = NULL; |
320ae51f JA |
2430 | } |
2431 | ||
2432 | /* | |
4b855ad3 | 2433 | * Map software to hardware queues. |
4412efec ML |
2434 | * |
2435 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 2436 | */ |
20e4d813 | 2437 | for_each_possible_cpu(i) { |
7d76f856 | 2438 | hctx_idx = set->map[HCTX_TYPE_DEFAULT].mq_map[i]; |
4412efec ML |
2439 | /* unmapped hw queue can be remapped after CPU topo changed */ |
2440 | if (!set->tags[hctx_idx] && | |
2441 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
2442 | /* | |
2443 | * If tags initialization fail for some hctx, | |
2444 | * that hctx won't be brought online. In this | |
2445 | * case, remap the current ctx to hctx[0] which | |
2446 | * is guaranteed to always have tags allocated | |
2447 | */ | |
7d76f856 | 2448 | set->map[HCTX_TYPE_DEFAULT].mq_map[i] = 0; |
4412efec ML |
2449 | } |
2450 | ||
897bb0c7 | 2451 | ctx = per_cpu_ptr(q->queue_ctx, i); |
b3c661b1 | 2452 | for (j = 0; j < set->nr_maps; j++) { |
bb94aea1 JW |
2453 | if (!set->map[j].nr_queues) { |
2454 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
2455 | HCTX_TYPE_DEFAULT, i); | |
e5edd5f2 | 2456 | continue; |
bb94aea1 | 2457 | } |
e5edd5f2 | 2458 | |
b3c661b1 | 2459 | hctx = blk_mq_map_queue_type(q, j, i); |
8ccdf4a3 | 2460 | ctx->hctxs[j] = hctx; |
b3c661b1 JA |
2461 | /* |
2462 | * If the CPU is already set in the mask, then we've | |
2463 | * mapped this one already. This can happen if | |
2464 | * devices share queues across queue maps. | |
2465 | */ | |
2466 | if (cpumask_test_cpu(i, hctx->cpumask)) | |
2467 | continue; | |
2468 | ||
2469 | cpumask_set_cpu(i, hctx->cpumask); | |
2470 | hctx->type = j; | |
2471 | ctx->index_hw[hctx->type] = hctx->nr_ctx; | |
2472 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2473 | ||
2474 | /* | |
2475 | * If the nr_ctx type overflows, we have exceeded the | |
2476 | * amount of sw queues we can support. | |
2477 | */ | |
2478 | BUG_ON(!hctx->nr_ctx); | |
2479 | } | |
bb94aea1 JW |
2480 | |
2481 | for (; j < HCTX_MAX_TYPES; j++) | |
2482 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
2483 | HCTX_TYPE_DEFAULT, i); | |
320ae51f | 2484 | } |
506e931f | 2485 | |
60de074b AM |
2486 | mutex_unlock(&q->sysfs_lock); |
2487 | ||
506e931f | 2488 | queue_for_each_hw_ctx(q, hctx, i) { |
4412efec ML |
2489 | /* |
2490 | * If no software queues are mapped to this hardware queue, | |
2491 | * disable it and free the request entries. | |
2492 | */ | |
2493 | if (!hctx->nr_ctx) { | |
2494 | /* Never unmap queue 0. We need it as a | |
2495 | * fallback in case of a new remap fails | |
2496 | * allocation | |
2497 | */ | |
2498 | if (i && set->tags[i]) | |
2499 | blk_mq_free_map_and_requests(set, i); | |
2500 | ||
2501 | hctx->tags = NULL; | |
2502 | continue; | |
2503 | } | |
484b4061 | 2504 | |
2a34c087 ML |
2505 | hctx->tags = set->tags[i]; |
2506 | WARN_ON(!hctx->tags); | |
2507 | ||
889fa31f CY |
2508 | /* |
2509 | * Set the map size to the number of mapped software queues. | |
2510 | * This is more accurate and more efficient than looping | |
2511 | * over all possibly mapped software queues. | |
2512 | */ | |
88459642 | 2513 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2514 | |
484b4061 JA |
2515 | /* |
2516 | * Initialize batch roundrobin counts | |
2517 | */ | |
f82ddf19 | 2518 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
2519 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2520 | } | |
320ae51f JA |
2521 | } |
2522 | ||
8e8320c9 JA |
2523 | /* |
2524 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
2525 | * the queue isn't live yet. | |
2526 | */ | |
2404e607 | 2527 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2528 | { |
2529 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2530 | int i; |
2531 | ||
2404e607 | 2532 | queue_for_each_hw_ctx(q, hctx, i) { |
97889f9a | 2533 | if (shared) |
2404e607 | 2534 | hctx->flags |= BLK_MQ_F_TAG_SHARED; |
97889f9a | 2535 | else |
2404e607 JM |
2536 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; |
2537 | } | |
2538 | } | |
2539 | ||
8e8320c9 JA |
2540 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, |
2541 | bool shared) | |
2404e607 JM |
2542 | { |
2543 | struct request_queue *q; | |
0d2602ca | 2544 | |
705cda97 BVA |
2545 | lockdep_assert_held(&set->tag_list_lock); |
2546 | ||
0d2602ca JA |
2547 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2548 | blk_mq_freeze_queue(q); | |
2404e607 | 2549 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2550 | blk_mq_unfreeze_queue(q); |
2551 | } | |
2552 | } | |
2553 | ||
2554 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2555 | { | |
2556 | struct blk_mq_tag_set *set = q->tag_set; | |
2557 | ||
0d2602ca | 2558 | mutex_lock(&set->tag_list_lock); |
705cda97 | 2559 | list_del_rcu(&q->tag_set_list); |
2404e607 JM |
2560 | if (list_is_singular(&set->tag_list)) { |
2561 | /* just transitioned to unshared */ | |
2562 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2563 | /* update existing queue */ | |
2564 | blk_mq_update_tag_set_depth(set, false); | |
2565 | } | |
0d2602ca | 2566 | mutex_unlock(&set->tag_list_lock); |
a347c7ad | 2567 | INIT_LIST_HEAD(&q->tag_set_list); |
0d2602ca JA |
2568 | } |
2569 | ||
2570 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2571 | struct request_queue *q) | |
2572 | { | |
0d2602ca | 2573 | mutex_lock(&set->tag_list_lock); |
2404e607 | 2574 | |
ff821d27 JA |
2575 | /* |
2576 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
2577 | */ | |
2578 | if (!list_empty(&set->tag_list) && | |
2579 | !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2404e607 JM |
2580 | set->flags |= BLK_MQ_F_TAG_SHARED; |
2581 | /* update existing queue */ | |
2582 | blk_mq_update_tag_set_depth(set, true); | |
2583 | } | |
2584 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2585 | queue_set_hctx_shared(q, true); | |
705cda97 | 2586 | list_add_tail_rcu(&q->tag_set_list, &set->tag_list); |
2404e607 | 2587 | |
0d2602ca JA |
2588 | mutex_unlock(&set->tag_list_lock); |
2589 | } | |
2590 | ||
1db4909e ML |
2591 | /* All allocations will be freed in release handler of q->mq_kobj */ |
2592 | static int blk_mq_alloc_ctxs(struct request_queue *q) | |
2593 | { | |
2594 | struct blk_mq_ctxs *ctxs; | |
2595 | int cpu; | |
2596 | ||
2597 | ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL); | |
2598 | if (!ctxs) | |
2599 | return -ENOMEM; | |
2600 | ||
2601 | ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx); | |
2602 | if (!ctxs->queue_ctx) | |
2603 | goto fail; | |
2604 | ||
2605 | for_each_possible_cpu(cpu) { | |
2606 | struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu); | |
2607 | ctx->ctxs = ctxs; | |
2608 | } | |
2609 | ||
2610 | q->mq_kobj = &ctxs->kobj; | |
2611 | q->queue_ctx = ctxs->queue_ctx; | |
2612 | ||
2613 | return 0; | |
2614 | fail: | |
2615 | kfree(ctxs); | |
2616 | return -ENOMEM; | |
2617 | } | |
2618 | ||
e09aae7e ML |
2619 | /* |
2620 | * It is the actual release handler for mq, but we do it from | |
2621 | * request queue's release handler for avoiding use-after-free | |
2622 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2623 | * but we can't group ctx/kctx kobj without it. | |
2624 | */ | |
2625 | void blk_mq_release(struct request_queue *q) | |
2626 | { | |
2627 | struct blk_mq_hw_ctx *hctx; | |
2628 | unsigned int i; | |
2629 | ||
2630 | /* hctx kobj stays in hctx */ | |
c3b4afca ML |
2631 | queue_for_each_hw_ctx(q, hctx, i) { |
2632 | if (!hctx) | |
2633 | continue; | |
6c8b232e | 2634 | kobject_put(&hctx->kobj); |
c3b4afca | 2635 | } |
e09aae7e ML |
2636 | |
2637 | kfree(q->queue_hw_ctx); | |
2638 | ||
7ea5fe31 ML |
2639 | /* |
2640 | * release .mq_kobj and sw queue's kobject now because | |
2641 | * both share lifetime with request queue. | |
2642 | */ | |
2643 | blk_mq_sysfs_deinit(q); | |
e09aae7e ML |
2644 | } |
2645 | ||
24d2f903 | 2646 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2647 | { |
2648 | struct request_queue *uninit_q, *q; | |
2649 | ||
6d469642 | 2650 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); |
b62c21b7 MS |
2651 | if (!uninit_q) |
2652 | return ERR_PTR(-ENOMEM); | |
2653 | ||
2654 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2655 | if (IS_ERR(q)) | |
2656 | blk_cleanup_queue(uninit_q); | |
2657 | ||
2658 | return q; | |
2659 | } | |
2660 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2661 | ||
9316a9ed JA |
2662 | /* |
2663 | * Helper for setting up a queue with mq ops, given queue depth, and | |
2664 | * the passed in mq ops flags. | |
2665 | */ | |
2666 | struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set, | |
2667 | const struct blk_mq_ops *ops, | |
2668 | unsigned int queue_depth, | |
2669 | unsigned int set_flags) | |
2670 | { | |
2671 | struct request_queue *q; | |
2672 | int ret; | |
2673 | ||
2674 | memset(set, 0, sizeof(*set)); | |
2675 | set->ops = ops; | |
2676 | set->nr_hw_queues = 1; | |
b3c661b1 | 2677 | set->nr_maps = 1; |
9316a9ed JA |
2678 | set->queue_depth = queue_depth; |
2679 | set->numa_node = NUMA_NO_NODE; | |
2680 | set->flags = set_flags; | |
2681 | ||
2682 | ret = blk_mq_alloc_tag_set(set); | |
2683 | if (ret) | |
2684 | return ERR_PTR(ret); | |
2685 | ||
2686 | q = blk_mq_init_queue(set); | |
2687 | if (IS_ERR(q)) { | |
2688 | blk_mq_free_tag_set(set); | |
2689 | return q; | |
2690 | } | |
2691 | ||
2692 | return q; | |
2693 | } | |
2694 | EXPORT_SYMBOL(blk_mq_init_sq_queue); | |
2695 | ||
07319678 BVA |
2696 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) |
2697 | { | |
2698 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2699 | ||
05707b64 | 2700 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu), |
07319678 BVA |
2701 | __alignof__(struct blk_mq_hw_ctx)) != |
2702 | sizeof(struct blk_mq_hw_ctx)); | |
2703 | ||
2704 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2705 | hw_ctx_size += sizeof(struct srcu_struct); | |
2706 | ||
2707 | return hw_ctx_size; | |
2708 | } | |
2709 | ||
34d11ffa JW |
2710 | static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx( |
2711 | struct blk_mq_tag_set *set, struct request_queue *q, | |
2712 | int hctx_idx, int node) | |
2713 | { | |
2714 | struct blk_mq_hw_ctx *hctx; | |
2715 | ||
2716 | hctx = kzalloc_node(blk_mq_hw_ctx_size(set), | |
2717 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
2718 | node); | |
2719 | if (!hctx) | |
2720 | return NULL; | |
2721 | ||
2722 | if (!zalloc_cpumask_var_node(&hctx->cpumask, | |
2723 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
2724 | node)) { | |
2725 | kfree(hctx); | |
2726 | return NULL; | |
2727 | } | |
2728 | ||
2729 | atomic_set(&hctx->nr_active, 0); | |
2730 | hctx->numa_node = node; | |
2731 | hctx->queue_num = hctx_idx; | |
2732 | ||
2733 | if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) { | |
2734 | free_cpumask_var(hctx->cpumask); | |
2735 | kfree(hctx); | |
2736 | return NULL; | |
2737 | } | |
2738 | blk_mq_hctx_kobj_init(hctx); | |
2739 | ||
2740 | return hctx; | |
2741 | } | |
2742 | ||
868f2f0b KB |
2743 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2744 | struct request_queue *q) | |
320ae51f | 2745 | { |
e01ad46d | 2746 | int i, j, end; |
868f2f0b | 2747 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; |
f14bbe77 | 2748 | |
fb350e0a ML |
2749 | /* protect against switching io scheduler */ |
2750 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 2751 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2752 | int node; |
34d11ffa | 2753 | struct blk_mq_hw_ctx *hctx; |
868f2f0b | 2754 | |
7d76f856 | 2755 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i); |
34d11ffa JW |
2756 | /* |
2757 | * If the hw queue has been mapped to another numa node, | |
2758 | * we need to realloc the hctx. If allocation fails, fallback | |
2759 | * to use the previous one. | |
2760 | */ | |
2761 | if (hctxs[i] && (hctxs[i]->numa_node == node)) | |
2762 | continue; | |
868f2f0b | 2763 | |
34d11ffa JW |
2764 | hctx = blk_mq_alloc_and_init_hctx(set, q, i, node); |
2765 | if (hctx) { | |
2766 | if (hctxs[i]) { | |
2767 | blk_mq_exit_hctx(q, set, hctxs[i], i); | |
2768 | kobject_put(&hctxs[i]->kobj); | |
2769 | } | |
2770 | hctxs[i] = hctx; | |
2771 | } else { | |
2772 | if (hctxs[i]) | |
2773 | pr_warn("Allocate new hctx on node %d fails,\ | |
2774 | fallback to previous one on node %d\n", | |
2775 | node, hctxs[i]->numa_node); | |
2776 | else | |
2777 | break; | |
868f2f0b | 2778 | } |
320ae51f | 2779 | } |
e01ad46d JW |
2780 | /* |
2781 | * Increasing nr_hw_queues fails. Free the newly allocated | |
2782 | * hctxs and keep the previous q->nr_hw_queues. | |
2783 | */ | |
2784 | if (i != set->nr_hw_queues) { | |
2785 | j = q->nr_hw_queues; | |
2786 | end = i; | |
2787 | } else { | |
2788 | j = i; | |
2789 | end = q->nr_hw_queues; | |
2790 | q->nr_hw_queues = set->nr_hw_queues; | |
2791 | } | |
34d11ffa | 2792 | |
e01ad46d | 2793 | for (; j < end; j++) { |
868f2f0b KB |
2794 | struct blk_mq_hw_ctx *hctx = hctxs[j]; |
2795 | ||
2796 | if (hctx) { | |
cc71a6f4 JA |
2797 | if (hctx->tags) |
2798 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 2799 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 2800 | kobject_put(&hctx->kobj); |
868f2f0b KB |
2801 | hctxs[j] = NULL; |
2802 | ||
2803 | } | |
2804 | } | |
fb350e0a | 2805 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
2806 | } |
2807 | ||
392546ae JA |
2808 | /* |
2809 | * Maximum number of hardware queues we support. For single sets, we'll never | |
2810 | * have more than the CPUs (software queues). For multiple sets, the tag_set | |
2811 | * user may have set ->nr_hw_queues larger. | |
2812 | */ | |
2813 | static unsigned int nr_hw_queues(struct blk_mq_tag_set *set) | |
2814 | { | |
2815 | if (set->nr_maps == 1) | |
2816 | return nr_cpu_ids; | |
2817 | ||
2818 | return max(set->nr_hw_queues, nr_cpu_ids); | |
2819 | } | |
2820 | ||
868f2f0b KB |
2821 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, |
2822 | struct request_queue *q) | |
2823 | { | |
66841672 ML |
2824 | /* mark the queue as mq asap */ |
2825 | q->mq_ops = set->ops; | |
2826 | ||
34dbad5d | 2827 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
2828 | blk_mq_poll_stats_bkt, |
2829 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
2830 | if (!q->poll_cb) |
2831 | goto err_exit; | |
2832 | ||
1db4909e | 2833 | if (blk_mq_alloc_ctxs(q)) |
c7de5726 | 2834 | goto err_exit; |
868f2f0b | 2835 | |
737f98cf ML |
2836 | /* init q->mq_kobj and sw queues' kobjects */ |
2837 | blk_mq_sysfs_init(q); | |
2838 | ||
392546ae JA |
2839 | q->nr_queues = nr_hw_queues(set); |
2840 | q->queue_hw_ctx = kcalloc_node(q->nr_queues, sizeof(*(q->queue_hw_ctx)), | |
868f2f0b KB |
2841 | GFP_KERNEL, set->numa_node); |
2842 | if (!q->queue_hw_ctx) | |
1db4909e | 2843 | goto err_sys_init; |
868f2f0b | 2844 | |
868f2f0b KB |
2845 | blk_mq_realloc_hw_ctxs(set, q); |
2846 | if (!q->nr_hw_queues) | |
2847 | goto err_hctxs; | |
320ae51f | 2848 | |
287922eb | 2849 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2850 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f | 2851 | |
a8908939 | 2852 | q->tag_set = set; |
320ae51f | 2853 | |
94eddfbe | 2854 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
cd19181b ML |
2855 | if (set->nr_maps > HCTX_TYPE_POLL && |
2856 | set->map[HCTX_TYPE_POLL].nr_queues) | |
6544d229 | 2857 | blk_queue_flag_set(QUEUE_FLAG_POLL, q); |
320ae51f | 2858 | |
1be036e9 CH |
2859 | q->sg_reserved_size = INT_MAX; |
2860 | ||
2849450a | 2861 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2862 | INIT_LIST_HEAD(&q->requeue_list); |
2863 | spin_lock_init(&q->requeue_lock); | |
2864 | ||
254d259d | 2865 | blk_queue_make_request(q, blk_mq_make_request); |
07068d5b | 2866 | |
eba71768 JA |
2867 | /* |
2868 | * Do this after blk_queue_make_request() overrides it... | |
2869 | */ | |
2870 | q->nr_requests = set->queue_depth; | |
2871 | ||
64f1c21e JA |
2872 | /* |
2873 | * Default to classic polling | |
2874 | */ | |
29ece8b4 | 2875 | q->poll_nsec = BLK_MQ_POLL_CLASSIC; |
64f1c21e | 2876 | |
24d2f903 | 2877 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 2878 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 2879 | blk_mq_map_swqueue(q); |
4593fdbe | 2880 | |
d3484991 JA |
2881 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { |
2882 | int ret; | |
2883 | ||
131d08e1 | 2884 | ret = elevator_init_mq(q); |
d3484991 JA |
2885 | if (ret) |
2886 | return ERR_PTR(ret); | |
2887 | } | |
2888 | ||
320ae51f | 2889 | return q; |
18741986 | 2890 | |
320ae51f | 2891 | err_hctxs: |
868f2f0b | 2892 | kfree(q->queue_hw_ctx); |
1db4909e ML |
2893 | err_sys_init: |
2894 | blk_mq_sysfs_deinit(q); | |
c7de5726 ML |
2895 | err_exit: |
2896 | q->mq_ops = NULL; | |
320ae51f JA |
2897 | return ERR_PTR(-ENOMEM); |
2898 | } | |
b62c21b7 | 2899 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2900 | |
2901 | void blk_mq_free_queue(struct request_queue *q) | |
2902 | { | |
624dbe47 | 2903 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2904 | |
0d2602ca | 2905 | blk_mq_del_queue_tag_set(q); |
624dbe47 | 2906 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
320ae51f | 2907 | } |
320ae51f | 2908 | |
a5164405 JA |
2909 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2910 | { | |
2911 | int i; | |
2912 | ||
cc71a6f4 JA |
2913 | for (i = 0; i < set->nr_hw_queues; i++) |
2914 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 2915 | goto out_unwind; |
a5164405 JA |
2916 | |
2917 | return 0; | |
2918 | ||
2919 | out_unwind: | |
2920 | while (--i >= 0) | |
cc71a6f4 | 2921 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 2922 | |
a5164405 JA |
2923 | return -ENOMEM; |
2924 | } | |
2925 | ||
2926 | /* | |
2927 | * Allocate the request maps associated with this tag_set. Note that this | |
2928 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2929 | * will be updated to reflect the allocated depth. | |
2930 | */ | |
2931 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2932 | { | |
2933 | unsigned int depth; | |
2934 | int err; | |
2935 | ||
2936 | depth = set->queue_depth; | |
2937 | do { | |
2938 | err = __blk_mq_alloc_rq_maps(set); | |
2939 | if (!err) | |
2940 | break; | |
2941 | ||
2942 | set->queue_depth >>= 1; | |
2943 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2944 | err = -ENOMEM; | |
2945 | break; | |
2946 | } | |
2947 | } while (set->queue_depth); | |
2948 | ||
2949 | if (!set->queue_depth || err) { | |
2950 | pr_err("blk-mq: failed to allocate request map\n"); | |
2951 | return -ENOMEM; | |
2952 | } | |
2953 | ||
2954 | if (depth != set->queue_depth) | |
2955 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2956 | depth, set->queue_depth); | |
2957 | ||
2958 | return 0; | |
2959 | } | |
2960 | ||
ebe8bddb OS |
2961 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
2962 | { | |
59388702 | 2963 | if (set->ops->map_queues && !is_kdump_kernel()) { |
b3c661b1 JA |
2964 | int i; |
2965 | ||
7d4901a9 ML |
2966 | /* |
2967 | * transport .map_queues is usually done in the following | |
2968 | * way: | |
2969 | * | |
2970 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
2971 | * mask = get_cpu_mask(queue) | |
2972 | * for_each_cpu(cpu, mask) | |
b3c661b1 | 2973 | * set->map[x].mq_map[cpu] = queue; |
7d4901a9 ML |
2974 | * } |
2975 | * | |
2976 | * When we need to remap, the table has to be cleared for | |
2977 | * killing stale mapping since one CPU may not be mapped | |
2978 | * to any hw queue. | |
2979 | */ | |
b3c661b1 JA |
2980 | for (i = 0; i < set->nr_maps; i++) |
2981 | blk_mq_clear_mq_map(&set->map[i]); | |
7d4901a9 | 2982 | |
ebe8bddb | 2983 | return set->ops->map_queues(set); |
b3c661b1 JA |
2984 | } else { |
2985 | BUG_ON(set->nr_maps > 1); | |
7d76f856 | 2986 | return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
b3c661b1 | 2987 | } |
ebe8bddb OS |
2988 | } |
2989 | ||
a4391c64 JA |
2990 | /* |
2991 | * Alloc a tag set to be associated with one or more request queues. | |
2992 | * May fail with EINVAL for various error conditions. May adjust the | |
c018c84f | 2993 | * requested depth down, if it's too large. In that case, the set |
a4391c64 JA |
2994 | * value will be stored in set->queue_depth. |
2995 | */ | |
24d2f903 CH |
2996 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2997 | { | |
b3c661b1 | 2998 | int i, ret; |
da695ba2 | 2999 | |
205fb5f5 BVA |
3000 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
3001 | ||
24d2f903 CH |
3002 | if (!set->nr_hw_queues) |
3003 | return -EINVAL; | |
a4391c64 | 3004 | if (!set->queue_depth) |
24d2f903 CH |
3005 | return -EINVAL; |
3006 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
3007 | return -EINVAL; | |
3008 | ||
7d7e0f90 | 3009 | if (!set->ops->queue_rq) |
24d2f903 CH |
3010 | return -EINVAL; |
3011 | ||
de148297 ML |
3012 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
3013 | return -EINVAL; | |
3014 | ||
a4391c64 JA |
3015 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
3016 | pr_info("blk-mq: reduced tag depth to %u\n", | |
3017 | BLK_MQ_MAX_DEPTH); | |
3018 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
3019 | } | |
24d2f903 | 3020 | |
b3c661b1 JA |
3021 | if (!set->nr_maps) |
3022 | set->nr_maps = 1; | |
3023 | else if (set->nr_maps > HCTX_MAX_TYPES) | |
3024 | return -EINVAL; | |
3025 | ||
6637fadf SL |
3026 | /* |
3027 | * If a crashdump is active, then we are potentially in a very | |
3028 | * memory constrained environment. Limit us to 1 queue and | |
3029 | * 64 tags to prevent using too much memory. | |
3030 | */ | |
3031 | if (is_kdump_kernel()) { | |
3032 | set->nr_hw_queues = 1; | |
59388702 | 3033 | set->nr_maps = 1; |
6637fadf SL |
3034 | set->queue_depth = min(64U, set->queue_depth); |
3035 | } | |
868f2f0b | 3036 | /* |
392546ae JA |
3037 | * There is no use for more h/w queues than cpus if we just have |
3038 | * a single map | |
868f2f0b | 3039 | */ |
392546ae | 3040 | if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids) |
868f2f0b | 3041 | set->nr_hw_queues = nr_cpu_ids; |
6637fadf | 3042 | |
392546ae | 3043 | set->tags = kcalloc_node(nr_hw_queues(set), sizeof(struct blk_mq_tags *), |
24d2f903 CH |
3044 | GFP_KERNEL, set->numa_node); |
3045 | if (!set->tags) | |
a5164405 | 3046 | return -ENOMEM; |
24d2f903 | 3047 | |
da695ba2 | 3048 | ret = -ENOMEM; |
b3c661b1 JA |
3049 | for (i = 0; i < set->nr_maps; i++) { |
3050 | set->map[i].mq_map = kcalloc_node(nr_cpu_ids, | |
07b35eb5 | 3051 | sizeof(set->map[i].mq_map[0]), |
b3c661b1 JA |
3052 | GFP_KERNEL, set->numa_node); |
3053 | if (!set->map[i].mq_map) | |
3054 | goto out_free_mq_map; | |
59388702 | 3055 | set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues; |
b3c661b1 | 3056 | } |
bdd17e75 | 3057 | |
ebe8bddb | 3058 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
3059 | if (ret) |
3060 | goto out_free_mq_map; | |
3061 | ||
3062 | ret = blk_mq_alloc_rq_maps(set); | |
3063 | if (ret) | |
bdd17e75 | 3064 | goto out_free_mq_map; |
24d2f903 | 3065 | |
0d2602ca JA |
3066 | mutex_init(&set->tag_list_lock); |
3067 | INIT_LIST_HEAD(&set->tag_list); | |
3068 | ||
24d2f903 | 3069 | return 0; |
bdd17e75 CH |
3070 | |
3071 | out_free_mq_map: | |
b3c661b1 JA |
3072 | for (i = 0; i < set->nr_maps; i++) { |
3073 | kfree(set->map[i].mq_map); | |
3074 | set->map[i].mq_map = NULL; | |
3075 | } | |
5676e7b6 RE |
3076 | kfree(set->tags); |
3077 | set->tags = NULL; | |
da695ba2 | 3078 | return ret; |
24d2f903 CH |
3079 | } |
3080 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
3081 | ||
3082 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
3083 | { | |
b3c661b1 | 3084 | int i, j; |
24d2f903 | 3085 | |
392546ae | 3086 | for (i = 0; i < nr_hw_queues(set); i++) |
cc71a6f4 | 3087 | blk_mq_free_map_and_requests(set, i); |
484b4061 | 3088 | |
b3c661b1 JA |
3089 | for (j = 0; j < set->nr_maps; j++) { |
3090 | kfree(set->map[j].mq_map); | |
3091 | set->map[j].mq_map = NULL; | |
3092 | } | |
bdd17e75 | 3093 | |
981bd189 | 3094 | kfree(set->tags); |
5676e7b6 | 3095 | set->tags = NULL; |
24d2f903 CH |
3096 | } |
3097 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
3098 | ||
e3a2b3f9 JA |
3099 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
3100 | { | |
3101 | struct blk_mq_tag_set *set = q->tag_set; | |
3102 | struct blk_mq_hw_ctx *hctx; | |
3103 | int i, ret; | |
3104 | ||
bd166ef1 | 3105 | if (!set) |
e3a2b3f9 JA |
3106 | return -EINVAL; |
3107 | ||
e5fa8140 AZ |
3108 | if (q->nr_requests == nr) |
3109 | return 0; | |
3110 | ||
70f36b60 | 3111 | blk_mq_freeze_queue(q); |
24f5a90f | 3112 | blk_mq_quiesce_queue(q); |
70f36b60 | 3113 | |
e3a2b3f9 JA |
3114 | ret = 0; |
3115 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
3116 | if (!hctx->tags) |
3117 | continue; | |
bd166ef1 JA |
3118 | /* |
3119 | * If we're using an MQ scheduler, just update the scheduler | |
3120 | * queue depth. This is similar to what the old code would do. | |
3121 | */ | |
70f36b60 | 3122 | if (!hctx->sched_tags) { |
c2e82a23 | 3123 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, |
70f36b60 JA |
3124 | false); |
3125 | } else { | |
3126 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
3127 | nr, true); | |
3128 | } | |
e3a2b3f9 JA |
3129 | if (ret) |
3130 | break; | |
3131 | } | |
3132 | ||
3133 | if (!ret) | |
3134 | q->nr_requests = nr; | |
3135 | ||
24f5a90f | 3136 | blk_mq_unquiesce_queue(q); |
70f36b60 | 3137 | blk_mq_unfreeze_queue(q); |
70f36b60 | 3138 | |
e3a2b3f9 JA |
3139 | return ret; |
3140 | } | |
3141 | ||
d48ece20 JW |
3142 | /* |
3143 | * request_queue and elevator_type pair. | |
3144 | * It is just used by __blk_mq_update_nr_hw_queues to cache | |
3145 | * the elevator_type associated with a request_queue. | |
3146 | */ | |
3147 | struct blk_mq_qe_pair { | |
3148 | struct list_head node; | |
3149 | struct request_queue *q; | |
3150 | struct elevator_type *type; | |
3151 | }; | |
3152 | ||
3153 | /* | |
3154 | * Cache the elevator_type in qe pair list and switch the | |
3155 | * io scheduler to 'none' | |
3156 | */ | |
3157 | static bool blk_mq_elv_switch_none(struct list_head *head, | |
3158 | struct request_queue *q) | |
3159 | { | |
3160 | struct blk_mq_qe_pair *qe; | |
3161 | ||
3162 | if (!q->elevator) | |
3163 | return true; | |
3164 | ||
3165 | qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); | |
3166 | if (!qe) | |
3167 | return false; | |
3168 | ||
3169 | INIT_LIST_HEAD(&qe->node); | |
3170 | qe->q = q; | |
3171 | qe->type = q->elevator->type; | |
3172 | list_add(&qe->node, head); | |
3173 | ||
3174 | mutex_lock(&q->sysfs_lock); | |
3175 | /* | |
3176 | * After elevator_switch_mq, the previous elevator_queue will be | |
3177 | * released by elevator_release. The reference of the io scheduler | |
3178 | * module get by elevator_get will also be put. So we need to get | |
3179 | * a reference of the io scheduler module here to prevent it to be | |
3180 | * removed. | |
3181 | */ | |
3182 | __module_get(qe->type->elevator_owner); | |
3183 | elevator_switch_mq(q, NULL); | |
3184 | mutex_unlock(&q->sysfs_lock); | |
3185 | ||
3186 | return true; | |
3187 | } | |
3188 | ||
3189 | static void blk_mq_elv_switch_back(struct list_head *head, | |
3190 | struct request_queue *q) | |
3191 | { | |
3192 | struct blk_mq_qe_pair *qe; | |
3193 | struct elevator_type *t = NULL; | |
3194 | ||
3195 | list_for_each_entry(qe, head, node) | |
3196 | if (qe->q == q) { | |
3197 | t = qe->type; | |
3198 | break; | |
3199 | } | |
3200 | ||
3201 | if (!t) | |
3202 | return; | |
3203 | ||
3204 | list_del(&qe->node); | |
3205 | kfree(qe); | |
3206 | ||
3207 | mutex_lock(&q->sysfs_lock); | |
3208 | elevator_switch_mq(q, t); | |
3209 | mutex_unlock(&q->sysfs_lock); | |
3210 | } | |
3211 | ||
e4dc2b32 KB |
3212 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
3213 | int nr_hw_queues) | |
868f2f0b KB |
3214 | { |
3215 | struct request_queue *q; | |
d48ece20 | 3216 | LIST_HEAD(head); |
e01ad46d | 3217 | int prev_nr_hw_queues; |
868f2f0b | 3218 | |
705cda97 BVA |
3219 | lockdep_assert_held(&set->tag_list_lock); |
3220 | ||
392546ae | 3221 | if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids) |
868f2f0b KB |
3222 | nr_hw_queues = nr_cpu_ids; |
3223 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
3224 | return; | |
3225 | ||
3226 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3227 | blk_mq_freeze_queue(q); | |
f5bbbbe4 JW |
3228 | /* |
3229 | * Sync with blk_mq_queue_tag_busy_iter. | |
3230 | */ | |
3231 | synchronize_rcu(); | |
d48ece20 JW |
3232 | /* |
3233 | * Switch IO scheduler to 'none', cleaning up the data associated | |
3234 | * with the previous scheduler. We will switch back once we are done | |
3235 | * updating the new sw to hw queue mappings. | |
3236 | */ | |
3237 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3238 | if (!blk_mq_elv_switch_none(&head, q)) | |
3239 | goto switch_back; | |
868f2f0b | 3240 | |
477e19de JW |
3241 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3242 | blk_mq_debugfs_unregister_hctxs(q); | |
3243 | blk_mq_sysfs_unregister(q); | |
3244 | } | |
3245 | ||
e01ad46d | 3246 | prev_nr_hw_queues = set->nr_hw_queues; |
868f2f0b | 3247 | set->nr_hw_queues = nr_hw_queues; |
ebe8bddb | 3248 | blk_mq_update_queue_map(set); |
e01ad46d | 3249 | fallback: |
868f2f0b KB |
3250 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3251 | blk_mq_realloc_hw_ctxs(set, q); | |
e01ad46d JW |
3252 | if (q->nr_hw_queues != set->nr_hw_queues) { |
3253 | pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n", | |
3254 | nr_hw_queues, prev_nr_hw_queues); | |
3255 | set->nr_hw_queues = prev_nr_hw_queues; | |
7d76f856 | 3256 | blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
e01ad46d JW |
3257 | goto fallback; |
3258 | } | |
477e19de JW |
3259 | blk_mq_map_swqueue(q); |
3260 | } | |
3261 | ||
3262 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
3263 | blk_mq_sysfs_register(q); | |
3264 | blk_mq_debugfs_register_hctxs(q); | |
868f2f0b KB |
3265 | } |
3266 | ||
d48ece20 JW |
3267 | switch_back: |
3268 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3269 | blk_mq_elv_switch_back(&head, q); | |
3270 | ||
868f2f0b KB |
3271 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
3272 | blk_mq_unfreeze_queue(q); | |
3273 | } | |
e4dc2b32 KB |
3274 | |
3275 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
3276 | { | |
3277 | mutex_lock(&set->tag_list_lock); | |
3278 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
3279 | mutex_unlock(&set->tag_list_lock); | |
3280 | } | |
868f2f0b KB |
3281 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
3282 | ||
34dbad5d OS |
3283 | /* Enable polling stats and return whether they were already enabled. */ |
3284 | static bool blk_poll_stats_enable(struct request_queue *q) | |
3285 | { | |
3286 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
7dfdbc73 | 3287 | blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q)) |
34dbad5d OS |
3288 | return true; |
3289 | blk_stat_add_callback(q, q->poll_cb); | |
3290 | return false; | |
3291 | } | |
3292 | ||
3293 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
3294 | { | |
3295 | /* | |
3296 | * We don't arm the callback if polling stats are not enabled or the | |
3297 | * callback is already active. | |
3298 | */ | |
3299 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
3300 | blk_stat_is_active(q->poll_cb)) | |
3301 | return; | |
3302 | ||
3303 | blk_stat_activate_msecs(q->poll_cb, 100); | |
3304 | } | |
3305 | ||
3306 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
3307 | { | |
3308 | struct request_queue *q = cb->data; | |
720b8ccc | 3309 | int bucket; |
34dbad5d | 3310 | |
720b8ccc SB |
3311 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
3312 | if (cb->stat[bucket].nr_samples) | |
3313 | q->poll_stat[bucket] = cb->stat[bucket]; | |
3314 | } | |
34dbad5d OS |
3315 | } |
3316 | ||
64f1c21e JA |
3317 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
3318 | struct blk_mq_hw_ctx *hctx, | |
3319 | struct request *rq) | |
3320 | { | |
64f1c21e | 3321 | unsigned long ret = 0; |
720b8ccc | 3322 | int bucket; |
64f1c21e JA |
3323 | |
3324 | /* | |
3325 | * If stats collection isn't on, don't sleep but turn it on for | |
3326 | * future users | |
3327 | */ | |
34dbad5d | 3328 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
3329 | return 0; |
3330 | ||
64f1c21e JA |
3331 | /* |
3332 | * As an optimistic guess, use half of the mean service time | |
3333 | * for this type of request. We can (and should) make this smarter. | |
3334 | * For instance, if the completion latencies are tight, we can | |
3335 | * get closer than just half the mean. This is especially | |
3336 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
3337 | * than ~10 usec. We do use the stats for the relevant IO size |
3338 | * if available which does lead to better estimates. | |
64f1c21e | 3339 | */ |
720b8ccc SB |
3340 | bucket = blk_mq_poll_stats_bkt(rq); |
3341 | if (bucket < 0) | |
3342 | return ret; | |
3343 | ||
3344 | if (q->poll_stat[bucket].nr_samples) | |
3345 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
3346 | |
3347 | return ret; | |
3348 | } | |
3349 | ||
06426adf | 3350 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 3351 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
3352 | struct request *rq) |
3353 | { | |
3354 | struct hrtimer_sleeper hs; | |
3355 | enum hrtimer_mode mode; | |
64f1c21e | 3356 | unsigned int nsecs; |
06426adf JA |
3357 | ktime_t kt; |
3358 | ||
76a86f9d | 3359 | if (rq->rq_flags & RQF_MQ_POLL_SLEPT) |
64f1c21e JA |
3360 | return false; |
3361 | ||
3362 | /* | |
1052b8ac | 3363 | * If we get here, hybrid polling is enabled. Hence poll_nsec can be: |
64f1c21e | 3364 | * |
64f1c21e JA |
3365 | * 0: use half of prev avg |
3366 | * >0: use this specific value | |
3367 | */ | |
1052b8ac | 3368 | if (q->poll_nsec > 0) |
64f1c21e JA |
3369 | nsecs = q->poll_nsec; |
3370 | else | |
3371 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
3372 | ||
3373 | if (!nsecs) | |
06426adf JA |
3374 | return false; |
3375 | ||
76a86f9d | 3376 | rq->rq_flags |= RQF_MQ_POLL_SLEPT; |
06426adf JA |
3377 | |
3378 | /* | |
3379 | * This will be replaced with the stats tracking code, using | |
3380 | * 'avg_completion_time / 2' as the pre-sleep target. | |
3381 | */ | |
8b0e1953 | 3382 | kt = nsecs; |
06426adf JA |
3383 | |
3384 | mode = HRTIMER_MODE_REL; | |
3385 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
3386 | hrtimer_set_expires(&hs.timer, kt); | |
3387 | ||
3388 | hrtimer_init_sleeper(&hs, current); | |
3389 | do { | |
5a61c363 | 3390 | if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE) |
06426adf JA |
3391 | break; |
3392 | set_current_state(TASK_UNINTERRUPTIBLE); | |
3393 | hrtimer_start_expires(&hs.timer, mode); | |
3394 | if (hs.task) | |
3395 | io_schedule(); | |
3396 | hrtimer_cancel(&hs.timer); | |
3397 | mode = HRTIMER_MODE_ABS; | |
3398 | } while (hs.task && !signal_pending(current)); | |
3399 | ||
3400 | __set_current_state(TASK_RUNNING); | |
3401 | destroy_hrtimer_on_stack(&hs.timer); | |
3402 | return true; | |
3403 | } | |
3404 | ||
1052b8ac JA |
3405 | static bool blk_mq_poll_hybrid(struct request_queue *q, |
3406 | struct blk_mq_hw_ctx *hctx, blk_qc_t cookie) | |
bbd7bb70 | 3407 | { |
1052b8ac JA |
3408 | struct request *rq; |
3409 | ||
29ece8b4 | 3410 | if (q->poll_nsec == BLK_MQ_POLL_CLASSIC) |
1052b8ac JA |
3411 | return false; |
3412 | ||
3413 | if (!blk_qc_t_is_internal(cookie)) | |
3414 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
3415 | else { | |
3416 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); | |
3417 | /* | |
3418 | * With scheduling, if the request has completed, we'll | |
3419 | * get a NULL return here, as we clear the sched tag when | |
3420 | * that happens. The request still remains valid, like always, | |
3421 | * so we should be safe with just the NULL check. | |
3422 | */ | |
3423 | if (!rq) | |
3424 | return false; | |
3425 | } | |
3426 | ||
3427 | return blk_mq_poll_hybrid_sleep(q, hctx, rq); | |
3428 | } | |
3429 | ||
529262d5 CH |
3430 | /** |
3431 | * blk_poll - poll for IO completions | |
3432 | * @q: the queue | |
3433 | * @cookie: cookie passed back at IO submission time | |
3434 | * @spin: whether to spin for completions | |
3435 | * | |
3436 | * Description: | |
3437 | * Poll for completions on the passed in queue. Returns number of | |
3438 | * completed entries found. If @spin is true, then blk_poll will continue | |
3439 | * looping until at least one completion is found, unless the task is | |
3440 | * otherwise marked running (or we need to reschedule). | |
3441 | */ | |
3442 | int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin) | |
1052b8ac JA |
3443 | { |
3444 | struct blk_mq_hw_ctx *hctx; | |
bbd7bb70 JA |
3445 | long state; |
3446 | ||
529262d5 CH |
3447 | if (!blk_qc_t_valid(cookie) || |
3448 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
1052b8ac JA |
3449 | return 0; |
3450 | ||
529262d5 CH |
3451 | if (current->plug) |
3452 | blk_flush_plug_list(current->plug, false); | |
3453 | ||
1052b8ac JA |
3454 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; |
3455 | ||
06426adf JA |
3456 | /* |
3457 | * If we sleep, have the caller restart the poll loop to reset | |
3458 | * the state. Like for the other success return cases, the | |
3459 | * caller is responsible for checking if the IO completed. If | |
3460 | * the IO isn't complete, we'll get called again and will go | |
3461 | * straight to the busy poll loop. | |
3462 | */ | |
1052b8ac | 3463 | if (blk_mq_poll_hybrid(q, hctx, cookie)) |
85f4d4b6 | 3464 | return 1; |
06426adf | 3465 | |
bbd7bb70 JA |
3466 | hctx->poll_considered++; |
3467 | ||
3468 | state = current->state; | |
aa61bec3 | 3469 | do { |
bbd7bb70 JA |
3470 | int ret; |
3471 | ||
3472 | hctx->poll_invoked++; | |
3473 | ||
9743139c | 3474 | ret = q->mq_ops->poll(hctx); |
bbd7bb70 JA |
3475 | if (ret > 0) { |
3476 | hctx->poll_success++; | |
849a3700 | 3477 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 3478 | return ret; |
bbd7bb70 JA |
3479 | } |
3480 | ||
3481 | if (signal_pending_state(state, current)) | |
849a3700 | 3482 | __set_current_state(TASK_RUNNING); |
bbd7bb70 JA |
3483 | |
3484 | if (current->state == TASK_RUNNING) | |
85f4d4b6 | 3485 | return 1; |
0a1b8b87 | 3486 | if (ret < 0 || !spin) |
bbd7bb70 JA |
3487 | break; |
3488 | cpu_relax(); | |
aa61bec3 | 3489 | } while (!need_resched()); |
bbd7bb70 | 3490 | |
67b4110f | 3491 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 3492 | return 0; |
bbd7bb70 | 3493 | } |
529262d5 | 3494 | EXPORT_SYMBOL_GPL(blk_poll); |
bbd7bb70 | 3495 | |
9cf2bab6 JA |
3496 | unsigned int blk_mq_rq_cpu(struct request *rq) |
3497 | { | |
3498 | return rq->mq_ctx->cpu; | |
3499 | } | |
3500 | EXPORT_SYMBOL(blk_mq_rq_cpu); | |
3501 | ||
320ae51f JA |
3502 | static int __init blk_mq_init(void) |
3503 | { | |
9467f859 TG |
3504 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
3505 | blk_mq_hctx_notify_dead); | |
320ae51f JA |
3506 | return 0; |
3507 | } | |
3508 | subsys_initcall(blk_mq_init); |