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 JA |
61 | /* |
62 | * Check if any of the ctx's have pending work in this hardware queue | |
63 | */ | |
79f720a7 | 64 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 65 | { |
79f720a7 JA |
66 | return !list_empty_careful(&hctx->dispatch) || |
67 | sbitmap_any_bit_set(&hctx->ctx_map) || | |
bd166ef1 | 68 | blk_mq_sched_has_work(hctx); |
1429d7c9 JA |
69 | } |
70 | ||
320ae51f JA |
71 | /* |
72 | * Mark this ctx as having pending work in this hardware queue | |
73 | */ | |
74 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
75 | struct blk_mq_ctx *ctx) | |
76 | { | |
f31967f0 JA |
77 | const int bit = ctx->index_hw[hctx->type]; |
78 | ||
79 | if (!sbitmap_test_bit(&hctx->ctx_map, bit)) | |
80 | sbitmap_set_bit(&hctx->ctx_map, bit); | |
1429d7c9 JA |
81 | } |
82 | ||
83 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
84 | struct blk_mq_ctx *ctx) | |
85 | { | |
f31967f0 JA |
86 | const int bit = ctx->index_hw[hctx->type]; |
87 | ||
88 | sbitmap_clear_bit(&hctx->ctx_map, bit); | |
320ae51f JA |
89 | } |
90 | ||
f299b7c7 JA |
91 | struct mq_inflight { |
92 | struct hd_struct *part; | |
93 | unsigned int *inflight; | |
94 | }; | |
95 | ||
7baa8572 | 96 | static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx, |
f299b7c7 JA |
97 | struct request *rq, void *priv, |
98 | bool reserved) | |
99 | { | |
100 | struct mq_inflight *mi = priv; | |
101 | ||
6131837b | 102 | /* |
e016b782 | 103 | * index[0] counts the specific partition that was asked for. |
6131837b OS |
104 | */ |
105 | if (rq->part == mi->part) | |
106 | mi->inflight[0]++; | |
7baa8572 JA |
107 | |
108 | return true; | |
f299b7c7 JA |
109 | } |
110 | ||
e016b782 | 111 | unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part) |
f299b7c7 | 112 | { |
e016b782 | 113 | unsigned inflight[2]; |
f299b7c7 JA |
114 | struct mq_inflight mi = { .part = part, .inflight = inflight, }; |
115 | ||
b8d62b3a | 116 | inflight[0] = inflight[1] = 0; |
f299b7c7 | 117 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
e016b782 MP |
118 | |
119 | return inflight[0]; | |
f299b7c7 JA |
120 | } |
121 | ||
7baa8572 | 122 | static bool blk_mq_check_inflight_rw(struct blk_mq_hw_ctx *hctx, |
bf0ddaba OS |
123 | struct request *rq, void *priv, |
124 | bool reserved) | |
125 | { | |
126 | struct mq_inflight *mi = priv; | |
127 | ||
128 | if (rq->part == mi->part) | |
129 | mi->inflight[rq_data_dir(rq)]++; | |
7baa8572 JA |
130 | |
131 | return true; | |
bf0ddaba OS |
132 | } |
133 | ||
134 | void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part, | |
135 | unsigned int inflight[2]) | |
136 | { | |
137 | struct mq_inflight mi = { .part = part, .inflight = inflight, }; | |
138 | ||
139 | inflight[0] = inflight[1] = 0; | |
140 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight_rw, &mi); | |
141 | } | |
142 | ||
1671d522 | 143 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 144 | { |
4ecd4fef | 145 | int freeze_depth; |
cddd5d17 | 146 | |
4ecd4fef CH |
147 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
148 | if (freeze_depth == 1) { | |
3ef28e83 | 149 | percpu_ref_kill(&q->q_usage_counter); |
344e9ffc | 150 | if (queue_is_mq(q)) |
055f6e18 | 151 | blk_mq_run_hw_queues(q, false); |
cddd5d17 | 152 | } |
f3af020b | 153 | } |
1671d522 | 154 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
f3af020b | 155 | |
6bae363e | 156 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 157 | { |
3ef28e83 | 158 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 159 | } |
6bae363e | 160 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 161 | |
f91328c4 KB |
162 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
163 | unsigned long timeout) | |
164 | { | |
165 | return wait_event_timeout(q->mq_freeze_wq, | |
166 | percpu_ref_is_zero(&q->q_usage_counter), | |
167 | timeout); | |
168 | } | |
169 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 170 | |
f3af020b TH |
171 | /* |
172 | * Guarantee no request is in use, so we can change any data structure of | |
173 | * the queue afterward. | |
174 | */ | |
3ef28e83 | 175 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 176 | { |
3ef28e83 DW |
177 | /* |
178 | * In the !blk_mq case we are only calling this to kill the | |
179 | * q_usage_counter, otherwise this increases the freeze depth | |
180 | * and waits for it to return to zero. For this reason there is | |
181 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
182 | * exported to drivers as the only user for unfreeze is blk_mq. | |
183 | */ | |
1671d522 | 184 | blk_freeze_queue_start(q); |
f3af020b TH |
185 | blk_mq_freeze_queue_wait(q); |
186 | } | |
3ef28e83 DW |
187 | |
188 | void blk_mq_freeze_queue(struct request_queue *q) | |
189 | { | |
190 | /* | |
191 | * ...just an alias to keep freeze and unfreeze actions balanced | |
192 | * in the blk_mq_* namespace | |
193 | */ | |
194 | blk_freeze_queue(q); | |
195 | } | |
c761d96b | 196 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 197 | |
b4c6a028 | 198 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 199 | { |
4ecd4fef | 200 | int freeze_depth; |
320ae51f | 201 | |
4ecd4fef CH |
202 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
203 | WARN_ON_ONCE(freeze_depth < 0); | |
204 | if (!freeze_depth) { | |
bdd63160 | 205 | percpu_ref_resurrect(&q->q_usage_counter); |
320ae51f | 206 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 207 | } |
320ae51f | 208 | } |
b4c6a028 | 209 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 210 | |
852ec809 BVA |
211 | /* |
212 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
213 | * mpt3sas driver such that this function can be removed. | |
214 | */ | |
215 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
216 | { | |
8814ce8a | 217 | blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q); |
852ec809 BVA |
218 | } |
219 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
220 | ||
6a83e74d | 221 | /** |
69e07c4a | 222 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished |
6a83e74d BVA |
223 | * @q: request queue. |
224 | * | |
225 | * Note: this function does not prevent that the struct request end_io() | |
69e07c4a ML |
226 | * callback function is invoked. Once this function is returned, we make |
227 | * sure no dispatch can happen until the queue is unquiesced via | |
228 | * blk_mq_unquiesce_queue(). | |
6a83e74d BVA |
229 | */ |
230 | void blk_mq_quiesce_queue(struct request_queue *q) | |
231 | { | |
232 | struct blk_mq_hw_ctx *hctx; | |
233 | unsigned int i; | |
234 | bool rcu = false; | |
235 | ||
1d9e9bc6 | 236 | blk_mq_quiesce_queue_nowait(q); |
f4560ffe | 237 | |
6a83e74d BVA |
238 | queue_for_each_hw_ctx(q, hctx, i) { |
239 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
05707b64 | 240 | synchronize_srcu(hctx->srcu); |
6a83e74d BVA |
241 | else |
242 | rcu = true; | |
243 | } | |
244 | if (rcu) | |
245 | synchronize_rcu(); | |
246 | } | |
247 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
248 | ||
e4e73913 ML |
249 | /* |
250 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
251 | * @q: request queue. | |
252 | * | |
253 | * This function recovers queue into the state before quiescing | |
254 | * which is done by blk_mq_quiesce_queue. | |
255 | */ | |
256 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
257 | { | |
8814ce8a | 258 | blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q); |
f4560ffe | 259 | |
1d9e9bc6 ML |
260 | /* dispatch requests which are inserted during quiescing */ |
261 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
262 | } |
263 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
264 | ||
aed3ea94 JA |
265 | void blk_mq_wake_waiters(struct request_queue *q) |
266 | { | |
267 | struct blk_mq_hw_ctx *hctx; | |
268 | unsigned int i; | |
269 | ||
270 | queue_for_each_hw_ctx(q, hctx, i) | |
271 | if (blk_mq_hw_queue_mapped(hctx)) | |
272 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
273 | } | |
274 | ||
320ae51f JA |
275 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
276 | { | |
277 | return blk_mq_has_free_tags(hctx->tags); | |
278 | } | |
279 | EXPORT_SYMBOL(blk_mq_can_queue); | |
280 | ||
fe1f4526 JA |
281 | /* |
282 | * Only need start/end time stamping if we have stats enabled, or using | |
283 | * an IO scheduler. | |
284 | */ | |
285 | static inline bool blk_mq_need_time_stamp(struct request *rq) | |
286 | { | |
287 | return (rq->rq_flags & RQF_IO_STAT) || rq->q->elevator; | |
288 | } | |
289 | ||
e4cdf1a1 CH |
290 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
291 | unsigned int tag, unsigned int op) | |
320ae51f | 292 | { |
e4cdf1a1 CH |
293 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
294 | struct request *rq = tags->static_rqs[tag]; | |
bf9ae8c5 | 295 | req_flags_t rq_flags = 0; |
c3a148d2 | 296 | |
e4cdf1a1 CH |
297 | if (data->flags & BLK_MQ_REQ_INTERNAL) { |
298 | rq->tag = -1; | |
299 | rq->internal_tag = tag; | |
300 | } else { | |
d263ed99 | 301 | if (data->hctx->flags & BLK_MQ_F_TAG_SHARED) { |
bf9ae8c5 | 302 | rq_flags = RQF_MQ_INFLIGHT; |
e4cdf1a1 CH |
303 | atomic_inc(&data->hctx->nr_active); |
304 | } | |
305 | rq->tag = tag; | |
306 | rq->internal_tag = -1; | |
307 | data->hctx->tags->rqs[rq->tag] = rq; | |
308 | } | |
309 | ||
af76e555 | 310 | /* csd/requeue_work/fifo_time is initialized before use */ |
e4cdf1a1 CH |
311 | rq->q = data->q; |
312 | rq->mq_ctx = data->ctx; | |
ea4f995e | 313 | rq->mq_hctx = data->hctx; |
bf9ae8c5 | 314 | rq->rq_flags = rq_flags; |
ef295ecf | 315 | rq->cmd_flags = op; |
1b6d65a0 BVA |
316 | if (data->flags & BLK_MQ_REQ_PREEMPT) |
317 | rq->rq_flags |= RQF_PREEMPT; | |
e4cdf1a1 | 318 | if (blk_queue_io_stat(data->q)) |
e8064021 | 319 | rq->rq_flags |= RQF_IO_STAT; |
7c3fb70f | 320 | INIT_LIST_HEAD(&rq->queuelist); |
af76e555 CH |
321 | INIT_HLIST_NODE(&rq->hash); |
322 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
323 | rq->rq_disk = NULL; |
324 | rq->part = NULL; | |
fe1f4526 JA |
325 | if (blk_mq_need_time_stamp(rq)) |
326 | rq->start_time_ns = ktime_get_ns(); | |
327 | else | |
328 | rq->start_time_ns = 0; | |
544ccc8d | 329 | rq->io_start_time_ns = 0; |
af76e555 CH |
330 | rq->nr_phys_segments = 0; |
331 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
332 | rq->nr_integrity_segments = 0; | |
333 | #endif | |
af76e555 CH |
334 | rq->special = NULL; |
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; | |
343 | rq->next_rq = NULL; | |
344 | ||
e4cdf1a1 | 345 | data->ctx->rq_dispatched[op_is_sync(op)]++; |
12f5b931 | 346 | refcount_set(&rq->ref, 1); |
e4cdf1a1 | 347 | return rq; |
5dee8577 CH |
348 | } |
349 | ||
d2c0d383 | 350 | static struct request *blk_mq_get_request(struct request_queue *q, |
f9afca4d JA |
351 | struct bio *bio, |
352 | struct blk_mq_alloc_data *data) | |
d2c0d383 CH |
353 | { |
354 | struct elevator_queue *e = q->elevator; | |
355 | struct request *rq; | |
e4cdf1a1 | 356 | unsigned int tag; |
21e768b4 | 357 | bool put_ctx_on_error = false; |
d2c0d383 CH |
358 | |
359 | blk_queue_enter_live(q); | |
360 | data->q = q; | |
21e768b4 BVA |
361 | if (likely(!data->ctx)) { |
362 | data->ctx = blk_mq_get_ctx(q); | |
363 | put_ctx_on_error = true; | |
364 | } | |
d2c0d383 | 365 | if (likely(!data->hctx)) |
f9afca4d JA |
366 | data->hctx = blk_mq_map_queue(q, data->cmd_flags, |
367 | data->ctx->cpu); | |
368 | if (data->cmd_flags & REQ_NOWAIT) | |
03a07c92 | 369 | data->flags |= BLK_MQ_REQ_NOWAIT; |
d2c0d383 CH |
370 | |
371 | if (e) { | |
372 | data->flags |= BLK_MQ_REQ_INTERNAL; | |
373 | ||
374 | /* | |
375 | * Flush requests are special and go directly to the | |
17a51199 JA |
376 | * dispatch list. Don't include reserved tags in the |
377 | * limiting, as it isn't useful. | |
d2c0d383 | 378 | */ |
f9afca4d JA |
379 | if (!op_is_flush(data->cmd_flags) && |
380 | e->type->ops.limit_depth && | |
17a51199 | 381 | !(data->flags & BLK_MQ_REQ_RESERVED)) |
f9afca4d | 382 | e->type->ops.limit_depth(data->cmd_flags, data); |
d263ed99 JW |
383 | } else { |
384 | blk_mq_tag_busy(data->hctx); | |
d2c0d383 CH |
385 | } |
386 | ||
e4cdf1a1 CH |
387 | tag = blk_mq_get_tag(data); |
388 | if (tag == BLK_MQ_TAG_FAIL) { | |
21e768b4 BVA |
389 | if (put_ctx_on_error) { |
390 | blk_mq_put_ctx(data->ctx); | |
1ad43c00 ML |
391 | data->ctx = NULL; |
392 | } | |
037cebb8 CH |
393 | blk_queue_exit(q); |
394 | return NULL; | |
d2c0d383 CH |
395 | } |
396 | ||
f9afca4d JA |
397 | rq = blk_mq_rq_ctx_init(data, tag, data->cmd_flags); |
398 | if (!op_is_flush(data->cmd_flags)) { | |
037cebb8 | 399 | rq->elv.icq = NULL; |
f9cd4bfe | 400 | if (e && e->type->ops.prepare_request) { |
e2b3fa5a DLM |
401 | if (e->type->icq_cache) |
402 | blk_mq_sched_assign_ioc(rq); | |
44e8c2bf | 403 | |
f9cd4bfe | 404 | e->type->ops.prepare_request(rq, bio); |
5bbf4e5a | 405 | rq->rq_flags |= RQF_ELVPRIV; |
44e8c2bf | 406 | } |
037cebb8 CH |
407 | } |
408 | data->hctx->queued++; | |
409 | return rq; | |
d2c0d383 CH |
410 | } |
411 | ||
cd6ce148 | 412 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, |
9a95e4ef | 413 | blk_mq_req_flags_t flags) |
320ae51f | 414 | { |
f9afca4d | 415 | struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op }; |
bd166ef1 | 416 | struct request *rq; |
a492f075 | 417 | int ret; |
320ae51f | 418 | |
3a0a5299 | 419 | ret = blk_queue_enter(q, flags); |
a492f075 JL |
420 | if (ret) |
421 | return ERR_PTR(ret); | |
320ae51f | 422 | |
f9afca4d | 423 | rq = blk_mq_get_request(q, NULL, &alloc_data); |
3280d66a | 424 | blk_queue_exit(q); |
841bac2c | 425 | |
bd166ef1 | 426 | if (!rq) |
a492f075 | 427 | return ERR_PTR(-EWOULDBLOCK); |
0c4de0f3 | 428 | |
1ad43c00 | 429 | blk_mq_put_ctx(alloc_data.ctx); |
1ad43c00 | 430 | |
0c4de0f3 CH |
431 | rq->__data_len = 0; |
432 | rq->__sector = (sector_t) -1; | |
433 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
434 | return rq; |
435 | } | |
4bb659b1 | 436 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 437 | |
cd6ce148 | 438 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
9a95e4ef | 439 | unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 440 | { |
f9afca4d | 441 | struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op }; |
1f5bd336 | 442 | struct request *rq; |
6d2809d5 | 443 | unsigned int cpu; |
1f5bd336 ML |
444 | int ret; |
445 | ||
446 | /* | |
447 | * If the tag allocator sleeps we could get an allocation for a | |
448 | * different hardware context. No need to complicate the low level | |
449 | * allocator for this for the rare use case of a command tied to | |
450 | * a specific queue. | |
451 | */ | |
452 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
453 | return ERR_PTR(-EINVAL); | |
454 | ||
455 | if (hctx_idx >= q->nr_hw_queues) | |
456 | return ERR_PTR(-EIO); | |
457 | ||
3a0a5299 | 458 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
459 | if (ret) |
460 | return ERR_PTR(ret); | |
461 | ||
c8712c6a CH |
462 | /* |
463 | * Check if the hardware context is actually mapped to anything. | |
464 | * If not tell the caller that it should skip this queue. | |
465 | */ | |
6d2809d5 OS |
466 | alloc_data.hctx = q->queue_hw_ctx[hctx_idx]; |
467 | if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) { | |
468 | blk_queue_exit(q); | |
469 | return ERR_PTR(-EXDEV); | |
c8712c6a | 470 | } |
20e4d813 | 471 | cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask); |
6d2809d5 | 472 | alloc_data.ctx = __blk_mq_get_ctx(q, cpu); |
1f5bd336 | 473 | |
f9afca4d | 474 | rq = blk_mq_get_request(q, NULL, &alloc_data); |
3280d66a | 475 | blk_queue_exit(q); |
c8712c6a | 476 | |
6d2809d5 OS |
477 | if (!rq) |
478 | return ERR_PTR(-EWOULDBLOCK); | |
479 | ||
480 | return rq; | |
1f5bd336 ML |
481 | } |
482 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
483 | ||
12f5b931 KB |
484 | static void __blk_mq_free_request(struct request *rq) |
485 | { | |
486 | struct request_queue *q = rq->q; | |
487 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 488 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
12f5b931 KB |
489 | const int sched_tag = rq->internal_tag; |
490 | ||
986d413b | 491 | blk_pm_mark_last_busy(rq); |
ea4f995e | 492 | rq->mq_hctx = NULL; |
12f5b931 KB |
493 | if (rq->tag != -1) |
494 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
495 | if (sched_tag != -1) | |
496 | blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag); | |
497 | blk_mq_sched_restart(hctx); | |
498 | blk_queue_exit(q); | |
499 | } | |
500 | ||
6af54051 | 501 | void blk_mq_free_request(struct request *rq) |
320ae51f | 502 | { |
320ae51f | 503 | struct request_queue *q = rq->q; |
6af54051 CH |
504 | struct elevator_queue *e = q->elevator; |
505 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 506 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
6af54051 | 507 | |
5bbf4e5a | 508 | if (rq->rq_flags & RQF_ELVPRIV) { |
f9cd4bfe JA |
509 | if (e && e->type->ops.finish_request) |
510 | e->type->ops.finish_request(rq); | |
6af54051 CH |
511 | if (rq->elv.icq) { |
512 | put_io_context(rq->elv.icq->ioc); | |
513 | rq->elv.icq = NULL; | |
514 | } | |
515 | } | |
320ae51f | 516 | |
6af54051 | 517 | ctx->rq_completed[rq_is_sync(rq)]++; |
e8064021 | 518 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 519 | atomic_dec(&hctx->nr_active); |
87760e5e | 520 | |
7beb2f84 JA |
521 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
522 | laptop_io_completion(q->backing_dev_info); | |
523 | ||
a7905043 | 524 | rq_qos_done(q, rq); |
0d2602ca | 525 | |
12f5b931 KB |
526 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
527 | if (refcount_dec_and_test(&rq->ref)) | |
528 | __blk_mq_free_request(rq); | |
320ae51f | 529 | } |
1a3b595a | 530 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 531 | |
2a842aca | 532 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
320ae51f | 533 | { |
fe1f4526 JA |
534 | u64 now = 0; |
535 | ||
536 | if (blk_mq_need_time_stamp(rq)) | |
537 | now = ktime_get_ns(); | |
522a7775 | 538 | |
4bc6339a OS |
539 | if (rq->rq_flags & RQF_STATS) { |
540 | blk_mq_poll_stats_start(rq->q); | |
522a7775 | 541 | blk_stat_add(rq, now); |
4bc6339a OS |
542 | } |
543 | ||
ed88660a OS |
544 | if (rq->internal_tag != -1) |
545 | blk_mq_sched_completed_request(rq, now); | |
546 | ||
522a7775 | 547 | blk_account_io_done(rq, now); |
0d11e6ac | 548 | |
91b63639 | 549 | if (rq->end_io) { |
a7905043 | 550 | rq_qos_done(rq->q, rq); |
320ae51f | 551 | rq->end_io(rq, error); |
91b63639 CH |
552 | } else { |
553 | if (unlikely(blk_bidi_rq(rq))) | |
554 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 555 | blk_mq_free_request(rq); |
91b63639 | 556 | } |
320ae51f | 557 | } |
c8a446ad | 558 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 559 | |
2a842aca | 560 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
561 | { |
562 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
563 | BUG(); | |
c8a446ad | 564 | __blk_mq_end_request(rq, error); |
63151a44 | 565 | } |
c8a446ad | 566 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 567 | |
30a91cb4 | 568 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 569 | { |
3d6efbf6 | 570 | struct request *rq = data; |
c7bb9ad1 | 571 | struct request_queue *q = rq->q; |
320ae51f | 572 | |
c7bb9ad1 | 573 | q->mq_ops->complete(rq); |
320ae51f | 574 | } |
320ae51f | 575 | |
453f8341 | 576 | static void __blk_mq_complete_request(struct request *rq) |
320ae51f JA |
577 | { |
578 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
c7bb9ad1 | 579 | struct request_queue *q = rq->q; |
38535201 | 580 | bool shared = false; |
320ae51f JA |
581 | int cpu; |
582 | ||
af78ff7c | 583 | WRITE_ONCE(rq->state, MQ_RQ_COMPLETE); |
36e76539 ML |
584 | /* |
585 | * Most of single queue controllers, there is only one irq vector | |
586 | * for handling IO completion, and the only irq's affinity is set | |
587 | * as all possible CPUs. On most of ARCHs, this affinity means the | |
588 | * irq is handled on one specific CPU. | |
589 | * | |
590 | * So complete IO reqeust in softirq context in case of single queue | |
591 | * for not degrading IO performance by irqsoff latency. | |
592 | */ | |
c7bb9ad1 | 593 | if (q->nr_hw_queues == 1) { |
36e76539 ML |
594 | __blk_complete_request(rq); |
595 | return; | |
596 | } | |
597 | ||
4ab32bf3 JA |
598 | /* |
599 | * For a polled request, always complete locallly, it's pointless | |
600 | * to redirect the completion. | |
601 | */ | |
602 | if ((rq->cmd_flags & REQ_HIPRI) || | |
603 | !test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags)) { | |
c7bb9ad1 | 604 | q->mq_ops->complete(rq); |
30a91cb4 CH |
605 | return; |
606 | } | |
320ae51f JA |
607 | |
608 | cpu = get_cpu(); | |
c7bb9ad1 | 609 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &q->queue_flags)) |
38535201 CH |
610 | shared = cpus_share_cache(cpu, ctx->cpu); |
611 | ||
612 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 613 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
614 | rq->csd.info = rq; |
615 | rq->csd.flags = 0; | |
c46fff2a | 616 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 617 | } else { |
c7bb9ad1 | 618 | q->mq_ops->complete(rq); |
3d6efbf6 | 619 | } |
320ae51f JA |
620 | put_cpu(); |
621 | } | |
30a91cb4 | 622 | |
04ced159 | 623 | static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx) |
b7435db8 | 624 | __releases(hctx->srcu) |
04ced159 JA |
625 | { |
626 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
627 | rcu_read_unlock(); | |
628 | else | |
05707b64 | 629 | srcu_read_unlock(hctx->srcu, srcu_idx); |
04ced159 JA |
630 | } |
631 | ||
632 | static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx) | |
b7435db8 | 633 | __acquires(hctx->srcu) |
04ced159 | 634 | { |
08b5a6e2 JA |
635 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
636 | /* shut up gcc false positive */ | |
637 | *srcu_idx = 0; | |
04ced159 | 638 | rcu_read_lock(); |
08b5a6e2 | 639 | } else |
05707b64 | 640 | *srcu_idx = srcu_read_lock(hctx->srcu); |
04ced159 JA |
641 | } |
642 | ||
30a91cb4 CH |
643 | /** |
644 | * blk_mq_complete_request - end I/O on a request | |
645 | * @rq: the request being processed | |
646 | * | |
647 | * Description: | |
648 | * Ends all I/O on a request. It does not handle partial completions. | |
649 | * The actual completion happens out-of-order, through a IPI handler. | |
650 | **/ | |
16c15eb1 | 651 | bool blk_mq_complete_request(struct request *rq) |
30a91cb4 | 652 | { |
12f5b931 | 653 | if (unlikely(blk_should_fake_timeout(rq->q))) |
16c15eb1 | 654 | return false; |
12f5b931 | 655 | __blk_mq_complete_request(rq); |
16c15eb1 | 656 | return true; |
30a91cb4 CH |
657 | } |
658 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 659 | |
973c0191 KB |
660 | int blk_mq_request_started(struct request *rq) |
661 | { | |
5a61c363 | 662 | return blk_mq_rq_state(rq) != MQ_RQ_IDLE; |
973c0191 KB |
663 | } |
664 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
665 | ||
e2490073 | 666 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
667 | { |
668 | struct request_queue *q = rq->q; | |
669 | ||
bd166ef1 JA |
670 | blk_mq_sched_started_request(rq); |
671 | ||
320ae51f JA |
672 | trace_block_rq_issue(q, rq); |
673 | ||
cf43e6be | 674 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
544ccc8d OS |
675 | rq->io_start_time_ns = ktime_get_ns(); |
676 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW | |
677 | rq->throtl_size = blk_rq_sectors(rq); | |
678 | #endif | |
cf43e6be | 679 | rq->rq_flags |= RQF_STATS; |
a7905043 | 680 | rq_qos_issue(q, rq); |
cf43e6be JA |
681 | } |
682 | ||
1d9bd516 | 683 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 684 | |
1d9bd516 | 685 | blk_add_timer(rq); |
12f5b931 | 686 | WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); |
49f5baa5 CH |
687 | |
688 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
689 | /* | |
690 | * Make sure space for the drain appears. We know we can do | |
691 | * this because max_hw_segments has been adjusted to be one | |
692 | * fewer than the device can handle. | |
693 | */ | |
694 | rq->nr_phys_segments++; | |
695 | } | |
320ae51f | 696 | } |
e2490073 | 697 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 698 | |
ed0791b2 | 699 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
700 | { |
701 | struct request_queue *q = rq->q; | |
702 | ||
923218f6 ML |
703 | blk_mq_put_driver_tag(rq); |
704 | ||
320ae51f | 705 | trace_block_rq_requeue(q, rq); |
a7905043 | 706 | rq_qos_requeue(q, rq); |
49f5baa5 | 707 | |
12f5b931 KB |
708 | if (blk_mq_request_started(rq)) { |
709 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); | |
da661267 | 710 | rq->rq_flags &= ~RQF_TIMED_OUT; |
e2490073 CH |
711 | if (q->dma_drain_size && blk_rq_bytes(rq)) |
712 | rq->nr_phys_segments--; | |
713 | } | |
320ae51f JA |
714 | } |
715 | ||
2b053aca | 716 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 717 | { |
ed0791b2 | 718 | __blk_mq_requeue_request(rq); |
ed0791b2 | 719 | |
105976f5 ML |
720 | /* this request will be re-inserted to io scheduler queue */ |
721 | blk_mq_sched_requeue_request(rq); | |
722 | ||
7d692330 | 723 | BUG_ON(!list_empty(&rq->queuelist)); |
2b053aca | 724 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
725 | } |
726 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
727 | ||
6fca6a61 CH |
728 | static void blk_mq_requeue_work(struct work_struct *work) |
729 | { | |
730 | struct request_queue *q = | |
2849450a | 731 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
732 | LIST_HEAD(rq_list); |
733 | struct request *rq, *next; | |
6fca6a61 | 734 | |
18e9781d | 735 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 736 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 737 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
738 | |
739 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
e8064021 | 740 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
6fca6a61 CH |
741 | continue; |
742 | ||
e8064021 | 743 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 744 | list_del_init(&rq->queuelist); |
9e97d295 | 745 | blk_mq_sched_insert_request(rq, true, false, false); |
6fca6a61 CH |
746 | } |
747 | ||
748 | while (!list_empty(&rq_list)) { | |
749 | rq = list_entry(rq_list.next, struct request, queuelist); | |
750 | list_del_init(&rq->queuelist); | |
9e97d295 | 751 | blk_mq_sched_insert_request(rq, false, false, false); |
6fca6a61 CH |
752 | } |
753 | ||
52d7f1b5 | 754 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
755 | } |
756 | ||
2b053aca BVA |
757 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
758 | bool kick_requeue_list) | |
6fca6a61 CH |
759 | { |
760 | struct request_queue *q = rq->q; | |
761 | unsigned long flags; | |
762 | ||
763 | /* | |
764 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
ff821d27 | 765 | * request head insertion from the workqueue. |
6fca6a61 | 766 | */ |
e8064021 | 767 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
768 | |
769 | spin_lock_irqsave(&q->requeue_lock, flags); | |
770 | if (at_head) { | |
e8064021 | 771 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
772 | list_add(&rq->queuelist, &q->requeue_list); |
773 | } else { | |
774 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
775 | } | |
776 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
777 | |
778 | if (kick_requeue_list) | |
779 | blk_mq_kick_requeue_list(q); | |
6fca6a61 CH |
780 | } |
781 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
782 | ||
783 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
784 | { | |
ae943d20 | 785 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
786 | } |
787 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
788 | ||
2849450a MS |
789 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
790 | unsigned long msecs) | |
791 | { | |
d4acf365 BVA |
792 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
793 | msecs_to_jiffies(msecs)); | |
2849450a MS |
794 | } |
795 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
796 | ||
0e62f51f JA |
797 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
798 | { | |
88c7b2b7 JA |
799 | if (tag < tags->nr_tags) { |
800 | prefetch(tags->rqs[tag]); | |
4ee86bab | 801 | return tags->rqs[tag]; |
88c7b2b7 | 802 | } |
4ee86bab HR |
803 | |
804 | return NULL; | |
24d2f903 CH |
805 | } |
806 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
807 | ||
3c94d83c JA |
808 | static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq, |
809 | void *priv, bool reserved) | |
ae879912 JA |
810 | { |
811 | /* | |
3c94d83c JA |
812 | * If we find a request that is inflight and the queue matches, |
813 | * we know the queue is busy. Return false to stop the iteration. | |
ae879912 | 814 | */ |
3c94d83c | 815 | if (rq->state == MQ_RQ_IN_FLIGHT && rq->q == hctx->queue) { |
ae879912 JA |
816 | bool *busy = priv; |
817 | ||
818 | *busy = true; | |
819 | return false; | |
820 | } | |
821 | ||
822 | return true; | |
823 | } | |
824 | ||
3c94d83c | 825 | bool blk_mq_queue_inflight(struct request_queue *q) |
ae879912 JA |
826 | { |
827 | bool busy = false; | |
828 | ||
3c94d83c | 829 | blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy); |
ae879912 JA |
830 | return busy; |
831 | } | |
3c94d83c | 832 | EXPORT_SYMBOL_GPL(blk_mq_queue_inflight); |
ae879912 | 833 | |
358f70da | 834 | static void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 835 | { |
da661267 | 836 | req->rq_flags |= RQF_TIMED_OUT; |
d1210d5a CH |
837 | if (req->q->mq_ops->timeout) { |
838 | enum blk_eh_timer_return ret; | |
839 | ||
840 | ret = req->q->mq_ops->timeout(req, reserved); | |
841 | if (ret == BLK_EH_DONE) | |
842 | return; | |
843 | WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); | |
46f92d42 | 844 | } |
d1210d5a CH |
845 | |
846 | blk_add_timer(req); | |
87ee7b11 | 847 | } |
5b3f25fc | 848 | |
12f5b931 | 849 | static bool blk_mq_req_expired(struct request *rq, unsigned long *next) |
81481eb4 | 850 | { |
12f5b931 | 851 | unsigned long deadline; |
87ee7b11 | 852 | |
12f5b931 KB |
853 | if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) |
854 | return false; | |
da661267 CH |
855 | if (rq->rq_flags & RQF_TIMED_OUT) |
856 | return false; | |
a7af0af3 | 857 | |
079076b3 | 858 | deadline = READ_ONCE(rq->deadline); |
12f5b931 KB |
859 | if (time_after_eq(jiffies, deadline)) |
860 | return true; | |
a7af0af3 | 861 | |
12f5b931 KB |
862 | if (*next == 0) |
863 | *next = deadline; | |
864 | else if (time_after(*next, deadline)) | |
865 | *next = deadline; | |
866 | return false; | |
87ee7b11 JA |
867 | } |
868 | ||
7baa8572 | 869 | static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
1d9bd516 TH |
870 | struct request *rq, void *priv, bool reserved) |
871 | { | |
12f5b931 KB |
872 | unsigned long *next = priv; |
873 | ||
874 | /* | |
875 | * Just do a quick check if it is expired before locking the request in | |
876 | * so we're not unnecessarilly synchronizing across CPUs. | |
877 | */ | |
878 | if (!blk_mq_req_expired(rq, next)) | |
7baa8572 | 879 | return true; |
12f5b931 KB |
880 | |
881 | /* | |
882 | * We have reason to believe the request may be expired. Take a | |
883 | * reference on the request to lock this request lifetime into its | |
884 | * currently allocated context to prevent it from being reallocated in | |
885 | * the event the completion by-passes this timeout handler. | |
886 | * | |
887 | * If the reference was already released, then the driver beat the | |
888 | * timeout handler to posting a natural completion. | |
889 | */ | |
890 | if (!refcount_inc_not_zero(&rq->ref)) | |
7baa8572 | 891 | return true; |
12f5b931 | 892 | |
1d9bd516 | 893 | /* |
12f5b931 KB |
894 | * The request is now locked and cannot be reallocated underneath the |
895 | * timeout handler's processing. Re-verify this exact request is truly | |
896 | * expired; if it is not expired, then the request was completed and | |
897 | * reallocated as a new request. | |
1d9bd516 | 898 | */ |
12f5b931 | 899 | if (blk_mq_req_expired(rq, next)) |
1d9bd516 | 900 | blk_mq_rq_timed_out(rq, reserved); |
12f5b931 KB |
901 | if (refcount_dec_and_test(&rq->ref)) |
902 | __blk_mq_free_request(rq); | |
7baa8572 JA |
903 | |
904 | return true; | |
1d9bd516 TH |
905 | } |
906 | ||
287922eb | 907 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 908 | { |
287922eb CH |
909 | struct request_queue *q = |
910 | container_of(work, struct request_queue, timeout_work); | |
12f5b931 | 911 | unsigned long next = 0; |
1d9bd516 | 912 | struct blk_mq_hw_ctx *hctx; |
81481eb4 | 913 | int i; |
320ae51f | 914 | |
71f79fb3 GKB |
915 | /* A deadlock might occur if a request is stuck requiring a |
916 | * timeout at the same time a queue freeze is waiting | |
917 | * completion, since the timeout code would not be able to | |
918 | * acquire the queue reference here. | |
919 | * | |
920 | * That's why we don't use blk_queue_enter here; instead, we use | |
921 | * percpu_ref_tryget directly, because we need to be able to | |
922 | * obtain a reference even in the short window between the queue | |
923 | * starting to freeze, by dropping the first reference in | |
1671d522 | 924 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
925 | * consumed, marked by the instant q_usage_counter reaches |
926 | * zero. | |
927 | */ | |
928 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
929 | return; |
930 | ||
12f5b931 | 931 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next); |
320ae51f | 932 | |
12f5b931 KB |
933 | if (next != 0) { |
934 | mod_timer(&q->timeout, next); | |
0d2602ca | 935 | } else { |
fcd36c36 BVA |
936 | /* |
937 | * Request timeouts are handled as a forward rolling timer. If | |
938 | * we end up here it means that no requests are pending and | |
939 | * also that no request has been pending for a while. Mark | |
940 | * each hctx as idle. | |
941 | */ | |
f054b56c ML |
942 | queue_for_each_hw_ctx(q, hctx, i) { |
943 | /* the hctx may be unmapped, so check it here */ | |
944 | if (blk_mq_hw_queue_mapped(hctx)) | |
945 | blk_mq_tag_idle(hctx); | |
946 | } | |
0d2602ca | 947 | } |
287922eb | 948 | blk_queue_exit(q); |
320ae51f JA |
949 | } |
950 | ||
88459642 OS |
951 | struct flush_busy_ctx_data { |
952 | struct blk_mq_hw_ctx *hctx; | |
953 | struct list_head *list; | |
954 | }; | |
955 | ||
956 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
957 | { | |
958 | struct flush_busy_ctx_data *flush_data = data; | |
959 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
960 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 961 | enum hctx_type type = hctx->type; |
88459642 | 962 | |
88459642 | 963 | spin_lock(&ctx->lock); |
c16d6b5a | 964 | list_splice_tail_init(&ctx->rq_lists[type], flush_data->list); |
e9a99a63 | 965 | sbitmap_clear_bit(sb, bitnr); |
88459642 OS |
966 | spin_unlock(&ctx->lock); |
967 | return true; | |
968 | } | |
969 | ||
1429d7c9 JA |
970 | /* |
971 | * Process software queues that have been marked busy, splicing them | |
972 | * to the for-dispatch | |
973 | */ | |
2c3ad667 | 974 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 975 | { |
88459642 OS |
976 | struct flush_busy_ctx_data data = { |
977 | .hctx = hctx, | |
978 | .list = list, | |
979 | }; | |
1429d7c9 | 980 | |
88459642 | 981 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 982 | } |
2c3ad667 | 983 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 984 | |
b347689f ML |
985 | struct dispatch_rq_data { |
986 | struct blk_mq_hw_ctx *hctx; | |
987 | struct request *rq; | |
988 | }; | |
989 | ||
990 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
991 | void *data) | |
992 | { | |
993 | struct dispatch_rq_data *dispatch_data = data; | |
994 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
995 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 996 | enum hctx_type type = hctx->type; |
b347689f ML |
997 | |
998 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
999 | if (!list_empty(&ctx->rq_lists[type])) { |
1000 | dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next); | |
b347689f | 1001 | list_del_init(&dispatch_data->rq->queuelist); |
c16d6b5a | 1002 | if (list_empty(&ctx->rq_lists[type])) |
b347689f ML |
1003 | sbitmap_clear_bit(sb, bitnr); |
1004 | } | |
1005 | spin_unlock(&ctx->lock); | |
1006 | ||
1007 | return !dispatch_data->rq; | |
1008 | } | |
1009 | ||
1010 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
1011 | struct blk_mq_ctx *start) | |
1012 | { | |
f31967f0 | 1013 | unsigned off = start ? start->index_hw[hctx->type] : 0; |
b347689f ML |
1014 | struct dispatch_rq_data data = { |
1015 | .hctx = hctx, | |
1016 | .rq = NULL, | |
1017 | }; | |
1018 | ||
1019 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1020 | dispatch_rq_from_ctx, &data); | |
1021 | ||
1022 | return data.rq; | |
1023 | } | |
1024 | ||
703fd1c0 JA |
1025 | static inline unsigned int queued_to_index(unsigned int queued) |
1026 | { | |
1027 | if (!queued) | |
1028 | return 0; | |
1429d7c9 | 1029 | |
703fd1c0 | 1030 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
1031 | } |
1032 | ||
8ab6bb9e | 1033 | bool blk_mq_get_driver_tag(struct request *rq) |
bd166ef1 JA |
1034 | { |
1035 | struct blk_mq_alloc_data data = { | |
1036 | .q = rq->q, | |
ea4f995e | 1037 | .hctx = rq->mq_hctx, |
8ab6bb9e | 1038 | .flags = BLK_MQ_REQ_NOWAIT, |
f9afca4d | 1039 | .cmd_flags = rq->cmd_flags, |
bd166ef1 | 1040 | }; |
d263ed99 | 1041 | bool shared; |
5feeacdd | 1042 | |
81380ca1 OS |
1043 | if (rq->tag != -1) |
1044 | goto done; | |
bd166ef1 | 1045 | |
415b806d SG |
1046 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) |
1047 | data.flags |= BLK_MQ_REQ_RESERVED; | |
1048 | ||
d263ed99 | 1049 | shared = blk_mq_tag_busy(data.hctx); |
bd166ef1 JA |
1050 | rq->tag = blk_mq_get_tag(&data); |
1051 | if (rq->tag >= 0) { | |
d263ed99 | 1052 | if (shared) { |
200e86b3 JA |
1053 | rq->rq_flags |= RQF_MQ_INFLIGHT; |
1054 | atomic_inc(&data.hctx->nr_active); | |
1055 | } | |
bd166ef1 | 1056 | data.hctx->tags->rqs[rq->tag] = rq; |
bd166ef1 JA |
1057 | } |
1058 | ||
81380ca1 | 1059 | done: |
81380ca1 | 1060 | return rq->tag != -1; |
bd166ef1 JA |
1061 | } |
1062 | ||
eb619fdb JA |
1063 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1064 | int flags, void *key) | |
da55f2cc OS |
1065 | { |
1066 | struct blk_mq_hw_ctx *hctx; | |
1067 | ||
1068 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1069 | ||
5815839b | 1070 | spin_lock(&hctx->dispatch_wait_lock); |
eb619fdb | 1071 | list_del_init(&wait->entry); |
5815839b ML |
1072 | spin_unlock(&hctx->dispatch_wait_lock); |
1073 | ||
da55f2cc OS |
1074 | blk_mq_run_hw_queue(hctx, true); |
1075 | return 1; | |
1076 | } | |
1077 | ||
f906a6a0 JA |
1078 | /* |
1079 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1080 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1081 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1082 | * marking us as waiting. |
1083 | */ | |
2278d69f | 1084 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, |
f906a6a0 | 1085 | struct request *rq) |
da55f2cc | 1086 | { |
5815839b | 1087 | struct wait_queue_head *wq; |
f906a6a0 JA |
1088 | wait_queue_entry_t *wait; |
1089 | bool ret; | |
da55f2cc | 1090 | |
2278d69f ML |
1091 | if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) { |
1092 | if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
1093 | set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state); | |
f906a6a0 | 1094 | |
c27d53fb BVA |
1095 | /* |
1096 | * It's possible that a tag was freed in the window between the | |
1097 | * allocation failure and adding the hardware queue to the wait | |
1098 | * queue. | |
1099 | * | |
1100 | * Don't clear RESTART here, someone else could have set it. | |
1101 | * At most this will cost an extra queue run. | |
1102 | */ | |
8ab6bb9e | 1103 | return blk_mq_get_driver_tag(rq); |
eb619fdb | 1104 | } |
eb619fdb | 1105 | |
2278d69f | 1106 | wait = &hctx->dispatch_wait; |
c27d53fb BVA |
1107 | if (!list_empty_careful(&wait->entry)) |
1108 | return false; | |
1109 | ||
5815839b ML |
1110 | wq = &bt_wait_ptr(&hctx->tags->bitmap_tags, hctx)->wait; |
1111 | ||
1112 | spin_lock_irq(&wq->lock); | |
1113 | spin_lock(&hctx->dispatch_wait_lock); | |
c27d53fb | 1114 | if (!list_empty(&wait->entry)) { |
5815839b ML |
1115 | spin_unlock(&hctx->dispatch_wait_lock); |
1116 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1117 | return false; |
eb619fdb JA |
1118 | } |
1119 | ||
5815839b ML |
1120 | wait->flags &= ~WQ_FLAG_EXCLUSIVE; |
1121 | __add_wait_queue(wq, wait); | |
c27d53fb | 1122 | |
da55f2cc | 1123 | /* |
eb619fdb JA |
1124 | * It's possible that a tag was freed in the window between the |
1125 | * allocation failure and adding the hardware queue to the wait | |
1126 | * queue. | |
da55f2cc | 1127 | */ |
8ab6bb9e | 1128 | ret = blk_mq_get_driver_tag(rq); |
c27d53fb | 1129 | if (!ret) { |
5815839b ML |
1130 | spin_unlock(&hctx->dispatch_wait_lock); |
1131 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1132 | return false; |
eb619fdb | 1133 | } |
c27d53fb BVA |
1134 | |
1135 | /* | |
1136 | * We got a tag, remove ourselves from the wait queue to ensure | |
1137 | * someone else gets the wakeup. | |
1138 | */ | |
c27d53fb | 1139 | list_del_init(&wait->entry); |
5815839b ML |
1140 | spin_unlock(&hctx->dispatch_wait_lock); |
1141 | spin_unlock_irq(&wq->lock); | |
c27d53fb BVA |
1142 | |
1143 | return true; | |
da55f2cc OS |
1144 | } |
1145 | ||
6e768717 ML |
1146 | #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 |
1147 | #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 | |
1148 | /* | |
1149 | * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): | |
1150 | * - EWMA is one simple way to compute running average value | |
1151 | * - weight(7/8 and 1/8) is applied so that it can decrease exponentially | |
1152 | * - take 4 as factor for avoiding to get too small(0) result, and this | |
1153 | * factor doesn't matter because EWMA decreases exponentially | |
1154 | */ | |
1155 | static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) | |
1156 | { | |
1157 | unsigned int ewma; | |
1158 | ||
1159 | if (hctx->queue->elevator) | |
1160 | return; | |
1161 | ||
1162 | ewma = hctx->dispatch_busy; | |
1163 | ||
1164 | if (!ewma && !busy) | |
1165 | return; | |
1166 | ||
1167 | ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; | |
1168 | if (busy) | |
1169 | ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; | |
1170 | ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; | |
1171 | ||
1172 | hctx->dispatch_busy = ewma; | |
1173 | } | |
1174 | ||
86ff7c2a ML |
1175 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
1176 | ||
1f57f8d4 JA |
1177 | /* |
1178 | * Returns true if we did some work AND can potentially do more. | |
1179 | */ | |
de148297 | 1180 | bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list, |
eb619fdb | 1181 | bool got_budget) |
320ae51f | 1182 | { |
81380ca1 | 1183 | struct blk_mq_hw_ctx *hctx; |
6d6f167c | 1184 | struct request *rq, *nxt; |
eb619fdb | 1185 | bool no_tag = false; |
fc17b653 | 1186 | int errors, queued; |
86ff7c2a | 1187 | blk_status_t ret = BLK_STS_OK; |
320ae51f | 1188 | |
81380ca1 OS |
1189 | if (list_empty(list)) |
1190 | return false; | |
1191 | ||
de148297 ML |
1192 | WARN_ON(!list_is_singular(list) && got_budget); |
1193 | ||
320ae51f JA |
1194 | /* |
1195 | * Now process all the entries, sending them to the driver. | |
1196 | */ | |
93efe981 | 1197 | errors = queued = 0; |
81380ca1 | 1198 | do { |
74c45052 | 1199 | struct blk_mq_queue_data bd; |
320ae51f | 1200 | |
f04c3df3 | 1201 | rq = list_first_entry(list, struct request, queuelist); |
0bca799b | 1202 | |
ea4f995e | 1203 | hctx = rq->mq_hctx; |
0bca799b ML |
1204 | if (!got_budget && !blk_mq_get_dispatch_budget(hctx)) |
1205 | break; | |
1206 | ||
8ab6bb9e | 1207 | if (!blk_mq_get_driver_tag(rq)) { |
3c782d67 | 1208 | /* |
da55f2cc | 1209 | * The initial allocation attempt failed, so we need to |
eb619fdb JA |
1210 | * rerun the hardware queue when a tag is freed. The |
1211 | * waitqueue takes care of that. If the queue is run | |
1212 | * before we add this entry back on the dispatch list, | |
1213 | * we'll re-run it below. | |
3c782d67 | 1214 | */ |
2278d69f | 1215 | if (!blk_mq_mark_tag_wait(hctx, rq)) { |
0bca799b | 1216 | blk_mq_put_dispatch_budget(hctx); |
f906a6a0 JA |
1217 | /* |
1218 | * For non-shared tags, the RESTART check | |
1219 | * will suffice. | |
1220 | */ | |
1221 | if (hctx->flags & BLK_MQ_F_TAG_SHARED) | |
1222 | no_tag = true; | |
de148297 ML |
1223 | break; |
1224 | } | |
1225 | } | |
1226 | ||
320ae51f | 1227 | list_del_init(&rq->queuelist); |
320ae51f | 1228 | |
74c45052 | 1229 | bd.rq = rq; |
113285b4 JA |
1230 | |
1231 | /* | |
1232 | * Flag last if we have no more requests, or if we have more | |
1233 | * but can't assign a driver tag to it. | |
1234 | */ | |
1235 | if (list_empty(list)) | |
1236 | bd.last = true; | |
1237 | else { | |
113285b4 | 1238 | nxt = list_first_entry(list, struct request, queuelist); |
8ab6bb9e | 1239 | bd.last = !blk_mq_get_driver_tag(nxt); |
113285b4 | 1240 | } |
74c45052 JA |
1241 | |
1242 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
86ff7c2a | 1243 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) { |
6d6f167c JW |
1244 | /* |
1245 | * If an I/O scheduler has been configured and we got a | |
ff821d27 JA |
1246 | * driver tag for the next request already, free it |
1247 | * again. | |
6d6f167c JW |
1248 | */ |
1249 | if (!list_empty(list)) { | |
1250 | nxt = list_first_entry(list, struct request, queuelist); | |
1251 | blk_mq_put_driver_tag(nxt); | |
1252 | } | |
f04c3df3 | 1253 | list_add(&rq->queuelist, list); |
ed0791b2 | 1254 | __blk_mq_requeue_request(rq); |
320ae51f | 1255 | break; |
fc17b653 CH |
1256 | } |
1257 | ||
1258 | if (unlikely(ret != BLK_STS_OK)) { | |
93efe981 | 1259 | errors++; |
2a842aca | 1260 | blk_mq_end_request(rq, BLK_STS_IOERR); |
fc17b653 | 1261 | continue; |
320ae51f JA |
1262 | } |
1263 | ||
fc17b653 | 1264 | queued++; |
81380ca1 | 1265 | } while (!list_empty(list)); |
320ae51f | 1266 | |
703fd1c0 | 1267 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
1268 | |
1269 | /* | |
1270 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1271 | * that is where we will continue on next queue run. | |
1272 | */ | |
f04c3df3 | 1273 | if (!list_empty(list)) { |
86ff7c2a ML |
1274 | bool needs_restart; |
1275 | ||
d666ba98 JA |
1276 | /* |
1277 | * If we didn't flush the entire list, we could have told | |
1278 | * the driver there was more coming, but that turned out to | |
1279 | * be a lie. | |
1280 | */ | |
1281 | if (q->mq_ops->commit_rqs) | |
1282 | q->mq_ops->commit_rqs(hctx); | |
1283 | ||
320ae51f | 1284 | spin_lock(&hctx->lock); |
c13660a0 | 1285 | list_splice_init(list, &hctx->dispatch); |
320ae51f | 1286 | spin_unlock(&hctx->lock); |
f04c3df3 | 1287 | |
9ba52e58 | 1288 | /* |
710c785f BVA |
1289 | * If SCHED_RESTART was set by the caller of this function and |
1290 | * it is no longer set that means that it was cleared by another | |
1291 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1292 | * |
eb619fdb JA |
1293 | * If 'no_tag' is set, that means that we failed getting |
1294 | * a driver tag with an I/O scheduler attached. If our dispatch | |
1295 | * waitqueue is no longer active, ensure that we run the queue | |
1296 | * AFTER adding our entries back to the list. | |
bd166ef1 | 1297 | * |
710c785f BVA |
1298 | * If no I/O scheduler has been configured it is possible that |
1299 | * the hardware queue got stopped and restarted before requests | |
1300 | * were pushed back onto the dispatch list. Rerun the queue to | |
1301 | * avoid starvation. Notes: | |
1302 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1303 | * been stopped before rerunning a queue. | |
1304 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1305 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1306 | * and dm-rq. |
86ff7c2a ML |
1307 | * |
1308 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
1309 | * bit is set, run queue after a delay to avoid IO stalls | |
1310 | * that could otherwise occur if the queue is idle. | |
bd166ef1 | 1311 | */ |
86ff7c2a ML |
1312 | needs_restart = blk_mq_sched_needs_restart(hctx); |
1313 | if (!needs_restart || | |
eb619fdb | 1314 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
bd166ef1 | 1315 | blk_mq_run_hw_queue(hctx, true); |
86ff7c2a ML |
1316 | else if (needs_restart && (ret == BLK_STS_RESOURCE)) |
1317 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); | |
1f57f8d4 | 1318 | |
6e768717 | 1319 | blk_mq_update_dispatch_busy(hctx, true); |
1f57f8d4 | 1320 | return false; |
6e768717 ML |
1321 | } else |
1322 | blk_mq_update_dispatch_busy(hctx, false); | |
f04c3df3 | 1323 | |
1f57f8d4 JA |
1324 | /* |
1325 | * If the host/device is unable to accept more work, inform the | |
1326 | * caller of that. | |
1327 | */ | |
1328 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) | |
1329 | return false; | |
1330 | ||
93efe981 | 1331 | return (queued + errors) != 0; |
f04c3df3 JA |
1332 | } |
1333 | ||
6a83e74d BVA |
1334 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1335 | { | |
1336 | int srcu_idx; | |
1337 | ||
b7a71e66 JA |
1338 | /* |
1339 | * We should be running this queue from one of the CPUs that | |
1340 | * are mapped to it. | |
7df938fb ML |
1341 | * |
1342 | * There are at least two related races now between setting | |
1343 | * hctx->next_cpu from blk_mq_hctx_next_cpu() and running | |
1344 | * __blk_mq_run_hw_queue(): | |
1345 | * | |
1346 | * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(), | |
1347 | * but later it becomes online, then this warning is harmless | |
1348 | * at all | |
1349 | * | |
1350 | * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(), | |
1351 | * but later it becomes offline, then the warning can't be | |
1352 | * triggered, and we depend on blk-mq timeout handler to | |
1353 | * handle dispatched requests to this hctx | |
b7a71e66 | 1354 | */ |
7df938fb ML |
1355 | if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && |
1356 | cpu_online(hctx->next_cpu)) { | |
1357 | printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n", | |
1358 | raw_smp_processor_id(), | |
1359 | cpumask_empty(hctx->cpumask) ? "inactive": "active"); | |
1360 | dump_stack(); | |
1361 | } | |
6a83e74d | 1362 | |
b7a71e66 JA |
1363 | /* |
1364 | * We can't run the queue inline with ints disabled. Ensure that | |
1365 | * we catch bad users of this early. | |
1366 | */ | |
1367 | WARN_ON_ONCE(in_interrupt()); | |
1368 | ||
04ced159 | 1369 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1370 | |
04ced159 JA |
1371 | hctx_lock(hctx, &srcu_idx); |
1372 | blk_mq_sched_dispatch_requests(hctx); | |
1373 | hctx_unlock(hctx, srcu_idx); | |
6a83e74d BVA |
1374 | } |
1375 | ||
f82ddf19 ML |
1376 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
1377 | { | |
1378 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
1379 | ||
1380 | if (cpu >= nr_cpu_ids) | |
1381 | cpu = cpumask_first(hctx->cpumask); | |
1382 | return cpu; | |
1383 | } | |
1384 | ||
506e931f JA |
1385 | /* |
1386 | * It'd be great if the workqueue API had a way to pass | |
1387 | * in a mask and had some smarts for more clever placement. | |
1388 | * For now we just round-robin here, switching for every | |
1389 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1390 | */ | |
1391 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1392 | { | |
7bed4595 | 1393 | bool tried = false; |
476f8c98 | 1394 | int next_cpu = hctx->next_cpu; |
7bed4595 | 1395 | |
b657d7e6 CH |
1396 | if (hctx->queue->nr_hw_queues == 1) |
1397 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1398 | |
1399 | if (--hctx->next_cpu_batch <= 0) { | |
7bed4595 | 1400 | select_cpu: |
476f8c98 | 1401 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
20e4d813 | 1402 | cpu_online_mask); |
506e931f | 1403 | if (next_cpu >= nr_cpu_ids) |
f82ddf19 | 1404 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
1405 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
1406 | } | |
1407 | ||
7bed4595 ML |
1408 | /* |
1409 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
1410 | * and it should only happen in the path of handling CPU DEAD. | |
1411 | */ | |
476f8c98 | 1412 | if (!cpu_online(next_cpu)) { |
7bed4595 ML |
1413 | if (!tried) { |
1414 | tried = true; | |
1415 | goto select_cpu; | |
1416 | } | |
1417 | ||
1418 | /* | |
1419 | * Make sure to re-select CPU next time once after CPUs | |
1420 | * in hctx->cpumask become online again. | |
1421 | */ | |
476f8c98 | 1422 | hctx->next_cpu = next_cpu; |
7bed4595 ML |
1423 | hctx->next_cpu_batch = 1; |
1424 | return WORK_CPU_UNBOUND; | |
1425 | } | |
476f8c98 ML |
1426 | |
1427 | hctx->next_cpu = next_cpu; | |
1428 | return next_cpu; | |
506e931f JA |
1429 | } |
1430 | ||
7587a5ae BVA |
1431 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1432 | unsigned long msecs) | |
320ae51f | 1433 | { |
5435c023 | 1434 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f JA |
1435 | return; |
1436 | ||
1b792f2f | 1437 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1438 | int cpu = get_cpu(); |
1439 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1440 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1441 | put_cpu(); |
398205b8 PB |
1442 | return; |
1443 | } | |
e4043dcf | 1444 | |
2a90d4aa | 1445 | put_cpu(); |
e4043dcf | 1446 | } |
398205b8 | 1447 | |
ae943d20 BVA |
1448 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
1449 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
1450 | } |
1451 | ||
1452 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1453 | { | |
1454 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1455 | } | |
1456 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1457 | ||
79f720a7 | 1458 | bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 1459 | { |
24f5a90f ML |
1460 | int srcu_idx; |
1461 | bool need_run; | |
1462 | ||
1463 | /* | |
1464 | * When queue is quiesced, we may be switching io scheduler, or | |
1465 | * updating nr_hw_queues, or other things, and we can't run queue | |
1466 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
1467 | * | |
1468 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
1469 | * quiesced. | |
1470 | */ | |
04ced159 JA |
1471 | hctx_lock(hctx, &srcu_idx); |
1472 | need_run = !blk_queue_quiesced(hctx->queue) && | |
1473 | blk_mq_hctx_has_pending(hctx); | |
1474 | hctx_unlock(hctx, srcu_idx); | |
24f5a90f ML |
1475 | |
1476 | if (need_run) { | |
79f720a7 JA |
1477 | __blk_mq_delay_run_hw_queue(hctx, async, 0); |
1478 | return true; | |
1479 | } | |
1480 | ||
1481 | return false; | |
320ae51f | 1482 | } |
5b727272 | 1483 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1484 | |
b94ec296 | 1485 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1486 | { |
1487 | struct blk_mq_hw_ctx *hctx; | |
1488 | int i; | |
1489 | ||
1490 | queue_for_each_hw_ctx(q, hctx, i) { | |
79f720a7 | 1491 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1492 | continue; |
1493 | ||
b94ec296 | 1494 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1495 | } |
1496 | } | |
b94ec296 | 1497 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1498 | |
fd001443 BVA |
1499 | /** |
1500 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1501 | * @q: request queue. | |
1502 | * | |
1503 | * The caller is responsible for serializing this function against | |
1504 | * blk_mq_{start,stop}_hw_queue(). | |
1505 | */ | |
1506 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1507 | { | |
1508 | struct blk_mq_hw_ctx *hctx; | |
1509 | int i; | |
1510 | ||
1511 | queue_for_each_hw_ctx(q, hctx, i) | |
1512 | if (blk_mq_hctx_stopped(hctx)) | |
1513 | return true; | |
1514 | ||
1515 | return false; | |
1516 | } | |
1517 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1518 | ||
39a70c76 ML |
1519 | /* |
1520 | * This function is often used for pausing .queue_rq() by driver when | |
1521 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1522 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1523 | * |
1524 | * We do not guarantee that dispatch can be drained or blocked | |
1525 | * after blk_mq_stop_hw_queue() returns. Please use | |
1526 | * blk_mq_quiesce_queue() for that requirement. | |
1527 | */ | |
2719aa21 JA |
1528 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1529 | { | |
641a9ed6 | 1530 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 1531 | |
641a9ed6 | 1532 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 1533 | } |
641a9ed6 | 1534 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 1535 | |
39a70c76 ML |
1536 | /* |
1537 | * This function is often used for pausing .queue_rq() by driver when | |
1538 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1539 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1540 | * |
1541 | * We do not guarantee that dispatch can be drained or blocked | |
1542 | * after blk_mq_stop_hw_queues() returns. Please use | |
1543 | * blk_mq_quiesce_queue() for that requirement. | |
1544 | */ | |
2719aa21 JA |
1545 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1546 | { | |
641a9ed6 ML |
1547 | struct blk_mq_hw_ctx *hctx; |
1548 | int i; | |
1549 | ||
1550 | queue_for_each_hw_ctx(q, hctx, i) | |
1551 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
1552 | } |
1553 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1554 | ||
320ae51f JA |
1555 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1556 | { | |
1557 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1558 | |
0ffbce80 | 1559 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1560 | } |
1561 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1562 | ||
2f268556 CH |
1563 | void blk_mq_start_hw_queues(struct request_queue *q) |
1564 | { | |
1565 | struct blk_mq_hw_ctx *hctx; | |
1566 | int i; | |
1567 | ||
1568 | queue_for_each_hw_ctx(q, hctx, i) | |
1569 | blk_mq_start_hw_queue(hctx); | |
1570 | } | |
1571 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1572 | ||
ae911c5e JA |
1573 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1574 | { | |
1575 | if (!blk_mq_hctx_stopped(hctx)) | |
1576 | return; | |
1577 | ||
1578 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1579 | blk_mq_run_hw_queue(hctx, async); | |
1580 | } | |
1581 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1582 | ||
1b4a3258 | 1583 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1584 | { |
1585 | struct blk_mq_hw_ctx *hctx; | |
1586 | int i; | |
1587 | ||
ae911c5e JA |
1588 | queue_for_each_hw_ctx(q, hctx, i) |
1589 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1590 | } |
1591 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1592 | ||
70f4db63 | 1593 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1594 | { |
1595 | struct blk_mq_hw_ctx *hctx; | |
1596 | ||
9f993737 | 1597 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 1598 | |
21c6e939 | 1599 | /* |
15fe8a90 | 1600 | * If we are stopped, don't run the queue. |
21c6e939 | 1601 | */ |
15fe8a90 | 1602 | if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
0196d6b4 | 1603 | return; |
7587a5ae BVA |
1604 | |
1605 | __blk_mq_run_hw_queue(hctx); | |
1606 | } | |
1607 | ||
cfd0c552 | 1608 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1609 | struct request *rq, |
1610 | bool at_head) | |
320ae51f | 1611 | { |
e57690fe | 1612 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
c16d6b5a | 1613 | enum hctx_type type = hctx->type; |
e57690fe | 1614 | |
7b607814 BVA |
1615 | lockdep_assert_held(&ctx->lock); |
1616 | ||
01b983c9 JA |
1617 | trace_block_rq_insert(hctx->queue, rq); |
1618 | ||
72a0a36e | 1619 | if (at_head) |
c16d6b5a | 1620 | list_add(&rq->queuelist, &ctx->rq_lists[type]); |
72a0a36e | 1621 | else |
c16d6b5a | 1622 | list_add_tail(&rq->queuelist, &ctx->rq_lists[type]); |
cfd0c552 | 1623 | } |
4bb659b1 | 1624 | |
2c3ad667 JA |
1625 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1626 | bool at_head) | |
cfd0c552 ML |
1627 | { |
1628 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1629 | ||
7b607814 BVA |
1630 | lockdep_assert_held(&ctx->lock); |
1631 | ||
e57690fe | 1632 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1633 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1634 | } |
1635 | ||
157f377b JA |
1636 | /* |
1637 | * Should only be used carefully, when the caller knows we want to | |
1638 | * bypass a potential IO scheduler on the target device. | |
1639 | */ | |
b0850297 | 1640 | void blk_mq_request_bypass_insert(struct request *rq, bool run_queue) |
157f377b | 1641 | { |
ea4f995e | 1642 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
157f377b JA |
1643 | |
1644 | spin_lock(&hctx->lock); | |
1645 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
1646 | spin_unlock(&hctx->lock); | |
1647 | ||
b0850297 ML |
1648 | if (run_queue) |
1649 | blk_mq_run_hw_queue(hctx, false); | |
157f377b JA |
1650 | } |
1651 | ||
bd166ef1 JA |
1652 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1653 | struct list_head *list) | |
320ae51f JA |
1654 | |
1655 | { | |
3f0cedc7 | 1656 | struct request *rq; |
c16d6b5a | 1657 | enum hctx_type type = hctx->type; |
3f0cedc7 | 1658 | |
320ae51f JA |
1659 | /* |
1660 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1661 | * offline now | |
1662 | */ | |
3f0cedc7 | 1663 | list_for_each_entry(rq, list, queuelist) { |
e57690fe | 1664 | BUG_ON(rq->mq_ctx != ctx); |
3f0cedc7 | 1665 | trace_block_rq_insert(hctx->queue, rq); |
320ae51f | 1666 | } |
3f0cedc7 ML |
1667 | |
1668 | spin_lock(&ctx->lock); | |
c16d6b5a | 1669 | list_splice_tail_init(list, &ctx->rq_lists[type]); |
cfd0c552 | 1670 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1671 | spin_unlock(&ctx->lock); |
320ae51f JA |
1672 | } |
1673 | ||
3110fc79 | 1674 | static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b) |
320ae51f JA |
1675 | { |
1676 | struct request *rqa = container_of(a, struct request, queuelist); | |
1677 | struct request *rqb = container_of(b, struct request, queuelist); | |
1678 | ||
3110fc79 JA |
1679 | if (rqa->mq_ctx < rqb->mq_ctx) |
1680 | return -1; | |
1681 | else if (rqa->mq_ctx > rqb->mq_ctx) | |
1682 | return 1; | |
1683 | else if (rqa->mq_hctx < rqb->mq_hctx) | |
1684 | return -1; | |
1685 | else if (rqa->mq_hctx > rqb->mq_hctx) | |
1686 | return 1; | |
1687 | ||
1688 | return blk_rq_pos(rqa) > blk_rq_pos(rqb); | |
320ae51f JA |
1689 | } |
1690 | ||
1691 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1692 | { | |
67cae4c9 | 1693 | struct blk_mq_hw_ctx *this_hctx; |
320ae51f JA |
1694 | struct blk_mq_ctx *this_ctx; |
1695 | struct request_queue *this_q; | |
1696 | struct request *rq; | |
1697 | LIST_HEAD(list); | |
67cae4c9 | 1698 | LIST_HEAD(rq_list); |
320ae51f JA |
1699 | unsigned int depth; |
1700 | ||
1701 | list_splice_init(&plug->mq_list, &list); | |
5f0ed774 | 1702 | plug->rq_count = 0; |
320ae51f | 1703 | |
ce5b009c JA |
1704 | if (plug->rq_count > 2 && plug->multiple_queues) |
1705 | list_sort(NULL, &list, plug_rq_cmp); | |
320ae51f JA |
1706 | |
1707 | this_q = NULL; | |
67cae4c9 | 1708 | this_hctx = NULL; |
320ae51f JA |
1709 | this_ctx = NULL; |
1710 | depth = 0; | |
1711 | ||
1712 | while (!list_empty(&list)) { | |
1713 | rq = list_entry_rq(list.next); | |
1714 | list_del_init(&rq->queuelist); | |
1715 | BUG_ON(!rq->q); | |
67cae4c9 JA |
1716 | if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) { |
1717 | if (this_hctx) { | |
587562d0 | 1718 | trace_block_unplug(this_q, depth, !from_schedule); |
67cae4c9 JA |
1719 | blk_mq_sched_insert_requests(this_hctx, this_ctx, |
1720 | &rq_list, | |
bd166ef1 | 1721 | from_schedule); |
320ae51f JA |
1722 | } |
1723 | ||
320ae51f | 1724 | this_q = rq->q; |
67cae4c9 JA |
1725 | this_ctx = rq->mq_ctx; |
1726 | this_hctx = rq->mq_hctx; | |
320ae51f JA |
1727 | depth = 0; |
1728 | } | |
1729 | ||
1730 | depth++; | |
67cae4c9 | 1731 | list_add_tail(&rq->queuelist, &rq_list); |
320ae51f JA |
1732 | } |
1733 | ||
1734 | /* | |
67cae4c9 JA |
1735 | * If 'this_hctx' is set, we know we have entries to complete |
1736 | * on 'rq_list'. Do those. | |
320ae51f | 1737 | */ |
67cae4c9 | 1738 | if (this_hctx) { |
587562d0 | 1739 | trace_block_unplug(this_q, depth, !from_schedule); |
67cae4c9 | 1740 | blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list, |
bd166ef1 | 1741 | from_schedule); |
320ae51f JA |
1742 | } |
1743 | } | |
1744 | ||
1745 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1746 | { | |
da8d7f07 | 1747 | blk_init_request_from_bio(rq, bio); |
4b570521 | 1748 | |
6e85eaf3 | 1749 | blk_account_io_start(rq, true); |
320ae51f JA |
1750 | } |
1751 | ||
fd2d3326 JA |
1752 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) |
1753 | { | |
bd166ef1 JA |
1754 | if (rq->tag != -1) |
1755 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); | |
1756 | ||
1757 | return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true); | |
fd2d3326 JA |
1758 | } |
1759 | ||
0f95549c MS |
1760 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
1761 | struct request *rq, | |
be94f058 | 1762 | blk_qc_t *cookie, bool last) |
f984df1f | 1763 | { |
f984df1f | 1764 | struct request_queue *q = rq->q; |
f984df1f SL |
1765 | struct blk_mq_queue_data bd = { |
1766 | .rq = rq, | |
be94f058 | 1767 | .last = last, |
f984df1f | 1768 | }; |
bd166ef1 | 1769 | blk_qc_t new_cookie; |
f06345ad | 1770 | blk_status_t ret; |
0f95549c MS |
1771 | |
1772 | new_cookie = request_to_qc_t(hctx, rq); | |
1773 | ||
1774 | /* | |
1775 | * For OK queue, we are done. For error, caller may kill it. | |
1776 | * Any other error (busy), just add it to our list as we | |
1777 | * previously would have done. | |
1778 | */ | |
1779 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
1780 | switch (ret) { | |
1781 | case BLK_STS_OK: | |
6ce3dd6e | 1782 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
1783 | *cookie = new_cookie; |
1784 | break; | |
1785 | case BLK_STS_RESOURCE: | |
86ff7c2a | 1786 | case BLK_STS_DEV_RESOURCE: |
6ce3dd6e | 1787 | blk_mq_update_dispatch_busy(hctx, true); |
0f95549c MS |
1788 | __blk_mq_requeue_request(rq); |
1789 | break; | |
1790 | default: | |
6ce3dd6e | 1791 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
1792 | *cookie = BLK_QC_T_NONE; |
1793 | break; | |
1794 | } | |
1795 | ||
1796 | return ret; | |
1797 | } | |
1798 | ||
d6a51a97 | 1799 | blk_status_t blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
0f95549c | 1800 | struct request *rq, |
396eaf21 | 1801 | blk_qc_t *cookie, |
7f556a44 | 1802 | bool bypass, bool last) |
0f95549c MS |
1803 | { |
1804 | struct request_queue *q = rq->q; | |
d964f04a | 1805 | bool run_queue = true; |
7f556a44 JW |
1806 | blk_status_t ret = BLK_STS_RESOURCE; |
1807 | int srcu_idx; | |
1808 | bool force = false; | |
d964f04a | 1809 | |
7f556a44 | 1810 | hctx_lock(hctx, &srcu_idx); |
23d4ee19 | 1811 | /* |
7f556a44 | 1812 | * hctx_lock is needed before checking quiesced flag. |
23d4ee19 | 1813 | * |
7f556a44 JW |
1814 | * When queue is stopped or quiesced, ignore 'bypass', insert |
1815 | * and return BLK_STS_OK to caller, and avoid driver to try to | |
1816 | * dispatch again. | |
23d4ee19 | 1817 | */ |
7f556a44 | 1818 | if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q))) { |
d964f04a | 1819 | run_queue = false; |
7f556a44 JW |
1820 | bypass = false; |
1821 | goto out_unlock; | |
d964f04a | 1822 | } |
f984df1f | 1823 | |
7f556a44 JW |
1824 | if (unlikely(q->elevator && !bypass)) |
1825 | goto out_unlock; | |
2253efc8 | 1826 | |
0bca799b | 1827 | if (!blk_mq_get_dispatch_budget(hctx)) |
7f556a44 | 1828 | goto out_unlock; |
bd166ef1 | 1829 | |
8ab6bb9e | 1830 | if (!blk_mq_get_driver_tag(rq)) { |
0bca799b | 1831 | blk_mq_put_dispatch_budget(hctx); |
7f556a44 | 1832 | goto out_unlock; |
88022d72 | 1833 | } |
de148297 | 1834 | |
7f556a44 JW |
1835 | /* |
1836 | * Always add a request that has been through | |
1837 | *.queue_rq() to the hardware dispatch list. | |
1838 | */ | |
1839 | force = true; | |
1840 | ret = __blk_mq_issue_directly(hctx, rq, cookie, last); | |
1841 | out_unlock: | |
04ced159 | 1842 | hctx_unlock(hctx, srcu_idx); |
7f556a44 JW |
1843 | switch (ret) { |
1844 | case BLK_STS_OK: | |
1845 | break; | |
1846 | case BLK_STS_DEV_RESOURCE: | |
1847 | case BLK_STS_RESOURCE: | |
1848 | if (force) { | |
1849 | blk_mq_request_bypass_insert(rq, run_queue); | |
1850 | /* | |
1851 | * We have to return BLK_STS_OK for the DM | |
1852 | * to avoid livelock. Otherwise, we return | |
1853 | * the real result to indicate whether the | |
1854 | * request is direct-issued successfully. | |
1855 | */ | |
1856 | ret = bypass ? BLK_STS_OK : ret; | |
1857 | } else if (!bypass) { | |
1858 | blk_mq_sched_insert_request(rq, false, | |
1859 | run_queue, false); | |
1860 | } | |
1861 | break; | |
1862 | default: | |
1863 | if (!bypass) | |
1864 | blk_mq_end_request(rq, ret); | |
1865 | break; | |
1866 | } | |
1867 | ||
1868 | return ret; | |
5eb6126e CH |
1869 | } |
1870 | ||
6ce3dd6e ML |
1871 | void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
1872 | struct list_head *list) | |
1873 | { | |
5b7a6f12 JW |
1874 | blk_qc_t unused; |
1875 | blk_status_t ret = BLK_STS_OK; | |
1876 | ||
6ce3dd6e | 1877 | while (!list_empty(list)) { |
6ce3dd6e ML |
1878 | struct request *rq = list_first_entry(list, struct request, |
1879 | queuelist); | |
1880 | ||
1881 | list_del_init(&rq->queuelist); | |
5b7a6f12 JW |
1882 | if (ret == BLK_STS_OK) |
1883 | ret = blk_mq_try_issue_directly(hctx, rq, &unused, | |
1884 | false, | |
c616cbee | 1885 | list_empty(list)); |
5b7a6f12 JW |
1886 | else |
1887 | blk_mq_sched_insert_request(rq, false, true, false); | |
6ce3dd6e | 1888 | } |
d666ba98 JA |
1889 | |
1890 | /* | |
1891 | * If we didn't flush the entire list, we could have told | |
1892 | * the driver there was more coming, but that turned out to | |
1893 | * be a lie. | |
1894 | */ | |
5b7a6f12 | 1895 | if (ret != BLK_STS_OK && hctx->queue->mq_ops->commit_rqs) |
d666ba98 | 1896 | hctx->queue->mq_ops->commit_rqs(hctx); |
6ce3dd6e ML |
1897 | } |
1898 | ||
ce5b009c JA |
1899 | static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq) |
1900 | { | |
1901 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1902 | plug->rq_count++; | |
1903 | if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) { | |
1904 | struct request *tmp; | |
1905 | ||
1906 | tmp = list_first_entry(&plug->mq_list, struct request, | |
1907 | queuelist); | |
1908 | if (tmp->q != rq->q) | |
1909 | plug->multiple_queues = true; | |
1910 | } | |
1911 | } | |
1912 | ||
dece1635 | 1913 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1914 | { |
ef295ecf | 1915 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1916 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
f9afca4d | 1917 | struct blk_mq_alloc_data data = { .flags = 0, .cmd_flags = bio->bi_opf }; |
07068d5b | 1918 | struct request *rq; |
f984df1f | 1919 | struct blk_plug *plug; |
5b3f341f | 1920 | struct request *same_queue_rq = NULL; |
7b371636 | 1921 | blk_qc_t cookie; |
07068d5b JA |
1922 | |
1923 | blk_queue_bounce(q, &bio); | |
1924 | ||
af67c31f | 1925 | blk_queue_split(q, &bio); |
f36ea50c | 1926 | |
e23947bd | 1927 | if (!bio_integrity_prep(bio)) |
dece1635 | 1928 | return BLK_QC_T_NONE; |
07068d5b | 1929 | |
87c279e6 | 1930 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
5f0ed774 | 1931 | blk_attempt_plug_merge(q, bio, &same_queue_rq)) |
87c279e6 | 1932 | return BLK_QC_T_NONE; |
f984df1f | 1933 | |
bd166ef1 JA |
1934 | if (blk_mq_sched_bio_merge(q, bio)) |
1935 | return BLK_QC_T_NONE; | |
1936 | ||
d5337560 | 1937 | rq_qos_throttle(q, bio); |
87760e5e | 1938 | |
f9afca4d | 1939 | rq = blk_mq_get_request(q, bio, &data); |
87760e5e | 1940 | if (unlikely(!rq)) { |
c1c80384 | 1941 | rq_qos_cleanup(q, bio); |
03a07c92 GR |
1942 | if (bio->bi_opf & REQ_NOWAIT) |
1943 | bio_wouldblock_error(bio); | |
dece1635 | 1944 | return BLK_QC_T_NONE; |
87760e5e JA |
1945 | } |
1946 | ||
d6f1dda2 XW |
1947 | trace_block_getrq(q, bio, bio->bi_opf); |
1948 | ||
c1c80384 | 1949 | rq_qos_track(q, rq, bio); |
07068d5b | 1950 | |
fd2d3326 | 1951 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b | 1952 | |
f984df1f | 1953 | plug = current->plug; |
07068d5b | 1954 | if (unlikely(is_flush_fua)) { |
f984df1f | 1955 | blk_mq_put_ctx(data.ctx); |
07068d5b | 1956 | blk_mq_bio_to_request(rq, bio); |
923218f6 ML |
1957 | |
1958 | /* bypass scheduler for flush rq */ | |
1959 | blk_insert_flush(rq); | |
1960 | blk_mq_run_hw_queue(data.hctx, true); | |
b2c5d16b JA |
1961 | } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs)) { |
1962 | /* | |
1963 | * Use plugging if we have a ->commit_rqs() hook as well, as | |
1964 | * we know the driver uses bd->last in a smart fashion. | |
1965 | */ | |
5f0ed774 | 1966 | unsigned int request_count = plug->rq_count; |
600271d9 SL |
1967 | struct request *last = NULL; |
1968 | ||
b00c53e8 | 1969 | blk_mq_put_ctx(data.ctx); |
e6c4438b | 1970 | blk_mq_bio_to_request(rq, bio); |
0a6219a9 | 1971 | |
676d0607 | 1972 | if (!request_count) |
e6c4438b | 1973 | trace_block_plug(q); |
600271d9 SL |
1974 | else |
1975 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c | 1976 | |
600271d9 SL |
1977 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1978 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1979 | blk_flush_plug_list(plug, false); |
1980 | trace_block_plug(q); | |
320ae51f | 1981 | } |
b094f89c | 1982 | |
ce5b009c | 1983 | blk_add_rq_to_plug(plug, rq); |
2299722c | 1984 | } else if (plug && !blk_queue_nomerges(q)) { |
bd166ef1 | 1985 | blk_mq_bio_to_request(rq, bio); |
07068d5b | 1986 | |
07068d5b | 1987 | /* |
6a83e74d | 1988 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1989 | * Otherwise the existing request in the plug list will be |
1990 | * issued. So the plug list will have one request at most | |
2299722c CH |
1991 | * The plug list might get flushed before this. If that happens, |
1992 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 1993 | */ |
2299722c CH |
1994 | if (list_empty(&plug->mq_list)) |
1995 | same_queue_rq = NULL; | |
4711b573 | 1996 | if (same_queue_rq) { |
2299722c | 1997 | list_del_init(&same_queue_rq->queuelist); |
4711b573 JA |
1998 | plug->rq_count--; |
1999 | } | |
ce5b009c | 2000 | blk_add_rq_to_plug(plug, rq); |
2299722c | 2001 | |
bf4907c0 JA |
2002 | blk_mq_put_ctx(data.ctx); |
2003 | ||
dad7a3be | 2004 | if (same_queue_rq) { |
ea4f995e | 2005 | data.hctx = same_queue_rq->mq_hctx; |
2299722c | 2006 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, |
7f556a44 | 2007 | &cookie, false, true); |
dad7a3be | 2008 | } |
6ce3dd6e ML |
2009 | } else if ((q->nr_hw_queues > 1 && is_sync) || (!q->elevator && |
2010 | !data.hctx->dispatch_busy)) { | |
bf4907c0 | 2011 | blk_mq_put_ctx(data.ctx); |
2299722c | 2012 | blk_mq_bio_to_request(rq, bio); |
7f556a44 | 2013 | blk_mq_try_issue_directly(data.hctx, rq, &cookie, false, true); |
ab42f35d | 2014 | } else { |
b00c53e8 | 2015 | blk_mq_put_ctx(data.ctx); |
ab42f35d | 2016 | blk_mq_bio_to_request(rq, bio); |
8fa9f556 | 2017 | blk_mq_sched_insert_request(rq, false, true, true); |
ab42f35d | 2018 | } |
320ae51f | 2019 | |
7b371636 | 2020 | return cookie; |
320ae51f JA |
2021 | } |
2022 | ||
cc71a6f4 JA |
2023 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2024 | unsigned int hctx_idx) | |
95363efd | 2025 | { |
e9b267d9 | 2026 | struct page *page; |
320ae51f | 2027 | |
24d2f903 | 2028 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 2029 | int i; |
320ae51f | 2030 | |
24d2f903 | 2031 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
2032 | struct request *rq = tags->static_rqs[i]; |
2033 | ||
2034 | if (!rq) | |
e9b267d9 | 2035 | continue; |
d6296d39 | 2036 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 2037 | tags->static_rqs[i] = NULL; |
e9b267d9 | 2038 | } |
320ae51f | 2039 | } |
320ae51f | 2040 | |
24d2f903 CH |
2041 | while (!list_empty(&tags->page_list)) { |
2042 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 2043 | list_del_init(&page->lru); |
f75782e4 CM |
2044 | /* |
2045 | * Remove kmemleak object previously allocated in | |
2046 | * blk_mq_init_rq_map(). | |
2047 | */ | |
2048 | kmemleak_free(page_address(page)); | |
320ae51f JA |
2049 | __free_pages(page, page->private); |
2050 | } | |
cc71a6f4 | 2051 | } |
320ae51f | 2052 | |
cc71a6f4 JA |
2053 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
2054 | { | |
24d2f903 | 2055 | kfree(tags->rqs); |
cc71a6f4 | 2056 | tags->rqs = NULL; |
2af8cbe3 JA |
2057 | kfree(tags->static_rqs); |
2058 | tags->static_rqs = NULL; | |
320ae51f | 2059 | |
24d2f903 | 2060 | blk_mq_free_tags(tags); |
320ae51f JA |
2061 | } |
2062 | ||
cc71a6f4 JA |
2063 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
2064 | unsigned int hctx_idx, | |
2065 | unsigned int nr_tags, | |
2066 | unsigned int reserved_tags) | |
320ae51f | 2067 | { |
24d2f903 | 2068 | struct blk_mq_tags *tags; |
59f082e4 | 2069 | int node; |
320ae51f | 2070 | |
ed76e329 | 2071 | node = blk_mq_hw_queue_to_node(&set->map[0], hctx_idx); |
59f082e4 SL |
2072 | if (node == NUMA_NO_NODE) |
2073 | node = set->numa_node; | |
2074 | ||
2075 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, | |
24391c0d | 2076 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
24d2f903 CH |
2077 | if (!tags) |
2078 | return NULL; | |
320ae51f | 2079 | |
590b5b7d | 2080 | tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
36e1f3d1 | 2081 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 2082 | node); |
24d2f903 CH |
2083 | if (!tags->rqs) { |
2084 | blk_mq_free_tags(tags); | |
2085 | return NULL; | |
2086 | } | |
320ae51f | 2087 | |
590b5b7d KC |
2088 | tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
2089 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
2090 | node); | |
2af8cbe3 JA |
2091 | if (!tags->static_rqs) { |
2092 | kfree(tags->rqs); | |
2093 | blk_mq_free_tags(tags); | |
2094 | return NULL; | |
2095 | } | |
2096 | ||
cc71a6f4 JA |
2097 | return tags; |
2098 | } | |
2099 | ||
2100 | static size_t order_to_size(unsigned int order) | |
2101 | { | |
2102 | return (size_t)PAGE_SIZE << order; | |
2103 | } | |
2104 | ||
1d9bd516 TH |
2105 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
2106 | unsigned int hctx_idx, int node) | |
2107 | { | |
2108 | int ret; | |
2109 | ||
2110 | if (set->ops->init_request) { | |
2111 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
2112 | if (ret) | |
2113 | return ret; | |
2114 | } | |
2115 | ||
12f5b931 | 2116 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
1d9bd516 TH |
2117 | return 0; |
2118 | } | |
2119 | ||
cc71a6f4 JA |
2120 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2121 | unsigned int hctx_idx, unsigned int depth) | |
2122 | { | |
2123 | unsigned int i, j, entries_per_page, max_order = 4; | |
2124 | size_t rq_size, left; | |
59f082e4 SL |
2125 | int node; |
2126 | ||
ed76e329 | 2127 | node = blk_mq_hw_queue_to_node(&set->map[0], hctx_idx); |
59f082e4 SL |
2128 | if (node == NUMA_NO_NODE) |
2129 | node = set->numa_node; | |
cc71a6f4 JA |
2130 | |
2131 | INIT_LIST_HEAD(&tags->page_list); | |
2132 | ||
320ae51f JA |
2133 | /* |
2134 | * rq_size is the size of the request plus driver payload, rounded | |
2135 | * to the cacheline size | |
2136 | */ | |
24d2f903 | 2137 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 2138 | cache_line_size()); |
cc71a6f4 | 2139 | left = rq_size * depth; |
320ae51f | 2140 | |
cc71a6f4 | 2141 | for (i = 0; i < depth; ) { |
320ae51f JA |
2142 | int this_order = max_order; |
2143 | struct page *page; | |
2144 | int to_do; | |
2145 | void *p; | |
2146 | ||
b3a834b1 | 2147 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
2148 | this_order--; |
2149 | ||
2150 | do { | |
59f082e4 | 2151 | page = alloc_pages_node(node, |
36e1f3d1 | 2152 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 2153 | this_order); |
320ae51f JA |
2154 | if (page) |
2155 | break; | |
2156 | if (!this_order--) | |
2157 | break; | |
2158 | if (order_to_size(this_order) < rq_size) | |
2159 | break; | |
2160 | } while (1); | |
2161 | ||
2162 | if (!page) | |
24d2f903 | 2163 | goto fail; |
320ae51f JA |
2164 | |
2165 | page->private = this_order; | |
24d2f903 | 2166 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
2167 | |
2168 | p = page_address(page); | |
f75782e4 CM |
2169 | /* |
2170 | * Allow kmemleak to scan these pages as they contain pointers | |
2171 | * to additional allocations like via ops->init_request(). | |
2172 | */ | |
36e1f3d1 | 2173 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 2174 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 2175 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
2176 | left -= to_do * rq_size; |
2177 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
2178 | struct request *rq = p; |
2179 | ||
2180 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
2181 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
2182 | tags->static_rqs[i] = NULL; | |
2183 | goto fail; | |
e9b267d9 CH |
2184 | } |
2185 | ||
320ae51f JA |
2186 | p += rq_size; |
2187 | i++; | |
2188 | } | |
2189 | } | |
cc71a6f4 | 2190 | return 0; |
320ae51f | 2191 | |
24d2f903 | 2192 | fail: |
cc71a6f4 JA |
2193 | blk_mq_free_rqs(set, tags, hctx_idx); |
2194 | return -ENOMEM; | |
320ae51f JA |
2195 | } |
2196 | ||
e57690fe JA |
2197 | /* |
2198 | * 'cpu' is going away. splice any existing rq_list entries from this | |
2199 | * software queue to the hw queue dispatch list, and ensure that it | |
2200 | * gets run. | |
2201 | */ | |
9467f859 | 2202 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 2203 | { |
9467f859 | 2204 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
2205 | struct blk_mq_ctx *ctx; |
2206 | LIST_HEAD(tmp); | |
c16d6b5a | 2207 | enum hctx_type type; |
484b4061 | 2208 | |
9467f859 | 2209 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 2210 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
c16d6b5a | 2211 | type = hctx->type; |
484b4061 JA |
2212 | |
2213 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
2214 | if (!list_empty(&ctx->rq_lists[type])) { |
2215 | list_splice_init(&ctx->rq_lists[type], &tmp); | |
484b4061 JA |
2216 | blk_mq_hctx_clear_pending(hctx, ctx); |
2217 | } | |
2218 | spin_unlock(&ctx->lock); | |
2219 | ||
2220 | if (list_empty(&tmp)) | |
9467f859 | 2221 | return 0; |
484b4061 | 2222 | |
e57690fe JA |
2223 | spin_lock(&hctx->lock); |
2224 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
2225 | spin_unlock(&hctx->lock); | |
484b4061 JA |
2226 | |
2227 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 2228 | return 0; |
484b4061 JA |
2229 | } |
2230 | ||
9467f859 | 2231 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 2232 | { |
9467f859 TG |
2233 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
2234 | &hctx->cpuhp_dead); | |
484b4061 JA |
2235 | } |
2236 | ||
c3b4afca | 2237 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
2238 | static void blk_mq_exit_hctx(struct request_queue *q, |
2239 | struct blk_mq_tag_set *set, | |
2240 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
2241 | { | |
8ab0b7dc ML |
2242 | if (blk_mq_hw_queue_mapped(hctx)) |
2243 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 2244 | |
f70ced09 | 2245 | if (set->ops->exit_request) |
d6296d39 | 2246 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); |
f70ced09 | 2247 | |
08e98fc6 ML |
2248 | if (set->ops->exit_hctx) |
2249 | set->ops->exit_hctx(hctx, hctx_idx); | |
2250 | ||
6a83e74d | 2251 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2252 | cleanup_srcu_struct(hctx->srcu); |
6a83e74d | 2253 | |
9467f859 | 2254 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 2255 | blk_free_flush_queue(hctx->fq); |
88459642 | 2256 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2257 | } |
2258 | ||
624dbe47 ML |
2259 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
2260 | struct blk_mq_tag_set *set, int nr_queue) | |
2261 | { | |
2262 | struct blk_mq_hw_ctx *hctx; | |
2263 | unsigned int i; | |
2264 | ||
2265 | queue_for_each_hw_ctx(q, hctx, i) { | |
2266 | if (i == nr_queue) | |
2267 | break; | |
477e19de | 2268 | blk_mq_debugfs_unregister_hctx(hctx); |
08e98fc6 | 2269 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 2270 | } |
624dbe47 ML |
2271 | } |
2272 | ||
08e98fc6 ML |
2273 | static int blk_mq_init_hctx(struct request_queue *q, |
2274 | struct blk_mq_tag_set *set, | |
2275 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 2276 | { |
08e98fc6 ML |
2277 | int node; |
2278 | ||
2279 | node = hctx->numa_node; | |
2280 | if (node == NUMA_NO_NODE) | |
2281 | node = hctx->numa_node = set->numa_node; | |
2282 | ||
9f993737 | 2283 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
2284 | spin_lock_init(&hctx->lock); |
2285 | INIT_LIST_HEAD(&hctx->dispatch); | |
2286 | hctx->queue = q; | |
2404e607 | 2287 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 2288 | |
9467f859 | 2289 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
2290 | |
2291 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
2292 | |
2293 | /* | |
08e98fc6 ML |
2294 | * Allocate space for all possible cpus to avoid allocation at |
2295 | * runtime | |
320ae51f | 2296 | */ |
d904bfa7 | 2297 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
5b202853 | 2298 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, node); |
08e98fc6 ML |
2299 | if (!hctx->ctxs) |
2300 | goto unregister_cpu_notifier; | |
320ae51f | 2301 | |
5b202853 JW |
2302 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), |
2303 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, node)) | |
08e98fc6 | 2304 | goto free_ctxs; |
320ae51f | 2305 | |
08e98fc6 | 2306 | hctx->nr_ctx = 0; |
320ae51f | 2307 | |
5815839b | 2308 | spin_lock_init(&hctx->dispatch_wait_lock); |
eb619fdb JA |
2309 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
2310 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
2311 | ||
08e98fc6 ML |
2312 | if (set->ops->init_hctx && |
2313 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
2314 | goto free_bitmap; | |
320ae51f | 2315 | |
5b202853 JW |
2316 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size, |
2317 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); | |
f70ced09 | 2318 | if (!hctx->fq) |
d48ece20 | 2319 | goto exit_hctx; |
320ae51f | 2320 | |
1d9bd516 | 2321 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node)) |
f70ced09 | 2322 | goto free_fq; |
320ae51f | 2323 | |
6a83e74d | 2324 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2325 | init_srcu_struct(hctx->srcu); |
6a83e74d | 2326 | |
08e98fc6 | 2327 | return 0; |
320ae51f | 2328 | |
f70ced09 ML |
2329 | free_fq: |
2330 | kfree(hctx->fq); | |
2331 | exit_hctx: | |
2332 | if (set->ops->exit_hctx) | |
2333 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 2334 | free_bitmap: |
88459642 | 2335 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2336 | free_ctxs: |
2337 | kfree(hctx->ctxs); | |
2338 | unregister_cpu_notifier: | |
9467f859 | 2339 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
2340 | return -1; |
2341 | } | |
320ae51f | 2342 | |
320ae51f JA |
2343 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
2344 | unsigned int nr_hw_queues) | |
2345 | { | |
b3c661b1 JA |
2346 | struct blk_mq_tag_set *set = q->tag_set; |
2347 | unsigned int i, j; | |
320ae51f JA |
2348 | |
2349 | for_each_possible_cpu(i) { | |
2350 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2351 | struct blk_mq_hw_ctx *hctx; | |
c16d6b5a | 2352 | int k; |
320ae51f | 2353 | |
320ae51f JA |
2354 | __ctx->cpu = i; |
2355 | spin_lock_init(&__ctx->lock); | |
c16d6b5a ML |
2356 | for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++) |
2357 | INIT_LIST_HEAD(&__ctx->rq_lists[k]); | |
2358 | ||
320ae51f JA |
2359 | __ctx->queue = q; |
2360 | ||
320ae51f JA |
2361 | /* |
2362 | * Set local node, IFF we have more than one hw queue. If | |
2363 | * not, we remain on the home node of the device | |
2364 | */ | |
b3c661b1 JA |
2365 | for (j = 0; j < set->nr_maps; j++) { |
2366 | hctx = blk_mq_map_queue_type(q, j, i); | |
2367 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
2368 | hctx->numa_node = local_memory_node(cpu_to_node(i)); | |
2369 | } | |
320ae51f JA |
2370 | } |
2371 | } | |
2372 | ||
cc71a6f4 JA |
2373 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
2374 | { | |
2375 | int ret = 0; | |
2376 | ||
2377 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
2378 | set->queue_depth, set->reserved_tags); | |
2379 | if (!set->tags[hctx_idx]) | |
2380 | return false; | |
2381 | ||
2382 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2383 | set->queue_depth); | |
2384 | if (!ret) | |
2385 | return true; | |
2386 | ||
2387 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2388 | set->tags[hctx_idx] = NULL; | |
2389 | return false; | |
2390 | } | |
2391 | ||
2392 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2393 | unsigned int hctx_idx) | |
2394 | { | |
4e6db0f2 | 2395 | if (set->tags && set->tags[hctx_idx]) { |
bd166ef1 JA |
2396 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); |
2397 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2398 | set->tags[hctx_idx] = NULL; | |
2399 | } | |
cc71a6f4 JA |
2400 | } |
2401 | ||
4b855ad3 | 2402 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 2403 | { |
b3c661b1 | 2404 | unsigned int i, j, hctx_idx; |
320ae51f JA |
2405 | struct blk_mq_hw_ctx *hctx; |
2406 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2407 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2408 | |
60de074b AM |
2409 | /* |
2410 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
2411 | */ | |
2412 | mutex_lock(&q->sysfs_lock); | |
2413 | ||
320ae51f | 2414 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 2415 | cpumask_clear(hctx->cpumask); |
320ae51f | 2416 | hctx->nr_ctx = 0; |
d416c92c | 2417 | hctx->dispatch_from = NULL; |
320ae51f JA |
2418 | } |
2419 | ||
2420 | /* | |
4b855ad3 | 2421 | * Map software to hardware queues. |
4412efec ML |
2422 | * |
2423 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 2424 | */ |
20e4d813 | 2425 | for_each_possible_cpu(i) { |
ed76e329 | 2426 | hctx_idx = set->map[0].mq_map[i]; |
4412efec ML |
2427 | /* unmapped hw queue can be remapped after CPU topo changed */ |
2428 | if (!set->tags[hctx_idx] && | |
2429 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
2430 | /* | |
2431 | * If tags initialization fail for some hctx, | |
2432 | * that hctx won't be brought online. In this | |
2433 | * case, remap the current ctx to hctx[0] which | |
2434 | * is guaranteed to always have tags allocated | |
2435 | */ | |
ed76e329 | 2436 | set->map[0].mq_map[i] = 0; |
4412efec ML |
2437 | } |
2438 | ||
897bb0c7 | 2439 | ctx = per_cpu_ptr(q->queue_ctx, i); |
b3c661b1 | 2440 | for (j = 0; j < set->nr_maps; j++) { |
e5edd5f2 ML |
2441 | if (!set->map[j].nr_queues) |
2442 | continue; | |
2443 | ||
b3c661b1 | 2444 | hctx = blk_mq_map_queue_type(q, j, i); |
f31967f0 | 2445 | |
b3c661b1 JA |
2446 | /* |
2447 | * If the CPU is already set in the mask, then we've | |
2448 | * mapped this one already. This can happen if | |
2449 | * devices share queues across queue maps. | |
2450 | */ | |
2451 | if (cpumask_test_cpu(i, hctx->cpumask)) | |
2452 | continue; | |
2453 | ||
2454 | cpumask_set_cpu(i, hctx->cpumask); | |
2455 | hctx->type = j; | |
2456 | ctx->index_hw[hctx->type] = hctx->nr_ctx; | |
2457 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2458 | ||
2459 | /* | |
2460 | * If the nr_ctx type overflows, we have exceeded the | |
2461 | * amount of sw queues we can support. | |
2462 | */ | |
2463 | BUG_ON(!hctx->nr_ctx); | |
2464 | } | |
320ae51f | 2465 | } |
506e931f | 2466 | |
60de074b AM |
2467 | mutex_unlock(&q->sysfs_lock); |
2468 | ||
506e931f | 2469 | queue_for_each_hw_ctx(q, hctx, i) { |
4412efec ML |
2470 | /* |
2471 | * If no software queues are mapped to this hardware queue, | |
2472 | * disable it and free the request entries. | |
2473 | */ | |
2474 | if (!hctx->nr_ctx) { | |
2475 | /* Never unmap queue 0. We need it as a | |
2476 | * fallback in case of a new remap fails | |
2477 | * allocation | |
2478 | */ | |
2479 | if (i && set->tags[i]) | |
2480 | blk_mq_free_map_and_requests(set, i); | |
2481 | ||
2482 | hctx->tags = NULL; | |
2483 | continue; | |
2484 | } | |
484b4061 | 2485 | |
2a34c087 ML |
2486 | hctx->tags = set->tags[i]; |
2487 | WARN_ON(!hctx->tags); | |
2488 | ||
889fa31f CY |
2489 | /* |
2490 | * Set the map size to the number of mapped software queues. | |
2491 | * This is more accurate and more efficient than looping | |
2492 | * over all possibly mapped software queues. | |
2493 | */ | |
88459642 | 2494 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2495 | |
484b4061 JA |
2496 | /* |
2497 | * Initialize batch roundrobin counts | |
2498 | */ | |
f82ddf19 | 2499 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
2500 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2501 | } | |
320ae51f JA |
2502 | } |
2503 | ||
8e8320c9 JA |
2504 | /* |
2505 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
2506 | * the queue isn't live yet. | |
2507 | */ | |
2404e607 | 2508 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2509 | { |
2510 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2511 | int i; |
2512 | ||
2404e607 | 2513 | queue_for_each_hw_ctx(q, hctx, i) { |
97889f9a | 2514 | if (shared) |
2404e607 | 2515 | hctx->flags |= BLK_MQ_F_TAG_SHARED; |
97889f9a | 2516 | else |
2404e607 JM |
2517 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; |
2518 | } | |
2519 | } | |
2520 | ||
8e8320c9 JA |
2521 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, |
2522 | bool shared) | |
2404e607 JM |
2523 | { |
2524 | struct request_queue *q; | |
0d2602ca | 2525 | |
705cda97 BVA |
2526 | lockdep_assert_held(&set->tag_list_lock); |
2527 | ||
0d2602ca JA |
2528 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2529 | blk_mq_freeze_queue(q); | |
2404e607 | 2530 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2531 | blk_mq_unfreeze_queue(q); |
2532 | } | |
2533 | } | |
2534 | ||
2535 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2536 | { | |
2537 | struct blk_mq_tag_set *set = q->tag_set; | |
2538 | ||
0d2602ca | 2539 | mutex_lock(&set->tag_list_lock); |
705cda97 | 2540 | list_del_rcu(&q->tag_set_list); |
2404e607 JM |
2541 | if (list_is_singular(&set->tag_list)) { |
2542 | /* just transitioned to unshared */ | |
2543 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2544 | /* update existing queue */ | |
2545 | blk_mq_update_tag_set_depth(set, false); | |
2546 | } | |
0d2602ca | 2547 | mutex_unlock(&set->tag_list_lock); |
a347c7ad | 2548 | INIT_LIST_HEAD(&q->tag_set_list); |
0d2602ca JA |
2549 | } |
2550 | ||
2551 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2552 | struct request_queue *q) | |
2553 | { | |
0d2602ca | 2554 | mutex_lock(&set->tag_list_lock); |
2404e607 | 2555 | |
ff821d27 JA |
2556 | /* |
2557 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
2558 | */ | |
2559 | if (!list_empty(&set->tag_list) && | |
2560 | !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2404e607 JM |
2561 | set->flags |= BLK_MQ_F_TAG_SHARED; |
2562 | /* update existing queue */ | |
2563 | blk_mq_update_tag_set_depth(set, true); | |
2564 | } | |
2565 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2566 | queue_set_hctx_shared(q, true); | |
705cda97 | 2567 | list_add_tail_rcu(&q->tag_set_list, &set->tag_list); |
2404e607 | 2568 | |
0d2602ca JA |
2569 | mutex_unlock(&set->tag_list_lock); |
2570 | } | |
2571 | ||
1db4909e ML |
2572 | /* All allocations will be freed in release handler of q->mq_kobj */ |
2573 | static int blk_mq_alloc_ctxs(struct request_queue *q) | |
2574 | { | |
2575 | struct blk_mq_ctxs *ctxs; | |
2576 | int cpu; | |
2577 | ||
2578 | ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL); | |
2579 | if (!ctxs) | |
2580 | return -ENOMEM; | |
2581 | ||
2582 | ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx); | |
2583 | if (!ctxs->queue_ctx) | |
2584 | goto fail; | |
2585 | ||
2586 | for_each_possible_cpu(cpu) { | |
2587 | struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu); | |
2588 | ctx->ctxs = ctxs; | |
2589 | } | |
2590 | ||
2591 | q->mq_kobj = &ctxs->kobj; | |
2592 | q->queue_ctx = ctxs->queue_ctx; | |
2593 | ||
2594 | return 0; | |
2595 | fail: | |
2596 | kfree(ctxs); | |
2597 | return -ENOMEM; | |
2598 | } | |
2599 | ||
e09aae7e ML |
2600 | /* |
2601 | * It is the actual release handler for mq, but we do it from | |
2602 | * request queue's release handler for avoiding use-after-free | |
2603 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2604 | * but we can't group ctx/kctx kobj without it. | |
2605 | */ | |
2606 | void blk_mq_release(struct request_queue *q) | |
2607 | { | |
2608 | struct blk_mq_hw_ctx *hctx; | |
2609 | unsigned int i; | |
2610 | ||
2611 | /* hctx kobj stays in hctx */ | |
c3b4afca ML |
2612 | queue_for_each_hw_ctx(q, hctx, i) { |
2613 | if (!hctx) | |
2614 | continue; | |
6c8b232e | 2615 | kobject_put(&hctx->kobj); |
c3b4afca | 2616 | } |
e09aae7e ML |
2617 | |
2618 | kfree(q->queue_hw_ctx); | |
2619 | ||
7ea5fe31 ML |
2620 | /* |
2621 | * release .mq_kobj and sw queue's kobject now because | |
2622 | * both share lifetime with request queue. | |
2623 | */ | |
2624 | blk_mq_sysfs_deinit(q); | |
e09aae7e ML |
2625 | } |
2626 | ||
24d2f903 | 2627 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2628 | { |
2629 | struct request_queue *uninit_q, *q; | |
2630 | ||
6d469642 | 2631 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); |
b62c21b7 MS |
2632 | if (!uninit_q) |
2633 | return ERR_PTR(-ENOMEM); | |
2634 | ||
2635 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2636 | if (IS_ERR(q)) | |
2637 | blk_cleanup_queue(uninit_q); | |
2638 | ||
2639 | return q; | |
2640 | } | |
2641 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2642 | ||
9316a9ed JA |
2643 | /* |
2644 | * Helper for setting up a queue with mq ops, given queue depth, and | |
2645 | * the passed in mq ops flags. | |
2646 | */ | |
2647 | struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set, | |
2648 | const struct blk_mq_ops *ops, | |
2649 | unsigned int queue_depth, | |
2650 | unsigned int set_flags) | |
2651 | { | |
2652 | struct request_queue *q; | |
2653 | int ret; | |
2654 | ||
2655 | memset(set, 0, sizeof(*set)); | |
2656 | set->ops = ops; | |
2657 | set->nr_hw_queues = 1; | |
b3c661b1 | 2658 | set->nr_maps = 1; |
9316a9ed JA |
2659 | set->queue_depth = queue_depth; |
2660 | set->numa_node = NUMA_NO_NODE; | |
2661 | set->flags = set_flags; | |
2662 | ||
2663 | ret = blk_mq_alloc_tag_set(set); | |
2664 | if (ret) | |
2665 | return ERR_PTR(ret); | |
2666 | ||
2667 | q = blk_mq_init_queue(set); | |
2668 | if (IS_ERR(q)) { | |
2669 | blk_mq_free_tag_set(set); | |
2670 | return q; | |
2671 | } | |
2672 | ||
2673 | return q; | |
2674 | } | |
2675 | EXPORT_SYMBOL(blk_mq_init_sq_queue); | |
2676 | ||
07319678 BVA |
2677 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) |
2678 | { | |
2679 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2680 | ||
05707b64 | 2681 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu), |
07319678 BVA |
2682 | __alignof__(struct blk_mq_hw_ctx)) != |
2683 | sizeof(struct blk_mq_hw_ctx)); | |
2684 | ||
2685 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2686 | hw_ctx_size += sizeof(struct srcu_struct); | |
2687 | ||
2688 | return hw_ctx_size; | |
2689 | } | |
2690 | ||
34d11ffa JW |
2691 | static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx( |
2692 | struct blk_mq_tag_set *set, struct request_queue *q, | |
2693 | int hctx_idx, int node) | |
2694 | { | |
2695 | struct blk_mq_hw_ctx *hctx; | |
2696 | ||
2697 | hctx = kzalloc_node(blk_mq_hw_ctx_size(set), | |
2698 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
2699 | node); | |
2700 | if (!hctx) | |
2701 | return NULL; | |
2702 | ||
2703 | if (!zalloc_cpumask_var_node(&hctx->cpumask, | |
2704 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
2705 | node)) { | |
2706 | kfree(hctx); | |
2707 | return NULL; | |
2708 | } | |
2709 | ||
2710 | atomic_set(&hctx->nr_active, 0); | |
2711 | hctx->numa_node = node; | |
2712 | hctx->queue_num = hctx_idx; | |
2713 | ||
2714 | if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) { | |
2715 | free_cpumask_var(hctx->cpumask); | |
2716 | kfree(hctx); | |
2717 | return NULL; | |
2718 | } | |
2719 | blk_mq_hctx_kobj_init(hctx); | |
2720 | ||
2721 | return hctx; | |
2722 | } | |
2723 | ||
868f2f0b KB |
2724 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2725 | struct request_queue *q) | |
320ae51f | 2726 | { |
e01ad46d | 2727 | int i, j, end; |
868f2f0b | 2728 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; |
f14bbe77 | 2729 | |
fb350e0a ML |
2730 | /* protect against switching io scheduler */ |
2731 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 2732 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2733 | int node; |
34d11ffa | 2734 | struct blk_mq_hw_ctx *hctx; |
868f2f0b | 2735 | |
ed76e329 | 2736 | node = blk_mq_hw_queue_to_node(&set->map[0], i); |
34d11ffa JW |
2737 | /* |
2738 | * If the hw queue has been mapped to another numa node, | |
2739 | * we need to realloc the hctx. If allocation fails, fallback | |
2740 | * to use the previous one. | |
2741 | */ | |
2742 | if (hctxs[i] && (hctxs[i]->numa_node == node)) | |
2743 | continue; | |
868f2f0b | 2744 | |
34d11ffa JW |
2745 | hctx = blk_mq_alloc_and_init_hctx(set, q, i, node); |
2746 | if (hctx) { | |
2747 | if (hctxs[i]) { | |
2748 | blk_mq_exit_hctx(q, set, hctxs[i], i); | |
2749 | kobject_put(&hctxs[i]->kobj); | |
2750 | } | |
2751 | hctxs[i] = hctx; | |
2752 | } else { | |
2753 | if (hctxs[i]) | |
2754 | pr_warn("Allocate new hctx on node %d fails,\ | |
2755 | fallback to previous one on node %d\n", | |
2756 | node, hctxs[i]->numa_node); | |
2757 | else | |
2758 | break; | |
868f2f0b | 2759 | } |
320ae51f | 2760 | } |
e01ad46d JW |
2761 | /* |
2762 | * Increasing nr_hw_queues fails. Free the newly allocated | |
2763 | * hctxs and keep the previous q->nr_hw_queues. | |
2764 | */ | |
2765 | if (i != set->nr_hw_queues) { | |
2766 | j = q->nr_hw_queues; | |
2767 | end = i; | |
2768 | } else { | |
2769 | j = i; | |
2770 | end = q->nr_hw_queues; | |
2771 | q->nr_hw_queues = set->nr_hw_queues; | |
2772 | } | |
34d11ffa | 2773 | |
e01ad46d | 2774 | for (; j < end; j++) { |
868f2f0b KB |
2775 | struct blk_mq_hw_ctx *hctx = hctxs[j]; |
2776 | ||
2777 | if (hctx) { | |
cc71a6f4 JA |
2778 | if (hctx->tags) |
2779 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 2780 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 2781 | kobject_put(&hctx->kobj); |
868f2f0b KB |
2782 | hctxs[j] = NULL; |
2783 | ||
2784 | } | |
2785 | } | |
fb350e0a | 2786 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
2787 | } |
2788 | ||
392546ae JA |
2789 | /* |
2790 | * Maximum number of hardware queues we support. For single sets, we'll never | |
2791 | * have more than the CPUs (software queues). For multiple sets, the tag_set | |
2792 | * user may have set ->nr_hw_queues larger. | |
2793 | */ | |
2794 | static unsigned int nr_hw_queues(struct blk_mq_tag_set *set) | |
2795 | { | |
2796 | if (set->nr_maps == 1) | |
2797 | return nr_cpu_ids; | |
2798 | ||
2799 | return max(set->nr_hw_queues, nr_cpu_ids); | |
2800 | } | |
2801 | ||
868f2f0b KB |
2802 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, |
2803 | struct request_queue *q) | |
2804 | { | |
66841672 ML |
2805 | /* mark the queue as mq asap */ |
2806 | q->mq_ops = set->ops; | |
2807 | ||
34dbad5d | 2808 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
2809 | blk_mq_poll_stats_bkt, |
2810 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
2811 | if (!q->poll_cb) |
2812 | goto err_exit; | |
2813 | ||
1db4909e | 2814 | if (blk_mq_alloc_ctxs(q)) |
c7de5726 | 2815 | goto err_exit; |
868f2f0b | 2816 | |
737f98cf ML |
2817 | /* init q->mq_kobj and sw queues' kobjects */ |
2818 | blk_mq_sysfs_init(q); | |
2819 | ||
392546ae JA |
2820 | q->nr_queues = nr_hw_queues(set); |
2821 | q->queue_hw_ctx = kcalloc_node(q->nr_queues, sizeof(*(q->queue_hw_ctx)), | |
868f2f0b KB |
2822 | GFP_KERNEL, set->numa_node); |
2823 | if (!q->queue_hw_ctx) | |
1db4909e | 2824 | goto err_sys_init; |
868f2f0b | 2825 | |
868f2f0b KB |
2826 | blk_mq_realloc_hw_ctxs(set, q); |
2827 | if (!q->nr_hw_queues) | |
2828 | goto err_hctxs; | |
320ae51f | 2829 | |
287922eb | 2830 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2831 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f | 2832 | |
a8908939 | 2833 | q->tag_set = set; |
320ae51f | 2834 | |
94eddfbe | 2835 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
6544d229 CH |
2836 | if (set->nr_maps > HCTX_TYPE_POLL) |
2837 | blk_queue_flag_set(QUEUE_FLAG_POLL, q); | |
320ae51f | 2838 | |
05f1dd53 | 2839 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
57d74df9 | 2840 | blk_queue_flag_set(QUEUE_FLAG_NO_SG_MERGE, q); |
05f1dd53 | 2841 | |
1be036e9 CH |
2842 | q->sg_reserved_size = INT_MAX; |
2843 | ||
2849450a | 2844 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2845 | INIT_LIST_HEAD(&q->requeue_list); |
2846 | spin_lock_init(&q->requeue_lock); | |
2847 | ||
254d259d | 2848 | blk_queue_make_request(q, blk_mq_make_request); |
07068d5b | 2849 | |
eba71768 JA |
2850 | /* |
2851 | * Do this after blk_queue_make_request() overrides it... | |
2852 | */ | |
2853 | q->nr_requests = set->queue_depth; | |
2854 | ||
64f1c21e JA |
2855 | /* |
2856 | * Default to classic polling | |
2857 | */ | |
2858 | q->poll_nsec = -1; | |
2859 | ||
24d2f903 | 2860 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 2861 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 2862 | blk_mq_map_swqueue(q); |
4593fdbe | 2863 | |
d3484991 JA |
2864 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { |
2865 | int ret; | |
2866 | ||
131d08e1 | 2867 | ret = elevator_init_mq(q); |
d3484991 JA |
2868 | if (ret) |
2869 | return ERR_PTR(ret); | |
2870 | } | |
2871 | ||
320ae51f | 2872 | return q; |
18741986 | 2873 | |
320ae51f | 2874 | err_hctxs: |
868f2f0b | 2875 | kfree(q->queue_hw_ctx); |
1db4909e ML |
2876 | err_sys_init: |
2877 | blk_mq_sysfs_deinit(q); | |
c7de5726 ML |
2878 | err_exit: |
2879 | q->mq_ops = NULL; | |
320ae51f JA |
2880 | return ERR_PTR(-ENOMEM); |
2881 | } | |
b62c21b7 | 2882 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2883 | |
2884 | void blk_mq_free_queue(struct request_queue *q) | |
2885 | { | |
624dbe47 | 2886 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2887 | |
0d2602ca | 2888 | blk_mq_del_queue_tag_set(q); |
624dbe47 | 2889 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
320ae51f | 2890 | } |
320ae51f | 2891 | |
a5164405 JA |
2892 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2893 | { | |
2894 | int i; | |
2895 | ||
cc71a6f4 JA |
2896 | for (i = 0; i < set->nr_hw_queues; i++) |
2897 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 2898 | goto out_unwind; |
a5164405 JA |
2899 | |
2900 | return 0; | |
2901 | ||
2902 | out_unwind: | |
2903 | while (--i >= 0) | |
cc71a6f4 | 2904 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 2905 | |
a5164405 JA |
2906 | return -ENOMEM; |
2907 | } | |
2908 | ||
2909 | /* | |
2910 | * Allocate the request maps associated with this tag_set. Note that this | |
2911 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2912 | * will be updated to reflect the allocated depth. | |
2913 | */ | |
2914 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2915 | { | |
2916 | unsigned int depth; | |
2917 | int err; | |
2918 | ||
2919 | depth = set->queue_depth; | |
2920 | do { | |
2921 | err = __blk_mq_alloc_rq_maps(set); | |
2922 | if (!err) | |
2923 | break; | |
2924 | ||
2925 | set->queue_depth >>= 1; | |
2926 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2927 | err = -ENOMEM; | |
2928 | break; | |
2929 | } | |
2930 | } while (set->queue_depth); | |
2931 | ||
2932 | if (!set->queue_depth || err) { | |
2933 | pr_err("blk-mq: failed to allocate request map\n"); | |
2934 | return -ENOMEM; | |
2935 | } | |
2936 | ||
2937 | if (depth != set->queue_depth) | |
2938 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2939 | depth, set->queue_depth); | |
2940 | ||
2941 | return 0; | |
2942 | } | |
2943 | ||
ebe8bddb OS |
2944 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
2945 | { | |
59388702 | 2946 | if (set->ops->map_queues && !is_kdump_kernel()) { |
b3c661b1 JA |
2947 | int i; |
2948 | ||
7d4901a9 ML |
2949 | /* |
2950 | * transport .map_queues is usually done in the following | |
2951 | * way: | |
2952 | * | |
2953 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
2954 | * mask = get_cpu_mask(queue) | |
2955 | * for_each_cpu(cpu, mask) | |
b3c661b1 | 2956 | * set->map[x].mq_map[cpu] = queue; |
7d4901a9 ML |
2957 | * } |
2958 | * | |
2959 | * When we need to remap, the table has to be cleared for | |
2960 | * killing stale mapping since one CPU may not be mapped | |
2961 | * to any hw queue. | |
2962 | */ | |
b3c661b1 JA |
2963 | for (i = 0; i < set->nr_maps; i++) |
2964 | blk_mq_clear_mq_map(&set->map[i]); | |
7d4901a9 | 2965 | |
ebe8bddb | 2966 | return set->ops->map_queues(set); |
b3c661b1 JA |
2967 | } else { |
2968 | BUG_ON(set->nr_maps > 1); | |
ed76e329 | 2969 | return blk_mq_map_queues(&set->map[0]); |
b3c661b1 | 2970 | } |
ebe8bddb OS |
2971 | } |
2972 | ||
a4391c64 JA |
2973 | /* |
2974 | * Alloc a tag set to be associated with one or more request queues. | |
2975 | * May fail with EINVAL for various error conditions. May adjust the | |
c018c84f | 2976 | * requested depth down, if it's too large. In that case, the set |
a4391c64 JA |
2977 | * value will be stored in set->queue_depth. |
2978 | */ | |
24d2f903 CH |
2979 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2980 | { | |
b3c661b1 | 2981 | int i, ret; |
da695ba2 | 2982 | |
205fb5f5 BVA |
2983 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2984 | ||
24d2f903 CH |
2985 | if (!set->nr_hw_queues) |
2986 | return -EINVAL; | |
a4391c64 | 2987 | if (!set->queue_depth) |
24d2f903 CH |
2988 | return -EINVAL; |
2989 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2990 | return -EINVAL; | |
2991 | ||
7d7e0f90 | 2992 | if (!set->ops->queue_rq) |
24d2f903 CH |
2993 | return -EINVAL; |
2994 | ||
de148297 ML |
2995 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
2996 | return -EINVAL; | |
2997 | ||
a4391c64 JA |
2998 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2999 | pr_info("blk-mq: reduced tag depth to %u\n", | |
3000 | BLK_MQ_MAX_DEPTH); | |
3001 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
3002 | } | |
24d2f903 | 3003 | |
b3c661b1 JA |
3004 | if (!set->nr_maps) |
3005 | set->nr_maps = 1; | |
3006 | else if (set->nr_maps > HCTX_MAX_TYPES) | |
3007 | return -EINVAL; | |
3008 | ||
6637fadf SL |
3009 | /* |
3010 | * If a crashdump is active, then we are potentially in a very | |
3011 | * memory constrained environment. Limit us to 1 queue and | |
3012 | * 64 tags to prevent using too much memory. | |
3013 | */ | |
3014 | if (is_kdump_kernel()) { | |
3015 | set->nr_hw_queues = 1; | |
59388702 | 3016 | set->nr_maps = 1; |
6637fadf SL |
3017 | set->queue_depth = min(64U, set->queue_depth); |
3018 | } | |
868f2f0b | 3019 | /* |
392546ae JA |
3020 | * There is no use for more h/w queues than cpus if we just have |
3021 | * a single map | |
868f2f0b | 3022 | */ |
392546ae | 3023 | if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids) |
868f2f0b | 3024 | set->nr_hw_queues = nr_cpu_ids; |
6637fadf | 3025 | |
392546ae | 3026 | set->tags = kcalloc_node(nr_hw_queues(set), sizeof(struct blk_mq_tags *), |
24d2f903 CH |
3027 | GFP_KERNEL, set->numa_node); |
3028 | if (!set->tags) | |
a5164405 | 3029 | return -ENOMEM; |
24d2f903 | 3030 | |
da695ba2 | 3031 | ret = -ENOMEM; |
b3c661b1 JA |
3032 | for (i = 0; i < set->nr_maps; i++) { |
3033 | set->map[i].mq_map = kcalloc_node(nr_cpu_ids, | |
07b35eb5 | 3034 | sizeof(set->map[i].mq_map[0]), |
b3c661b1 JA |
3035 | GFP_KERNEL, set->numa_node); |
3036 | if (!set->map[i].mq_map) | |
3037 | goto out_free_mq_map; | |
59388702 | 3038 | set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues; |
b3c661b1 | 3039 | } |
bdd17e75 | 3040 | |
ebe8bddb | 3041 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
3042 | if (ret) |
3043 | goto out_free_mq_map; | |
3044 | ||
3045 | ret = blk_mq_alloc_rq_maps(set); | |
3046 | if (ret) | |
bdd17e75 | 3047 | goto out_free_mq_map; |
24d2f903 | 3048 | |
0d2602ca JA |
3049 | mutex_init(&set->tag_list_lock); |
3050 | INIT_LIST_HEAD(&set->tag_list); | |
3051 | ||
24d2f903 | 3052 | return 0; |
bdd17e75 CH |
3053 | |
3054 | out_free_mq_map: | |
b3c661b1 JA |
3055 | for (i = 0; i < set->nr_maps; i++) { |
3056 | kfree(set->map[i].mq_map); | |
3057 | set->map[i].mq_map = NULL; | |
3058 | } | |
5676e7b6 RE |
3059 | kfree(set->tags); |
3060 | set->tags = NULL; | |
da695ba2 | 3061 | return ret; |
24d2f903 CH |
3062 | } |
3063 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
3064 | ||
3065 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
3066 | { | |
b3c661b1 | 3067 | int i, j; |
24d2f903 | 3068 | |
392546ae | 3069 | for (i = 0; i < nr_hw_queues(set); i++) |
cc71a6f4 | 3070 | blk_mq_free_map_and_requests(set, i); |
484b4061 | 3071 | |
b3c661b1 JA |
3072 | for (j = 0; j < set->nr_maps; j++) { |
3073 | kfree(set->map[j].mq_map); | |
3074 | set->map[j].mq_map = NULL; | |
3075 | } | |
bdd17e75 | 3076 | |
981bd189 | 3077 | kfree(set->tags); |
5676e7b6 | 3078 | set->tags = NULL; |
24d2f903 CH |
3079 | } |
3080 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
3081 | ||
e3a2b3f9 JA |
3082 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
3083 | { | |
3084 | struct blk_mq_tag_set *set = q->tag_set; | |
3085 | struct blk_mq_hw_ctx *hctx; | |
3086 | int i, ret; | |
3087 | ||
bd166ef1 | 3088 | if (!set) |
e3a2b3f9 JA |
3089 | return -EINVAL; |
3090 | ||
70f36b60 | 3091 | blk_mq_freeze_queue(q); |
24f5a90f | 3092 | blk_mq_quiesce_queue(q); |
70f36b60 | 3093 | |
e3a2b3f9 JA |
3094 | ret = 0; |
3095 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
3096 | if (!hctx->tags) |
3097 | continue; | |
bd166ef1 JA |
3098 | /* |
3099 | * If we're using an MQ scheduler, just update the scheduler | |
3100 | * queue depth. This is similar to what the old code would do. | |
3101 | */ | |
70f36b60 | 3102 | if (!hctx->sched_tags) { |
c2e82a23 | 3103 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, |
70f36b60 JA |
3104 | false); |
3105 | } else { | |
3106 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
3107 | nr, true); | |
3108 | } | |
e3a2b3f9 JA |
3109 | if (ret) |
3110 | break; | |
3111 | } | |
3112 | ||
3113 | if (!ret) | |
3114 | q->nr_requests = nr; | |
3115 | ||
24f5a90f | 3116 | blk_mq_unquiesce_queue(q); |
70f36b60 | 3117 | blk_mq_unfreeze_queue(q); |
70f36b60 | 3118 | |
e3a2b3f9 JA |
3119 | return ret; |
3120 | } | |
3121 | ||
d48ece20 JW |
3122 | /* |
3123 | * request_queue and elevator_type pair. | |
3124 | * It is just used by __blk_mq_update_nr_hw_queues to cache | |
3125 | * the elevator_type associated with a request_queue. | |
3126 | */ | |
3127 | struct blk_mq_qe_pair { | |
3128 | struct list_head node; | |
3129 | struct request_queue *q; | |
3130 | struct elevator_type *type; | |
3131 | }; | |
3132 | ||
3133 | /* | |
3134 | * Cache the elevator_type in qe pair list and switch the | |
3135 | * io scheduler to 'none' | |
3136 | */ | |
3137 | static bool blk_mq_elv_switch_none(struct list_head *head, | |
3138 | struct request_queue *q) | |
3139 | { | |
3140 | struct blk_mq_qe_pair *qe; | |
3141 | ||
3142 | if (!q->elevator) | |
3143 | return true; | |
3144 | ||
3145 | qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); | |
3146 | if (!qe) | |
3147 | return false; | |
3148 | ||
3149 | INIT_LIST_HEAD(&qe->node); | |
3150 | qe->q = q; | |
3151 | qe->type = q->elevator->type; | |
3152 | list_add(&qe->node, head); | |
3153 | ||
3154 | mutex_lock(&q->sysfs_lock); | |
3155 | /* | |
3156 | * After elevator_switch_mq, the previous elevator_queue will be | |
3157 | * released by elevator_release. The reference of the io scheduler | |
3158 | * module get by elevator_get will also be put. So we need to get | |
3159 | * a reference of the io scheduler module here to prevent it to be | |
3160 | * removed. | |
3161 | */ | |
3162 | __module_get(qe->type->elevator_owner); | |
3163 | elevator_switch_mq(q, NULL); | |
3164 | mutex_unlock(&q->sysfs_lock); | |
3165 | ||
3166 | return true; | |
3167 | } | |
3168 | ||
3169 | static void blk_mq_elv_switch_back(struct list_head *head, | |
3170 | struct request_queue *q) | |
3171 | { | |
3172 | struct blk_mq_qe_pair *qe; | |
3173 | struct elevator_type *t = NULL; | |
3174 | ||
3175 | list_for_each_entry(qe, head, node) | |
3176 | if (qe->q == q) { | |
3177 | t = qe->type; | |
3178 | break; | |
3179 | } | |
3180 | ||
3181 | if (!t) | |
3182 | return; | |
3183 | ||
3184 | list_del(&qe->node); | |
3185 | kfree(qe); | |
3186 | ||
3187 | mutex_lock(&q->sysfs_lock); | |
3188 | elevator_switch_mq(q, t); | |
3189 | mutex_unlock(&q->sysfs_lock); | |
3190 | } | |
3191 | ||
e4dc2b32 KB |
3192 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
3193 | int nr_hw_queues) | |
868f2f0b KB |
3194 | { |
3195 | struct request_queue *q; | |
d48ece20 | 3196 | LIST_HEAD(head); |
e01ad46d | 3197 | int prev_nr_hw_queues; |
868f2f0b | 3198 | |
705cda97 BVA |
3199 | lockdep_assert_held(&set->tag_list_lock); |
3200 | ||
392546ae | 3201 | if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids) |
868f2f0b KB |
3202 | nr_hw_queues = nr_cpu_ids; |
3203 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
3204 | return; | |
3205 | ||
3206 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3207 | blk_mq_freeze_queue(q); | |
f5bbbbe4 JW |
3208 | /* |
3209 | * Sync with blk_mq_queue_tag_busy_iter. | |
3210 | */ | |
3211 | synchronize_rcu(); | |
d48ece20 JW |
3212 | /* |
3213 | * Switch IO scheduler to 'none', cleaning up the data associated | |
3214 | * with the previous scheduler. We will switch back once we are done | |
3215 | * updating the new sw to hw queue mappings. | |
3216 | */ | |
3217 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3218 | if (!blk_mq_elv_switch_none(&head, q)) | |
3219 | goto switch_back; | |
868f2f0b | 3220 | |
477e19de JW |
3221 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3222 | blk_mq_debugfs_unregister_hctxs(q); | |
3223 | blk_mq_sysfs_unregister(q); | |
3224 | } | |
3225 | ||
e01ad46d | 3226 | prev_nr_hw_queues = set->nr_hw_queues; |
868f2f0b | 3227 | set->nr_hw_queues = nr_hw_queues; |
ebe8bddb | 3228 | blk_mq_update_queue_map(set); |
e01ad46d | 3229 | fallback: |
868f2f0b KB |
3230 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3231 | blk_mq_realloc_hw_ctxs(set, q); | |
e01ad46d JW |
3232 | if (q->nr_hw_queues != set->nr_hw_queues) { |
3233 | pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n", | |
3234 | nr_hw_queues, prev_nr_hw_queues); | |
3235 | set->nr_hw_queues = prev_nr_hw_queues; | |
ed76e329 | 3236 | blk_mq_map_queues(&set->map[0]); |
e01ad46d JW |
3237 | goto fallback; |
3238 | } | |
477e19de JW |
3239 | blk_mq_map_swqueue(q); |
3240 | } | |
3241 | ||
3242 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
3243 | blk_mq_sysfs_register(q); | |
3244 | blk_mq_debugfs_register_hctxs(q); | |
868f2f0b KB |
3245 | } |
3246 | ||
d48ece20 JW |
3247 | switch_back: |
3248 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3249 | blk_mq_elv_switch_back(&head, q); | |
3250 | ||
868f2f0b KB |
3251 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
3252 | blk_mq_unfreeze_queue(q); | |
3253 | } | |
e4dc2b32 KB |
3254 | |
3255 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
3256 | { | |
3257 | mutex_lock(&set->tag_list_lock); | |
3258 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
3259 | mutex_unlock(&set->tag_list_lock); | |
3260 | } | |
868f2f0b KB |
3261 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
3262 | ||
34dbad5d OS |
3263 | /* Enable polling stats and return whether they were already enabled. */ |
3264 | static bool blk_poll_stats_enable(struct request_queue *q) | |
3265 | { | |
3266 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
7dfdbc73 | 3267 | blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q)) |
34dbad5d OS |
3268 | return true; |
3269 | blk_stat_add_callback(q, q->poll_cb); | |
3270 | return false; | |
3271 | } | |
3272 | ||
3273 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
3274 | { | |
3275 | /* | |
3276 | * We don't arm the callback if polling stats are not enabled or the | |
3277 | * callback is already active. | |
3278 | */ | |
3279 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
3280 | blk_stat_is_active(q->poll_cb)) | |
3281 | return; | |
3282 | ||
3283 | blk_stat_activate_msecs(q->poll_cb, 100); | |
3284 | } | |
3285 | ||
3286 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
3287 | { | |
3288 | struct request_queue *q = cb->data; | |
720b8ccc | 3289 | int bucket; |
34dbad5d | 3290 | |
720b8ccc SB |
3291 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
3292 | if (cb->stat[bucket].nr_samples) | |
3293 | q->poll_stat[bucket] = cb->stat[bucket]; | |
3294 | } | |
34dbad5d OS |
3295 | } |
3296 | ||
64f1c21e JA |
3297 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
3298 | struct blk_mq_hw_ctx *hctx, | |
3299 | struct request *rq) | |
3300 | { | |
64f1c21e | 3301 | unsigned long ret = 0; |
720b8ccc | 3302 | int bucket; |
64f1c21e JA |
3303 | |
3304 | /* | |
3305 | * If stats collection isn't on, don't sleep but turn it on for | |
3306 | * future users | |
3307 | */ | |
34dbad5d | 3308 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
3309 | return 0; |
3310 | ||
64f1c21e JA |
3311 | /* |
3312 | * As an optimistic guess, use half of the mean service time | |
3313 | * for this type of request. We can (and should) make this smarter. | |
3314 | * For instance, if the completion latencies are tight, we can | |
3315 | * get closer than just half the mean. This is especially | |
3316 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
3317 | * than ~10 usec. We do use the stats for the relevant IO size |
3318 | * if available which does lead to better estimates. | |
64f1c21e | 3319 | */ |
720b8ccc SB |
3320 | bucket = blk_mq_poll_stats_bkt(rq); |
3321 | if (bucket < 0) | |
3322 | return ret; | |
3323 | ||
3324 | if (q->poll_stat[bucket].nr_samples) | |
3325 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
3326 | |
3327 | return ret; | |
3328 | } | |
3329 | ||
06426adf | 3330 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 3331 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
3332 | struct request *rq) |
3333 | { | |
3334 | struct hrtimer_sleeper hs; | |
3335 | enum hrtimer_mode mode; | |
64f1c21e | 3336 | unsigned int nsecs; |
06426adf JA |
3337 | ktime_t kt; |
3338 | ||
76a86f9d | 3339 | if (rq->rq_flags & RQF_MQ_POLL_SLEPT) |
64f1c21e JA |
3340 | return false; |
3341 | ||
3342 | /* | |
1052b8ac | 3343 | * If we get here, hybrid polling is enabled. Hence poll_nsec can be: |
64f1c21e | 3344 | * |
64f1c21e JA |
3345 | * 0: use half of prev avg |
3346 | * >0: use this specific value | |
3347 | */ | |
1052b8ac | 3348 | if (q->poll_nsec > 0) |
64f1c21e JA |
3349 | nsecs = q->poll_nsec; |
3350 | else | |
3351 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
3352 | ||
3353 | if (!nsecs) | |
06426adf JA |
3354 | return false; |
3355 | ||
76a86f9d | 3356 | rq->rq_flags |= RQF_MQ_POLL_SLEPT; |
06426adf JA |
3357 | |
3358 | /* | |
3359 | * This will be replaced with the stats tracking code, using | |
3360 | * 'avg_completion_time / 2' as the pre-sleep target. | |
3361 | */ | |
8b0e1953 | 3362 | kt = nsecs; |
06426adf JA |
3363 | |
3364 | mode = HRTIMER_MODE_REL; | |
3365 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
3366 | hrtimer_set_expires(&hs.timer, kt); | |
3367 | ||
3368 | hrtimer_init_sleeper(&hs, current); | |
3369 | do { | |
5a61c363 | 3370 | if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE) |
06426adf JA |
3371 | break; |
3372 | set_current_state(TASK_UNINTERRUPTIBLE); | |
3373 | hrtimer_start_expires(&hs.timer, mode); | |
3374 | if (hs.task) | |
3375 | io_schedule(); | |
3376 | hrtimer_cancel(&hs.timer); | |
3377 | mode = HRTIMER_MODE_ABS; | |
3378 | } while (hs.task && !signal_pending(current)); | |
3379 | ||
3380 | __set_current_state(TASK_RUNNING); | |
3381 | destroy_hrtimer_on_stack(&hs.timer); | |
3382 | return true; | |
3383 | } | |
3384 | ||
1052b8ac JA |
3385 | static bool blk_mq_poll_hybrid(struct request_queue *q, |
3386 | struct blk_mq_hw_ctx *hctx, blk_qc_t cookie) | |
bbd7bb70 | 3387 | { |
1052b8ac JA |
3388 | struct request *rq; |
3389 | ||
3390 | if (q->poll_nsec == -1) | |
3391 | return false; | |
3392 | ||
3393 | if (!blk_qc_t_is_internal(cookie)) | |
3394 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
3395 | else { | |
3396 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); | |
3397 | /* | |
3398 | * With scheduling, if the request has completed, we'll | |
3399 | * get a NULL return here, as we clear the sched tag when | |
3400 | * that happens. The request still remains valid, like always, | |
3401 | * so we should be safe with just the NULL check. | |
3402 | */ | |
3403 | if (!rq) | |
3404 | return false; | |
3405 | } | |
3406 | ||
3407 | return blk_mq_poll_hybrid_sleep(q, hctx, rq); | |
3408 | } | |
3409 | ||
529262d5 CH |
3410 | /** |
3411 | * blk_poll - poll for IO completions | |
3412 | * @q: the queue | |
3413 | * @cookie: cookie passed back at IO submission time | |
3414 | * @spin: whether to spin for completions | |
3415 | * | |
3416 | * Description: | |
3417 | * Poll for completions on the passed in queue. Returns number of | |
3418 | * completed entries found. If @spin is true, then blk_poll will continue | |
3419 | * looping until at least one completion is found, unless the task is | |
3420 | * otherwise marked running (or we need to reschedule). | |
3421 | */ | |
3422 | int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin) | |
1052b8ac JA |
3423 | { |
3424 | struct blk_mq_hw_ctx *hctx; | |
bbd7bb70 JA |
3425 | long state; |
3426 | ||
529262d5 CH |
3427 | if (!blk_qc_t_valid(cookie) || |
3428 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
1052b8ac JA |
3429 | return 0; |
3430 | ||
529262d5 CH |
3431 | if (current->plug) |
3432 | blk_flush_plug_list(current->plug, false); | |
3433 | ||
1052b8ac JA |
3434 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; |
3435 | ||
06426adf JA |
3436 | /* |
3437 | * If we sleep, have the caller restart the poll loop to reset | |
3438 | * the state. Like for the other success return cases, the | |
3439 | * caller is responsible for checking if the IO completed. If | |
3440 | * the IO isn't complete, we'll get called again and will go | |
3441 | * straight to the busy poll loop. | |
3442 | */ | |
1052b8ac | 3443 | if (blk_mq_poll_hybrid(q, hctx, cookie)) |
85f4d4b6 | 3444 | return 1; |
06426adf | 3445 | |
bbd7bb70 JA |
3446 | hctx->poll_considered++; |
3447 | ||
3448 | state = current->state; | |
aa61bec3 | 3449 | do { |
bbd7bb70 JA |
3450 | int ret; |
3451 | ||
3452 | hctx->poll_invoked++; | |
3453 | ||
9743139c | 3454 | ret = q->mq_ops->poll(hctx); |
bbd7bb70 JA |
3455 | if (ret > 0) { |
3456 | hctx->poll_success++; | |
849a3700 | 3457 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 3458 | return ret; |
bbd7bb70 JA |
3459 | } |
3460 | ||
3461 | if (signal_pending_state(state, current)) | |
849a3700 | 3462 | __set_current_state(TASK_RUNNING); |
bbd7bb70 JA |
3463 | |
3464 | if (current->state == TASK_RUNNING) | |
85f4d4b6 | 3465 | return 1; |
0a1b8b87 | 3466 | if (ret < 0 || !spin) |
bbd7bb70 JA |
3467 | break; |
3468 | cpu_relax(); | |
aa61bec3 | 3469 | } while (!need_resched()); |
bbd7bb70 | 3470 | |
67b4110f | 3471 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 3472 | return 0; |
bbd7bb70 | 3473 | } |
529262d5 | 3474 | EXPORT_SYMBOL_GPL(blk_poll); |
bbd7bb70 | 3475 | |
9cf2bab6 JA |
3476 | unsigned int blk_mq_rq_cpu(struct request *rq) |
3477 | { | |
3478 | return rq->mq_ctx->cpu; | |
3479 | } | |
3480 | EXPORT_SYMBOL(blk_mq_rq_cpu); | |
3481 | ||
320ae51f JA |
3482 | static int __init blk_mq_init(void) |
3483 | { | |
9467f859 TG |
3484 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
3485 | blk_mq_hctx_notify_dead); | |
320ae51f JA |
3486 | return 0; |
3487 | } | |
3488 | subsys_initcall(blk_mq_init); |